Files
llmx/codex-rs/core/src/config.rs

1789 lines
63 KiB
Rust
Raw Normal View History

use crate::config_profile::ConfigProfile;
use crate::config_types::History;
use crate::config_types::McpServerConfig;
use crate::config_types::ReasoningSummaryFormat;
use crate::config_types::SandboxWorkspaceWrite;
feat: introduce support for shell_environment_policy in config.toml (#1061) To date, when handling `shell` and `local_shell` tool calls, we were spawning new processes using the environment inherited from the Codex process itself. This means that the sensitive `OPENAI_API_KEY` that Codex needs to talk to OpenAI models was made available to everything run by `shell` and `local_shell`. While there are cases where that might be useful, it does not seem like a good default. This PR introduces a complex `shell_environment_policy` config option to control the `env` used with these tool calls. It is inevitably a bit complex so that it is possible to override individual components of the policy so without having to restate the entire thing. Details are in the updated `README.md` in this PR, but here is the relevant bit that explains the individual fields of `shell_environment_policy`: | Field | Type | Default | Description | | ------------------------- | -------------------------- | ------- | ----------------------------------------------------------------------------------------------------------------------------------------------- | | `inherit` | string | `core` | Starting template for the environment:<br>`core` (`HOME`, `PATH`, `USER`, …), `all` (clone full parent env), or `none` (start empty). | | `ignore_default_excludes` | boolean | `false` | When `false`, Codex removes any var whose **name** contains `KEY`, `SECRET`, or `TOKEN` (case-insensitive) before other rules run. | | `exclude` | array&lt;string&gt; | `[]` | Case-insensitive glob patterns to drop after the default filter.<br>Examples: `"AWS_*"`, `"AZURE_*"`. | | `set` | table&lt;string,string&gt; | `{}` | Explicit key/value overrides or additions – always win over inherited values. | | `include_only` | array&lt;string&gt; | `[]` | If non-empty, a whitelist of patterns; only variables that match _one_ pattern survive the final step. (Generally used with `inherit = "all"`.) | In particular, note that the default is `inherit = "core"`, so: * if you have extra env variables that you want to inherit from the parent process, use `inherit = "all"` and then specify `include_only` * if you have extra env variables where you want to hardcode the values, the default `inherit = "core"` will work fine, but then you need to specify `set` This configuration is not battle-tested, so we will probably still have to play with it a bit. `core/src/exec_env.rs` has the critical business logic as well as unit tests. Though if nothing else, previous to this change: ``` $ cargo run --bin codex -- debug seatbelt -- printenv OPENAI_API_KEY # ...prints OPENAI_API_KEY... ``` But after this change it does not print anything (as desired). One final thing to call out about this PR is that the `configure_command!` macro we use in `core/src/exec.rs` has to do some complex logic with respect to how it builds up the `env` for the process being spawned under Landlock/seccomp. Specifically, doing `cmd.env_clear()` followed by `cmd.envs(&$env_map)` (which is arguably the most intuitive way to do it) caused the Landlock unit tests to fail because the processes spawned by the unit tests started failing in unexpected ways! If we forgo `env_clear()` in favor of updating env vars one at a time, the tests still pass. The comment in the code talks about this a bit, and while I would like to investigate this more, I need to move on for the moment, but I do plan to come back to it to fully understand what is going on. For example, this suggests that we might not be able to spawn a C program that calls `env_clear()`, which would be...weird. We may still have to fiddle with our Landlock config if that is the case.
2025-05-22 09:51:19 -07:00
use crate::config_types::ShellEnvironmentPolicy;
use crate::config_types::ShellEnvironmentPolicyToml;
use crate::config_types::Tui;
use crate::config_types::UriBasedFileOpener;
use crate::git_info::resolve_root_git_project_for_trust;
use crate::model_family::ModelFamily;
use crate::model_family::find_family_for_model;
use crate::model_provider_info::ModelProviderInfo;
use crate::model_provider_info::built_in_model_providers;
feat: show number of tokens remaining in UI (#1388) When using the OpenAI Responses API, we now record the `usage` field for a `"response.completed"` event, which includes metrics about the number of tokens consumed. We also introduce `openai_model_info.rs`, which includes current data about the most common OpenAI models available via the API (specifically `context_window` and `max_output_tokens`). If Codex does not recognize the model, you can set `model_context_window` and `model_max_output_tokens` explicitly in `config.toml`. When then introduce a new event type to `protocol.rs`, `TokenCount`, which includes the `TokenUsage` for the most recent turn. Finally, we update the TUI to record the running sum of tokens used so the percentage of available context window remaining can be reported via the placeholder text for the composer: ![Screenshot 2025-06-25 at 11 20 55 PM](https://github.com/user-attachments/assets/6fd6982f-7247-4f14-84b2-2e600cb1fd49) We could certainly get much fancier with this (such as reporting the estimated cost of the conversation), but for now, we are just trying to achieve feature parity with the TypeScript CLI. Though arguably this improves upon the TypeScript CLI, as the TypeScript CLI uses heuristics to estimate the number of tokens used rather than using the `usage` information directly: https://github.com/openai/codex/blob/296996d74e345b1b05d8c3451a06ace21c5ada96/codex-cli/src/utils/approximate-tokens-used.ts#L3-L16 Fixes https://github.com/openai/codex/issues/1242
2025-06-25 23:31:11 -07:00
use crate::openai_model_info::get_model_info;
use crate::protocol::AskForApproval;
use crate::protocol::SandboxPolicy;
use anyhow::Context;
use codex_protocol::config_types::ReasoningEffort;
use codex_protocol::config_types::ReasoningSummary;
use codex_protocol::config_types::SandboxMode;
use codex_protocol::config_types::Verbosity;
use codex_protocol::mcp_protocol::Tools;
use codex_protocol::mcp_protocol::UserSavedConfig;
use dirs::home_dir;
use serde::Deserialize;
feat: support mcp_servers in config.toml (#829) This adds initial support for MCP servers in the style of Claude Desktop and Cursor. Note this PR is the bare minimum to get things working end to end: all configured MCP servers are launched every time Codex is run, there is no recovery for MCP servers that crash, etc. (Also, I took some shortcuts to change some fields of `Session` to be `pub(crate)`, which also means there are circular deps between `codex.rs` and `mcp_tool_call.rs`, but I will clean that up in a subsequent PR.) `codex-rs/README.md` is updated as part of this PR to explain how to use this feature. There is a bit of plumbing to route the new settings from `Config` to the business logic in `codex.rs`. The most significant chunks for new code are in `mcp_connection_manager.rs` (which defines the `McpConnectionManager` struct) and `mcp_tool_call.rs`, which is responsible for tool calls. This PR also introduces new `McpToolCallBegin` and `McpToolCallEnd` event types to the protocol, but does not add any handlers for them. (See https://github.com/openai/codex/pull/836 for initial usage.) To test, I added the following to my `~/.codex/config.toml`: ```toml # Local build of https://github.com/hideya/mcp-server-weather-js [mcp_servers.weather] command = "/Users/mbolin/code/mcp-server-weather-js/dist/index.js" args = [] ``` And then I ran the following: ``` codex-rs$ cargo run --bin codex exec 'what is the weather in san francisco' [2025-05-06T22:40:05] Task started: 1 [2025-05-06T22:40:18] Agent message: Here’s the latest National Weather Service forecast for San Francisco (downtown, near 37.77° N, 122.42° W): This Afternoon (Tue): • Sunny, high near 69 °F • West-southwest wind around 12 mph Tonight: • Partly cloudy, low around 52 °F • SW wind 7–10 mph ... ``` Note that Codex itself is not able to make network calls, so it would not normally be able to get live weather information like this. However, the weather MCP is [currently] not run under the Codex sandbox, so it is able to hit `api.weather.gov` and fetch current weather information. --- [//]: # (BEGIN SAPLING FOOTER) Stack created with [Sapling](https://sapling-scm.com). Best reviewed with [ReviewStack](https://reviewstack.dev/openai/codex/pull/829). * #836 * __->__ #829
2025-05-06 15:47:59 -07:00
use std::collections::HashMap;
use std::path::Path;
use std::path::PathBuf;
use tempfile::NamedTempFile;
feat: add support for -c/--config to override individual config items (#1137) This PR introduces support for `-c`/`--config` so users can override individual config values on the command line using `--config name=value`. Example: ``` codex --config model=o4-mini ``` Making it possible to set arbitrary config values on the command line results in a more flexible configuration scheme and makes it easier to provide single-line examples that can be copy-pasted from documentation. Effectively, it means there are four levels of configuration for some values: - Default value (e.g., `model` currently defaults to `o4-mini`) - Value in `config.toml` (e.g., user could override the default to be `model = "o3"` in their `config.toml`) - Specifying `-c` or `--config` to override `model` (e.g., user can include `-c model=o3` in their list of args to Codex) - If available, a config-specific flag can be used, which takes precedence over `-c` (e.g., user can specify `--model o3` in their list of args to Codex) Now that it is possible to specify anything that could be configured in `config.toml` on the command line using `-c`, we do not need to have a custom flag for every possible config option (which can clutter the output of `--help`). To that end, as part of this PR, we drop support for the `--disable-response-storage` flag, as users can now specify `-c disable_response_storage=true` to get the equivalent functionality. Under the hood, this works by loading the `config.toml` into a `toml::Value`. Then for each `key=value`, we create a small synthetic TOML file with `value` so that we can run the TOML parser to get the equivalent `toml::Value`. We then parse `key` to determine the point in the original `toml::Value` to do the insert/replace. Once all of the overrides from `-c` args have been applied, the `toml::Value` is deserialized into a `ConfigToml` and then the `ConfigOverrides` are applied, as before.
2025-05-27 23:11:44 -07:00
use toml::Value as TomlValue;
use toml_edit::DocumentMut;
const OPENAI_DEFAULT_MODEL: &str = "gpt-5";
Review Mode (Core) (#3401) ## 📝 Review Mode -- Core This PR introduces the Core implementation for Review mode: - New op `Op::Review { prompt: String }:` spawns a child review task with isolated context, a review‑specific system prompt, and a `Config.review_model`. - `EnteredReviewMode`: emitted when the child review session starts. Every event from this point onwards reflects the review session. - `ExitedReviewMode(Option<ReviewOutputEvent>)`: emitted when the review finishes or is interrupted, with optional structured findings: ```json { "findings": [ { "title": "<≤ 80 chars, imperative>", "body": "<valid Markdown explaining *why* this is a problem; cite files/lines/functions>", "confidence_score": <float 0.0-1.0>, "priority": <int 0-3>, "code_location": { "absolute_file_path": "<file path>", "line_range": {"start": <int>, "end": <int>} } } ], "overall_correctness": "patch is correct" | "patch is incorrect", "overall_explanation": "<1-3 sentence explanation justifying the overall_correctness verdict>", "overall_confidence_score": <float 0.0-1.0> } ``` ## Questions ### Why separate out its own message history? We want the review thread to match the training of our review models as much as possible -- that means using a custom prompt, removing user instructions, and starting a clean chat history. We also want to make sure the review thread doesn't leak into the parent thread. ### Why do this as a mode, vs. sub-agents? 1. We want review to be a synchronous task, so it's fine for now to do a bespoke implementation. 2. We're still unclear about the final structure for sub-agents. We'd prefer to land this quickly and then refactor into sub-agents without rushing that implementation.
2025-09-12 16:25:10 -07:00
const OPENAI_DEFAULT_REVIEW_MODEL: &str = "gpt-5";
pub const GPT5_HIGH_MODEL: &str = "gpt-5-high-new";
/// Maximum number of bytes of the documentation that will be embedded. Larger
/// files are *silently truncated* to this size so we do not take up too much of
/// the context window.
pub(crate) const PROJECT_DOC_MAX_BYTES: usize = 32 * 1024; // 32 KiB
pub(crate) const CONFIG_TOML_FILE: &str = "config.toml";
/// Application configuration loaded from disk and merged with overrides.
#[derive(Debug, Clone, PartialEq)]
pub struct Config {
/// Optional override of model selection.
pub model: String,
fix: overhaul SandboxPolicy and config loading in Rust (#732) Previous to this PR, `SandboxPolicy` was a bit difficult to work with: https://github.com/openai/codex/blob/237f8a11e11fdcc793a09e787e48215676d9b95b/codex-rs/core/src/protocol.rs#L98-L108 Specifically: * It was an `enum` and therefore options were mutually exclusive as opposed to additive. * It defined things in terms of what the agent _could not_ do as opposed to what they _could_ do. This made things hard to support because we would prefer to build up a sandbox config by starting with something extremely restrictive and only granting permissions for things the user as explicitly allowed. This PR changes things substantially by redefining the policy in terms of two concepts: * A `SandboxPermission` enum that defines permissions that can be granted to the agent/sandbox. * A `SandboxPolicy` that internally stores a `Vec<SandboxPermission>`, but externally exposes a simpler API that can be used to configure Seatbelt/Landlock. Previous to this PR, we supported a `--sandbox` flag that effectively mapped to an enum value in `SandboxPolicy`. Though now that `SandboxPolicy` is a wrapper around `Vec<SandboxPermission>`, the single `--sandbox` flag no longer makes sense. While I could have turned it into a flag that the user can specify multiple times, I think the current values to use with such a flag are long and potentially messy, so for the moment, I have dropped support for `--sandbox` altogether and we can bring it back once we have figured out the naming thing. Since `--sandbox` is gone, users now have to specify `--full-auto` to get a sandbox that allows writes in `cwd`. Admittedly, there is no clean way to specify the equivalent of `--full-auto` in your `config.toml` right now, so we will have to revisit that, as well. Because `Config` presents a `SandboxPolicy` field and `SandboxPolicy` changed considerably, I had to overhaul how config loading works, as well. There are now two distinct concepts, `ConfigToml` and `Config`: * `ConfigToml` is the deserialization of `~/.codex/config.toml`. As one might expect, every field is `Optional` and it is `#[derive(Deserialize, Default)]`. Consistent use of `Optional` makes it clear what the user has specified explicitly. * `Config` is the "normalized config" and is produced by merging `ConfigToml` with `ConfigOverrides`. Where `ConfigToml` contains a raw `Option<Vec<SandboxPermission>>`, `Config` presents only the final `SandboxPolicy`. The changes to `core/src/exec.rs` and `core/src/linux.rs` merit extra special attention to ensure we are faithfully mapping the `SandboxPolicy` to the Seatbelt and Landlock configs, respectively. Also, take note that `core/src/seatbelt_readonly_policy.sbpl` has been renamed to `codex-rs/core/src/seatbelt_base_policy.sbpl` and that `(allow file-read*)` has been removed from the `.sbpl` file as now this is added to the policy in `core/src/exec.rs` when `sandbox_policy.has_full_disk_read_access()` is `true`.
2025-04-29 15:01:16 -07:00
Review Mode (Core) (#3401) ## 📝 Review Mode -- Core This PR introduces the Core implementation for Review mode: - New op `Op::Review { prompt: String }:` spawns a child review task with isolated context, a review‑specific system prompt, and a `Config.review_model`. - `EnteredReviewMode`: emitted when the child review session starts. Every event from this point onwards reflects the review session. - `ExitedReviewMode(Option<ReviewOutputEvent>)`: emitted when the review finishes or is interrupted, with optional structured findings: ```json { "findings": [ { "title": "<≤ 80 chars, imperative>", "body": "<valid Markdown explaining *why* this is a problem; cite files/lines/functions>", "confidence_score": <float 0.0-1.0>, "priority": <int 0-3>, "code_location": { "absolute_file_path": "<file path>", "line_range": {"start": <int>, "end": <int>} } } ], "overall_correctness": "patch is correct" | "patch is incorrect", "overall_explanation": "<1-3 sentence explanation justifying the overall_correctness verdict>", "overall_confidence_score": <float 0.0-1.0> } ``` ## Questions ### Why separate out its own message history? We want the review thread to match the training of our review models as much as possible -- that means using a custom prompt, removing user instructions, and starting a clean chat history. We also want to make sure the review thread doesn't leak into the parent thread. ### Why do this as a mode, vs. sub-agents? 1. We want review to be a synchronous task, so it's fine for now to do a bespoke implementation. 2. We're still unclear about the final structure for sub-agents. We'd prefer to land this quickly and then refactor into sub-agents without rushing that implementation.
2025-09-12 16:25:10 -07:00
/// Model used specifically for review sessions. Defaults to "gpt-5".
pub review_model: String,
pub model_family: ModelFamily,
feat: show number of tokens remaining in UI (#1388) When using the OpenAI Responses API, we now record the `usage` field for a `"response.completed"` event, which includes metrics about the number of tokens consumed. We also introduce `openai_model_info.rs`, which includes current data about the most common OpenAI models available via the API (specifically `context_window` and `max_output_tokens`). If Codex does not recognize the model, you can set `model_context_window` and `model_max_output_tokens` explicitly in `config.toml`. When then introduce a new event type to `protocol.rs`, `TokenCount`, which includes the `TokenUsage` for the most recent turn. Finally, we update the TUI to record the running sum of tokens used so the percentage of available context window remaining can be reported via the placeholder text for the composer: ![Screenshot 2025-06-25 at 11 20 55 PM](https://github.com/user-attachments/assets/6fd6982f-7247-4f14-84b2-2e600cb1fd49) We could certainly get much fancier with this (such as reporting the estimated cost of the conversation), but for now, we are just trying to achieve feature parity with the TypeScript CLI. Though arguably this improves upon the TypeScript CLI, as the TypeScript CLI uses heuristics to estimate the number of tokens used rather than using the `usage` information directly: https://github.com/openai/codex/blob/296996d74e345b1b05d8c3451a06ace21c5ada96/codex-cli/src/utils/approximate-tokens-used.ts#L3-L16 Fixes https://github.com/openai/codex/issues/1242
2025-06-25 23:31:11 -07:00
/// Size of the context window for the model, in tokens.
pub model_context_window: Option<u64>,
/// Maximum number of output tokens.
pub model_max_output_tokens: Option<u64>,
/// Token usage threshold triggering auto-compaction of conversation history.
pub model_auto_compact_token_limit: Option<i64>,
feat: support the chat completions API in the Rust CLI (#862) This is a substantial PR to add support for the chat completions API, which in turn makes it possible to use non-OpenAI model providers (just like in the TypeScript CLI): * It moves a number of structs from `client.rs` to `client_common.rs` so they can be shared. * It introduces support for the chat completions API in `chat_completions.rs`. * It updates `ModelProviderInfo` so that `env_key` is `Option<String>` instead of `String` (for e.g., ollama) and adds a `wire_api` field * It updates `client.rs` to choose between `stream_responses()` and `stream_chat_completions()` based on the `wire_api` for the `ModelProviderInfo` * It updates the `exec` and TUI CLIs to no longer fail if the `OPENAI_API_KEY` environment variable is not set * It updates the TUI so that `EventMsg::Error` is displayed more prominently when it occurs, particularly now that it is important to alert users to the `CodexErr::EnvVar` variant. * `CodexErr::EnvVar` was updated to include an optional `instructions` field so we can preserve the behavior where we direct users to https://platform.openai.com if `OPENAI_API_KEY` is not set. * Cleaned up the "welcome message" in the TUI to ensure the model provider is displayed. * Updated the docs in `codex-rs/README.md`. To exercise the chat completions API from OpenAI models, I added the following to my `config.toml`: ```toml model = "gpt-4o" model_provider = "openai-chat-completions" [model_providers.openai-chat-completions] name = "OpenAI using Chat Completions" base_url = "https://api.openai.com/v1" env_key = "OPENAI_API_KEY" wire_api = "chat" ``` Though to test a non-OpenAI provider, I installed ollama with mistral locally on my Mac because ChatGPT said that would be a good match for my hardware: ```shell brew install ollama ollama serve ollama pull mistral ``` Then I added the following to my `~/.codex/config.toml`: ```toml model = "mistral" model_provider = "ollama" ``` Note this code could certainly use more test coverage, but I want to get this in so folks can start playing with it. For reference, I believe https://github.com/openai/codex/pull/247 was roughly the comparable PR on the TypeScript side.
2025-05-08 21:46:06 -07:00
/// Key into the model_providers map that specifies which provider to use.
pub model_provider_id: String,
/// Info needed to make an API request to the model.
pub model_provider: ModelProviderInfo,
fix: overhaul SandboxPolicy and config loading in Rust (#732) Previous to this PR, `SandboxPolicy` was a bit difficult to work with: https://github.com/openai/codex/blob/237f8a11e11fdcc793a09e787e48215676d9b95b/codex-rs/core/src/protocol.rs#L98-L108 Specifically: * It was an `enum` and therefore options were mutually exclusive as opposed to additive. * It defined things in terms of what the agent _could not_ do as opposed to what they _could_ do. This made things hard to support because we would prefer to build up a sandbox config by starting with something extremely restrictive and only granting permissions for things the user as explicitly allowed. This PR changes things substantially by redefining the policy in terms of two concepts: * A `SandboxPermission` enum that defines permissions that can be granted to the agent/sandbox. * A `SandboxPolicy` that internally stores a `Vec<SandboxPermission>`, but externally exposes a simpler API that can be used to configure Seatbelt/Landlock. Previous to this PR, we supported a `--sandbox` flag that effectively mapped to an enum value in `SandboxPolicy`. Though now that `SandboxPolicy` is a wrapper around `Vec<SandboxPermission>`, the single `--sandbox` flag no longer makes sense. While I could have turned it into a flag that the user can specify multiple times, I think the current values to use with such a flag are long and potentially messy, so for the moment, I have dropped support for `--sandbox` altogether and we can bring it back once we have figured out the naming thing. Since `--sandbox` is gone, users now have to specify `--full-auto` to get a sandbox that allows writes in `cwd`. Admittedly, there is no clean way to specify the equivalent of `--full-auto` in your `config.toml` right now, so we will have to revisit that, as well. Because `Config` presents a `SandboxPolicy` field and `SandboxPolicy` changed considerably, I had to overhaul how config loading works, as well. There are now two distinct concepts, `ConfigToml` and `Config`: * `ConfigToml` is the deserialization of `~/.codex/config.toml`. As one might expect, every field is `Optional` and it is `#[derive(Deserialize, Default)]`. Consistent use of `Optional` makes it clear what the user has specified explicitly. * `Config` is the "normalized config" and is produced by merging `ConfigToml` with `ConfigOverrides`. Where `ConfigToml` contains a raw `Option<Vec<SandboxPermission>>`, `Config` presents only the final `SandboxPolicy`. The changes to `core/src/exec.rs` and `core/src/linux.rs` merit extra special attention to ensure we are faithfully mapping the `SandboxPolicy` to the Seatbelt and Landlock configs, respectively. Also, take note that `core/src/seatbelt_readonly_policy.sbpl` has been renamed to `codex-rs/core/src/seatbelt_base_policy.sbpl` and that `(allow file-read*)` has been removed from the `.sbpl` file as now this is added to the policy in `core/src/exec.rs` when `sandbox_policy.has_full_disk_read_access()` is `true`.
2025-04-29 15:01:16 -07:00
/// Approval policy for executing commands.
pub approval_policy: AskForApproval,
fix: overhaul SandboxPolicy and config loading in Rust (#732) Previous to this PR, `SandboxPolicy` was a bit difficult to work with: https://github.com/openai/codex/blob/237f8a11e11fdcc793a09e787e48215676d9b95b/codex-rs/core/src/protocol.rs#L98-L108 Specifically: * It was an `enum` and therefore options were mutually exclusive as opposed to additive. * It defined things in terms of what the agent _could not_ do as opposed to what they _could_ do. This made things hard to support because we would prefer to build up a sandbox config by starting with something extremely restrictive and only granting permissions for things the user as explicitly allowed. This PR changes things substantially by redefining the policy in terms of two concepts: * A `SandboxPermission` enum that defines permissions that can be granted to the agent/sandbox. * A `SandboxPolicy` that internally stores a `Vec<SandboxPermission>`, but externally exposes a simpler API that can be used to configure Seatbelt/Landlock. Previous to this PR, we supported a `--sandbox` flag that effectively mapped to an enum value in `SandboxPolicy`. Though now that `SandboxPolicy` is a wrapper around `Vec<SandboxPermission>`, the single `--sandbox` flag no longer makes sense. While I could have turned it into a flag that the user can specify multiple times, I think the current values to use with such a flag are long and potentially messy, so for the moment, I have dropped support for `--sandbox` altogether and we can bring it back once we have figured out the naming thing. Since `--sandbox` is gone, users now have to specify `--full-auto` to get a sandbox that allows writes in `cwd`. Admittedly, there is no clean way to specify the equivalent of `--full-auto` in your `config.toml` right now, so we will have to revisit that, as well. Because `Config` presents a `SandboxPolicy` field and `SandboxPolicy` changed considerably, I had to overhaul how config loading works, as well. There are now two distinct concepts, `ConfigToml` and `Config`: * `ConfigToml` is the deserialization of `~/.codex/config.toml`. As one might expect, every field is `Optional` and it is `#[derive(Deserialize, Default)]`. Consistent use of `Optional` makes it clear what the user has specified explicitly. * `Config` is the "normalized config" and is produced by merging `ConfigToml` with `ConfigOverrides`. Where `ConfigToml` contains a raw `Option<Vec<SandboxPermission>>`, `Config` presents only the final `SandboxPolicy`. The changes to `core/src/exec.rs` and `core/src/linux.rs` merit extra special attention to ensure we are faithfully mapping the `SandboxPolicy` to the Seatbelt and Landlock configs, respectively. Also, take note that `core/src/seatbelt_readonly_policy.sbpl` has been renamed to `codex-rs/core/src/seatbelt_base_policy.sbpl` and that `(allow file-read*)` has been removed from the `.sbpl` file as now this is added to the policy in `core/src/exec.rs` when `sandbox_policy.has_full_disk_read_access()` is `true`.
2025-04-29 15:01:16 -07:00
pub sandbox_policy: SandboxPolicy,
feat: introduce support for shell_environment_policy in config.toml (#1061) To date, when handling `shell` and `local_shell` tool calls, we were spawning new processes using the environment inherited from the Codex process itself. This means that the sensitive `OPENAI_API_KEY` that Codex needs to talk to OpenAI models was made available to everything run by `shell` and `local_shell`. While there are cases where that might be useful, it does not seem like a good default. This PR introduces a complex `shell_environment_policy` config option to control the `env` used with these tool calls. It is inevitably a bit complex so that it is possible to override individual components of the policy so without having to restate the entire thing. Details are in the updated `README.md` in this PR, but here is the relevant bit that explains the individual fields of `shell_environment_policy`: | Field | Type | Default | Description | | ------------------------- | -------------------------- | ------- | ----------------------------------------------------------------------------------------------------------------------------------------------- | | `inherit` | string | `core` | Starting template for the environment:<br>`core` (`HOME`, `PATH`, `USER`, …), `all` (clone full parent env), or `none` (start empty). | | `ignore_default_excludes` | boolean | `false` | When `false`, Codex removes any var whose **name** contains `KEY`, `SECRET`, or `TOKEN` (case-insensitive) before other rules run. | | `exclude` | array&lt;string&gt; | `[]` | Case-insensitive glob patterns to drop after the default filter.<br>Examples: `"AWS_*"`, `"AZURE_*"`. | | `set` | table&lt;string,string&gt; | `{}` | Explicit key/value overrides or additions – always win over inherited values. | | `include_only` | array&lt;string&gt; | `[]` | If non-empty, a whitelist of patterns; only variables that match _one_ pattern survive the final step. (Generally used with `inherit = "all"`.) | In particular, note that the default is `inherit = "core"`, so: * if you have extra env variables that you want to inherit from the parent process, use `inherit = "all"` and then specify `include_only` * if you have extra env variables where you want to hardcode the values, the default `inherit = "core"` will work fine, but then you need to specify `set` This configuration is not battle-tested, so we will probably still have to play with it a bit. `core/src/exec_env.rs` has the critical business logic as well as unit tests. Though if nothing else, previous to this change: ``` $ cargo run --bin codex -- debug seatbelt -- printenv OPENAI_API_KEY # ...prints OPENAI_API_KEY... ``` But after this change it does not print anything (as desired). One final thing to call out about this PR is that the `configure_command!` macro we use in `core/src/exec.rs` has to do some complex logic with respect to how it builds up the `env` for the process being spawned under Landlock/seccomp. Specifically, doing `cmd.env_clear()` followed by `cmd.envs(&$env_map)` (which is arguably the most intuitive way to do it) caused the Landlock unit tests to fail because the processes spawned by the unit tests started failing in unexpected ways! If we forgo `env_clear()` in favor of updating env vars one at a time, the tests still pass. The comment in the code talks about this a bit, and while I would like to investigate this more, I need to move on for the moment, but I do plan to come back to it to fully understand what is going on. For example, this suggests that we might not be able to spawn a C program that calls `env_clear()`, which would be...weird. We may still have to fiddle with our Landlock config if that is the case.
2025-05-22 09:51:19 -07:00
pub shell_environment_policy: ShellEnvironmentPolicy,
/// When `true`, `AgentReasoning` events emitted by the backend will be
/// suppressed from the frontend output. This can reduce visual noise when
/// users are only interested in the final agent responses.
pub hide_agent_reasoning: bool,
/// When set to `true`, `AgentReasoningRawContentEvent` events will be shown in the UI/output.
/// Defaults to `false`.
pub show_raw_agent_reasoning: bool,
/// User-provided instructions from AGENTS.md.
pub user_instructions: Option<String>,
/// Base instructions override.
pub base_instructions: Option<String>,
feat: configurable notifications in the Rust CLI (#793) With this change, you can specify a program that will be executed to get notified about events generated by Codex. The notification info will be packaged as a JSON object. The supported notification types are defined by the `UserNotification` enum introduced in this PR. Initially, it contains only one variant, `AgentTurnComplete`: ```rust pub(crate) enum UserNotification { #[serde(rename_all = "kebab-case")] AgentTurnComplete { turn_id: String, /// Messages that the user sent to the agent to initiate the turn. input_messages: Vec<String>, /// The last message sent by the assistant in the turn. last_assistant_message: Option<String>, }, } ``` This is intended to support the common case when a "turn" ends, which often means it is now your chance to give Codex further instructions. For example, I have the following in my `~/.codex/config.toml`: ```toml notify = ["python3", "/Users/mbolin/.codex/notify.py"] ``` I created my own custom notifier script that calls out to [terminal-notifier](https://github.com/julienXX/terminal-notifier) to show a desktop push notification on macOS. Contents of `notify.py`: ```python #!/usr/bin/env python3 import json import subprocess import sys def main() -> int: if len(sys.argv) != 2: print("Usage: notify.py <NOTIFICATION_JSON>") return 1 try: notification = json.loads(sys.argv[1]) except json.JSONDecodeError: return 1 match notification_type := notification.get("type"): case "agent-turn-complete": assistant_message = notification.get("last-assistant-message") if assistant_message: title = f"Codex: {assistant_message}" else: title = "Codex: Turn Complete!" input_messages = notification.get("input_messages", []) message = " ".join(input_messages) title += message case _: print(f"not sending a push notification for: {notification_type}") return 0 subprocess.check_output( [ "terminal-notifier", "-title", title, "-message", message, "-group", "codex", "-ignoreDnD", "-activate", "com.googlecode.iterm2", ] ) return 0 if __name__ == "__main__": sys.exit(main()) ``` For reference, here are related PRs that tried to add this functionality to the TypeScript version of the Codex CLI: * https://github.com/openai/codex/pull/160 * https://github.com/openai/codex/pull/498
2025-05-02 19:48:13 -07:00
/// Optional external notifier command. When set, Codex will spawn this
/// program after each completed *turn* (i.e. when the agent finishes
/// processing a user submission). The value must be the full command
/// broken into argv tokens **without** the trailing JSON argument - Codex
/// appends one extra argument containing a JSON payload describing the
/// event.
///
/// Example `~/.codex/config.toml` snippet:
///
/// ```toml
/// notify = ["notify-send", "Codex"]
/// ```
///
/// which will be invoked as:
///
/// ```shell
/// notify-send Codex '{"type":"agent-turn-complete","turn-id":"12345"}'
/// ```
///
/// If unset the feature is disabled.
pub notify: Option<Vec<String>>,
/// The directory that should be treated as the current working directory
/// for the session. All relative paths inside the business-logic layer are
/// resolved against this path.
pub cwd: PathBuf,
feat: support mcp_servers in config.toml (#829) This adds initial support for MCP servers in the style of Claude Desktop and Cursor. Note this PR is the bare minimum to get things working end to end: all configured MCP servers are launched every time Codex is run, there is no recovery for MCP servers that crash, etc. (Also, I took some shortcuts to change some fields of `Session` to be `pub(crate)`, which also means there are circular deps between `codex.rs` and `mcp_tool_call.rs`, but I will clean that up in a subsequent PR.) `codex-rs/README.md` is updated as part of this PR to explain how to use this feature. There is a bit of plumbing to route the new settings from `Config` to the business logic in `codex.rs`. The most significant chunks for new code are in `mcp_connection_manager.rs` (which defines the `McpConnectionManager` struct) and `mcp_tool_call.rs`, which is responsible for tool calls. This PR also introduces new `McpToolCallBegin` and `McpToolCallEnd` event types to the protocol, but does not add any handlers for them. (See https://github.com/openai/codex/pull/836 for initial usage.) To test, I added the following to my `~/.codex/config.toml`: ```toml # Local build of https://github.com/hideya/mcp-server-weather-js [mcp_servers.weather] command = "/Users/mbolin/code/mcp-server-weather-js/dist/index.js" args = [] ``` And then I ran the following: ``` codex-rs$ cargo run --bin codex exec 'what is the weather in san francisco' [2025-05-06T22:40:05] Task started: 1 [2025-05-06T22:40:18] Agent message: Here’s the latest National Weather Service forecast for San Francisco (downtown, near 37.77° N, 122.42° W): This Afternoon (Tue): • Sunny, high near 69 °F • West-southwest wind around 12 mph Tonight: • Partly cloudy, low around 52 °F • SW wind 7–10 mph ... ``` Note that Codex itself is not able to make network calls, so it would not normally be able to get live weather information like this. However, the weather MCP is [currently] not run under the Codex sandbox, so it is able to hit `api.weather.gov` and fetch current weather information. --- [//]: # (BEGIN SAPLING FOOTER) Stack created with [Sapling](https://sapling-scm.com). Best reviewed with [ReviewStack](https://reviewstack.dev/openai/codex/pull/829). * #836 * __->__ #829
2025-05-06 15:47:59 -07:00
/// Definition for MCP servers that Codex can reach out to for tool calls.
pub mcp_servers: HashMap<String, McpServerConfig>,
/// Combined provider map (defaults merged with user-defined overrides).
pub model_providers: HashMap<String, ModelProviderInfo>,
/// Maximum number of bytes to include from an AGENTS.md project doc file.
pub project_doc_max_bytes: usize,
/// Directory containing all Codex state (defaults to `~/.codex` but can be
/// overridden by the `CODEX_HOME` environment variable).
pub codex_home: PathBuf,
feat: record messages from user in ~/.codex/history.jsonl (#939) This is a large change to support a "history" feature like you would expect in a shell like Bash. History events are recorded in `$CODEX_HOME/history.jsonl`. Because it is a JSONL file, it is straightforward to append new entries (as opposed to the TypeScript file that uses `$CODEX_HOME/history.json`, so to be valid JSON, each new entry entails rewriting the entire file). Because it is possible for there to be multiple instances of Codex CLI writing to `history.jsonl` at once, we use advisory file locking when working with `history.jsonl` in `codex-rs/core/src/message_history.rs`. Because we believe history is a sufficiently useful feature, we enable it by default. Though to provide some safety, we set the file permissions of `history.jsonl` to be `o600` so that other users on the system cannot read the user's history. We do not yet support a default list of `SENSITIVE_PATTERNS` as the TypeScript CLI does: https://github.com/openai/codex/blob/3fdf9df1335ac9501e3fb0e61715359145711e8b/codex-cli/src/utils/storage/command-history.ts#L10-L17 We are going to take a more conservative approach to this list in the Rust CLI. For example, while `/\b[A-Za-z0-9-_]{20,}\b/` might exclude sensitive information like API tokens, it would also exclude valuable information such as references to Git commits. As noted in the updated documentation, users can opt-out of history by adding the following to `config.toml`: ```toml [history] persistence = "none" ``` Because `history.jsonl` could, in theory, be quite large, we take a[n arguably overly pedantic] approach in reading history entries into memory. Specifically, we start by telling the client the current number of entries in the history file (`history_entry_count`) as well as the inode (`history_log_id`) of `history.jsonl` (see the new fields on `SessionConfiguredEvent`). The client is responsible for keeping new entries in memory to create a "local history," but if the user hits up enough times to go "past" the end of local history, then the client should use the new `GetHistoryEntryRequest` in the protocol to fetch older entries. Specifically, it should pass the `history_log_id` it was given originally and work backwards from `history_entry_count`. (It should really fetch history in batches rather than one-at-a-time, but that is something we can improve upon in subsequent PRs.) The motivation behind this crazy scheme is that it is designed to defend against: * The `history.jsonl` being truncated during the session such that the index into the history is no longer consistent with what had been read up to that point. We do not yet have logic to enforce a `max_bytes` for `history.jsonl`, but once we do, we will aspire to implement it in a way that should result in a new inode for the file on most systems. * New items from concurrent Codex CLI sessions amending to the history. Because, in absence of truncation, `history.jsonl` is an append-only log, so long as the client reads backwards from `history_entry_count`, it should always get a consistent view of history. (That said, it will not be able to read _new_ commands from concurrent sessions, but perhaps we will introduce a `/` command to reload latest history or something down the road.) Admittedly, my testing of this feature thus far has been fairly light. I expect we will find bugs and introduce enhancements/fixes going forward.
2025-05-15 16:26:23 -07:00
/// Settings that govern if and what will be written to `~/.codex/history.jsonl`.
pub history: History,
/// Optional URI-based file opener. If set, citations to files in the model
/// output will be hyperlinked using the specified URI scheme.
pub file_opener: UriBasedFileOpener,
/// Path to the `codex-linux-sandbox` executable. This must be set if
/// [`crate::exec::SandboxType::LinuxSeccomp`] is used. Note that this
/// cannot be set in the config file: it must be set in code via
/// [`ConfigOverrides`].
///
/// When this program is invoked, arg0 will be set to `codex-linux-sandbox`.
pub codex_linux_sandbox_exe: Option<PathBuf>,
/// Value to use for `reasoning.effort` when making a request using the
/// Responses API.
pub model_reasoning_effort: Option<ReasoningEffort>,
/// If not "none", the value to use for `reasoning.summary` when making a
/// request using the Responses API.
pub model_reasoning_summary: ReasoningSummary,
feat(gpt5): add model_verbosity for GPT‑5 via Responses API (#2108) **Summary** - Adds `model_verbosity` config (values: low, medium, high). - Sends `text.verbosity` only for GPT‑5 family models via the Responses API. - Updates docs and adds serialization tests. **Motivation** - GPT‑5 introduces a verbosity control to steer output length/detail without pro mpt surgery. - Exposing it as a config knob keeps prompts stable and makes behavior explicit and repeatable. **Changes** - Config: - Added `Verbosity` enum (low|medium|high). - Added optional `model_verbosity` to `ConfigToml`, `Config`, and `ConfigProfi le`. - Request wiring: - Extended `ResponsesApiRequest` with optional `text` object. - Populates `text.verbosity` only when model family is `gpt-5`; omitted otherw ise. - Tests: - Verifies `text.verbosity` serializes when set and is omitted when not set. - Docs: - Added “GPT‑5 Verbosity” section in `codex-rs/README.md`. - Added `model_verbosity` section to `codex-rs/config.md`. **Usage** - In `~/.codex/config.toml`: - `model = "gpt-5"` - `model_verbosity = "low"` (or `"medium"` default, `"high"`) - CLI override example: - `codex -c model="gpt-5" -c model_verbosity="high"` **API Impact** - Requests to GPT‑5 via Responses API include: `text: { verbosity: "low|medium|h igh" }` when configured. - For legacy models or Chat Completions providers, `text` is omitted. **Backward Compatibility** - Default behavior unchanged when `model_verbosity` is not set (server default “ medium”). **Testing** - Added unit tests for serialization/omission of `text.verbosity`. - Ran `cargo fmt` and `cargo test --all-features` (all green). **Docs** - `README.md`: new “GPT‑5 Verbosity” note under Config with example. - `config.md`: new `model_verbosity` section. **Out of Scope** - No changes to temperature/top_p or other GPT‑5 parameters. - No changes to Chat Completions wiring. **Risks / Notes** - If OpenAI changes the wire shape for verbosity, we may need to update `Respons esApiRequest`. - Behavior gated to `gpt-5` model family to avoid unexpected effects elsewhere. **Checklist** - [x] Code gated to GPT‑5 family only - [x] Docs updated (`README.md`, `config.md`) - [x] Tests added and passing - [x] Formatting applied Release note: Add `model_verbosity` config to control GPT‑5 output verbosity via the Responses API (low|medium|high).
2025-08-22 17:12:10 +01:00
/// Optional verbosity control for GPT-5 models (Responses API `text.verbosity`).
pub model_verbosity: Option<Verbosity>,
/// Base URL for requests to ChatGPT (as opposed to the OpenAI API).
pub chatgpt_base_url: String,
/// Experimental rollout resume path (absolute path to .jsonl; undocumented).
pub experimental_resume: Option<PathBuf>,
/// Include an experimental plan tool that the model can use to update its current plan and status of each step.
pub include_plan_tool: bool,
/// Include the `apply_patch` tool for models that benefit from invoking
/// file edits as a structured tool call. When unset, this falls back to the
/// model family's default preference.
pub include_apply_patch_tool: bool,
pub tools_web_search_request: bool,
pub use_experimental_streamable_shell_tool: bool,
/// If set to `true`, used only the experimental unified exec tool.
pub use_experimental_unified_exec_tool: bool,
/// Include the `view_image` tool that lets the agent attach a local image path to context.
pub include_view_image_tool: bool,
/// The active profile name used to derive this `Config` (if any).
pub active_profile: Option<String>,
/// When true, disables burst-paste detection for typed input entirely.
/// All characters are inserted as they are received, and no buffering
/// or placeholder replacement will occur for fast keypress bursts.
pub disable_paste_burst: bool,
feat: record messages from user in ~/.codex/history.jsonl (#939) This is a large change to support a "history" feature like you would expect in a shell like Bash. History events are recorded in `$CODEX_HOME/history.jsonl`. Because it is a JSONL file, it is straightforward to append new entries (as opposed to the TypeScript file that uses `$CODEX_HOME/history.json`, so to be valid JSON, each new entry entails rewriting the entire file). Because it is possible for there to be multiple instances of Codex CLI writing to `history.jsonl` at once, we use advisory file locking when working with `history.jsonl` in `codex-rs/core/src/message_history.rs`. Because we believe history is a sufficiently useful feature, we enable it by default. Though to provide some safety, we set the file permissions of `history.jsonl` to be `o600` so that other users on the system cannot read the user's history. We do not yet support a default list of `SENSITIVE_PATTERNS` as the TypeScript CLI does: https://github.com/openai/codex/blob/3fdf9df1335ac9501e3fb0e61715359145711e8b/codex-cli/src/utils/storage/command-history.ts#L10-L17 We are going to take a more conservative approach to this list in the Rust CLI. For example, while `/\b[A-Za-z0-9-_]{20,}\b/` might exclude sensitive information like API tokens, it would also exclude valuable information such as references to Git commits. As noted in the updated documentation, users can opt-out of history by adding the following to `config.toml`: ```toml [history] persistence = "none" ``` Because `history.jsonl` could, in theory, be quite large, we take a[n arguably overly pedantic] approach in reading history entries into memory. Specifically, we start by telling the client the current number of entries in the history file (`history_entry_count`) as well as the inode (`history_log_id`) of `history.jsonl` (see the new fields on `SessionConfiguredEvent`). The client is responsible for keeping new entries in memory to create a "local history," but if the user hits up enough times to go "past" the end of local history, then the client should use the new `GetHistoryEntryRequest` in the protocol to fetch older entries. Specifically, it should pass the `history_log_id` it was given originally and work backwards from `history_entry_count`. (It should really fetch history in batches rather than one-at-a-time, but that is something we can improve upon in subsequent PRs.) The motivation behind this crazy scheme is that it is designed to defend against: * The `history.jsonl` being truncated during the session such that the index into the history is no longer consistent with what had been read up to that point. We do not yet have logic to enforce a `max_bytes` for `history.jsonl`, but once we do, we will aspire to implement it in a way that should result in a new inode for the file on most systems. * New items from concurrent Codex CLI sessions amending to the history. Because, in absence of truncation, `history.jsonl` is an append-only log, so long as the client reads backwards from `history_entry_count`, it should always get a consistent view of history. (That said, it will not be able to read _new_ commands from concurrent sessions, but perhaps we will introduce a `/` command to reload latest history or something down the road.) Admittedly, my testing of this feature thus far has been fairly light. I expect we will find bugs and introduce enhancements/fixes going forward.
2025-05-15 16:26:23 -07:00
}
feat: add support for -c/--config to override individual config items (#1137) This PR introduces support for `-c`/`--config` so users can override individual config values on the command line using `--config name=value`. Example: ``` codex --config model=o4-mini ``` Making it possible to set arbitrary config values on the command line results in a more flexible configuration scheme and makes it easier to provide single-line examples that can be copy-pasted from documentation. Effectively, it means there are four levels of configuration for some values: - Default value (e.g., `model` currently defaults to `o4-mini`) - Value in `config.toml` (e.g., user could override the default to be `model = "o3"` in their `config.toml`) - Specifying `-c` or `--config` to override `model` (e.g., user can include `-c model=o3` in their list of args to Codex) - If available, a config-specific flag can be used, which takes precedence over `-c` (e.g., user can specify `--model o3` in their list of args to Codex) Now that it is possible to specify anything that could be configured in `config.toml` on the command line using `-c`, we do not need to have a custom flag for every possible config option (which can clutter the output of `--help`). To that end, as part of this PR, we drop support for the `--disable-response-storage` flag, as users can now specify `-c disable_response_storage=true` to get the equivalent functionality. Under the hood, this works by loading the `config.toml` into a `toml::Value`. Then for each `key=value`, we create a small synthetic TOML file with `value` so that we can run the TOML parser to get the equivalent `toml::Value`. We then parse `key` to determine the point in the original `toml::Value` to do the insert/replace. Once all of the overrides from `-c` args have been applied, the `toml::Value` is deserialized into a `ConfigToml` and then the `ConfigOverrides` are applied, as before.
2025-05-27 23:11:44 -07:00
impl Config {
/// Load configuration with *generic* CLI overrides (`-c key=value`) applied
/// **in between** the values parsed from `config.toml` and the
/// strongly-typed overrides specified via [`ConfigOverrides`].
///
/// The precedence order is therefore: `config.toml` < `-c` overrides <
/// `ConfigOverrides`.
pub fn load_with_cli_overrides(
cli_overrides: Vec<(String, TomlValue)>,
overrides: ConfigOverrides,
) -> std::io::Result<Self> {
// Resolve the directory that stores Codex state (e.g. ~/.codex or the
// value of $CODEX_HOME) so we can embed it into the resulting
// `Config` instance.
let codex_home = find_codex_home()?;
// Step 1: parse `config.toml` into a generic JSON value.
let mut root_value = load_config_as_toml(&codex_home)?;
// Step 2: apply the `-c` overrides.
for (path, value) in cli_overrides.into_iter() {
apply_toml_override(&mut root_value, &path, value);
}
// Step 3: deserialize into `ConfigToml` so that Serde can enforce the
// correct types.
let cfg: ConfigToml = root_value.try_into().map_err(|e| {
tracing::error!("Failed to deserialize overridden config: {e}");
std::io::Error::new(std::io::ErrorKind::InvalidData, e)
})?;
// Step 4: merge with the strongly-typed overrides.
Self::load_from_base_config_with_overrides(cfg, overrides, codex_home)
}
}
pub fn load_config_as_toml_with_cli_overrides(
codex_home: &Path,
cli_overrides: Vec<(String, TomlValue)>,
) -> std::io::Result<ConfigToml> {
let mut root_value = load_config_as_toml(codex_home)?;
for (path, value) in cli_overrides.into_iter() {
apply_toml_override(&mut root_value, &path, value);
}
let cfg: ConfigToml = root_value.try_into().map_err(|e| {
tracing::error!("Failed to deserialize overridden config: {e}");
std::io::Error::new(std::io::ErrorKind::InvalidData, e)
})?;
Ok(cfg)
}
feat: add support for -c/--config to override individual config items (#1137) This PR introduces support for `-c`/`--config` so users can override individual config values on the command line using `--config name=value`. Example: ``` codex --config model=o4-mini ``` Making it possible to set arbitrary config values on the command line results in a more flexible configuration scheme and makes it easier to provide single-line examples that can be copy-pasted from documentation. Effectively, it means there are four levels of configuration for some values: - Default value (e.g., `model` currently defaults to `o4-mini`) - Value in `config.toml` (e.g., user could override the default to be `model = "o3"` in their `config.toml`) - Specifying `-c` or `--config` to override `model` (e.g., user can include `-c model=o3` in their list of args to Codex) - If available, a config-specific flag can be used, which takes precedence over `-c` (e.g., user can specify `--model o3` in their list of args to Codex) Now that it is possible to specify anything that could be configured in `config.toml` on the command line using `-c`, we do not need to have a custom flag for every possible config option (which can clutter the output of `--help`). To that end, as part of this PR, we drop support for the `--disable-response-storage` flag, as users can now specify `-c disable_response_storage=true` to get the equivalent functionality. Under the hood, this works by loading the `config.toml` into a `toml::Value`. Then for each `key=value`, we create a small synthetic TOML file with `value` so that we can run the TOML parser to get the equivalent `toml::Value`. We then parse `key` to determine the point in the original `toml::Value` to do the insert/replace. Once all of the overrides from `-c` args have been applied, the `toml::Value` is deserialized into a `ConfigToml` and then the `ConfigOverrides` are applied, as before.
2025-05-27 23:11:44 -07:00
/// Read `CODEX_HOME/config.toml` and return it as a generic TOML value. Returns
/// an empty TOML table when the file does not exist.
pub fn load_config_as_toml(codex_home: &Path) -> std::io::Result<TomlValue> {
let config_path = codex_home.join(CONFIG_TOML_FILE);
feat: add support for -c/--config to override individual config items (#1137) This PR introduces support for `-c`/`--config` so users can override individual config values on the command line using `--config name=value`. Example: ``` codex --config model=o4-mini ``` Making it possible to set arbitrary config values on the command line results in a more flexible configuration scheme and makes it easier to provide single-line examples that can be copy-pasted from documentation. Effectively, it means there are four levels of configuration for some values: - Default value (e.g., `model` currently defaults to `o4-mini`) - Value in `config.toml` (e.g., user could override the default to be `model = "o3"` in their `config.toml`) - Specifying `-c` or `--config` to override `model` (e.g., user can include `-c model=o3` in their list of args to Codex) - If available, a config-specific flag can be used, which takes precedence over `-c` (e.g., user can specify `--model o3` in their list of args to Codex) Now that it is possible to specify anything that could be configured in `config.toml` on the command line using `-c`, we do not need to have a custom flag for every possible config option (which can clutter the output of `--help`). To that end, as part of this PR, we drop support for the `--disable-response-storage` flag, as users can now specify `-c disable_response_storage=true` to get the equivalent functionality. Under the hood, this works by loading the `config.toml` into a `toml::Value`. Then for each `key=value`, we create a small synthetic TOML file with `value` so that we can run the TOML parser to get the equivalent `toml::Value`. We then parse `key` to determine the point in the original `toml::Value` to do the insert/replace. Once all of the overrides from `-c` args have been applied, the `toml::Value` is deserialized into a `ConfigToml` and then the `ConfigOverrides` are applied, as before.
2025-05-27 23:11:44 -07:00
match std::fs::read_to_string(&config_path) {
Ok(contents) => match toml::from_str::<TomlValue>(&contents) {
Ok(val) => Ok(val),
Err(e) => {
tracing::error!("Failed to parse config.toml: {e}");
Err(std::io::Error::new(std::io::ErrorKind::InvalidData, e))
}
},
Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
tracing::info!("config.toml not found, using defaults");
Ok(TomlValue::Table(Default::default()))
}
Err(e) => {
tracing::error!("Failed to read config.toml: {e}");
Err(e)
}
}
}
fn set_project_trusted_inner(doc: &mut DocumentMut, project_path: &Path) -> anyhow::Result<()> {
// Ensure we render a human-friendly structure:
//
// [projects]
// [projects."/path/to/project"]
// trust_level = "trusted"
//
// rather than inline tables like:
//
// [projects]
// "/path/to/project" = { trust_level = "trusted" }
let project_key = project_path.to_string_lossy().to_string();
// Ensure top-level `projects` exists as a non-inline, explicit table. If it
// exists but was previously represented as a non-table (e.g., inline),
// replace it with an explicit table.
{
let root = doc.as_table_mut();
// If `projects` exists but isn't a standard table (e.g., it's an inline table),
// convert it to an explicit table while preserving existing entries.
let existing_projects = root.get("projects").cloned();
if existing_projects.as_ref().is_none_or(|i| !i.is_table()) {
let mut projects_tbl = toml_edit::Table::new();
projects_tbl.set_implicit(true);
// If there was an existing inline table, migrate its entries to explicit tables.
if let Some(inline_tbl) = existing_projects.as_ref().and_then(|i| i.as_inline_table()) {
for (k, v) in inline_tbl.iter() {
if let Some(inner_tbl) = v.as_inline_table() {
let new_tbl = inner_tbl.clone().into_table();
projects_tbl.insert(k, toml_edit::Item::Table(new_tbl));
}
}
}
root.insert("projects", toml_edit::Item::Table(projects_tbl));
}
}
let Some(projects_tbl) = doc["projects"].as_table_mut() else {
return Err(anyhow::anyhow!(
"projects table missing after initialization"
));
};
// Ensure the per-project entry is its own explicit table. If it exists but
// is not a table (e.g., an inline table), replace it with an explicit table.
let needs_proj_table = !projects_tbl.contains_key(project_key.as_str())
|| projects_tbl
.get(project_key.as_str())
.and_then(|i| i.as_table())
.is_none();
if needs_proj_table {
projects_tbl.insert(project_key.as_str(), toml_edit::table());
}
let Some(proj_tbl) = projects_tbl
.get_mut(project_key.as_str())
.and_then(|i| i.as_table_mut())
else {
return Err(anyhow::anyhow!("project table missing for {}", project_key));
};
proj_tbl.set_implicit(false);
proj_tbl["trust_level"] = toml_edit::value("trusted");
Ok(())
}
/// Patch `CODEX_HOME/config.toml` project state.
/// Use with caution.
pub fn set_project_trusted(codex_home: &Path, project_path: &Path) -> anyhow::Result<()> {
let config_path = codex_home.join(CONFIG_TOML_FILE);
// Parse existing config if present; otherwise start a new document.
let mut doc = match std::fs::read_to_string(config_path.clone()) {
Ok(s) => s.parse::<DocumentMut>()?,
Err(e) if e.kind() == std::io::ErrorKind::NotFound => DocumentMut::new(),
Err(e) => return Err(e.into()),
};
set_project_trusted_inner(&mut doc, project_path)?;
// ensure codex_home exists
std::fs::create_dir_all(codex_home)?;
// create a tmp_file
let tmp_file = NamedTempFile::new_in(codex_home)?;
std::fs::write(tmp_file.path(), doc.to_string())?;
// atomically move the tmp file into config.toml
tmp_file.persist(config_path)?;
Ok(())
}
fn ensure_profile_table<'a>(
doc: &'a mut DocumentMut,
profile_name: &str,
) -> anyhow::Result<&'a mut toml_edit::Table> {
let mut created_profiles_table = false;
{
let root = doc.as_table_mut();
let needs_table = !root.contains_key("profiles")
|| root
.get("profiles")
.and_then(|item| item.as_table())
.is_none();
if needs_table {
root.insert("profiles", toml_edit::table());
created_profiles_table = true;
}
}
let Some(profiles_table) = doc["profiles"].as_table_mut() else {
return Err(anyhow::anyhow!(
"profiles table missing after initialization"
));
};
if created_profiles_table {
profiles_table.set_implicit(true);
}
let needs_profile_table = !profiles_table.contains_key(profile_name)
|| profiles_table
.get(profile_name)
.and_then(|item| item.as_table())
.is_none();
if needs_profile_table {
profiles_table.insert(profile_name, toml_edit::table());
}
let Some(profile_table) = profiles_table
.get_mut(profile_name)
.and_then(|item| item.as_table_mut())
else {
return Err(anyhow::anyhow!(format!(
"profile table missing for {profile_name}"
)));
};
profile_table.set_implicit(false);
Ok(profile_table)
}
// TODO(jif) refactor config persistence.
pub async fn persist_model_selection(
codex_home: &Path,
active_profile: Option<&str>,
model: &str,
effort: Option<ReasoningEffort>,
) -> anyhow::Result<()> {
let config_path = codex_home.join(CONFIG_TOML_FILE);
let serialized = match tokio::fs::read_to_string(&config_path).await {
Ok(contents) => contents,
Err(err) if err.kind() == std::io::ErrorKind::NotFound => String::new(),
Err(err) => return Err(err.into()),
};
let mut doc = if serialized.is_empty() {
DocumentMut::new()
} else {
serialized.parse::<DocumentMut>()?
};
if let Some(profile_name) = active_profile {
let profile_table = ensure_profile_table(&mut doc, profile_name)?;
profile_table["model"] = toml_edit::value(model);
match effort {
Some(effort) => {
profile_table["model_reasoning_effort"] = toml_edit::value(effort.to_string());
}
None => {
profile_table.remove("model_reasoning_effort");
}
}
} else {
let table = doc.as_table_mut();
table["model"] = toml_edit::value(model);
match effort {
Some(effort) => {
table["model_reasoning_effort"] = toml_edit::value(effort.to_string());
}
None => {
table.remove("model_reasoning_effort");
}
}
}
// TODO(jif) refactor the home creation
tokio::fs::create_dir_all(codex_home)
.await
.with_context(|| {
format!(
"failed to create Codex home directory at {}",
codex_home.display()
)
})?;
tokio::fs::write(&config_path, doc.to_string())
.await
.with_context(|| format!("failed to persist config.toml at {}", config_path.display()))?;
Ok(())
}
feat: add support for -c/--config to override individual config items (#1137) This PR introduces support for `-c`/`--config` so users can override individual config values on the command line using `--config name=value`. Example: ``` codex --config model=o4-mini ``` Making it possible to set arbitrary config values on the command line results in a more flexible configuration scheme and makes it easier to provide single-line examples that can be copy-pasted from documentation. Effectively, it means there are four levels of configuration for some values: - Default value (e.g., `model` currently defaults to `o4-mini`) - Value in `config.toml` (e.g., user could override the default to be `model = "o3"` in their `config.toml`) - Specifying `-c` or `--config` to override `model` (e.g., user can include `-c model=o3` in their list of args to Codex) - If available, a config-specific flag can be used, which takes precedence over `-c` (e.g., user can specify `--model o3` in their list of args to Codex) Now that it is possible to specify anything that could be configured in `config.toml` on the command line using `-c`, we do not need to have a custom flag for every possible config option (which can clutter the output of `--help`). To that end, as part of this PR, we drop support for the `--disable-response-storage` flag, as users can now specify `-c disable_response_storage=true` to get the equivalent functionality. Under the hood, this works by loading the `config.toml` into a `toml::Value`. Then for each `key=value`, we create a small synthetic TOML file with `value` so that we can run the TOML parser to get the equivalent `toml::Value`. We then parse `key` to determine the point in the original `toml::Value` to do the insert/replace. Once all of the overrides from `-c` args have been applied, the `toml::Value` is deserialized into a `ConfigToml` and then the `ConfigOverrides` are applied, as before.
2025-05-27 23:11:44 -07:00
/// Apply a single dotted-path override onto a TOML value.
fn apply_toml_override(root: &mut TomlValue, path: &str, value: TomlValue) {
use toml::value::Table;
let segments: Vec<&str> = path.split('.').collect();
let mut current = root;
for (idx, segment) in segments.iter().enumerate() {
let is_last = idx == segments.len() - 1;
if is_last {
match current {
TomlValue::Table(table) => {
table.insert(segment.to_string(), value);
}
_ => {
let mut table = Table::new();
table.insert(segment.to_string(), value);
*current = TomlValue::Table(table);
}
}
return;
}
// Traverse or create intermediate object.
match current {
TomlValue::Table(table) => {
current = table
.entry(segment.to_string())
.or_insert_with(|| TomlValue::Table(Table::new()));
}
_ => {
*current = TomlValue::Table(Table::new());
if let TomlValue::Table(tbl) = current {
current = tbl
.entry(segment.to_string())
.or_insert_with(|| TomlValue::Table(Table::new()));
}
}
}
}
}
fix: overhaul SandboxPolicy and config loading in Rust (#732) Previous to this PR, `SandboxPolicy` was a bit difficult to work with: https://github.com/openai/codex/blob/237f8a11e11fdcc793a09e787e48215676d9b95b/codex-rs/core/src/protocol.rs#L98-L108 Specifically: * It was an `enum` and therefore options were mutually exclusive as opposed to additive. * It defined things in terms of what the agent _could not_ do as opposed to what they _could_ do. This made things hard to support because we would prefer to build up a sandbox config by starting with something extremely restrictive and only granting permissions for things the user as explicitly allowed. This PR changes things substantially by redefining the policy in terms of two concepts: * A `SandboxPermission` enum that defines permissions that can be granted to the agent/sandbox. * A `SandboxPolicy` that internally stores a `Vec<SandboxPermission>`, but externally exposes a simpler API that can be used to configure Seatbelt/Landlock. Previous to this PR, we supported a `--sandbox` flag that effectively mapped to an enum value in `SandboxPolicy`. Though now that `SandboxPolicy` is a wrapper around `Vec<SandboxPermission>`, the single `--sandbox` flag no longer makes sense. While I could have turned it into a flag that the user can specify multiple times, I think the current values to use with such a flag are long and potentially messy, so for the moment, I have dropped support for `--sandbox` altogether and we can bring it back once we have figured out the naming thing. Since `--sandbox` is gone, users now have to specify `--full-auto` to get a sandbox that allows writes in `cwd`. Admittedly, there is no clean way to specify the equivalent of `--full-auto` in your `config.toml` right now, so we will have to revisit that, as well. Because `Config` presents a `SandboxPolicy` field and `SandboxPolicy` changed considerably, I had to overhaul how config loading works, as well. There are now two distinct concepts, `ConfigToml` and `Config`: * `ConfigToml` is the deserialization of `~/.codex/config.toml`. As one might expect, every field is `Optional` and it is `#[derive(Deserialize, Default)]`. Consistent use of `Optional` makes it clear what the user has specified explicitly. * `Config` is the "normalized config" and is produced by merging `ConfigToml` with `ConfigOverrides`. Where `ConfigToml` contains a raw `Option<Vec<SandboxPermission>>`, `Config` presents only the final `SandboxPolicy`. The changes to `core/src/exec.rs` and `core/src/linux.rs` merit extra special attention to ensure we are faithfully mapping the `SandboxPolicy` to the Seatbelt and Landlock configs, respectively. Also, take note that `core/src/seatbelt_readonly_policy.sbpl` has been renamed to `codex-rs/core/src/seatbelt_base_policy.sbpl` and that `(allow file-read*)` has been removed from the `.sbpl` file as now this is added to the policy in `core/src/exec.rs` when `sandbox_policy.has_full_disk_read_access()` is `true`.
2025-04-29 15:01:16 -07:00
/// Base config deserialized from ~/.codex/config.toml.
#[derive(Deserialize, Debug, Clone, Default, PartialEq)]
fix: overhaul SandboxPolicy and config loading in Rust (#732) Previous to this PR, `SandboxPolicy` was a bit difficult to work with: https://github.com/openai/codex/blob/237f8a11e11fdcc793a09e787e48215676d9b95b/codex-rs/core/src/protocol.rs#L98-L108 Specifically: * It was an `enum` and therefore options were mutually exclusive as opposed to additive. * It defined things in terms of what the agent _could not_ do as opposed to what they _could_ do. This made things hard to support because we would prefer to build up a sandbox config by starting with something extremely restrictive and only granting permissions for things the user as explicitly allowed. This PR changes things substantially by redefining the policy in terms of two concepts: * A `SandboxPermission` enum that defines permissions that can be granted to the agent/sandbox. * A `SandboxPolicy` that internally stores a `Vec<SandboxPermission>`, but externally exposes a simpler API that can be used to configure Seatbelt/Landlock. Previous to this PR, we supported a `--sandbox` flag that effectively mapped to an enum value in `SandboxPolicy`. Though now that `SandboxPolicy` is a wrapper around `Vec<SandboxPermission>`, the single `--sandbox` flag no longer makes sense. While I could have turned it into a flag that the user can specify multiple times, I think the current values to use with such a flag are long and potentially messy, so for the moment, I have dropped support for `--sandbox` altogether and we can bring it back once we have figured out the naming thing. Since `--sandbox` is gone, users now have to specify `--full-auto` to get a sandbox that allows writes in `cwd`. Admittedly, there is no clean way to specify the equivalent of `--full-auto` in your `config.toml` right now, so we will have to revisit that, as well. Because `Config` presents a `SandboxPolicy` field and `SandboxPolicy` changed considerably, I had to overhaul how config loading works, as well. There are now two distinct concepts, `ConfigToml` and `Config`: * `ConfigToml` is the deserialization of `~/.codex/config.toml`. As one might expect, every field is `Optional` and it is `#[derive(Deserialize, Default)]`. Consistent use of `Optional` makes it clear what the user has specified explicitly. * `Config` is the "normalized config" and is produced by merging `ConfigToml` with `ConfigOverrides`. Where `ConfigToml` contains a raw `Option<Vec<SandboxPermission>>`, `Config` presents only the final `SandboxPolicy`. The changes to `core/src/exec.rs` and `core/src/linux.rs` merit extra special attention to ensure we are faithfully mapping the `SandboxPolicy` to the Seatbelt and Landlock configs, respectively. Also, take note that `core/src/seatbelt_readonly_policy.sbpl` has been renamed to `codex-rs/core/src/seatbelt_base_policy.sbpl` and that `(allow file-read*)` has been removed from the `.sbpl` file as now this is added to the policy in `core/src/exec.rs` when `sandbox_policy.has_full_disk_read_access()` is `true`.
2025-04-29 15:01:16 -07:00
pub struct ConfigToml {
/// Optional override of model selection.
pub model: Option<String>,
Review Mode (Core) (#3401) ## 📝 Review Mode -- Core This PR introduces the Core implementation for Review mode: - New op `Op::Review { prompt: String }:` spawns a child review task with isolated context, a review‑specific system prompt, and a `Config.review_model`. - `EnteredReviewMode`: emitted when the child review session starts. Every event from this point onwards reflects the review session. - `ExitedReviewMode(Option<ReviewOutputEvent>)`: emitted when the review finishes or is interrupted, with optional structured findings: ```json { "findings": [ { "title": "<≤ 80 chars, imperative>", "body": "<valid Markdown explaining *why* this is a problem; cite files/lines/functions>", "confidence_score": <float 0.0-1.0>, "priority": <int 0-3>, "code_location": { "absolute_file_path": "<file path>", "line_range": {"start": <int>, "end": <int>} } } ], "overall_correctness": "patch is correct" | "patch is incorrect", "overall_explanation": "<1-3 sentence explanation justifying the overall_correctness verdict>", "overall_confidence_score": <float 0.0-1.0> } ``` ## Questions ### Why separate out its own message history? We want the review thread to match the training of our review models as much as possible -- that means using a custom prompt, removing user instructions, and starting a clean chat history. We also want to make sure the review thread doesn't leak into the parent thread. ### Why do this as a mode, vs. sub-agents? 1. We want review to be a synchronous task, so it's fine for now to do a bespoke implementation. 2. We're still unclear about the final structure for sub-agents. We'd prefer to land this quickly and then refactor into sub-agents without rushing that implementation.
2025-09-12 16:25:10 -07:00
/// Review model override used by the `/review` feature.
pub review_model: Option<String>,
fix: overhaul SandboxPolicy and config loading in Rust (#732) Previous to this PR, `SandboxPolicy` was a bit difficult to work with: https://github.com/openai/codex/blob/237f8a11e11fdcc793a09e787e48215676d9b95b/codex-rs/core/src/protocol.rs#L98-L108 Specifically: * It was an `enum` and therefore options were mutually exclusive as opposed to additive. * It defined things in terms of what the agent _could not_ do as opposed to what they _could_ do. This made things hard to support because we would prefer to build up a sandbox config by starting with something extremely restrictive and only granting permissions for things the user as explicitly allowed. This PR changes things substantially by redefining the policy in terms of two concepts: * A `SandboxPermission` enum that defines permissions that can be granted to the agent/sandbox. * A `SandboxPolicy` that internally stores a `Vec<SandboxPermission>`, but externally exposes a simpler API that can be used to configure Seatbelt/Landlock. Previous to this PR, we supported a `--sandbox` flag that effectively mapped to an enum value in `SandboxPolicy`. Though now that `SandboxPolicy` is a wrapper around `Vec<SandboxPermission>`, the single `--sandbox` flag no longer makes sense. While I could have turned it into a flag that the user can specify multiple times, I think the current values to use with such a flag are long and potentially messy, so for the moment, I have dropped support for `--sandbox` altogether and we can bring it back once we have figured out the naming thing. Since `--sandbox` is gone, users now have to specify `--full-auto` to get a sandbox that allows writes in `cwd`. Admittedly, there is no clean way to specify the equivalent of `--full-auto` in your `config.toml` right now, so we will have to revisit that, as well. Because `Config` presents a `SandboxPolicy` field and `SandboxPolicy` changed considerably, I had to overhaul how config loading works, as well. There are now two distinct concepts, `ConfigToml` and `Config`: * `ConfigToml` is the deserialization of `~/.codex/config.toml`. As one might expect, every field is `Optional` and it is `#[derive(Deserialize, Default)]`. Consistent use of `Optional` makes it clear what the user has specified explicitly. * `Config` is the "normalized config" and is produced by merging `ConfigToml` with `ConfigOverrides`. Where `ConfigToml` contains a raw `Option<Vec<SandboxPermission>>`, `Config` presents only the final `SandboxPolicy`. The changes to `core/src/exec.rs` and `core/src/linux.rs` merit extra special attention to ensure we are faithfully mapping the `SandboxPolicy` to the Seatbelt and Landlock configs, respectively. Also, take note that `core/src/seatbelt_readonly_policy.sbpl` has been renamed to `codex-rs/core/src/seatbelt_base_policy.sbpl` and that `(allow file-read*)` has been removed from the `.sbpl` file as now this is added to the policy in `core/src/exec.rs` when `sandbox_policy.has_full_disk_read_access()` is `true`.
2025-04-29 15:01:16 -07:00
/// Provider to use from the model_providers map.
pub model_provider: Option<String>,
feat: show number of tokens remaining in UI (#1388) When using the OpenAI Responses API, we now record the `usage` field for a `"response.completed"` event, which includes metrics about the number of tokens consumed. We also introduce `openai_model_info.rs`, which includes current data about the most common OpenAI models available via the API (specifically `context_window` and `max_output_tokens`). If Codex does not recognize the model, you can set `model_context_window` and `model_max_output_tokens` explicitly in `config.toml`. When then introduce a new event type to `protocol.rs`, `TokenCount`, which includes the `TokenUsage` for the most recent turn. Finally, we update the TUI to record the running sum of tokens used so the percentage of available context window remaining can be reported via the placeholder text for the composer: ![Screenshot 2025-06-25 at 11 20 55 PM](https://github.com/user-attachments/assets/6fd6982f-7247-4f14-84b2-2e600cb1fd49) We could certainly get much fancier with this (such as reporting the estimated cost of the conversation), but for now, we are just trying to achieve feature parity with the TypeScript CLI. Though arguably this improves upon the TypeScript CLI, as the TypeScript CLI uses heuristics to estimate the number of tokens used rather than using the `usage` information directly: https://github.com/openai/codex/blob/296996d74e345b1b05d8c3451a06ace21c5ada96/codex-cli/src/utils/approximate-tokens-used.ts#L3-L16 Fixes https://github.com/openai/codex/issues/1242
2025-06-25 23:31:11 -07:00
/// Size of the context window for the model, in tokens.
pub model_context_window: Option<u64>,
/// Maximum number of output tokens.
pub model_max_output_tokens: Option<u64>,
/// Token usage threshold triggering auto-compaction of conversation history.
pub model_auto_compact_token_limit: Option<i64>,
fix: overhaul SandboxPolicy and config loading in Rust (#732) Previous to this PR, `SandboxPolicy` was a bit difficult to work with: https://github.com/openai/codex/blob/237f8a11e11fdcc793a09e787e48215676d9b95b/codex-rs/core/src/protocol.rs#L98-L108 Specifically: * It was an `enum` and therefore options were mutually exclusive as opposed to additive. * It defined things in terms of what the agent _could not_ do as opposed to what they _could_ do. This made things hard to support because we would prefer to build up a sandbox config by starting with something extremely restrictive and only granting permissions for things the user as explicitly allowed. This PR changes things substantially by redefining the policy in terms of two concepts: * A `SandboxPermission` enum that defines permissions that can be granted to the agent/sandbox. * A `SandboxPolicy` that internally stores a `Vec<SandboxPermission>`, but externally exposes a simpler API that can be used to configure Seatbelt/Landlock. Previous to this PR, we supported a `--sandbox` flag that effectively mapped to an enum value in `SandboxPolicy`. Though now that `SandboxPolicy` is a wrapper around `Vec<SandboxPermission>`, the single `--sandbox` flag no longer makes sense. While I could have turned it into a flag that the user can specify multiple times, I think the current values to use with such a flag are long and potentially messy, so for the moment, I have dropped support for `--sandbox` altogether and we can bring it back once we have figured out the naming thing. Since `--sandbox` is gone, users now have to specify `--full-auto` to get a sandbox that allows writes in `cwd`. Admittedly, there is no clean way to specify the equivalent of `--full-auto` in your `config.toml` right now, so we will have to revisit that, as well. Because `Config` presents a `SandboxPolicy` field and `SandboxPolicy` changed considerably, I had to overhaul how config loading works, as well. There are now two distinct concepts, `ConfigToml` and `Config`: * `ConfigToml` is the deserialization of `~/.codex/config.toml`. As one might expect, every field is `Optional` and it is `#[derive(Deserialize, Default)]`. Consistent use of `Optional` makes it clear what the user has specified explicitly. * `Config` is the "normalized config" and is produced by merging `ConfigToml` with `ConfigOverrides`. Where `ConfigToml` contains a raw `Option<Vec<SandboxPermission>>`, `Config` presents only the final `SandboxPolicy`. The changes to `core/src/exec.rs` and `core/src/linux.rs` merit extra special attention to ensure we are faithfully mapping the `SandboxPolicy` to the Seatbelt and Landlock configs, respectively. Also, take note that `core/src/seatbelt_readonly_policy.sbpl` has been renamed to `codex-rs/core/src/seatbelt_base_policy.sbpl` and that `(allow file-read*)` has been removed from the `.sbpl` file as now this is added to the policy in `core/src/exec.rs` when `sandbox_policy.has_full_disk_read_access()` is `true`.
2025-04-29 15:01:16 -07:00
/// Default approval policy for executing commands.
pub approval_policy: Option<AskForApproval>,
feat: introduce support for shell_environment_policy in config.toml (#1061) To date, when handling `shell` and `local_shell` tool calls, we were spawning new processes using the environment inherited from the Codex process itself. This means that the sensitive `OPENAI_API_KEY` that Codex needs to talk to OpenAI models was made available to everything run by `shell` and `local_shell`. While there are cases where that might be useful, it does not seem like a good default. This PR introduces a complex `shell_environment_policy` config option to control the `env` used with these tool calls. It is inevitably a bit complex so that it is possible to override individual components of the policy so without having to restate the entire thing. Details are in the updated `README.md` in this PR, but here is the relevant bit that explains the individual fields of `shell_environment_policy`: | Field | Type | Default | Description | | ------------------------- | -------------------------- | ------- | ----------------------------------------------------------------------------------------------------------------------------------------------- | | `inherit` | string | `core` | Starting template for the environment:<br>`core` (`HOME`, `PATH`, `USER`, …), `all` (clone full parent env), or `none` (start empty). | | `ignore_default_excludes` | boolean | `false` | When `false`, Codex removes any var whose **name** contains `KEY`, `SECRET`, or `TOKEN` (case-insensitive) before other rules run. | | `exclude` | array&lt;string&gt; | `[]` | Case-insensitive glob patterns to drop after the default filter.<br>Examples: `"AWS_*"`, `"AZURE_*"`. | | `set` | table&lt;string,string&gt; | `{}` | Explicit key/value overrides or additions – always win over inherited values. | | `include_only` | array&lt;string&gt; | `[]` | If non-empty, a whitelist of patterns; only variables that match _one_ pattern survive the final step. (Generally used with `inherit = "all"`.) | In particular, note that the default is `inherit = "core"`, so: * if you have extra env variables that you want to inherit from the parent process, use `inherit = "all"` and then specify `include_only` * if you have extra env variables where you want to hardcode the values, the default `inherit = "core"` will work fine, but then you need to specify `set` This configuration is not battle-tested, so we will probably still have to play with it a bit. `core/src/exec_env.rs` has the critical business logic as well as unit tests. Though if nothing else, previous to this change: ``` $ cargo run --bin codex -- debug seatbelt -- printenv OPENAI_API_KEY # ...prints OPENAI_API_KEY... ``` But after this change it does not print anything (as desired). One final thing to call out about this PR is that the `configure_command!` macro we use in `core/src/exec.rs` has to do some complex logic with respect to how it builds up the `env` for the process being spawned under Landlock/seccomp. Specifically, doing `cmd.env_clear()` followed by `cmd.envs(&$env_map)` (which is arguably the most intuitive way to do it) caused the Landlock unit tests to fail because the processes spawned by the unit tests started failing in unexpected ways! If we forgo `env_clear()` in favor of updating env vars one at a time, the tests still pass. The comment in the code talks about this a bit, and while I would like to investigate this more, I need to move on for the moment, but I do plan to come back to it to fully understand what is going on. For example, this suggests that we might not be able to spawn a C program that calls `env_clear()`, which would be...weird. We may still have to fiddle with our Landlock config if that is the case.
2025-05-22 09:51:19 -07:00
#[serde(default)]
pub shell_environment_policy: ShellEnvironmentPolicyToml,
feat: add support for --sandbox flag (#1476) On a high-level, we try to design `config.toml` so that you don't have to "comment out a lot of stuff" when testing different options. Previously, defining a sandbox policy was somewhat at odds with this principle because you would define the policy as attributes of `[sandbox]` like so: ```toml [sandbox] mode = "workspace-write" writable_roots = [ "/tmp" ] ``` but if you wanted to temporarily change to a read-only sandbox, you might feel compelled to modify your file to be: ```toml [sandbox] mode = "read-only" # mode = "workspace-write" # writable_roots = [ "/tmp" ] ``` Technically, commenting out `writable_roots` would not be strictly necessary, as `mode = "read-only"` would ignore `writable_roots`, but it's still a reasonable thing to do to keep things tidy. Currently, the various values for `mode` do not support that many attributes, so this is not that hard to maintain, but one could imagine this becoming more complex in the future. In this PR, we change Codex CLI so that it no longer recognizes `[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`, and `[sandbox_workspace_write]` is used to further configure the sandbox when when `sandbox_mode = "workspace-write"` is used: ```toml sandbox_mode = "workspace-write" [sandbox_workspace_write] writable_roots = [ "/tmp" ] ``` This feels a bit more future-proof in that it is less tedious to configure different sandboxes: ```toml sandbox_mode = "workspace-write" [sandbox_read_only] # read-only options here... [sandbox_workspace_write] writable_roots = [ "/tmp" ] [sandbox_danger_full_access] # danger-full-access options here... ``` In this scheme, you never need to comment out the configuration for an individual sandbox type: you only need to redefine `sandbox_mode`. Relatedly, previous to this change, a user had to do `-c sandbox.mode=read-only` to change the mode on the command line. With this change, things are arguably a bit cleaner because the equivalent option is `-c sandbox_mode=read-only` (and now `-c sandbox_workspace_write=...` can be set separately). Though more importantly, we introduce the `-s/--sandbox` option to the CLI, which maps directly to `sandbox_mode` in `config.toml`, making config override behavior easier to reason about. Moreover, as you can see in the updates to the various Markdown files, it is much easier to explain how to configure sandboxing when things like `--sandbox read-only` can be used as an example. Relatedly, this cleanup also made it straightforward to add support for a `sandbox` option for Codex when used as an MCP server (see the changes to `mcp-server/src/codex_tool_config.rs`). Fixes https://github.com/openai/codex/issues/1248.
2025-07-07 22:31:30 -07:00
/// Sandbox mode to use.
pub sandbox_mode: Option<SandboxMode>,
/// Sandbox configuration to apply if `sandbox` is `WorkspaceWrite`.
pub sandbox_workspace_write: Option<SandboxWorkspaceWrite>,
fix: overhaul SandboxPolicy and config loading in Rust (#732) Previous to this PR, `SandboxPolicy` was a bit difficult to work with: https://github.com/openai/codex/blob/237f8a11e11fdcc793a09e787e48215676d9b95b/codex-rs/core/src/protocol.rs#L98-L108 Specifically: * It was an `enum` and therefore options were mutually exclusive as opposed to additive. * It defined things in terms of what the agent _could not_ do as opposed to what they _could_ do. This made things hard to support because we would prefer to build up a sandbox config by starting with something extremely restrictive and only granting permissions for things the user as explicitly allowed. This PR changes things substantially by redefining the policy in terms of two concepts: * A `SandboxPermission` enum that defines permissions that can be granted to the agent/sandbox. * A `SandboxPolicy` that internally stores a `Vec<SandboxPermission>`, but externally exposes a simpler API that can be used to configure Seatbelt/Landlock. Previous to this PR, we supported a `--sandbox` flag that effectively mapped to an enum value in `SandboxPolicy`. Though now that `SandboxPolicy` is a wrapper around `Vec<SandboxPermission>`, the single `--sandbox` flag no longer makes sense. While I could have turned it into a flag that the user can specify multiple times, I think the current values to use with such a flag are long and potentially messy, so for the moment, I have dropped support for `--sandbox` altogether and we can bring it back once we have figured out the naming thing. Since `--sandbox` is gone, users now have to specify `--full-auto` to get a sandbox that allows writes in `cwd`. Admittedly, there is no clean way to specify the equivalent of `--full-auto` in your `config.toml` right now, so we will have to revisit that, as well. Because `Config` presents a `SandboxPolicy` field and `SandboxPolicy` changed considerably, I had to overhaul how config loading works, as well. There are now two distinct concepts, `ConfigToml` and `Config`: * `ConfigToml` is the deserialization of `~/.codex/config.toml`. As one might expect, every field is `Optional` and it is `#[derive(Deserialize, Default)]`. Consistent use of `Optional` makes it clear what the user has specified explicitly. * `Config` is the "normalized config" and is produced by merging `ConfigToml` with `ConfigOverrides`. Where `ConfigToml` contains a raw `Option<Vec<SandboxPermission>>`, `Config` presents only the final `SandboxPolicy`. The changes to `core/src/exec.rs` and `core/src/linux.rs` merit extra special attention to ensure we are faithfully mapping the `SandboxPolicy` to the Seatbelt and Landlock configs, respectively. Also, take note that `core/src/seatbelt_readonly_policy.sbpl` has been renamed to `codex-rs/core/src/seatbelt_base_policy.sbpl` and that `(allow file-read*)` has been removed from the `.sbpl` file as now this is added to the policy in `core/src/exec.rs` when `sandbox_policy.has_full_disk_read_access()` is `true`.
2025-04-29 15:01:16 -07:00
feat: configurable notifications in the Rust CLI (#793) With this change, you can specify a program that will be executed to get notified about events generated by Codex. The notification info will be packaged as a JSON object. The supported notification types are defined by the `UserNotification` enum introduced in this PR. Initially, it contains only one variant, `AgentTurnComplete`: ```rust pub(crate) enum UserNotification { #[serde(rename_all = "kebab-case")] AgentTurnComplete { turn_id: String, /// Messages that the user sent to the agent to initiate the turn. input_messages: Vec<String>, /// The last message sent by the assistant in the turn. last_assistant_message: Option<String>, }, } ``` This is intended to support the common case when a "turn" ends, which often means it is now your chance to give Codex further instructions. For example, I have the following in my `~/.codex/config.toml`: ```toml notify = ["python3", "/Users/mbolin/.codex/notify.py"] ``` I created my own custom notifier script that calls out to [terminal-notifier](https://github.com/julienXX/terminal-notifier) to show a desktop push notification on macOS. Contents of `notify.py`: ```python #!/usr/bin/env python3 import json import subprocess import sys def main() -> int: if len(sys.argv) != 2: print("Usage: notify.py <NOTIFICATION_JSON>") return 1 try: notification = json.loads(sys.argv[1]) except json.JSONDecodeError: return 1 match notification_type := notification.get("type"): case "agent-turn-complete": assistant_message = notification.get("last-assistant-message") if assistant_message: title = f"Codex: {assistant_message}" else: title = "Codex: Turn Complete!" input_messages = notification.get("input_messages", []) message = " ".join(input_messages) title += message case _: print(f"not sending a push notification for: {notification_type}") return 0 subprocess.check_output( [ "terminal-notifier", "-title", title, "-message", message, "-group", "codex", "-ignoreDnD", "-activate", "com.googlecode.iterm2", ] ) return 0 if __name__ == "__main__": sys.exit(main()) ``` For reference, here are related PRs that tried to add this functionality to the TypeScript version of the Codex CLI: * https://github.com/openai/codex/pull/160 * https://github.com/openai/codex/pull/498
2025-05-02 19:48:13 -07:00
/// Optional external command to spawn for end-user notifications.
#[serde(default)]
pub notify: Option<Vec<String>>,
fix: overhaul SandboxPolicy and config loading in Rust (#732) Previous to this PR, `SandboxPolicy` was a bit difficult to work with: https://github.com/openai/codex/blob/237f8a11e11fdcc793a09e787e48215676d9b95b/codex-rs/core/src/protocol.rs#L98-L108 Specifically: * It was an `enum` and therefore options were mutually exclusive as opposed to additive. * It defined things in terms of what the agent _could not_ do as opposed to what they _could_ do. This made things hard to support because we would prefer to build up a sandbox config by starting with something extremely restrictive and only granting permissions for things the user as explicitly allowed. This PR changes things substantially by redefining the policy in terms of two concepts: * A `SandboxPermission` enum that defines permissions that can be granted to the agent/sandbox. * A `SandboxPolicy` that internally stores a `Vec<SandboxPermission>`, but externally exposes a simpler API that can be used to configure Seatbelt/Landlock. Previous to this PR, we supported a `--sandbox` flag that effectively mapped to an enum value in `SandboxPolicy`. Though now that `SandboxPolicy` is a wrapper around `Vec<SandboxPermission>`, the single `--sandbox` flag no longer makes sense. While I could have turned it into a flag that the user can specify multiple times, I think the current values to use with such a flag are long and potentially messy, so for the moment, I have dropped support for `--sandbox` altogether and we can bring it back once we have figured out the naming thing. Since `--sandbox` is gone, users now have to specify `--full-auto` to get a sandbox that allows writes in `cwd`. Admittedly, there is no clean way to specify the equivalent of `--full-auto` in your `config.toml` right now, so we will have to revisit that, as well. Because `Config` presents a `SandboxPolicy` field and `SandboxPolicy` changed considerably, I had to overhaul how config loading works, as well. There are now two distinct concepts, `ConfigToml` and `Config`: * `ConfigToml` is the deserialization of `~/.codex/config.toml`. As one might expect, every field is `Optional` and it is `#[derive(Deserialize, Default)]`. Consistent use of `Optional` makes it clear what the user has specified explicitly. * `Config` is the "normalized config" and is produced by merging `ConfigToml` with `ConfigOverrides`. Where `ConfigToml` contains a raw `Option<Vec<SandboxPermission>>`, `Config` presents only the final `SandboxPolicy`. The changes to `core/src/exec.rs` and `core/src/linux.rs` merit extra special attention to ensure we are faithfully mapping the `SandboxPolicy` to the Seatbelt and Landlock configs, respectively. Also, take note that `core/src/seatbelt_readonly_policy.sbpl` has been renamed to `codex-rs/core/src/seatbelt_base_policy.sbpl` and that `(allow file-read*)` has been removed from the `.sbpl` file as now this is added to the policy in `core/src/exec.rs` when `sandbox_policy.has_full_disk_read_access()` is `true`.
2025-04-29 15:01:16 -07:00
/// System instructions.
pub instructions: Option<String>,
feat: support mcp_servers in config.toml (#829) This adds initial support for MCP servers in the style of Claude Desktop and Cursor. Note this PR is the bare minimum to get things working end to end: all configured MCP servers are launched every time Codex is run, there is no recovery for MCP servers that crash, etc. (Also, I took some shortcuts to change some fields of `Session` to be `pub(crate)`, which also means there are circular deps between `codex.rs` and `mcp_tool_call.rs`, but I will clean that up in a subsequent PR.) `codex-rs/README.md` is updated as part of this PR to explain how to use this feature. There is a bit of plumbing to route the new settings from `Config` to the business logic in `codex.rs`. The most significant chunks for new code are in `mcp_connection_manager.rs` (which defines the `McpConnectionManager` struct) and `mcp_tool_call.rs`, which is responsible for tool calls. This PR also introduces new `McpToolCallBegin` and `McpToolCallEnd` event types to the protocol, but does not add any handlers for them. (See https://github.com/openai/codex/pull/836 for initial usage.) To test, I added the following to my `~/.codex/config.toml`: ```toml # Local build of https://github.com/hideya/mcp-server-weather-js [mcp_servers.weather] command = "/Users/mbolin/code/mcp-server-weather-js/dist/index.js" args = [] ``` And then I ran the following: ``` codex-rs$ cargo run --bin codex exec 'what is the weather in san francisco' [2025-05-06T22:40:05] Task started: 1 [2025-05-06T22:40:18] Agent message: Here’s the latest National Weather Service forecast for San Francisco (downtown, near 37.77° N, 122.42° W): This Afternoon (Tue): • Sunny, high near 69 °F • West-southwest wind around 12 mph Tonight: • Partly cloudy, low around 52 °F • SW wind 7–10 mph ... ``` Note that Codex itself is not able to make network calls, so it would not normally be able to get live weather information like this. However, the weather MCP is [currently] not run under the Codex sandbox, so it is able to hit `api.weather.gov` and fetch current weather information. --- [//]: # (BEGIN SAPLING FOOTER) Stack created with [Sapling](https://sapling-scm.com). Best reviewed with [ReviewStack](https://reviewstack.dev/openai/codex/pull/829). * #836 * __->__ #829
2025-05-06 15:47:59 -07:00
/// Definition for MCP servers that Codex can reach out to for tool calls.
#[serde(default)]
pub mcp_servers: HashMap<String, McpServerConfig>,
/// User-defined provider entries that extend/override the built-in list.
#[serde(default)]
pub model_providers: HashMap<String, ModelProviderInfo>,
/// Maximum number of bytes to include from an AGENTS.md project doc file.
pub project_doc_max_bytes: Option<usize>,
/// Profile to use from the `profiles` map.
pub profile: Option<String>,
/// Named profiles to facilitate switching between different configurations.
#[serde(default)]
pub profiles: HashMap<String, ConfigProfile>,
feat: record messages from user in ~/.codex/history.jsonl (#939) This is a large change to support a "history" feature like you would expect in a shell like Bash. History events are recorded in `$CODEX_HOME/history.jsonl`. Because it is a JSONL file, it is straightforward to append new entries (as opposed to the TypeScript file that uses `$CODEX_HOME/history.json`, so to be valid JSON, each new entry entails rewriting the entire file). Because it is possible for there to be multiple instances of Codex CLI writing to `history.jsonl` at once, we use advisory file locking when working with `history.jsonl` in `codex-rs/core/src/message_history.rs`. Because we believe history is a sufficiently useful feature, we enable it by default. Though to provide some safety, we set the file permissions of `history.jsonl` to be `o600` so that other users on the system cannot read the user's history. We do not yet support a default list of `SENSITIVE_PATTERNS` as the TypeScript CLI does: https://github.com/openai/codex/blob/3fdf9df1335ac9501e3fb0e61715359145711e8b/codex-cli/src/utils/storage/command-history.ts#L10-L17 We are going to take a more conservative approach to this list in the Rust CLI. For example, while `/\b[A-Za-z0-9-_]{20,}\b/` might exclude sensitive information like API tokens, it would also exclude valuable information such as references to Git commits. As noted in the updated documentation, users can opt-out of history by adding the following to `config.toml`: ```toml [history] persistence = "none" ``` Because `history.jsonl` could, in theory, be quite large, we take a[n arguably overly pedantic] approach in reading history entries into memory. Specifically, we start by telling the client the current number of entries in the history file (`history_entry_count`) as well as the inode (`history_log_id`) of `history.jsonl` (see the new fields on `SessionConfiguredEvent`). The client is responsible for keeping new entries in memory to create a "local history," but if the user hits up enough times to go "past" the end of local history, then the client should use the new `GetHistoryEntryRequest` in the protocol to fetch older entries. Specifically, it should pass the `history_log_id` it was given originally and work backwards from `history_entry_count`. (It should really fetch history in batches rather than one-at-a-time, but that is something we can improve upon in subsequent PRs.) The motivation behind this crazy scheme is that it is designed to defend against: * The `history.jsonl` being truncated during the session such that the index into the history is no longer consistent with what had been read up to that point. We do not yet have logic to enforce a `max_bytes` for `history.jsonl`, but once we do, we will aspire to implement it in a way that should result in a new inode for the file on most systems. * New items from concurrent Codex CLI sessions amending to the history. Because, in absence of truncation, `history.jsonl` is an append-only log, so long as the client reads backwards from `history_entry_count`, it should always get a consistent view of history. (That said, it will not be able to read _new_ commands from concurrent sessions, but perhaps we will introduce a `/` command to reload latest history or something down the road.) Admittedly, my testing of this feature thus far has been fairly light. I expect we will find bugs and introduce enhancements/fixes going forward.
2025-05-15 16:26:23 -07:00
/// Settings that govern if and what will be written to `~/.codex/history.jsonl`.
#[serde(default)]
pub history: Option<History>,
/// Optional URI-based file opener. If set, citations to files in the model
/// output will be hyperlinked using the specified URI scheme.
pub file_opener: Option<UriBasedFileOpener>,
/// Collection of settings that are specific to the TUI.
pub tui: Option<Tui>,
/// When set to `true`, `AgentReasoning` events will be hidden from the
/// UI/output. Defaults to `false`.
pub hide_agent_reasoning: Option<bool>,
/// When set to `true`, `AgentReasoningRawContentEvent` events will be shown in the UI/output.
/// Defaults to `false`.
pub show_raw_agent_reasoning: Option<bool>,
pub model_reasoning_effort: Option<ReasoningEffort>,
pub model_reasoning_summary: Option<ReasoningSummary>,
feat(gpt5): add model_verbosity for GPT‑5 via Responses API (#2108) **Summary** - Adds `model_verbosity` config (values: low, medium, high). - Sends `text.verbosity` only for GPT‑5 family models via the Responses API. - Updates docs and adds serialization tests. **Motivation** - GPT‑5 introduces a verbosity control to steer output length/detail without pro mpt surgery. - Exposing it as a config knob keeps prompts stable and makes behavior explicit and repeatable. **Changes** - Config: - Added `Verbosity` enum (low|medium|high). - Added optional `model_verbosity` to `ConfigToml`, `Config`, and `ConfigProfi le`. - Request wiring: - Extended `ResponsesApiRequest` with optional `text` object. - Populates `text.verbosity` only when model family is `gpt-5`; omitted otherw ise. - Tests: - Verifies `text.verbosity` serializes when set and is omitted when not set. - Docs: - Added “GPT‑5 Verbosity” section in `codex-rs/README.md`. - Added `model_verbosity` section to `codex-rs/config.md`. **Usage** - In `~/.codex/config.toml`: - `model = "gpt-5"` - `model_verbosity = "low"` (or `"medium"` default, `"high"`) - CLI override example: - `codex -c model="gpt-5" -c model_verbosity="high"` **API Impact** - Requests to GPT‑5 via Responses API include: `text: { verbosity: "low|medium|h igh" }` when configured. - For legacy models or Chat Completions providers, `text` is omitted. **Backward Compatibility** - Default behavior unchanged when `model_verbosity` is not set (server default “ medium”). **Testing** - Added unit tests for serialization/omission of `text.verbosity`. - Ran `cargo fmt` and `cargo test --all-features` (all green). **Docs** - `README.md`: new “GPT‑5 Verbosity” note under Config with example. - `config.md`: new `model_verbosity` section. **Out of Scope** - No changes to temperature/top_p or other GPT‑5 parameters. - No changes to Chat Completions wiring. **Risks / Notes** - If OpenAI changes the wire shape for verbosity, we may need to update `Respons esApiRequest`. - Behavior gated to `gpt-5` model family to avoid unexpected effects elsewhere. **Checklist** - [x] Code gated to GPT‑5 family only - [x] Docs updated (`README.md`, `config.md`) - [x] Tests added and passing - [x] Formatting applied Release note: Add `model_verbosity` config to control GPT‑5 output verbosity via the Responses API (low|medium|high).
2025-08-22 17:12:10 +01:00
/// Optional verbosity control for GPT-5 models (Responses API `text.verbosity`).
pub model_verbosity: Option<Verbosity>,
/// Override to force-enable reasoning summaries for the configured model.
pub model_supports_reasoning_summaries: Option<bool>,
/// Override to force reasoning summary format for the configured model.
pub model_reasoning_summary_format: Option<ReasoningSummaryFormat>,
/// Base URL for requests to ChatGPT (as opposed to the OpenAI API).
pub chatgpt_base_url: Option<String>,
/// Experimental rollout resume path (absolute path to .jsonl; undocumented).
pub experimental_resume: Option<PathBuf>,
/// Experimental path to a file whose contents replace the built-in BASE_INSTRUCTIONS.
pub experimental_instructions_file: Option<PathBuf>,
pub experimental_use_exec_command_tool: Option<bool>,
pub experimental_use_unified_exec_tool: Option<bool>,
pub projects: Option<HashMap<String, ProjectConfig>>,
/// Nested tools section for feature toggles
pub tools: Option<ToolsToml>,
/// When true, disables burst-paste detection for typed input entirely.
/// All characters are inserted as they are received, and no buffering
/// or placeholder replacement will occur for fast keypress bursts.
pub disable_paste_burst: Option<bool>,
}
impl From<ConfigToml> for UserSavedConfig {
fn from(config_toml: ConfigToml) -> Self {
let profiles = config_toml
.profiles
.into_iter()
.map(|(k, v)| (k, v.into()))
.collect();
Self {
approval_policy: config_toml.approval_policy,
sandbox_mode: config_toml.sandbox_mode,
sandbox_settings: config_toml.sandbox_workspace_write.map(From::from),
model: config_toml.model,
model_reasoning_effort: config_toml.model_reasoning_effort,
model_reasoning_summary: config_toml.model_reasoning_summary,
model_verbosity: config_toml.model_verbosity,
tools: config_toml.tools.map(From::from),
profile: config_toml.profile,
profiles,
}
}
}
#[derive(Deserialize, Debug, Clone, PartialEq, Eq)]
pub struct ProjectConfig {
pub trust_level: Option<String>,
fix: overhaul SandboxPolicy and config loading in Rust (#732) Previous to this PR, `SandboxPolicy` was a bit difficult to work with: https://github.com/openai/codex/blob/237f8a11e11fdcc793a09e787e48215676d9b95b/codex-rs/core/src/protocol.rs#L98-L108 Specifically: * It was an `enum` and therefore options were mutually exclusive as opposed to additive. * It defined things in terms of what the agent _could not_ do as opposed to what they _could_ do. This made things hard to support because we would prefer to build up a sandbox config by starting with something extremely restrictive and only granting permissions for things the user as explicitly allowed. This PR changes things substantially by redefining the policy in terms of two concepts: * A `SandboxPermission` enum that defines permissions that can be granted to the agent/sandbox. * A `SandboxPolicy` that internally stores a `Vec<SandboxPermission>`, but externally exposes a simpler API that can be used to configure Seatbelt/Landlock. Previous to this PR, we supported a `--sandbox` flag that effectively mapped to an enum value in `SandboxPolicy`. Though now that `SandboxPolicy` is a wrapper around `Vec<SandboxPermission>`, the single `--sandbox` flag no longer makes sense. While I could have turned it into a flag that the user can specify multiple times, I think the current values to use with such a flag are long and potentially messy, so for the moment, I have dropped support for `--sandbox` altogether and we can bring it back once we have figured out the naming thing. Since `--sandbox` is gone, users now have to specify `--full-auto` to get a sandbox that allows writes in `cwd`. Admittedly, there is no clean way to specify the equivalent of `--full-auto` in your `config.toml` right now, so we will have to revisit that, as well. Because `Config` presents a `SandboxPolicy` field and `SandboxPolicy` changed considerably, I had to overhaul how config loading works, as well. There are now two distinct concepts, `ConfigToml` and `Config`: * `ConfigToml` is the deserialization of `~/.codex/config.toml`. As one might expect, every field is `Optional` and it is `#[derive(Deserialize, Default)]`. Consistent use of `Optional` makes it clear what the user has specified explicitly. * `Config` is the "normalized config" and is produced by merging `ConfigToml` with `ConfigOverrides`. Where `ConfigToml` contains a raw `Option<Vec<SandboxPermission>>`, `Config` presents only the final `SandboxPolicy`. The changes to `core/src/exec.rs` and `core/src/linux.rs` merit extra special attention to ensure we are faithfully mapping the `SandboxPolicy` to the Seatbelt and Landlock configs, respectively. Also, take note that `core/src/seatbelt_readonly_policy.sbpl` has been renamed to `codex-rs/core/src/seatbelt_base_policy.sbpl` and that `(allow file-read*)` has been removed from the `.sbpl` file as now this is added to the policy in `core/src/exec.rs` when `sandbox_policy.has_full_disk_read_access()` is `true`.
2025-04-29 15:01:16 -07:00
}
#[derive(Deserialize, Debug, Clone, Default, PartialEq)]
pub struct ToolsToml {
#[serde(default, alias = "web_search_request")]
pub web_search: Option<bool>,
/// Enable the `view_image` tool that lets the agent attach local images.
#[serde(default)]
pub view_image: Option<bool>,
}
impl From<ToolsToml> for Tools {
fn from(tools_toml: ToolsToml) -> Self {
Self {
web_search: tools_toml.web_search,
view_image: tools_toml.view_image,
}
}
}
feat: add support for --sandbox flag (#1476) On a high-level, we try to design `config.toml` so that you don't have to "comment out a lot of stuff" when testing different options. Previously, defining a sandbox policy was somewhat at odds with this principle because you would define the policy as attributes of `[sandbox]` like so: ```toml [sandbox] mode = "workspace-write" writable_roots = [ "/tmp" ] ``` but if you wanted to temporarily change to a read-only sandbox, you might feel compelled to modify your file to be: ```toml [sandbox] mode = "read-only" # mode = "workspace-write" # writable_roots = [ "/tmp" ] ``` Technically, commenting out `writable_roots` would not be strictly necessary, as `mode = "read-only"` would ignore `writable_roots`, but it's still a reasonable thing to do to keep things tidy. Currently, the various values for `mode` do not support that many attributes, so this is not that hard to maintain, but one could imagine this becoming more complex in the future. In this PR, we change Codex CLI so that it no longer recognizes `[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`, and `[sandbox_workspace_write]` is used to further configure the sandbox when when `sandbox_mode = "workspace-write"` is used: ```toml sandbox_mode = "workspace-write" [sandbox_workspace_write] writable_roots = [ "/tmp" ] ``` This feels a bit more future-proof in that it is less tedious to configure different sandboxes: ```toml sandbox_mode = "workspace-write" [sandbox_read_only] # read-only options here... [sandbox_workspace_write] writable_roots = [ "/tmp" ] [sandbox_danger_full_access] # danger-full-access options here... ``` In this scheme, you never need to comment out the configuration for an individual sandbox type: you only need to redefine `sandbox_mode`. Relatedly, previous to this change, a user had to do `-c sandbox.mode=read-only` to change the mode on the command line. With this change, things are arguably a bit cleaner because the equivalent option is `-c sandbox_mode=read-only` (and now `-c sandbox_workspace_write=...` can be set separately). Though more importantly, we introduce the `-s/--sandbox` option to the CLI, which maps directly to `sandbox_mode` in `config.toml`, making config override behavior easier to reason about. Moreover, as you can see in the updates to the various Markdown files, it is much easier to explain how to configure sandboxing when things like `--sandbox read-only` can be used as an example. Relatedly, this cleanup also made it straightforward to add support for a `sandbox` option for Codex when used as an MCP server (see the changes to `mcp-server/src/codex_tool_config.rs`). Fixes https://github.com/openai/codex/issues/1248.
2025-07-07 22:31:30 -07:00
impl ConfigToml {
/// Derive the effective sandbox policy from the configuration.
fn derive_sandbox_policy(&self, sandbox_mode_override: Option<SandboxMode>) -> SandboxPolicy {
let resolved_sandbox_mode = sandbox_mode_override
.or(self.sandbox_mode)
.unwrap_or_default();
match resolved_sandbox_mode {
SandboxMode::ReadOnly => SandboxPolicy::new_read_only_policy(),
SandboxMode::WorkspaceWrite => match self.sandbox_workspace_write.as_ref() {
Some(SandboxWorkspaceWrite {
writable_roots,
network_access,
exclude_tmpdir_env_var,
exclude_slash_tmp,
}) => SandboxPolicy::WorkspaceWrite {
writable_roots: writable_roots.clone(),
network_access: *network_access,
exclude_tmpdir_env_var: *exclude_tmpdir_env_var,
exclude_slash_tmp: *exclude_slash_tmp,
feat: add support for --sandbox flag (#1476) On a high-level, we try to design `config.toml` so that you don't have to "comment out a lot of stuff" when testing different options. Previously, defining a sandbox policy was somewhat at odds with this principle because you would define the policy as attributes of `[sandbox]` like so: ```toml [sandbox] mode = "workspace-write" writable_roots = [ "/tmp" ] ``` but if you wanted to temporarily change to a read-only sandbox, you might feel compelled to modify your file to be: ```toml [sandbox] mode = "read-only" # mode = "workspace-write" # writable_roots = [ "/tmp" ] ``` Technically, commenting out `writable_roots` would not be strictly necessary, as `mode = "read-only"` would ignore `writable_roots`, but it's still a reasonable thing to do to keep things tidy. Currently, the various values for `mode` do not support that many attributes, so this is not that hard to maintain, but one could imagine this becoming more complex in the future. In this PR, we change Codex CLI so that it no longer recognizes `[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`, and `[sandbox_workspace_write]` is used to further configure the sandbox when when `sandbox_mode = "workspace-write"` is used: ```toml sandbox_mode = "workspace-write" [sandbox_workspace_write] writable_roots = [ "/tmp" ] ``` This feels a bit more future-proof in that it is less tedious to configure different sandboxes: ```toml sandbox_mode = "workspace-write" [sandbox_read_only] # read-only options here... [sandbox_workspace_write] writable_roots = [ "/tmp" ] [sandbox_danger_full_access] # danger-full-access options here... ``` In this scheme, you never need to comment out the configuration for an individual sandbox type: you only need to redefine `sandbox_mode`. Relatedly, previous to this change, a user had to do `-c sandbox.mode=read-only` to change the mode on the command line. With this change, things are arguably a bit cleaner because the equivalent option is `-c sandbox_mode=read-only` (and now `-c sandbox_workspace_write=...` can be set separately). Though more importantly, we introduce the `-s/--sandbox` option to the CLI, which maps directly to `sandbox_mode` in `config.toml`, making config override behavior easier to reason about. Moreover, as you can see in the updates to the various Markdown files, it is much easier to explain how to configure sandboxing when things like `--sandbox read-only` can be used as an example. Relatedly, this cleanup also made it straightforward to add support for a `sandbox` option for Codex when used as an MCP server (see the changes to `mcp-server/src/codex_tool_config.rs`). Fixes https://github.com/openai/codex/issues/1248.
2025-07-07 22:31:30 -07:00
},
None => SandboxPolicy::new_workspace_write_policy(),
},
SandboxMode::DangerFullAccess => SandboxPolicy::DangerFullAccess,
}
}
pub fn is_cwd_trusted(&self, resolved_cwd: &Path) -> bool {
let projects = self.projects.clone().unwrap_or_default();
let is_path_trusted = |path: &Path| {
let path_str = path.to_string_lossy().to_string();
projects
.get(&path_str)
.map(|p| p.trust_level.as_deref() == Some("trusted"))
.unwrap_or(false)
};
// Fast path: exact cwd match
if is_path_trusted(resolved_cwd) {
return true;
}
// If cwd lives inside a git worktree, check whether the root git project
// (the primary repository working directory) is trusted. This lets
// worktrees inherit trust from the main project.
if let Some(root_project) = resolve_root_git_project_for_trust(resolved_cwd) {
return is_path_trusted(&root_project);
}
false
}
pub fn get_config_profile(
&self,
override_profile: Option<String>,
) -> Result<ConfigProfile, std::io::Error> {
let profile = override_profile.or_else(|| self.profile.clone());
match profile {
Some(key) => {
if let Some(profile) = self.profiles.get(key.as_str()) {
return Ok(profile.clone());
}
Err(std::io::Error::new(
std::io::ErrorKind::NotFound,
format!("config profile `{key}` not found"),
))
}
None => Ok(ConfigProfile::default()),
}
}
feat: add support for --sandbox flag (#1476) On a high-level, we try to design `config.toml` so that you don't have to "comment out a lot of stuff" when testing different options. Previously, defining a sandbox policy was somewhat at odds with this principle because you would define the policy as attributes of `[sandbox]` like so: ```toml [sandbox] mode = "workspace-write" writable_roots = [ "/tmp" ] ``` but if you wanted to temporarily change to a read-only sandbox, you might feel compelled to modify your file to be: ```toml [sandbox] mode = "read-only" # mode = "workspace-write" # writable_roots = [ "/tmp" ] ``` Technically, commenting out `writable_roots` would not be strictly necessary, as `mode = "read-only"` would ignore `writable_roots`, but it's still a reasonable thing to do to keep things tidy. Currently, the various values for `mode` do not support that many attributes, so this is not that hard to maintain, but one could imagine this becoming more complex in the future. In this PR, we change Codex CLI so that it no longer recognizes `[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`, and `[sandbox_workspace_write]` is used to further configure the sandbox when when `sandbox_mode = "workspace-write"` is used: ```toml sandbox_mode = "workspace-write" [sandbox_workspace_write] writable_roots = [ "/tmp" ] ``` This feels a bit more future-proof in that it is less tedious to configure different sandboxes: ```toml sandbox_mode = "workspace-write" [sandbox_read_only] # read-only options here... [sandbox_workspace_write] writable_roots = [ "/tmp" ] [sandbox_danger_full_access] # danger-full-access options here... ``` In this scheme, you never need to comment out the configuration for an individual sandbox type: you only need to redefine `sandbox_mode`. Relatedly, previous to this change, a user had to do `-c sandbox.mode=read-only` to change the mode on the command line. With this change, things are arguably a bit cleaner because the equivalent option is `-c sandbox_mode=read-only` (and now `-c sandbox_workspace_write=...` can be set separately). Though more importantly, we introduce the `-s/--sandbox` option to the CLI, which maps directly to `sandbox_mode` in `config.toml`, making config override behavior easier to reason about. Moreover, as you can see in the updates to the various Markdown files, it is much easier to explain how to configure sandboxing when things like `--sandbox read-only` can be used as an example. Relatedly, this cleanup also made it straightforward to add support for a `sandbox` option for Codex when used as an MCP server (see the changes to `mcp-server/src/codex_tool_config.rs`). Fixes https://github.com/openai/codex/issues/1248.
2025-07-07 22:31:30 -07:00
}
/// Optional overrides for user configuration (e.g., from CLI flags).
#[derive(Default, Debug, Clone)]
pub struct ConfigOverrides {
pub model: Option<String>,
Review Mode (Core) (#3401) ## 📝 Review Mode -- Core This PR introduces the Core implementation for Review mode: - New op `Op::Review { prompt: String }:` spawns a child review task with isolated context, a review‑specific system prompt, and a `Config.review_model`. - `EnteredReviewMode`: emitted when the child review session starts. Every event from this point onwards reflects the review session. - `ExitedReviewMode(Option<ReviewOutputEvent>)`: emitted when the review finishes or is interrupted, with optional structured findings: ```json { "findings": [ { "title": "<≤ 80 chars, imperative>", "body": "<valid Markdown explaining *why* this is a problem; cite files/lines/functions>", "confidence_score": <float 0.0-1.0>, "priority": <int 0-3>, "code_location": { "absolute_file_path": "<file path>", "line_range": {"start": <int>, "end": <int>} } } ], "overall_correctness": "patch is correct" | "patch is incorrect", "overall_explanation": "<1-3 sentence explanation justifying the overall_correctness verdict>", "overall_confidence_score": <float 0.0-1.0> } ``` ## Questions ### Why separate out its own message history? We want the review thread to match the training of our review models as much as possible -- that means using a custom prompt, removing user instructions, and starting a clean chat history. We also want to make sure the review thread doesn't leak into the parent thread. ### Why do this as a mode, vs. sub-agents? 1. We want review to be a synchronous task, so it's fine for now to do a bespoke implementation. 2. We're still unclear about the final structure for sub-agents. We'd prefer to land this quickly and then refactor into sub-agents without rushing that implementation.
2025-09-12 16:25:10 -07:00
pub review_model: Option<String>,
pub cwd: Option<PathBuf>,
pub approval_policy: Option<AskForApproval>,
feat: add support for --sandbox flag (#1476) On a high-level, we try to design `config.toml` so that you don't have to "comment out a lot of stuff" when testing different options. Previously, defining a sandbox policy was somewhat at odds with this principle because you would define the policy as attributes of `[sandbox]` like so: ```toml [sandbox] mode = "workspace-write" writable_roots = [ "/tmp" ] ``` but if you wanted to temporarily change to a read-only sandbox, you might feel compelled to modify your file to be: ```toml [sandbox] mode = "read-only" # mode = "workspace-write" # writable_roots = [ "/tmp" ] ``` Technically, commenting out `writable_roots` would not be strictly necessary, as `mode = "read-only"` would ignore `writable_roots`, but it's still a reasonable thing to do to keep things tidy. Currently, the various values for `mode` do not support that many attributes, so this is not that hard to maintain, but one could imagine this becoming more complex in the future. In this PR, we change Codex CLI so that it no longer recognizes `[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`, and `[sandbox_workspace_write]` is used to further configure the sandbox when when `sandbox_mode = "workspace-write"` is used: ```toml sandbox_mode = "workspace-write" [sandbox_workspace_write] writable_roots = [ "/tmp" ] ``` This feels a bit more future-proof in that it is less tedious to configure different sandboxes: ```toml sandbox_mode = "workspace-write" [sandbox_read_only] # read-only options here... [sandbox_workspace_write] writable_roots = [ "/tmp" ] [sandbox_danger_full_access] # danger-full-access options here... ``` In this scheme, you never need to comment out the configuration for an individual sandbox type: you only need to redefine `sandbox_mode`. Relatedly, previous to this change, a user had to do `-c sandbox.mode=read-only` to change the mode on the command line. With this change, things are arguably a bit cleaner because the equivalent option is `-c sandbox_mode=read-only` (and now `-c sandbox_workspace_write=...` can be set separately). Though more importantly, we introduce the `-s/--sandbox` option to the CLI, which maps directly to `sandbox_mode` in `config.toml`, making config override behavior easier to reason about. Moreover, as you can see in the updates to the various Markdown files, it is much easier to explain how to configure sandboxing when things like `--sandbox read-only` can be used as an example. Relatedly, this cleanup also made it straightforward to add support for a `sandbox` option for Codex when used as an MCP server (see the changes to `mcp-server/src/codex_tool_config.rs`). Fixes https://github.com/openai/codex/issues/1248.
2025-07-07 22:31:30 -07:00
pub sandbox_mode: Option<SandboxMode>,
pub model_provider: Option<String>,
pub config_profile: Option<String>,
pub codex_linux_sandbox_exe: Option<PathBuf>,
pub base_instructions: Option<String>,
pub include_plan_tool: Option<bool>,
pub include_apply_patch_tool: Option<bool>,
pub include_view_image_tool: Option<bool>,
pub show_raw_agent_reasoning: Option<bool>,
pub tools_web_search_request: Option<bool>,
}
impl Config {
/// Meant to be used exclusively for tests: `load_with_overrides()` should
/// be used in all other cases.
pub fn load_from_base_config_with_overrides(
cfg: ConfigToml,
overrides: ConfigOverrides,
codex_home: PathBuf,
) -> std::io::Result<Self> {
let user_instructions = Self::load_instructions(Some(&codex_home));
// Destructure ConfigOverrides fully to ensure all overrides are applied.
let ConfigOverrides {
model,
Review Mode (Core) (#3401) ## 📝 Review Mode -- Core This PR introduces the Core implementation for Review mode: - New op `Op::Review { prompt: String }:` spawns a child review task with isolated context, a review‑specific system prompt, and a `Config.review_model`. - `EnteredReviewMode`: emitted when the child review session starts. Every event from this point onwards reflects the review session. - `ExitedReviewMode(Option<ReviewOutputEvent>)`: emitted when the review finishes or is interrupted, with optional structured findings: ```json { "findings": [ { "title": "<≤ 80 chars, imperative>", "body": "<valid Markdown explaining *why* this is a problem; cite files/lines/functions>", "confidence_score": <float 0.0-1.0>, "priority": <int 0-3>, "code_location": { "absolute_file_path": "<file path>", "line_range": {"start": <int>, "end": <int>} } } ], "overall_correctness": "patch is correct" | "patch is incorrect", "overall_explanation": "<1-3 sentence explanation justifying the overall_correctness verdict>", "overall_confidence_score": <float 0.0-1.0> } ``` ## Questions ### Why separate out its own message history? We want the review thread to match the training of our review models as much as possible -- that means using a custom prompt, removing user instructions, and starting a clean chat history. We also want to make sure the review thread doesn't leak into the parent thread. ### Why do this as a mode, vs. sub-agents? 1. We want review to be a synchronous task, so it's fine for now to do a bespoke implementation. 2. We're still unclear about the final structure for sub-agents. We'd prefer to land this quickly and then refactor into sub-agents without rushing that implementation.
2025-09-12 16:25:10 -07:00
review_model: override_review_model,
cwd,
approval_policy,
feat: add support for --sandbox flag (#1476) On a high-level, we try to design `config.toml` so that you don't have to "comment out a lot of stuff" when testing different options. Previously, defining a sandbox policy was somewhat at odds with this principle because you would define the policy as attributes of `[sandbox]` like so: ```toml [sandbox] mode = "workspace-write" writable_roots = [ "/tmp" ] ``` but if you wanted to temporarily change to a read-only sandbox, you might feel compelled to modify your file to be: ```toml [sandbox] mode = "read-only" # mode = "workspace-write" # writable_roots = [ "/tmp" ] ``` Technically, commenting out `writable_roots` would not be strictly necessary, as `mode = "read-only"` would ignore `writable_roots`, but it's still a reasonable thing to do to keep things tidy. Currently, the various values for `mode` do not support that many attributes, so this is not that hard to maintain, but one could imagine this becoming more complex in the future. In this PR, we change Codex CLI so that it no longer recognizes `[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`, and `[sandbox_workspace_write]` is used to further configure the sandbox when when `sandbox_mode = "workspace-write"` is used: ```toml sandbox_mode = "workspace-write" [sandbox_workspace_write] writable_roots = [ "/tmp" ] ``` This feels a bit more future-proof in that it is less tedious to configure different sandboxes: ```toml sandbox_mode = "workspace-write" [sandbox_read_only] # read-only options here... [sandbox_workspace_write] writable_roots = [ "/tmp" ] [sandbox_danger_full_access] # danger-full-access options here... ``` In this scheme, you never need to comment out the configuration for an individual sandbox type: you only need to redefine `sandbox_mode`. Relatedly, previous to this change, a user had to do `-c sandbox.mode=read-only` to change the mode on the command line. With this change, things are arguably a bit cleaner because the equivalent option is `-c sandbox_mode=read-only` (and now `-c sandbox_workspace_write=...` can be set separately). Though more importantly, we introduce the `-s/--sandbox` option to the CLI, which maps directly to `sandbox_mode` in `config.toml`, making config override behavior easier to reason about. Moreover, as you can see in the updates to the various Markdown files, it is much easier to explain how to configure sandboxing when things like `--sandbox read-only` can be used as an example. Relatedly, this cleanup also made it straightforward to add support for a `sandbox` option for Codex when used as an MCP server (see the changes to `mcp-server/src/codex_tool_config.rs`). Fixes https://github.com/openai/codex/issues/1248.
2025-07-07 22:31:30 -07:00
sandbox_mode,
model_provider,
config_profile: config_profile_key,
codex_linux_sandbox_exe,
base_instructions,
include_plan_tool,
include_apply_patch_tool,
include_view_image_tool,
show_raw_agent_reasoning,
tools_web_search_request: override_tools_web_search_request,
} = overrides;
let active_profile_name = config_profile_key
.as_ref()
.or(cfg.profile.as_ref())
.cloned();
let config_profile = match active_profile_name.as_ref() {
Some(key) => cfg
.profiles
feat: add support for --sandbox flag (#1476) On a high-level, we try to design `config.toml` so that you don't have to "comment out a lot of stuff" when testing different options. Previously, defining a sandbox policy was somewhat at odds with this principle because you would define the policy as attributes of `[sandbox]` like so: ```toml [sandbox] mode = "workspace-write" writable_roots = [ "/tmp" ] ``` but if you wanted to temporarily change to a read-only sandbox, you might feel compelled to modify your file to be: ```toml [sandbox] mode = "read-only" # mode = "workspace-write" # writable_roots = [ "/tmp" ] ``` Technically, commenting out `writable_roots` would not be strictly necessary, as `mode = "read-only"` would ignore `writable_roots`, but it's still a reasonable thing to do to keep things tidy. Currently, the various values for `mode` do not support that many attributes, so this is not that hard to maintain, but one could imagine this becoming more complex in the future. In this PR, we change Codex CLI so that it no longer recognizes `[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`, and `[sandbox_workspace_write]` is used to further configure the sandbox when when `sandbox_mode = "workspace-write"` is used: ```toml sandbox_mode = "workspace-write" [sandbox_workspace_write] writable_roots = [ "/tmp" ] ``` This feels a bit more future-proof in that it is less tedious to configure different sandboxes: ```toml sandbox_mode = "workspace-write" [sandbox_read_only] # read-only options here... [sandbox_workspace_write] writable_roots = [ "/tmp" ] [sandbox_danger_full_access] # danger-full-access options here... ``` In this scheme, you never need to comment out the configuration for an individual sandbox type: you only need to redefine `sandbox_mode`. Relatedly, previous to this change, a user had to do `-c sandbox.mode=read-only` to change the mode on the command line. With this change, things are arguably a bit cleaner because the equivalent option is `-c sandbox_mode=read-only` (and now `-c sandbox_workspace_write=...` can be set separately). Though more importantly, we introduce the `-s/--sandbox` option to the CLI, which maps directly to `sandbox_mode` in `config.toml`, making config override behavior easier to reason about. Moreover, as you can see in the updates to the various Markdown files, it is much easier to explain how to configure sandboxing when things like `--sandbox read-only` can be used as an example. Relatedly, this cleanup also made it straightforward to add support for a `sandbox` option for Codex when used as an MCP server (see the changes to `mcp-server/src/codex_tool_config.rs`). Fixes https://github.com/openai/codex/issues/1248.
2025-07-07 22:31:30 -07:00
.get(key)
.ok_or_else(|| {
std::io::Error::new(
std::io::ErrorKind::NotFound,
format!("config profile `{key}` not found"),
)
})?
.clone(),
None => ConfigProfile::default(),
};
feat: add support for --sandbox flag (#1476) On a high-level, we try to design `config.toml` so that you don't have to "comment out a lot of stuff" when testing different options. Previously, defining a sandbox policy was somewhat at odds with this principle because you would define the policy as attributes of `[sandbox]` like so: ```toml [sandbox] mode = "workspace-write" writable_roots = [ "/tmp" ] ``` but if you wanted to temporarily change to a read-only sandbox, you might feel compelled to modify your file to be: ```toml [sandbox] mode = "read-only" # mode = "workspace-write" # writable_roots = [ "/tmp" ] ``` Technically, commenting out `writable_roots` would not be strictly necessary, as `mode = "read-only"` would ignore `writable_roots`, but it's still a reasonable thing to do to keep things tidy. Currently, the various values for `mode` do not support that many attributes, so this is not that hard to maintain, but one could imagine this becoming more complex in the future. In this PR, we change Codex CLI so that it no longer recognizes `[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`, and `[sandbox_workspace_write]` is used to further configure the sandbox when when `sandbox_mode = "workspace-write"` is used: ```toml sandbox_mode = "workspace-write" [sandbox_workspace_write] writable_roots = [ "/tmp" ] ``` This feels a bit more future-proof in that it is less tedious to configure different sandboxes: ```toml sandbox_mode = "workspace-write" [sandbox_read_only] # read-only options here... [sandbox_workspace_write] writable_roots = [ "/tmp" ] [sandbox_danger_full_access] # danger-full-access options here... ``` In this scheme, you never need to comment out the configuration for an individual sandbox type: you only need to redefine `sandbox_mode`. Relatedly, previous to this change, a user had to do `-c sandbox.mode=read-only` to change the mode on the command line. With this change, things are arguably a bit cleaner because the equivalent option is `-c sandbox_mode=read-only` (and now `-c sandbox_workspace_write=...` can be set separately). Though more importantly, we introduce the `-s/--sandbox` option to the CLI, which maps directly to `sandbox_mode` in `config.toml`, making config override behavior easier to reason about. Moreover, as you can see in the updates to the various Markdown files, it is much easier to explain how to configure sandboxing when things like `--sandbox read-only` can be used as an example. Relatedly, this cleanup also made it straightforward to add support for a `sandbox` option for Codex when used as an MCP server (see the changes to `mcp-server/src/codex_tool_config.rs`). Fixes https://github.com/openai/codex/issues/1248.
2025-07-07 22:31:30 -07:00
let sandbox_policy = cfg.derive_sandbox_policy(sandbox_mode);
let mut model_providers = built_in_model_providers();
// Merge user-defined providers into the built-in list.
for (key, provider) in cfg.model_providers.into_iter() {
model_providers.entry(key).or_insert(provider);
}
let model_provider_id = model_provider
.or(config_profile.model_provider)
.or(cfg.model_provider)
.unwrap_or_else(|| "openai".to_string());
let model_provider = model_providers
feat: support the chat completions API in the Rust CLI (#862) This is a substantial PR to add support for the chat completions API, which in turn makes it possible to use non-OpenAI model providers (just like in the TypeScript CLI): * It moves a number of structs from `client.rs` to `client_common.rs` so they can be shared. * It introduces support for the chat completions API in `chat_completions.rs`. * It updates `ModelProviderInfo` so that `env_key` is `Option<String>` instead of `String` (for e.g., ollama) and adds a `wire_api` field * It updates `client.rs` to choose between `stream_responses()` and `stream_chat_completions()` based on the `wire_api` for the `ModelProviderInfo` * It updates the `exec` and TUI CLIs to no longer fail if the `OPENAI_API_KEY` environment variable is not set * It updates the TUI so that `EventMsg::Error` is displayed more prominently when it occurs, particularly now that it is important to alert users to the `CodexErr::EnvVar` variant. * `CodexErr::EnvVar` was updated to include an optional `instructions` field so we can preserve the behavior where we direct users to https://platform.openai.com if `OPENAI_API_KEY` is not set. * Cleaned up the "welcome message" in the TUI to ensure the model provider is displayed. * Updated the docs in `codex-rs/README.md`. To exercise the chat completions API from OpenAI models, I added the following to my `config.toml`: ```toml model = "gpt-4o" model_provider = "openai-chat-completions" [model_providers.openai-chat-completions] name = "OpenAI using Chat Completions" base_url = "https://api.openai.com/v1" env_key = "OPENAI_API_KEY" wire_api = "chat" ``` Though to test a non-OpenAI provider, I installed ollama with mistral locally on my Mac because ChatGPT said that would be a good match for my hardware: ```shell brew install ollama ollama serve ollama pull mistral ``` Then I added the following to my `~/.codex/config.toml`: ```toml model = "mistral" model_provider = "ollama" ``` Note this code could certainly use more test coverage, but I want to get this in so folks can start playing with it. For reference, I believe https://github.com/openai/codex/pull/247 was roughly the comparable PR on the TypeScript side.
2025-05-08 21:46:06 -07:00
.get(&model_provider_id)
.ok_or_else(|| {
std::io::Error::new(
std::io::ErrorKind::NotFound,
feat: support the chat completions API in the Rust CLI (#862) This is a substantial PR to add support for the chat completions API, which in turn makes it possible to use non-OpenAI model providers (just like in the TypeScript CLI): * It moves a number of structs from `client.rs` to `client_common.rs` so they can be shared. * It introduces support for the chat completions API in `chat_completions.rs`. * It updates `ModelProviderInfo` so that `env_key` is `Option<String>` instead of `String` (for e.g., ollama) and adds a `wire_api` field * It updates `client.rs` to choose between `stream_responses()` and `stream_chat_completions()` based on the `wire_api` for the `ModelProviderInfo` * It updates the `exec` and TUI CLIs to no longer fail if the `OPENAI_API_KEY` environment variable is not set * It updates the TUI so that `EventMsg::Error` is displayed more prominently when it occurs, particularly now that it is important to alert users to the `CodexErr::EnvVar` variant. * `CodexErr::EnvVar` was updated to include an optional `instructions` field so we can preserve the behavior where we direct users to https://platform.openai.com if `OPENAI_API_KEY` is not set. * Cleaned up the "welcome message" in the TUI to ensure the model provider is displayed. * Updated the docs in `codex-rs/README.md`. To exercise the chat completions API from OpenAI models, I added the following to my `config.toml`: ```toml model = "gpt-4o" model_provider = "openai-chat-completions" [model_providers.openai-chat-completions] name = "OpenAI using Chat Completions" base_url = "https://api.openai.com/v1" env_key = "OPENAI_API_KEY" wire_api = "chat" ``` Though to test a non-OpenAI provider, I installed ollama with mistral locally on my Mac because ChatGPT said that would be a good match for my hardware: ```shell brew install ollama ollama serve ollama pull mistral ``` Then I added the following to my `~/.codex/config.toml`: ```toml model = "mistral" model_provider = "ollama" ``` Note this code could certainly use more test coverage, but I want to get this in so folks can start playing with it. For reference, I believe https://github.com/openai/codex/pull/247 was roughly the comparable PR on the TypeScript side.
2025-05-08 21:46:06 -07:00
format!("Model provider `{model_provider_id}` not found"),
)
})?
.clone();
let shell_environment_policy = cfg.shell_environment_policy.into();
feat: introduce support for shell_environment_policy in config.toml (#1061) To date, when handling `shell` and `local_shell` tool calls, we were spawning new processes using the environment inherited from the Codex process itself. This means that the sensitive `OPENAI_API_KEY` that Codex needs to talk to OpenAI models was made available to everything run by `shell` and `local_shell`. While there are cases where that might be useful, it does not seem like a good default. This PR introduces a complex `shell_environment_policy` config option to control the `env` used with these tool calls. It is inevitably a bit complex so that it is possible to override individual components of the policy so without having to restate the entire thing. Details are in the updated `README.md` in this PR, but here is the relevant bit that explains the individual fields of `shell_environment_policy`: | Field | Type | Default | Description | | ------------------------- | -------------------------- | ------- | ----------------------------------------------------------------------------------------------------------------------------------------------- | | `inherit` | string | `core` | Starting template for the environment:<br>`core` (`HOME`, `PATH`, `USER`, …), `all` (clone full parent env), or `none` (start empty). | | `ignore_default_excludes` | boolean | `false` | When `false`, Codex removes any var whose **name** contains `KEY`, `SECRET`, or `TOKEN` (case-insensitive) before other rules run. | | `exclude` | array&lt;string&gt; | `[]` | Case-insensitive glob patterns to drop after the default filter.<br>Examples: `"AWS_*"`, `"AZURE_*"`. | | `set` | table&lt;string,string&gt; | `{}` | Explicit key/value overrides or additions – always win over inherited values. | | `include_only` | array&lt;string&gt; | `[]` | If non-empty, a whitelist of patterns; only variables that match _one_ pattern survive the final step. (Generally used with `inherit = "all"`.) | In particular, note that the default is `inherit = "core"`, so: * if you have extra env variables that you want to inherit from the parent process, use `inherit = "all"` and then specify `include_only` * if you have extra env variables where you want to hardcode the values, the default `inherit = "core"` will work fine, but then you need to specify `set` This configuration is not battle-tested, so we will probably still have to play with it a bit. `core/src/exec_env.rs` has the critical business logic as well as unit tests. Though if nothing else, previous to this change: ``` $ cargo run --bin codex -- debug seatbelt -- printenv OPENAI_API_KEY # ...prints OPENAI_API_KEY... ``` But after this change it does not print anything (as desired). One final thing to call out about this PR is that the `configure_command!` macro we use in `core/src/exec.rs` has to do some complex logic with respect to how it builds up the `env` for the process being spawned under Landlock/seccomp. Specifically, doing `cmd.env_clear()` followed by `cmd.envs(&$env_map)` (which is arguably the most intuitive way to do it) caused the Landlock unit tests to fail because the processes spawned by the unit tests started failing in unexpected ways! If we forgo `env_clear()` in favor of updating env vars one at a time, the tests still pass. The comment in the code talks about this a bit, and while I would like to investigate this more, I need to move on for the moment, but I do plan to come back to it to fully understand what is going on. For example, this suggests that we might not be able to spawn a C program that calls `env_clear()`, which would be...weird. We may still have to fiddle with our Landlock config if that is the case.
2025-05-22 09:51:19 -07:00
let resolved_cwd = {
use std::env;
match cwd {
None => {
tracing::info!("cwd not set, using current dir");
env::current_dir()?
}
Some(p) if p.is_absolute() => p,
Some(p) => {
// Resolve relative path against the current working directory.
tracing::info!("cwd is relative, resolving against current dir");
let mut current = env::current_dir()?;
current.push(p);
current
}
}
};
let history = cfg.history.unwrap_or_default();
let tools_web_search_request = override_tools_web_search_request
.or(cfg.tools.as_ref().and_then(|t| t.web_search))
.unwrap_or(false);
feat: record messages from user in ~/.codex/history.jsonl (#939) This is a large change to support a "history" feature like you would expect in a shell like Bash. History events are recorded in `$CODEX_HOME/history.jsonl`. Because it is a JSONL file, it is straightforward to append new entries (as opposed to the TypeScript file that uses `$CODEX_HOME/history.json`, so to be valid JSON, each new entry entails rewriting the entire file). Because it is possible for there to be multiple instances of Codex CLI writing to `history.jsonl` at once, we use advisory file locking when working with `history.jsonl` in `codex-rs/core/src/message_history.rs`. Because we believe history is a sufficiently useful feature, we enable it by default. Though to provide some safety, we set the file permissions of `history.jsonl` to be `o600` so that other users on the system cannot read the user's history. We do not yet support a default list of `SENSITIVE_PATTERNS` as the TypeScript CLI does: https://github.com/openai/codex/blob/3fdf9df1335ac9501e3fb0e61715359145711e8b/codex-cli/src/utils/storage/command-history.ts#L10-L17 We are going to take a more conservative approach to this list in the Rust CLI. For example, while `/\b[A-Za-z0-9-_]{20,}\b/` might exclude sensitive information like API tokens, it would also exclude valuable information such as references to Git commits. As noted in the updated documentation, users can opt-out of history by adding the following to `config.toml`: ```toml [history] persistence = "none" ``` Because `history.jsonl` could, in theory, be quite large, we take a[n arguably overly pedantic] approach in reading history entries into memory. Specifically, we start by telling the client the current number of entries in the history file (`history_entry_count`) as well as the inode (`history_log_id`) of `history.jsonl` (see the new fields on `SessionConfiguredEvent`). The client is responsible for keeping new entries in memory to create a "local history," but if the user hits up enough times to go "past" the end of local history, then the client should use the new `GetHistoryEntryRequest` in the protocol to fetch older entries. Specifically, it should pass the `history_log_id` it was given originally and work backwards from `history_entry_count`. (It should really fetch history in batches rather than one-at-a-time, but that is something we can improve upon in subsequent PRs.) The motivation behind this crazy scheme is that it is designed to defend against: * The `history.jsonl` being truncated during the session such that the index into the history is no longer consistent with what had been read up to that point. We do not yet have logic to enforce a `max_bytes` for `history.jsonl`, but once we do, we will aspire to implement it in a way that should result in a new inode for the file on most systems. * New items from concurrent Codex CLI sessions amending to the history. Because, in absence of truncation, `history.jsonl` is an append-only log, so long as the client reads backwards from `history_entry_count`, it should always get a consistent view of history. (That said, it will not be able to read _new_ commands from concurrent sessions, but perhaps we will introduce a `/` command to reload latest history or something down the road.) Admittedly, my testing of this feature thus far has been fairly light. I expect we will find bugs and introduce enhancements/fixes going forward.
2025-05-15 16:26:23 -07:00
let include_view_image_tool = include_view_image_tool
.or(cfg.tools.as_ref().and_then(|t| t.view_image))
.unwrap_or(true);
feat: show number of tokens remaining in UI (#1388) When using the OpenAI Responses API, we now record the `usage` field for a `"response.completed"` event, which includes metrics about the number of tokens consumed. We also introduce `openai_model_info.rs`, which includes current data about the most common OpenAI models available via the API (specifically `context_window` and `max_output_tokens`). If Codex does not recognize the model, you can set `model_context_window` and `model_max_output_tokens` explicitly in `config.toml`. When then introduce a new event type to `protocol.rs`, `TokenCount`, which includes the `TokenUsage` for the most recent turn. Finally, we update the TUI to record the running sum of tokens used so the percentage of available context window remaining can be reported via the placeholder text for the composer: ![Screenshot 2025-06-25 at 11 20 55 PM](https://github.com/user-attachments/assets/6fd6982f-7247-4f14-84b2-2e600cb1fd49) We could certainly get much fancier with this (such as reporting the estimated cost of the conversation), but for now, we are just trying to achieve feature parity with the TypeScript CLI. Though arguably this improves upon the TypeScript CLI, as the TypeScript CLI uses heuristics to estimate the number of tokens used rather than using the `usage` information directly: https://github.com/openai/codex/blob/296996d74e345b1b05d8c3451a06ace21c5ada96/codex-cli/src/utils/approximate-tokens-used.ts#L3-L16 Fixes https://github.com/openai/codex/issues/1242
2025-06-25 23:31:11 -07:00
let model = model
.or(config_profile.model)
.or(cfg.model)
.unwrap_or_else(default_model);
let mut model_family = find_family_for_model(&model).unwrap_or_else(|| ModelFamily {
slug: model.clone(),
family: model.clone(),
needs_special_apply_patch_instructions: false,
supports_reasoning_summaries: false,
reasoning_summary_format: ReasoningSummaryFormat::None,
uses_local_shell_tool: false,
apply_patch_tool_type: None,
});
if let Some(supports_reasoning_summaries) = cfg.model_supports_reasoning_summaries {
model_family.supports_reasoning_summaries = supports_reasoning_summaries;
}
if let Some(model_reasoning_summary_format) = cfg.model_reasoning_summary_format {
model_family.reasoning_summary_format = model_reasoning_summary_format;
}
let openai_model_info = get_model_info(&model_family);
feat: show number of tokens remaining in UI (#1388) When using the OpenAI Responses API, we now record the `usage` field for a `"response.completed"` event, which includes metrics about the number of tokens consumed. We also introduce `openai_model_info.rs`, which includes current data about the most common OpenAI models available via the API (specifically `context_window` and `max_output_tokens`). If Codex does not recognize the model, you can set `model_context_window` and `model_max_output_tokens` explicitly in `config.toml`. When then introduce a new event type to `protocol.rs`, `TokenCount`, which includes the `TokenUsage` for the most recent turn. Finally, we update the TUI to record the running sum of tokens used so the percentage of available context window remaining can be reported via the placeholder text for the composer: ![Screenshot 2025-06-25 at 11 20 55 PM](https://github.com/user-attachments/assets/6fd6982f-7247-4f14-84b2-2e600cb1fd49) We could certainly get much fancier with this (such as reporting the estimated cost of the conversation), but for now, we are just trying to achieve feature parity with the TypeScript CLI. Though arguably this improves upon the TypeScript CLI, as the TypeScript CLI uses heuristics to estimate the number of tokens used rather than using the `usage` information directly: https://github.com/openai/codex/blob/296996d74e345b1b05d8c3451a06ace21c5ada96/codex-cli/src/utils/approximate-tokens-used.ts#L3-L16 Fixes https://github.com/openai/codex/issues/1242
2025-06-25 23:31:11 -07:00
let model_context_window = cfg
.model_context_window
.or_else(|| openai_model_info.as_ref().map(|info| info.context_window));
let model_max_output_tokens = cfg.model_max_output_tokens.or_else(|| {
openai_model_info
.as_ref()
.map(|info| info.max_output_tokens)
});
let model_auto_compact_token_limit = cfg.model_auto_compact_token_limit.or_else(|| {
openai_model_info
.as_ref()
.and_then(|info| info.auto_compact_token_limit)
});
let experimental_resume = cfg.experimental_resume;
// Load base instructions override from a file if specified. If the
// path is relative, resolve it against the effective cwd so the
// behaviour matches other path-like config values.
let experimental_instructions_path = config_profile
.experimental_instructions_file
.as_ref()
.or(cfg.experimental_instructions_file.as_ref());
let file_base_instructions =
Self::get_base_instructions(experimental_instructions_path, &resolved_cwd)?;
let base_instructions = base_instructions.or(file_base_instructions);
Review Mode (Core) (#3401) ## 📝 Review Mode -- Core This PR introduces the Core implementation for Review mode: - New op `Op::Review { prompt: String }:` spawns a child review task with isolated context, a review‑specific system prompt, and a `Config.review_model`. - `EnteredReviewMode`: emitted when the child review session starts. Every event from this point onwards reflects the review session. - `ExitedReviewMode(Option<ReviewOutputEvent>)`: emitted when the review finishes or is interrupted, with optional structured findings: ```json { "findings": [ { "title": "<≤ 80 chars, imperative>", "body": "<valid Markdown explaining *why* this is a problem; cite files/lines/functions>", "confidence_score": <float 0.0-1.0>, "priority": <int 0-3>, "code_location": { "absolute_file_path": "<file path>", "line_range": {"start": <int>, "end": <int>} } } ], "overall_correctness": "patch is correct" | "patch is incorrect", "overall_explanation": "<1-3 sentence explanation justifying the overall_correctness verdict>", "overall_confidence_score": <float 0.0-1.0> } ``` ## Questions ### Why separate out its own message history? We want the review thread to match the training of our review models as much as possible -- that means using a custom prompt, removing user instructions, and starting a clean chat history. We also want to make sure the review thread doesn't leak into the parent thread. ### Why do this as a mode, vs. sub-agents? 1. We want review to be a synchronous task, so it's fine for now to do a bespoke implementation. 2. We're still unclear about the final structure for sub-agents. We'd prefer to land this quickly and then refactor into sub-agents without rushing that implementation.
2025-09-12 16:25:10 -07:00
// Default review model when not set in config; allow CLI override to take precedence.
let review_model = override_review_model
.or(cfg.review_model)
.unwrap_or_else(default_review_model);
let config = Self {
feat: show number of tokens remaining in UI (#1388) When using the OpenAI Responses API, we now record the `usage` field for a `"response.completed"` event, which includes metrics about the number of tokens consumed. We also introduce `openai_model_info.rs`, which includes current data about the most common OpenAI models available via the API (specifically `context_window` and `max_output_tokens`). If Codex does not recognize the model, you can set `model_context_window` and `model_max_output_tokens` explicitly in `config.toml`. When then introduce a new event type to `protocol.rs`, `TokenCount`, which includes the `TokenUsage` for the most recent turn. Finally, we update the TUI to record the running sum of tokens used so the percentage of available context window remaining can be reported via the placeholder text for the composer: ![Screenshot 2025-06-25 at 11 20 55 PM](https://github.com/user-attachments/assets/6fd6982f-7247-4f14-84b2-2e600cb1fd49) We could certainly get much fancier with this (such as reporting the estimated cost of the conversation), but for now, we are just trying to achieve feature parity with the TypeScript CLI. Though arguably this improves upon the TypeScript CLI, as the TypeScript CLI uses heuristics to estimate the number of tokens used rather than using the `usage` information directly: https://github.com/openai/codex/blob/296996d74e345b1b05d8c3451a06ace21c5ada96/codex-cli/src/utils/approximate-tokens-used.ts#L3-L16 Fixes https://github.com/openai/codex/issues/1242
2025-06-25 23:31:11 -07:00
model,
Review Mode (Core) (#3401) ## 📝 Review Mode -- Core This PR introduces the Core implementation for Review mode: - New op `Op::Review { prompt: String }:` spawns a child review task with isolated context, a review‑specific system prompt, and a `Config.review_model`. - `EnteredReviewMode`: emitted when the child review session starts. Every event from this point onwards reflects the review session. - `ExitedReviewMode(Option<ReviewOutputEvent>)`: emitted when the review finishes or is interrupted, with optional structured findings: ```json { "findings": [ { "title": "<≤ 80 chars, imperative>", "body": "<valid Markdown explaining *why* this is a problem; cite files/lines/functions>", "confidence_score": <float 0.0-1.0>, "priority": <int 0-3>, "code_location": { "absolute_file_path": "<file path>", "line_range": {"start": <int>, "end": <int>} } } ], "overall_correctness": "patch is correct" | "patch is incorrect", "overall_explanation": "<1-3 sentence explanation justifying the overall_correctness verdict>", "overall_confidence_score": <float 0.0-1.0> } ``` ## Questions ### Why separate out its own message history? We want the review thread to match the training of our review models as much as possible -- that means using a custom prompt, removing user instructions, and starting a clean chat history. We also want to make sure the review thread doesn't leak into the parent thread. ### Why do this as a mode, vs. sub-agents? 1. We want review to be a synchronous task, so it's fine for now to do a bespoke implementation. 2. We're still unclear about the final structure for sub-agents. We'd prefer to land this quickly and then refactor into sub-agents without rushing that implementation.
2025-09-12 16:25:10 -07:00
review_model,
model_family,
feat: show number of tokens remaining in UI (#1388) When using the OpenAI Responses API, we now record the `usage` field for a `"response.completed"` event, which includes metrics about the number of tokens consumed. We also introduce `openai_model_info.rs`, which includes current data about the most common OpenAI models available via the API (specifically `context_window` and `max_output_tokens`). If Codex does not recognize the model, you can set `model_context_window` and `model_max_output_tokens` explicitly in `config.toml`. When then introduce a new event type to `protocol.rs`, `TokenCount`, which includes the `TokenUsage` for the most recent turn. Finally, we update the TUI to record the running sum of tokens used so the percentage of available context window remaining can be reported via the placeholder text for the composer: ![Screenshot 2025-06-25 at 11 20 55 PM](https://github.com/user-attachments/assets/6fd6982f-7247-4f14-84b2-2e600cb1fd49) We could certainly get much fancier with this (such as reporting the estimated cost of the conversation), but for now, we are just trying to achieve feature parity with the TypeScript CLI. Though arguably this improves upon the TypeScript CLI, as the TypeScript CLI uses heuristics to estimate the number of tokens used rather than using the `usage` information directly: https://github.com/openai/codex/blob/296996d74e345b1b05d8c3451a06ace21c5ada96/codex-cli/src/utils/approximate-tokens-used.ts#L3-L16 Fixes https://github.com/openai/codex/issues/1242
2025-06-25 23:31:11 -07:00
model_context_window,
model_max_output_tokens,
model_auto_compact_token_limit,
feat: support the chat completions API in the Rust CLI (#862) This is a substantial PR to add support for the chat completions API, which in turn makes it possible to use non-OpenAI model providers (just like in the TypeScript CLI): * It moves a number of structs from `client.rs` to `client_common.rs` so they can be shared. * It introduces support for the chat completions API in `chat_completions.rs`. * It updates `ModelProviderInfo` so that `env_key` is `Option<String>` instead of `String` (for e.g., ollama) and adds a `wire_api` field * It updates `client.rs` to choose between `stream_responses()` and `stream_chat_completions()` based on the `wire_api` for the `ModelProviderInfo` * It updates the `exec` and TUI CLIs to no longer fail if the `OPENAI_API_KEY` environment variable is not set * It updates the TUI so that `EventMsg::Error` is displayed more prominently when it occurs, particularly now that it is important to alert users to the `CodexErr::EnvVar` variant. * `CodexErr::EnvVar` was updated to include an optional `instructions` field so we can preserve the behavior where we direct users to https://platform.openai.com if `OPENAI_API_KEY` is not set. * Cleaned up the "welcome message" in the TUI to ensure the model provider is displayed. * Updated the docs in `codex-rs/README.md`. To exercise the chat completions API from OpenAI models, I added the following to my `config.toml`: ```toml model = "gpt-4o" model_provider = "openai-chat-completions" [model_providers.openai-chat-completions] name = "OpenAI using Chat Completions" base_url = "https://api.openai.com/v1" env_key = "OPENAI_API_KEY" wire_api = "chat" ``` Though to test a non-OpenAI provider, I installed ollama with mistral locally on my Mac because ChatGPT said that would be a good match for my hardware: ```shell brew install ollama ollama serve ollama pull mistral ``` Then I added the following to my `~/.codex/config.toml`: ```toml model = "mistral" model_provider = "ollama" ``` Note this code could certainly use more test coverage, but I want to get this in so folks can start playing with it. For reference, I believe https://github.com/openai/codex/pull/247 was roughly the comparable PR on the TypeScript side.
2025-05-08 21:46:06 -07:00
model_provider_id,
model_provider,
cwd: resolved_cwd,
fix: overhaul SandboxPolicy and config loading in Rust (#732) Previous to this PR, `SandboxPolicy` was a bit difficult to work with: https://github.com/openai/codex/blob/237f8a11e11fdcc793a09e787e48215676d9b95b/codex-rs/core/src/protocol.rs#L98-L108 Specifically: * It was an `enum` and therefore options were mutually exclusive as opposed to additive. * It defined things in terms of what the agent _could not_ do as opposed to what they _could_ do. This made things hard to support because we would prefer to build up a sandbox config by starting with something extremely restrictive and only granting permissions for things the user as explicitly allowed. This PR changes things substantially by redefining the policy in terms of two concepts: * A `SandboxPermission` enum that defines permissions that can be granted to the agent/sandbox. * A `SandboxPolicy` that internally stores a `Vec<SandboxPermission>`, but externally exposes a simpler API that can be used to configure Seatbelt/Landlock. Previous to this PR, we supported a `--sandbox` flag that effectively mapped to an enum value in `SandboxPolicy`. Though now that `SandboxPolicy` is a wrapper around `Vec<SandboxPermission>`, the single `--sandbox` flag no longer makes sense. While I could have turned it into a flag that the user can specify multiple times, I think the current values to use with such a flag are long and potentially messy, so for the moment, I have dropped support for `--sandbox` altogether and we can bring it back once we have figured out the naming thing. Since `--sandbox` is gone, users now have to specify `--full-auto` to get a sandbox that allows writes in `cwd`. Admittedly, there is no clean way to specify the equivalent of `--full-auto` in your `config.toml` right now, so we will have to revisit that, as well. Because `Config` presents a `SandboxPolicy` field and `SandboxPolicy` changed considerably, I had to overhaul how config loading works, as well. There are now two distinct concepts, `ConfigToml` and `Config`: * `ConfigToml` is the deserialization of `~/.codex/config.toml`. As one might expect, every field is `Optional` and it is `#[derive(Deserialize, Default)]`. Consistent use of `Optional` makes it clear what the user has specified explicitly. * `Config` is the "normalized config" and is produced by merging `ConfigToml` with `ConfigOverrides`. Where `ConfigToml` contains a raw `Option<Vec<SandboxPermission>>`, `Config` presents only the final `SandboxPolicy`. The changes to `core/src/exec.rs` and `core/src/linux.rs` merit extra special attention to ensure we are faithfully mapping the `SandboxPolicy` to the Seatbelt and Landlock configs, respectively. Also, take note that `core/src/seatbelt_readonly_policy.sbpl` has been renamed to `codex-rs/core/src/seatbelt_base_policy.sbpl` and that `(allow file-read*)` has been removed from the `.sbpl` file as now this is added to the policy in `core/src/exec.rs` when `sandbox_policy.has_full_disk_read_access()` is `true`.
2025-04-29 15:01:16 -07:00
approval_policy: approval_policy
.or(config_profile.approval_policy)
fix: overhaul SandboxPolicy and config loading in Rust (#732) Previous to this PR, `SandboxPolicy` was a bit difficult to work with: https://github.com/openai/codex/blob/237f8a11e11fdcc793a09e787e48215676d9b95b/codex-rs/core/src/protocol.rs#L98-L108 Specifically: * It was an `enum` and therefore options were mutually exclusive as opposed to additive. * It defined things in terms of what the agent _could not_ do as opposed to what they _could_ do. This made things hard to support because we would prefer to build up a sandbox config by starting with something extremely restrictive and only granting permissions for things the user as explicitly allowed. This PR changes things substantially by redefining the policy in terms of two concepts: * A `SandboxPermission` enum that defines permissions that can be granted to the agent/sandbox. * A `SandboxPolicy` that internally stores a `Vec<SandboxPermission>`, but externally exposes a simpler API that can be used to configure Seatbelt/Landlock. Previous to this PR, we supported a `--sandbox` flag that effectively mapped to an enum value in `SandboxPolicy`. Though now that `SandboxPolicy` is a wrapper around `Vec<SandboxPermission>`, the single `--sandbox` flag no longer makes sense. While I could have turned it into a flag that the user can specify multiple times, I think the current values to use with such a flag are long and potentially messy, so for the moment, I have dropped support for `--sandbox` altogether and we can bring it back once we have figured out the naming thing. Since `--sandbox` is gone, users now have to specify `--full-auto` to get a sandbox that allows writes in `cwd`. Admittedly, there is no clean way to specify the equivalent of `--full-auto` in your `config.toml` right now, so we will have to revisit that, as well. Because `Config` presents a `SandboxPolicy` field and `SandboxPolicy` changed considerably, I had to overhaul how config loading works, as well. There are now two distinct concepts, `ConfigToml` and `Config`: * `ConfigToml` is the deserialization of `~/.codex/config.toml`. As one might expect, every field is `Optional` and it is `#[derive(Deserialize, Default)]`. Consistent use of `Optional` makes it clear what the user has specified explicitly. * `Config` is the "normalized config" and is produced by merging `ConfigToml` with `ConfigOverrides`. Where `ConfigToml` contains a raw `Option<Vec<SandboxPermission>>`, `Config` presents only the final `SandboxPolicy`. The changes to `core/src/exec.rs` and `core/src/linux.rs` merit extra special attention to ensure we are faithfully mapping the `SandboxPolicy` to the Seatbelt and Landlock configs, respectively. Also, take note that `core/src/seatbelt_readonly_policy.sbpl` has been renamed to `codex-rs/core/src/seatbelt_base_policy.sbpl` and that `(allow file-read*)` has been removed from the `.sbpl` file as now this is added to the policy in `core/src/exec.rs` when `sandbox_policy.has_full_disk_read_access()` is `true`.
2025-04-29 15:01:16 -07:00
.or(cfg.approval_policy)
.unwrap_or_else(AskForApproval::default),
sandbox_policy,
feat: introduce support for shell_environment_policy in config.toml (#1061) To date, when handling `shell` and `local_shell` tool calls, we were spawning new processes using the environment inherited from the Codex process itself. This means that the sensitive `OPENAI_API_KEY` that Codex needs to talk to OpenAI models was made available to everything run by `shell` and `local_shell`. While there are cases where that might be useful, it does not seem like a good default. This PR introduces a complex `shell_environment_policy` config option to control the `env` used with these tool calls. It is inevitably a bit complex so that it is possible to override individual components of the policy so without having to restate the entire thing. Details are in the updated `README.md` in this PR, but here is the relevant bit that explains the individual fields of `shell_environment_policy`: | Field | Type | Default | Description | | ------------------------- | -------------------------- | ------- | ----------------------------------------------------------------------------------------------------------------------------------------------- | | `inherit` | string | `core` | Starting template for the environment:<br>`core` (`HOME`, `PATH`, `USER`, …), `all` (clone full parent env), or `none` (start empty). | | `ignore_default_excludes` | boolean | `false` | When `false`, Codex removes any var whose **name** contains `KEY`, `SECRET`, or `TOKEN` (case-insensitive) before other rules run. | | `exclude` | array&lt;string&gt; | `[]` | Case-insensitive glob patterns to drop after the default filter.<br>Examples: `"AWS_*"`, `"AZURE_*"`. | | `set` | table&lt;string,string&gt; | `{}` | Explicit key/value overrides or additions – always win over inherited values. | | `include_only` | array&lt;string&gt; | `[]` | If non-empty, a whitelist of patterns; only variables that match _one_ pattern survive the final step. (Generally used with `inherit = "all"`.) | In particular, note that the default is `inherit = "core"`, so: * if you have extra env variables that you want to inherit from the parent process, use `inherit = "all"` and then specify `include_only` * if you have extra env variables where you want to hardcode the values, the default `inherit = "core"` will work fine, but then you need to specify `set` This configuration is not battle-tested, so we will probably still have to play with it a bit. `core/src/exec_env.rs` has the critical business logic as well as unit tests. Though if nothing else, previous to this change: ``` $ cargo run --bin codex -- debug seatbelt -- printenv OPENAI_API_KEY # ...prints OPENAI_API_KEY... ``` But after this change it does not print anything (as desired). One final thing to call out about this PR is that the `configure_command!` macro we use in `core/src/exec.rs` has to do some complex logic with respect to how it builds up the `env` for the process being spawned under Landlock/seccomp. Specifically, doing `cmd.env_clear()` followed by `cmd.envs(&$env_map)` (which is arguably the most intuitive way to do it) caused the Landlock unit tests to fail because the processes spawned by the unit tests started failing in unexpected ways! If we forgo `env_clear()` in favor of updating env vars one at a time, the tests still pass. The comment in the code talks about this a bit, and while I would like to investigate this more, I need to move on for the moment, but I do plan to come back to it to fully understand what is going on. For example, this suggests that we might not be able to spawn a C program that calls `env_clear()`, which would be...weird. We may still have to fiddle with our Landlock config if that is the case.
2025-05-22 09:51:19 -07:00
shell_environment_policy,
feat: configurable notifications in the Rust CLI (#793) With this change, you can specify a program that will be executed to get notified about events generated by Codex. The notification info will be packaged as a JSON object. The supported notification types are defined by the `UserNotification` enum introduced in this PR. Initially, it contains only one variant, `AgentTurnComplete`: ```rust pub(crate) enum UserNotification { #[serde(rename_all = "kebab-case")] AgentTurnComplete { turn_id: String, /// Messages that the user sent to the agent to initiate the turn. input_messages: Vec<String>, /// The last message sent by the assistant in the turn. last_assistant_message: Option<String>, }, } ``` This is intended to support the common case when a "turn" ends, which often means it is now your chance to give Codex further instructions. For example, I have the following in my `~/.codex/config.toml`: ```toml notify = ["python3", "/Users/mbolin/.codex/notify.py"] ``` I created my own custom notifier script that calls out to [terminal-notifier](https://github.com/julienXX/terminal-notifier) to show a desktop push notification on macOS. Contents of `notify.py`: ```python #!/usr/bin/env python3 import json import subprocess import sys def main() -> int: if len(sys.argv) != 2: print("Usage: notify.py <NOTIFICATION_JSON>") return 1 try: notification = json.loads(sys.argv[1]) except json.JSONDecodeError: return 1 match notification_type := notification.get("type"): case "agent-turn-complete": assistant_message = notification.get("last-assistant-message") if assistant_message: title = f"Codex: {assistant_message}" else: title = "Codex: Turn Complete!" input_messages = notification.get("input_messages", []) message = " ".join(input_messages) title += message case _: print(f"not sending a push notification for: {notification_type}") return 0 subprocess.check_output( [ "terminal-notifier", "-title", title, "-message", message, "-group", "codex", "-ignoreDnD", "-activate", "com.googlecode.iterm2", ] ) return 0 if __name__ == "__main__": sys.exit(main()) ``` For reference, here are related PRs that tried to add this functionality to the TypeScript version of the Codex CLI: * https://github.com/openai/codex/pull/160 * https://github.com/openai/codex/pull/498
2025-05-02 19:48:13 -07:00
notify: cfg.notify,
user_instructions,
base_instructions,
feat: support mcp_servers in config.toml (#829) This adds initial support for MCP servers in the style of Claude Desktop and Cursor. Note this PR is the bare minimum to get things working end to end: all configured MCP servers are launched every time Codex is run, there is no recovery for MCP servers that crash, etc. (Also, I took some shortcuts to change some fields of `Session` to be `pub(crate)`, which also means there are circular deps between `codex.rs` and `mcp_tool_call.rs`, but I will clean that up in a subsequent PR.) `codex-rs/README.md` is updated as part of this PR to explain how to use this feature. There is a bit of plumbing to route the new settings from `Config` to the business logic in `codex.rs`. The most significant chunks for new code are in `mcp_connection_manager.rs` (which defines the `McpConnectionManager` struct) and `mcp_tool_call.rs`, which is responsible for tool calls. This PR also introduces new `McpToolCallBegin` and `McpToolCallEnd` event types to the protocol, but does not add any handlers for them. (See https://github.com/openai/codex/pull/836 for initial usage.) To test, I added the following to my `~/.codex/config.toml`: ```toml # Local build of https://github.com/hideya/mcp-server-weather-js [mcp_servers.weather] command = "/Users/mbolin/code/mcp-server-weather-js/dist/index.js" args = [] ``` And then I ran the following: ``` codex-rs$ cargo run --bin codex exec 'what is the weather in san francisco' [2025-05-06T22:40:05] Task started: 1 [2025-05-06T22:40:18] Agent message: Here’s the latest National Weather Service forecast for San Francisco (downtown, near 37.77° N, 122.42° W): This Afternoon (Tue): • Sunny, high near 69 °F • West-southwest wind around 12 mph Tonight: • Partly cloudy, low around 52 °F • SW wind 7–10 mph ... ``` Note that Codex itself is not able to make network calls, so it would not normally be able to get live weather information like this. However, the weather MCP is [currently] not run under the Codex sandbox, so it is able to hit `api.weather.gov` and fetch current weather information. --- [//]: # (BEGIN SAPLING FOOTER) Stack created with [Sapling](https://sapling-scm.com). Best reviewed with [ReviewStack](https://reviewstack.dev/openai/codex/pull/829). * #836 * __->__ #829
2025-05-06 15:47:59 -07:00
mcp_servers: cfg.mcp_servers,
model_providers,
project_doc_max_bytes: cfg.project_doc_max_bytes.unwrap_or(PROJECT_DOC_MAX_BYTES),
codex_home,
feat: record messages from user in ~/.codex/history.jsonl (#939) This is a large change to support a "history" feature like you would expect in a shell like Bash. History events are recorded in `$CODEX_HOME/history.jsonl`. Because it is a JSONL file, it is straightforward to append new entries (as opposed to the TypeScript file that uses `$CODEX_HOME/history.json`, so to be valid JSON, each new entry entails rewriting the entire file). Because it is possible for there to be multiple instances of Codex CLI writing to `history.jsonl` at once, we use advisory file locking when working with `history.jsonl` in `codex-rs/core/src/message_history.rs`. Because we believe history is a sufficiently useful feature, we enable it by default. Though to provide some safety, we set the file permissions of `history.jsonl` to be `o600` so that other users on the system cannot read the user's history. We do not yet support a default list of `SENSITIVE_PATTERNS` as the TypeScript CLI does: https://github.com/openai/codex/blob/3fdf9df1335ac9501e3fb0e61715359145711e8b/codex-cli/src/utils/storage/command-history.ts#L10-L17 We are going to take a more conservative approach to this list in the Rust CLI. For example, while `/\b[A-Za-z0-9-_]{20,}\b/` might exclude sensitive information like API tokens, it would also exclude valuable information such as references to Git commits. As noted in the updated documentation, users can opt-out of history by adding the following to `config.toml`: ```toml [history] persistence = "none" ``` Because `history.jsonl` could, in theory, be quite large, we take a[n arguably overly pedantic] approach in reading history entries into memory. Specifically, we start by telling the client the current number of entries in the history file (`history_entry_count`) as well as the inode (`history_log_id`) of `history.jsonl` (see the new fields on `SessionConfiguredEvent`). The client is responsible for keeping new entries in memory to create a "local history," but if the user hits up enough times to go "past" the end of local history, then the client should use the new `GetHistoryEntryRequest` in the protocol to fetch older entries. Specifically, it should pass the `history_log_id` it was given originally and work backwards from `history_entry_count`. (It should really fetch history in batches rather than one-at-a-time, but that is something we can improve upon in subsequent PRs.) The motivation behind this crazy scheme is that it is designed to defend against: * The `history.jsonl` being truncated during the session such that the index into the history is no longer consistent with what had been read up to that point. We do not yet have logic to enforce a `max_bytes` for `history.jsonl`, but once we do, we will aspire to implement it in a way that should result in a new inode for the file on most systems. * New items from concurrent Codex CLI sessions amending to the history. Because, in absence of truncation, `history.jsonl` is an append-only log, so long as the client reads backwards from `history_entry_count`, it should always get a consistent view of history. (That said, it will not be able to read _new_ commands from concurrent sessions, but perhaps we will introduce a `/` command to reload latest history or something down the road.) Admittedly, my testing of this feature thus far has been fairly light. I expect we will find bugs and introduce enhancements/fixes going forward.
2025-05-15 16:26:23 -07:00
history,
file_opener: cfg.file_opener.unwrap_or(UriBasedFileOpener::VsCode),
codex_linux_sandbox_exe,
hide_agent_reasoning: cfg.hide_agent_reasoning.unwrap_or(false),
show_raw_agent_reasoning: cfg
.show_raw_agent_reasoning
.or(show_raw_agent_reasoning)
.unwrap_or(false),
model_reasoning_effort: config_profile
.model_reasoning_effort
.or(cfg.model_reasoning_effort),
model_reasoning_summary: config_profile
.model_reasoning_summary
.or(cfg.model_reasoning_summary)
.unwrap_or_default(),
feat(gpt5): add model_verbosity for GPT‑5 via Responses API (#2108) **Summary** - Adds `model_verbosity` config (values: low, medium, high). - Sends `text.verbosity` only for GPT‑5 family models via the Responses API. - Updates docs and adds serialization tests. **Motivation** - GPT‑5 introduces a verbosity control to steer output length/detail without pro mpt surgery. - Exposing it as a config knob keeps prompts stable and makes behavior explicit and repeatable. **Changes** - Config: - Added `Verbosity` enum (low|medium|high). - Added optional `model_verbosity` to `ConfigToml`, `Config`, and `ConfigProfi le`. - Request wiring: - Extended `ResponsesApiRequest` with optional `text` object. - Populates `text.verbosity` only when model family is `gpt-5`; omitted otherw ise. - Tests: - Verifies `text.verbosity` serializes when set and is omitted when not set. - Docs: - Added “GPT‑5 Verbosity” section in `codex-rs/README.md`. - Added `model_verbosity` section to `codex-rs/config.md`. **Usage** - In `~/.codex/config.toml`: - `model = "gpt-5"` - `model_verbosity = "low"` (or `"medium"` default, `"high"`) - CLI override example: - `codex -c model="gpt-5" -c model_verbosity="high"` **API Impact** - Requests to GPT‑5 via Responses API include: `text: { verbosity: "low|medium|h igh" }` when configured. - For legacy models or Chat Completions providers, `text` is omitted. **Backward Compatibility** - Default behavior unchanged when `model_verbosity` is not set (server default “ medium”). **Testing** - Added unit tests for serialization/omission of `text.verbosity`. - Ran `cargo fmt` and `cargo test --all-features` (all green). **Docs** - `README.md`: new “GPT‑5 Verbosity” note under Config with example. - `config.md`: new `model_verbosity` section. **Out of Scope** - No changes to temperature/top_p or other GPT‑5 parameters. - No changes to Chat Completions wiring. **Risks / Notes** - If OpenAI changes the wire shape for verbosity, we may need to update `Respons esApiRequest`. - Behavior gated to `gpt-5` model family to avoid unexpected effects elsewhere. **Checklist** - [x] Code gated to GPT‑5 family only - [x] Docs updated (`README.md`, `config.md`) - [x] Tests added and passing - [x] Formatting applied Release note: Add `model_verbosity` config to control GPT‑5 output verbosity via the Responses API (low|medium|high).
2025-08-22 17:12:10 +01:00
model_verbosity: config_profile.model_verbosity.or(cfg.model_verbosity),
chatgpt_base_url: config_profile
.chatgpt_base_url
.or(cfg.chatgpt_base_url)
.unwrap_or("https://chatgpt.com/backend-api/".to_string()),
experimental_resume,
include_plan_tool: include_plan_tool.unwrap_or(false),
include_apply_patch_tool: include_apply_patch_tool.unwrap_or(false),
tools_web_search_request,
use_experimental_streamable_shell_tool: cfg
.experimental_use_exec_command_tool
.unwrap_or(false),
use_experimental_unified_exec_tool: cfg
.experimental_use_unified_exec_tool
.unwrap_or(false),
include_view_image_tool,
active_profile: active_profile_name,
disable_paste_burst: cfg.disable_paste_burst.unwrap_or(false),
};
Ok(config)
}
fn load_instructions(codex_dir: Option<&Path>) -> Option<String> {
let mut p = match codex_dir {
Some(p) => p.to_path_buf(),
None => return None,
};
p.push("AGENTS.md");
fix: agent instructions were not being included when ~/.codex/instructions.md was empty (#908) I had seen issues where `codex-rs` would not always write files without me pressuring it to do so, and between that and the report of https://github.com/openai/codex/issues/900, I decided to look into this further. I found two serious issues with agent instructions: (1) We were only sending agent instructions on the first turn, but looking at the TypeScript code, we should be sending them on every turn. (2) There was a serious issue where the agent instructions were frequently lost: * The TypeScript CLI appears to keep writing `~/.codex/instructions.md`: https://github.com/openai/codex/blob/55142e3e6caddd1e613b71bcb89385ce5cc708bf/codex-cli/src/utils/config.ts#L586 * If `instructions.md` is present, the Rust CLI uses the contents of it INSTEAD OF the default prompt, even if `instructions.md` is empty: https://github.com/openai/codex/blob/55142e3e6caddd1e613b71bcb89385ce5cc708bf/codex-rs/core/src/config.rs#L202-L203 The combination of these two things means that I have been using `codex-rs` without these key instructions: https://github.com/openai/codex/blob/main/codex-rs/core/prompt.md Looking at the TypeScript code, it appears we should be concatenating these three items every time (if they exist): * `prompt.md` * `~/.codex/instructions.md` * nearest `AGENTS.md` This PR fixes things so that: * `Config.instructions` is `None` if `instructions.md` is empty * `Payload.instructions` is now `&'a str` instead of `Option<&'a String>` because we should always have _something_ to send * `Prompt` now has a `get_full_instructions()` helper that returns a `Cow<str>` that will always include the agent instructions first.
2025-05-12 17:24:44 -07:00
std::fs::read_to_string(&p).ok().and_then(|s| {
let s = s.trim();
if s.is_empty() {
None
} else {
Some(s.to_string())
}
})
}
fn get_base_instructions(
path: Option<&PathBuf>,
cwd: &Path,
) -> std::io::Result<Option<String>> {
let p = match path.as_ref() {
None => return Ok(None),
Some(p) => p,
};
// Resolve relative paths against the provided cwd to make CLI
// overrides consistent regardless of where the process was launched
// from.
let full_path = if p.is_relative() {
cwd.join(p)
} else {
p.to_path_buf()
};
let contents = std::fs::read_to_string(&full_path).map_err(|e| {
std::io::Error::new(
e.kind(),
format!(
"failed to read experimental instructions file {}: {e}",
full_path.display()
),
)
})?;
let s = contents.trim().to_string();
if s.is_empty() {
Err(std::io::Error::new(
std::io::ErrorKind::InvalidData,
format!(
"experimental instructions file is empty: {}",
full_path.display()
),
))
} else {
Ok(Some(s))
}
}
}
fn default_model() -> String {
OPENAI_DEFAULT_MODEL.to_string()
}
Review Mode (Core) (#3401) ## 📝 Review Mode -- Core This PR introduces the Core implementation for Review mode: - New op `Op::Review { prompt: String }:` spawns a child review task with isolated context, a review‑specific system prompt, and a `Config.review_model`. - `EnteredReviewMode`: emitted when the child review session starts. Every event from this point onwards reflects the review session. - `ExitedReviewMode(Option<ReviewOutputEvent>)`: emitted when the review finishes or is interrupted, with optional structured findings: ```json { "findings": [ { "title": "<≤ 80 chars, imperative>", "body": "<valid Markdown explaining *why* this is a problem; cite files/lines/functions>", "confidence_score": <float 0.0-1.0>, "priority": <int 0-3>, "code_location": { "absolute_file_path": "<file path>", "line_range": {"start": <int>, "end": <int>} } } ], "overall_correctness": "patch is correct" | "patch is incorrect", "overall_explanation": "<1-3 sentence explanation justifying the overall_correctness verdict>", "overall_confidence_score": <float 0.0-1.0> } ``` ## Questions ### Why separate out its own message history? We want the review thread to match the training of our review models as much as possible -- that means using a custom prompt, removing user instructions, and starting a clean chat history. We also want to make sure the review thread doesn't leak into the parent thread. ### Why do this as a mode, vs. sub-agents? 1. We want review to be a synchronous task, so it's fine for now to do a bespoke implementation. 2. We're still unclear about the final structure for sub-agents. We'd prefer to land this quickly and then refactor into sub-agents without rushing that implementation.
2025-09-12 16:25:10 -07:00
fn default_review_model() -> String {
OPENAI_DEFAULT_REVIEW_MODEL.to_string()
}
/// Returns the path to the Codex configuration directory, which can be
/// specified by the `CODEX_HOME` environment variable. If not set, defaults to
/// `~/.codex`.
///
/// - If `CODEX_HOME` is set, the value will be canonicalized and this
/// function will Err if the path does not exist.
/// - If `CODEX_HOME` is not set, this function does not verify that the
/// directory exists.
pub fn find_codex_home() -> std::io::Result<PathBuf> {
// Honor the `CODEX_HOME` environment variable when it is set to allow users
// (and tests) to override the default location.
if let Ok(val) = std::env::var("CODEX_HOME")
&& !val.is_empty()
{
return PathBuf::from(val).canonicalize();
}
let mut p = home_dir().ok_or_else(|| {
std::io::Error::new(
std::io::ErrorKind::NotFound,
"Could not find home directory",
)
})?;
p.push(".codex");
Ok(p)
}
/// Returns the path to the folder where Codex logs are stored. Does not verify
/// that the directory exists.
pub fn log_dir(cfg: &Config) -> std::io::Result<PathBuf> {
let mut p = cfg.codex_home.clone();
p.push("log");
Ok(p)
}
#[cfg(test)]
mod tests {
use crate::config_types::HistoryPersistence;
use super::*;
use pretty_assertions::assert_eq;
use tempfile::TempDir;
feat: record messages from user in ~/.codex/history.jsonl (#939) This is a large change to support a "history" feature like you would expect in a shell like Bash. History events are recorded in `$CODEX_HOME/history.jsonl`. Because it is a JSONL file, it is straightforward to append new entries (as opposed to the TypeScript file that uses `$CODEX_HOME/history.json`, so to be valid JSON, each new entry entails rewriting the entire file). Because it is possible for there to be multiple instances of Codex CLI writing to `history.jsonl` at once, we use advisory file locking when working with `history.jsonl` in `codex-rs/core/src/message_history.rs`. Because we believe history is a sufficiently useful feature, we enable it by default. Though to provide some safety, we set the file permissions of `history.jsonl` to be `o600` so that other users on the system cannot read the user's history. We do not yet support a default list of `SENSITIVE_PATTERNS` as the TypeScript CLI does: https://github.com/openai/codex/blob/3fdf9df1335ac9501e3fb0e61715359145711e8b/codex-cli/src/utils/storage/command-history.ts#L10-L17 We are going to take a more conservative approach to this list in the Rust CLI. For example, while `/\b[A-Za-z0-9-_]{20,}\b/` might exclude sensitive information like API tokens, it would also exclude valuable information such as references to Git commits. As noted in the updated documentation, users can opt-out of history by adding the following to `config.toml`: ```toml [history] persistence = "none" ``` Because `history.jsonl` could, in theory, be quite large, we take a[n arguably overly pedantic] approach in reading history entries into memory. Specifically, we start by telling the client the current number of entries in the history file (`history_entry_count`) as well as the inode (`history_log_id`) of `history.jsonl` (see the new fields on `SessionConfiguredEvent`). The client is responsible for keeping new entries in memory to create a "local history," but if the user hits up enough times to go "past" the end of local history, then the client should use the new `GetHistoryEntryRequest` in the protocol to fetch older entries. Specifically, it should pass the `history_log_id` it was given originally and work backwards from `history_entry_count`. (It should really fetch history in batches rather than one-at-a-time, but that is something we can improve upon in subsequent PRs.) The motivation behind this crazy scheme is that it is designed to defend against: * The `history.jsonl` being truncated during the session such that the index into the history is no longer consistent with what had been read up to that point. We do not yet have logic to enforce a `max_bytes` for `history.jsonl`, but once we do, we will aspire to implement it in a way that should result in a new inode for the file on most systems. * New items from concurrent Codex CLI sessions amending to the history. Because, in absence of truncation, `history.jsonl` is an append-only log, so long as the client reads backwards from `history_entry_count`, it should always get a consistent view of history. (That said, it will not be able to read _new_ commands from concurrent sessions, but perhaps we will introduce a `/` command to reload latest history or something down the road.) Admittedly, my testing of this feature thus far has been fairly light. I expect we will find bugs and introduce enhancements/fixes going forward.
2025-05-15 16:26:23 -07:00
#[test]
fn test_toml_parsing() {
let history_with_persistence = r#"
[history]
persistence = "save-all"
"#;
let history_with_persistence_cfg = toml::from_str::<ConfigToml>(history_with_persistence)
.expect("TOML deserialization should succeed");
feat: record messages from user in ~/.codex/history.jsonl (#939) This is a large change to support a "history" feature like you would expect in a shell like Bash. History events are recorded in `$CODEX_HOME/history.jsonl`. Because it is a JSONL file, it is straightforward to append new entries (as opposed to the TypeScript file that uses `$CODEX_HOME/history.json`, so to be valid JSON, each new entry entails rewriting the entire file). Because it is possible for there to be multiple instances of Codex CLI writing to `history.jsonl` at once, we use advisory file locking when working with `history.jsonl` in `codex-rs/core/src/message_history.rs`. Because we believe history is a sufficiently useful feature, we enable it by default. Though to provide some safety, we set the file permissions of `history.jsonl` to be `o600` so that other users on the system cannot read the user's history. We do not yet support a default list of `SENSITIVE_PATTERNS` as the TypeScript CLI does: https://github.com/openai/codex/blob/3fdf9df1335ac9501e3fb0e61715359145711e8b/codex-cli/src/utils/storage/command-history.ts#L10-L17 We are going to take a more conservative approach to this list in the Rust CLI. For example, while `/\b[A-Za-z0-9-_]{20,}\b/` might exclude sensitive information like API tokens, it would also exclude valuable information such as references to Git commits. As noted in the updated documentation, users can opt-out of history by adding the following to `config.toml`: ```toml [history] persistence = "none" ``` Because `history.jsonl` could, in theory, be quite large, we take a[n arguably overly pedantic] approach in reading history entries into memory. Specifically, we start by telling the client the current number of entries in the history file (`history_entry_count`) as well as the inode (`history_log_id`) of `history.jsonl` (see the new fields on `SessionConfiguredEvent`). The client is responsible for keeping new entries in memory to create a "local history," but if the user hits up enough times to go "past" the end of local history, then the client should use the new `GetHistoryEntryRequest` in the protocol to fetch older entries. Specifically, it should pass the `history_log_id` it was given originally and work backwards from `history_entry_count`. (It should really fetch history in batches rather than one-at-a-time, but that is something we can improve upon in subsequent PRs.) The motivation behind this crazy scheme is that it is designed to defend against: * The `history.jsonl` being truncated during the session such that the index into the history is no longer consistent with what had been read up to that point. We do not yet have logic to enforce a `max_bytes` for `history.jsonl`, but once we do, we will aspire to implement it in a way that should result in a new inode for the file on most systems. * New items from concurrent Codex CLI sessions amending to the history. Because, in absence of truncation, `history.jsonl` is an append-only log, so long as the client reads backwards from `history_entry_count`, it should always get a consistent view of history. (That said, it will not be able to read _new_ commands from concurrent sessions, but perhaps we will introduce a `/` command to reload latest history or something down the road.) Admittedly, my testing of this feature thus far has been fairly light. I expect we will find bugs and introduce enhancements/fixes going forward.
2025-05-15 16:26:23 -07:00
assert_eq!(
Some(History {
persistence: HistoryPersistence::SaveAll,
max_bytes: None,
}),
history_with_persistence_cfg.history
);
let history_no_persistence = r#"
[history]
persistence = "none"
"#;
let history_no_persistence_cfg = toml::from_str::<ConfigToml>(history_no_persistence)
.expect("TOML deserialization should succeed");
feat: record messages from user in ~/.codex/history.jsonl (#939) This is a large change to support a "history" feature like you would expect in a shell like Bash. History events are recorded in `$CODEX_HOME/history.jsonl`. Because it is a JSONL file, it is straightforward to append new entries (as opposed to the TypeScript file that uses `$CODEX_HOME/history.json`, so to be valid JSON, each new entry entails rewriting the entire file). Because it is possible for there to be multiple instances of Codex CLI writing to `history.jsonl` at once, we use advisory file locking when working with `history.jsonl` in `codex-rs/core/src/message_history.rs`. Because we believe history is a sufficiently useful feature, we enable it by default. Though to provide some safety, we set the file permissions of `history.jsonl` to be `o600` so that other users on the system cannot read the user's history. We do not yet support a default list of `SENSITIVE_PATTERNS` as the TypeScript CLI does: https://github.com/openai/codex/blob/3fdf9df1335ac9501e3fb0e61715359145711e8b/codex-cli/src/utils/storage/command-history.ts#L10-L17 We are going to take a more conservative approach to this list in the Rust CLI. For example, while `/\b[A-Za-z0-9-_]{20,}\b/` might exclude sensitive information like API tokens, it would also exclude valuable information such as references to Git commits. As noted in the updated documentation, users can opt-out of history by adding the following to `config.toml`: ```toml [history] persistence = "none" ``` Because `history.jsonl` could, in theory, be quite large, we take a[n arguably overly pedantic] approach in reading history entries into memory. Specifically, we start by telling the client the current number of entries in the history file (`history_entry_count`) as well as the inode (`history_log_id`) of `history.jsonl` (see the new fields on `SessionConfiguredEvent`). The client is responsible for keeping new entries in memory to create a "local history," but if the user hits up enough times to go "past" the end of local history, then the client should use the new `GetHistoryEntryRequest` in the protocol to fetch older entries. Specifically, it should pass the `history_log_id` it was given originally and work backwards from `history_entry_count`. (It should really fetch history in batches rather than one-at-a-time, but that is something we can improve upon in subsequent PRs.) The motivation behind this crazy scheme is that it is designed to defend against: * The `history.jsonl` being truncated during the session such that the index into the history is no longer consistent with what had been read up to that point. We do not yet have logic to enforce a `max_bytes` for `history.jsonl`, but once we do, we will aspire to implement it in a way that should result in a new inode for the file on most systems. * New items from concurrent Codex CLI sessions amending to the history. Because, in absence of truncation, `history.jsonl` is an append-only log, so long as the client reads backwards from `history_entry_count`, it should always get a consistent view of history. (That said, it will not be able to read _new_ commands from concurrent sessions, but perhaps we will introduce a `/` command to reload latest history or something down the road.) Admittedly, my testing of this feature thus far has been fairly light. I expect we will find bugs and introduce enhancements/fixes going forward.
2025-05-15 16:26:23 -07:00
assert_eq!(
Some(History {
persistence: HistoryPersistence::None,
max_bytes: None,
}),
history_no_persistence_cfg.history
);
}
#[test]
fn test_sandbox_config_parsing() {
let sandbox_full_access = r#"
feat: add support for --sandbox flag (#1476) On a high-level, we try to design `config.toml` so that you don't have to "comment out a lot of stuff" when testing different options. Previously, defining a sandbox policy was somewhat at odds with this principle because you would define the policy as attributes of `[sandbox]` like so: ```toml [sandbox] mode = "workspace-write" writable_roots = [ "/tmp" ] ``` but if you wanted to temporarily change to a read-only sandbox, you might feel compelled to modify your file to be: ```toml [sandbox] mode = "read-only" # mode = "workspace-write" # writable_roots = [ "/tmp" ] ``` Technically, commenting out `writable_roots` would not be strictly necessary, as `mode = "read-only"` would ignore `writable_roots`, but it's still a reasonable thing to do to keep things tidy. Currently, the various values for `mode` do not support that many attributes, so this is not that hard to maintain, but one could imagine this becoming more complex in the future. In this PR, we change Codex CLI so that it no longer recognizes `[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`, and `[sandbox_workspace_write]` is used to further configure the sandbox when when `sandbox_mode = "workspace-write"` is used: ```toml sandbox_mode = "workspace-write" [sandbox_workspace_write] writable_roots = [ "/tmp" ] ``` This feels a bit more future-proof in that it is less tedious to configure different sandboxes: ```toml sandbox_mode = "workspace-write" [sandbox_read_only] # read-only options here... [sandbox_workspace_write] writable_roots = [ "/tmp" ] [sandbox_danger_full_access] # danger-full-access options here... ``` In this scheme, you never need to comment out the configuration for an individual sandbox type: you only need to redefine `sandbox_mode`. Relatedly, previous to this change, a user had to do `-c sandbox.mode=read-only` to change the mode on the command line. With this change, things are arguably a bit cleaner because the equivalent option is `-c sandbox_mode=read-only` (and now `-c sandbox_workspace_write=...` can be set separately). Though more importantly, we introduce the `-s/--sandbox` option to the CLI, which maps directly to `sandbox_mode` in `config.toml`, making config override behavior easier to reason about. Moreover, as you can see in the updates to the various Markdown files, it is much easier to explain how to configure sandboxing when things like `--sandbox read-only` can be used as an example. Relatedly, this cleanup also made it straightforward to add support for a `sandbox` option for Codex when used as an MCP server (see the changes to `mcp-server/src/codex_tool_config.rs`). Fixes https://github.com/openai/codex/issues/1248.
2025-07-07 22:31:30 -07:00
sandbox_mode = "danger-full-access"
[sandbox_workspace_write]
network_access = false # This should be ignored.
"#;
let sandbox_full_access_cfg = toml::from_str::<ConfigToml>(sandbox_full_access)
.expect("TOML deserialization should succeed");
feat: add support for --sandbox flag (#1476) On a high-level, we try to design `config.toml` so that you don't have to "comment out a lot of stuff" when testing different options. Previously, defining a sandbox policy was somewhat at odds with this principle because you would define the policy as attributes of `[sandbox]` like so: ```toml [sandbox] mode = "workspace-write" writable_roots = [ "/tmp" ] ``` but if you wanted to temporarily change to a read-only sandbox, you might feel compelled to modify your file to be: ```toml [sandbox] mode = "read-only" # mode = "workspace-write" # writable_roots = [ "/tmp" ] ``` Technically, commenting out `writable_roots` would not be strictly necessary, as `mode = "read-only"` would ignore `writable_roots`, but it's still a reasonable thing to do to keep things tidy. Currently, the various values for `mode` do not support that many attributes, so this is not that hard to maintain, but one could imagine this becoming more complex in the future. In this PR, we change Codex CLI so that it no longer recognizes `[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`, and `[sandbox_workspace_write]` is used to further configure the sandbox when when `sandbox_mode = "workspace-write"` is used: ```toml sandbox_mode = "workspace-write" [sandbox_workspace_write] writable_roots = [ "/tmp" ] ``` This feels a bit more future-proof in that it is less tedious to configure different sandboxes: ```toml sandbox_mode = "workspace-write" [sandbox_read_only] # read-only options here... [sandbox_workspace_write] writable_roots = [ "/tmp" ] [sandbox_danger_full_access] # danger-full-access options here... ``` In this scheme, you never need to comment out the configuration for an individual sandbox type: you only need to redefine `sandbox_mode`. Relatedly, previous to this change, a user had to do `-c sandbox.mode=read-only` to change the mode on the command line. With this change, things are arguably a bit cleaner because the equivalent option is `-c sandbox_mode=read-only` (and now `-c sandbox_workspace_write=...` can be set separately). Though more importantly, we introduce the `-s/--sandbox` option to the CLI, which maps directly to `sandbox_mode` in `config.toml`, making config override behavior easier to reason about. Moreover, as you can see in the updates to the various Markdown files, it is much easier to explain how to configure sandboxing when things like `--sandbox read-only` can be used as an example. Relatedly, this cleanup also made it straightforward to add support for a `sandbox` option for Codex when used as an MCP server (see the changes to `mcp-server/src/codex_tool_config.rs`). Fixes https://github.com/openai/codex/issues/1248.
2025-07-07 22:31:30 -07:00
let sandbox_mode_override = None;
assert_eq!(
feat: add support for --sandbox flag (#1476) On a high-level, we try to design `config.toml` so that you don't have to "comment out a lot of stuff" when testing different options. Previously, defining a sandbox policy was somewhat at odds with this principle because you would define the policy as attributes of `[sandbox]` like so: ```toml [sandbox] mode = "workspace-write" writable_roots = [ "/tmp" ] ``` but if you wanted to temporarily change to a read-only sandbox, you might feel compelled to modify your file to be: ```toml [sandbox] mode = "read-only" # mode = "workspace-write" # writable_roots = [ "/tmp" ] ``` Technically, commenting out `writable_roots` would not be strictly necessary, as `mode = "read-only"` would ignore `writable_roots`, but it's still a reasonable thing to do to keep things tidy. Currently, the various values for `mode` do not support that many attributes, so this is not that hard to maintain, but one could imagine this becoming more complex in the future. In this PR, we change Codex CLI so that it no longer recognizes `[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`, and `[sandbox_workspace_write]` is used to further configure the sandbox when when `sandbox_mode = "workspace-write"` is used: ```toml sandbox_mode = "workspace-write" [sandbox_workspace_write] writable_roots = [ "/tmp" ] ``` This feels a bit more future-proof in that it is less tedious to configure different sandboxes: ```toml sandbox_mode = "workspace-write" [sandbox_read_only] # read-only options here... [sandbox_workspace_write] writable_roots = [ "/tmp" ] [sandbox_danger_full_access] # danger-full-access options here... ``` In this scheme, you never need to comment out the configuration for an individual sandbox type: you only need to redefine `sandbox_mode`. Relatedly, previous to this change, a user had to do `-c sandbox.mode=read-only` to change the mode on the command line. With this change, things are arguably a bit cleaner because the equivalent option is `-c sandbox_mode=read-only` (and now `-c sandbox_workspace_write=...` can be set separately). Though more importantly, we introduce the `-s/--sandbox` option to the CLI, which maps directly to `sandbox_mode` in `config.toml`, making config override behavior easier to reason about. Moreover, as you can see in the updates to the various Markdown files, it is much easier to explain how to configure sandboxing when things like `--sandbox read-only` can be used as an example. Relatedly, this cleanup also made it straightforward to add support for a `sandbox` option for Codex when used as an MCP server (see the changes to `mcp-server/src/codex_tool_config.rs`). Fixes https://github.com/openai/codex/issues/1248.
2025-07-07 22:31:30 -07:00
SandboxPolicy::DangerFullAccess,
sandbox_full_access_cfg.derive_sandbox_policy(sandbox_mode_override)
);
let sandbox_read_only = r#"
feat: add support for --sandbox flag (#1476) On a high-level, we try to design `config.toml` so that you don't have to "comment out a lot of stuff" when testing different options. Previously, defining a sandbox policy was somewhat at odds with this principle because you would define the policy as attributes of `[sandbox]` like so: ```toml [sandbox] mode = "workspace-write" writable_roots = [ "/tmp" ] ``` but if you wanted to temporarily change to a read-only sandbox, you might feel compelled to modify your file to be: ```toml [sandbox] mode = "read-only" # mode = "workspace-write" # writable_roots = [ "/tmp" ] ``` Technically, commenting out `writable_roots` would not be strictly necessary, as `mode = "read-only"` would ignore `writable_roots`, but it's still a reasonable thing to do to keep things tidy. Currently, the various values for `mode` do not support that many attributes, so this is not that hard to maintain, but one could imagine this becoming more complex in the future. In this PR, we change Codex CLI so that it no longer recognizes `[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`, and `[sandbox_workspace_write]` is used to further configure the sandbox when when `sandbox_mode = "workspace-write"` is used: ```toml sandbox_mode = "workspace-write" [sandbox_workspace_write] writable_roots = [ "/tmp" ] ``` This feels a bit more future-proof in that it is less tedious to configure different sandboxes: ```toml sandbox_mode = "workspace-write" [sandbox_read_only] # read-only options here... [sandbox_workspace_write] writable_roots = [ "/tmp" ] [sandbox_danger_full_access] # danger-full-access options here... ``` In this scheme, you never need to comment out the configuration for an individual sandbox type: you only need to redefine `sandbox_mode`. Relatedly, previous to this change, a user had to do `-c sandbox.mode=read-only` to change the mode on the command line. With this change, things are arguably a bit cleaner because the equivalent option is `-c sandbox_mode=read-only` (and now `-c sandbox_workspace_write=...` can be set separately). Though more importantly, we introduce the `-s/--sandbox` option to the CLI, which maps directly to `sandbox_mode` in `config.toml`, making config override behavior easier to reason about. Moreover, as you can see in the updates to the various Markdown files, it is much easier to explain how to configure sandboxing when things like `--sandbox read-only` can be used as an example. Relatedly, this cleanup also made it straightforward to add support for a `sandbox` option for Codex when used as an MCP server (see the changes to `mcp-server/src/codex_tool_config.rs`). Fixes https://github.com/openai/codex/issues/1248.
2025-07-07 22:31:30 -07:00
sandbox_mode = "read-only"
[sandbox_workspace_write]
network_access = true # This should be ignored.
"#;
let sandbox_read_only_cfg = toml::from_str::<ConfigToml>(sandbox_read_only)
.expect("TOML deserialization should succeed");
feat: add support for --sandbox flag (#1476) On a high-level, we try to design `config.toml` so that you don't have to "comment out a lot of stuff" when testing different options. Previously, defining a sandbox policy was somewhat at odds with this principle because you would define the policy as attributes of `[sandbox]` like so: ```toml [sandbox] mode = "workspace-write" writable_roots = [ "/tmp" ] ``` but if you wanted to temporarily change to a read-only sandbox, you might feel compelled to modify your file to be: ```toml [sandbox] mode = "read-only" # mode = "workspace-write" # writable_roots = [ "/tmp" ] ``` Technically, commenting out `writable_roots` would not be strictly necessary, as `mode = "read-only"` would ignore `writable_roots`, but it's still a reasonable thing to do to keep things tidy. Currently, the various values for `mode` do not support that many attributes, so this is not that hard to maintain, but one could imagine this becoming more complex in the future. In this PR, we change Codex CLI so that it no longer recognizes `[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`, and `[sandbox_workspace_write]` is used to further configure the sandbox when when `sandbox_mode = "workspace-write"` is used: ```toml sandbox_mode = "workspace-write" [sandbox_workspace_write] writable_roots = [ "/tmp" ] ``` This feels a bit more future-proof in that it is less tedious to configure different sandboxes: ```toml sandbox_mode = "workspace-write" [sandbox_read_only] # read-only options here... [sandbox_workspace_write] writable_roots = [ "/tmp" ] [sandbox_danger_full_access] # danger-full-access options here... ``` In this scheme, you never need to comment out the configuration for an individual sandbox type: you only need to redefine `sandbox_mode`. Relatedly, previous to this change, a user had to do `-c sandbox.mode=read-only` to change the mode on the command line. With this change, things are arguably a bit cleaner because the equivalent option is `-c sandbox_mode=read-only` (and now `-c sandbox_workspace_write=...` can be set separately). Though more importantly, we introduce the `-s/--sandbox` option to the CLI, which maps directly to `sandbox_mode` in `config.toml`, making config override behavior easier to reason about. Moreover, as you can see in the updates to the various Markdown files, it is much easier to explain how to configure sandboxing when things like `--sandbox read-only` can be used as an example. Relatedly, this cleanup also made it straightforward to add support for a `sandbox` option for Codex when used as an MCP server (see the changes to `mcp-server/src/codex_tool_config.rs`). Fixes https://github.com/openai/codex/issues/1248.
2025-07-07 22:31:30 -07:00
let sandbox_mode_override = None;
assert_eq!(
SandboxPolicy::ReadOnly,
sandbox_read_only_cfg.derive_sandbox_policy(sandbox_mode_override)
);
let sandbox_workspace_write = r#"
feat: add support for --sandbox flag (#1476) On a high-level, we try to design `config.toml` so that you don't have to "comment out a lot of stuff" when testing different options. Previously, defining a sandbox policy was somewhat at odds with this principle because you would define the policy as attributes of `[sandbox]` like so: ```toml [sandbox] mode = "workspace-write" writable_roots = [ "/tmp" ] ``` but if you wanted to temporarily change to a read-only sandbox, you might feel compelled to modify your file to be: ```toml [sandbox] mode = "read-only" # mode = "workspace-write" # writable_roots = [ "/tmp" ] ``` Technically, commenting out `writable_roots` would not be strictly necessary, as `mode = "read-only"` would ignore `writable_roots`, but it's still a reasonable thing to do to keep things tidy. Currently, the various values for `mode` do not support that many attributes, so this is not that hard to maintain, but one could imagine this becoming more complex in the future. In this PR, we change Codex CLI so that it no longer recognizes `[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`, and `[sandbox_workspace_write]` is used to further configure the sandbox when when `sandbox_mode = "workspace-write"` is used: ```toml sandbox_mode = "workspace-write" [sandbox_workspace_write] writable_roots = [ "/tmp" ] ``` This feels a bit more future-proof in that it is less tedious to configure different sandboxes: ```toml sandbox_mode = "workspace-write" [sandbox_read_only] # read-only options here... [sandbox_workspace_write] writable_roots = [ "/tmp" ] [sandbox_danger_full_access] # danger-full-access options here... ``` In this scheme, you never need to comment out the configuration for an individual sandbox type: you only need to redefine `sandbox_mode`. Relatedly, previous to this change, a user had to do `-c sandbox.mode=read-only` to change the mode on the command line. With this change, things are arguably a bit cleaner because the equivalent option is `-c sandbox_mode=read-only` (and now `-c sandbox_workspace_write=...` can be set separately). Though more importantly, we introduce the `-s/--sandbox` option to the CLI, which maps directly to `sandbox_mode` in `config.toml`, making config override behavior easier to reason about. Moreover, as you can see in the updates to the various Markdown files, it is much easier to explain how to configure sandboxing when things like `--sandbox read-only` can be used as an example. Relatedly, this cleanup also made it straightforward to add support for a `sandbox` option for Codex when used as an MCP server (see the changes to `mcp-server/src/codex_tool_config.rs`). Fixes https://github.com/openai/codex/issues/1248.
2025-07-07 22:31:30 -07:00
sandbox_mode = "workspace-write"
[sandbox_workspace_write]
writable_roots = [
"/my/workspace",
]
exclude_tmpdir_env_var = true
exclude_slash_tmp = true
"#;
let sandbox_workspace_write_cfg = toml::from_str::<ConfigToml>(sandbox_workspace_write)
.expect("TOML deserialization should succeed");
feat: add support for --sandbox flag (#1476) On a high-level, we try to design `config.toml` so that you don't have to "comment out a lot of stuff" when testing different options. Previously, defining a sandbox policy was somewhat at odds with this principle because you would define the policy as attributes of `[sandbox]` like so: ```toml [sandbox] mode = "workspace-write" writable_roots = [ "/tmp" ] ``` but if you wanted to temporarily change to a read-only sandbox, you might feel compelled to modify your file to be: ```toml [sandbox] mode = "read-only" # mode = "workspace-write" # writable_roots = [ "/tmp" ] ``` Technically, commenting out `writable_roots` would not be strictly necessary, as `mode = "read-only"` would ignore `writable_roots`, but it's still a reasonable thing to do to keep things tidy. Currently, the various values for `mode` do not support that many attributes, so this is not that hard to maintain, but one could imagine this becoming more complex in the future. In this PR, we change Codex CLI so that it no longer recognizes `[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`, and `[sandbox_workspace_write]` is used to further configure the sandbox when when `sandbox_mode = "workspace-write"` is used: ```toml sandbox_mode = "workspace-write" [sandbox_workspace_write] writable_roots = [ "/tmp" ] ``` This feels a bit more future-proof in that it is less tedious to configure different sandboxes: ```toml sandbox_mode = "workspace-write" [sandbox_read_only] # read-only options here... [sandbox_workspace_write] writable_roots = [ "/tmp" ] [sandbox_danger_full_access] # danger-full-access options here... ``` In this scheme, you never need to comment out the configuration for an individual sandbox type: you only need to redefine `sandbox_mode`. Relatedly, previous to this change, a user had to do `-c sandbox.mode=read-only` to change the mode on the command line. With this change, things are arguably a bit cleaner because the equivalent option is `-c sandbox_mode=read-only` (and now `-c sandbox_workspace_write=...` can be set separately). Though more importantly, we introduce the `-s/--sandbox` option to the CLI, which maps directly to `sandbox_mode` in `config.toml`, making config override behavior easier to reason about. Moreover, as you can see in the updates to the various Markdown files, it is much easier to explain how to configure sandboxing when things like `--sandbox read-only` can be used as an example. Relatedly, this cleanup also made it straightforward to add support for a `sandbox` option for Codex when used as an MCP server (see the changes to `mcp-server/src/codex_tool_config.rs`). Fixes https://github.com/openai/codex/issues/1248.
2025-07-07 22:31:30 -07:00
let sandbox_mode_override = None;
assert_eq!(
feat: add support for --sandbox flag (#1476) On a high-level, we try to design `config.toml` so that you don't have to "comment out a lot of stuff" when testing different options. Previously, defining a sandbox policy was somewhat at odds with this principle because you would define the policy as attributes of `[sandbox]` like so: ```toml [sandbox] mode = "workspace-write" writable_roots = [ "/tmp" ] ``` but if you wanted to temporarily change to a read-only sandbox, you might feel compelled to modify your file to be: ```toml [sandbox] mode = "read-only" # mode = "workspace-write" # writable_roots = [ "/tmp" ] ``` Technically, commenting out `writable_roots` would not be strictly necessary, as `mode = "read-only"` would ignore `writable_roots`, but it's still a reasonable thing to do to keep things tidy. Currently, the various values for `mode` do not support that many attributes, so this is not that hard to maintain, but one could imagine this becoming more complex in the future. In this PR, we change Codex CLI so that it no longer recognizes `[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`, and `[sandbox_workspace_write]` is used to further configure the sandbox when when `sandbox_mode = "workspace-write"` is used: ```toml sandbox_mode = "workspace-write" [sandbox_workspace_write] writable_roots = [ "/tmp" ] ``` This feels a bit more future-proof in that it is less tedious to configure different sandboxes: ```toml sandbox_mode = "workspace-write" [sandbox_read_only] # read-only options here... [sandbox_workspace_write] writable_roots = [ "/tmp" ] [sandbox_danger_full_access] # danger-full-access options here... ``` In this scheme, you never need to comment out the configuration for an individual sandbox type: you only need to redefine `sandbox_mode`. Relatedly, previous to this change, a user had to do `-c sandbox.mode=read-only` to change the mode on the command line. With this change, things are arguably a bit cleaner because the equivalent option is `-c sandbox_mode=read-only` (and now `-c sandbox_workspace_write=...` can be set separately). Though more importantly, we introduce the `-s/--sandbox` option to the CLI, which maps directly to `sandbox_mode` in `config.toml`, making config override behavior easier to reason about. Moreover, as you can see in the updates to the various Markdown files, it is much easier to explain how to configure sandboxing when things like `--sandbox read-only` can be used as an example. Relatedly, this cleanup also made it straightforward to add support for a `sandbox` option for Codex when used as an MCP server (see the changes to `mcp-server/src/codex_tool_config.rs`). Fixes https://github.com/openai/codex/issues/1248.
2025-07-07 22:31:30 -07:00
SandboxPolicy::WorkspaceWrite {
writable_roots: vec![PathBuf::from("/my/workspace")],
feat: add support for --sandbox flag (#1476) On a high-level, we try to design `config.toml` so that you don't have to "comment out a lot of stuff" when testing different options. Previously, defining a sandbox policy was somewhat at odds with this principle because you would define the policy as attributes of `[sandbox]` like so: ```toml [sandbox] mode = "workspace-write" writable_roots = [ "/tmp" ] ``` but if you wanted to temporarily change to a read-only sandbox, you might feel compelled to modify your file to be: ```toml [sandbox] mode = "read-only" # mode = "workspace-write" # writable_roots = [ "/tmp" ] ``` Technically, commenting out `writable_roots` would not be strictly necessary, as `mode = "read-only"` would ignore `writable_roots`, but it's still a reasonable thing to do to keep things tidy. Currently, the various values for `mode` do not support that many attributes, so this is not that hard to maintain, but one could imagine this becoming more complex in the future. In this PR, we change Codex CLI so that it no longer recognizes `[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`, and `[sandbox_workspace_write]` is used to further configure the sandbox when when `sandbox_mode = "workspace-write"` is used: ```toml sandbox_mode = "workspace-write" [sandbox_workspace_write] writable_roots = [ "/tmp" ] ``` This feels a bit more future-proof in that it is less tedious to configure different sandboxes: ```toml sandbox_mode = "workspace-write" [sandbox_read_only] # read-only options here... [sandbox_workspace_write] writable_roots = [ "/tmp" ] [sandbox_danger_full_access] # danger-full-access options here... ``` In this scheme, you never need to comment out the configuration for an individual sandbox type: you only need to redefine `sandbox_mode`. Relatedly, previous to this change, a user had to do `-c sandbox.mode=read-only` to change the mode on the command line. With this change, things are arguably a bit cleaner because the equivalent option is `-c sandbox_mode=read-only` (and now `-c sandbox_workspace_write=...` can be set separately). Though more importantly, we introduce the `-s/--sandbox` option to the CLI, which maps directly to `sandbox_mode` in `config.toml`, making config override behavior easier to reason about. Moreover, as you can see in the updates to the various Markdown files, it is much easier to explain how to configure sandboxing when things like `--sandbox read-only` can be used as an example. Relatedly, this cleanup also made it straightforward to add support for a `sandbox` option for Codex when used as an MCP server (see the changes to `mcp-server/src/codex_tool_config.rs`). Fixes https://github.com/openai/codex/issues/1248.
2025-07-07 22:31:30 -07:00
network_access: false,
exclude_tmpdir_env_var: true,
exclude_slash_tmp: true,
feat: add support for --sandbox flag (#1476) On a high-level, we try to design `config.toml` so that you don't have to "comment out a lot of stuff" when testing different options. Previously, defining a sandbox policy was somewhat at odds with this principle because you would define the policy as attributes of `[sandbox]` like so: ```toml [sandbox] mode = "workspace-write" writable_roots = [ "/tmp" ] ``` but if you wanted to temporarily change to a read-only sandbox, you might feel compelled to modify your file to be: ```toml [sandbox] mode = "read-only" # mode = "workspace-write" # writable_roots = [ "/tmp" ] ``` Technically, commenting out `writable_roots` would not be strictly necessary, as `mode = "read-only"` would ignore `writable_roots`, but it's still a reasonable thing to do to keep things tidy. Currently, the various values for `mode` do not support that many attributes, so this is not that hard to maintain, but one could imagine this becoming more complex in the future. In this PR, we change Codex CLI so that it no longer recognizes `[sandbox]`. Instead, it introduces a top-level option, `sandbox_mode`, and `[sandbox_workspace_write]` is used to further configure the sandbox when when `sandbox_mode = "workspace-write"` is used: ```toml sandbox_mode = "workspace-write" [sandbox_workspace_write] writable_roots = [ "/tmp" ] ``` This feels a bit more future-proof in that it is less tedious to configure different sandboxes: ```toml sandbox_mode = "workspace-write" [sandbox_read_only] # read-only options here... [sandbox_workspace_write] writable_roots = [ "/tmp" ] [sandbox_danger_full_access] # danger-full-access options here... ``` In this scheme, you never need to comment out the configuration for an individual sandbox type: you only need to redefine `sandbox_mode`. Relatedly, previous to this change, a user had to do `-c sandbox.mode=read-only` to change the mode on the command line. With this change, things are arguably a bit cleaner because the equivalent option is `-c sandbox_mode=read-only` (and now `-c sandbox_workspace_write=...` can be set separately). Though more importantly, we introduce the `-s/--sandbox` option to the CLI, which maps directly to `sandbox_mode` in `config.toml`, making config override behavior easier to reason about. Moreover, as you can see in the updates to the various Markdown files, it is much easier to explain how to configure sandboxing when things like `--sandbox read-only` can be used as an example. Relatedly, this cleanup also made it straightforward to add support for a `sandbox` option for Codex when used as an MCP server (see the changes to `mcp-server/src/codex_tool_config.rs`). Fixes https://github.com/openai/codex/issues/1248.
2025-07-07 22:31:30 -07:00
},
sandbox_workspace_write_cfg.derive_sandbox_policy(sandbox_mode_override)
);
}
#[tokio::test]
async fn persist_model_selection_updates_defaults() -> anyhow::Result<()> {
let codex_home = TempDir::new()?;
persist_model_selection(
codex_home.path(),
None,
"gpt-5-high-new",
Some(ReasoningEffort::High),
)
.await?;
let serialized =
tokio::fs::read_to_string(codex_home.path().join(CONFIG_TOML_FILE)).await?;
let parsed: ConfigToml = toml::from_str(&serialized)?;
assert_eq!(parsed.model.as_deref(), Some("gpt-5-high-new"));
assert_eq!(parsed.model_reasoning_effort, Some(ReasoningEffort::High));
Ok(())
}
#[tokio::test]
async fn persist_model_selection_overwrites_existing_model() -> anyhow::Result<()> {
let codex_home = TempDir::new()?;
let config_path = codex_home.path().join(CONFIG_TOML_FILE);
tokio::fs::write(
&config_path,
r#"
model = "gpt-5"
model_reasoning_effort = "medium"
[profiles.dev]
model = "gpt-4.1"
"#,
)
.await?;
persist_model_selection(
codex_home.path(),
None,
"o4-mini",
Some(ReasoningEffort::High),
)
.await?;
let serialized = tokio::fs::read_to_string(config_path).await?;
let parsed: ConfigToml = toml::from_str(&serialized)?;
assert_eq!(parsed.model.as_deref(), Some("o4-mini"));
assert_eq!(parsed.model_reasoning_effort, Some(ReasoningEffort::High));
assert_eq!(
parsed
.profiles
.get("dev")
.and_then(|profile| profile.model.as_deref()),
Some("gpt-4.1"),
);
Ok(())
}
#[tokio::test]
async fn persist_model_selection_updates_profile() -> anyhow::Result<()> {
let codex_home = TempDir::new()?;
persist_model_selection(
codex_home.path(),
Some("dev"),
"gpt-5-high-new",
Some(ReasoningEffort::Low),
)
.await?;
let serialized =
tokio::fs::read_to_string(codex_home.path().join(CONFIG_TOML_FILE)).await?;
let parsed: ConfigToml = toml::from_str(&serialized)?;
let profile = parsed
.profiles
.get("dev")
.expect("profile should be created");
assert_eq!(profile.model.as_deref(), Some("gpt-5-high-new"));
assert_eq!(profile.model_reasoning_effort, Some(ReasoningEffort::Low));
Ok(())
}
#[tokio::test]
async fn persist_model_selection_updates_existing_profile() -> anyhow::Result<()> {
let codex_home = TempDir::new()?;
let config_path = codex_home.path().join(CONFIG_TOML_FILE);
tokio::fs::write(
&config_path,
r#"
[profiles.dev]
model = "gpt-4"
model_reasoning_effort = "medium"
[profiles.prod]
model = "gpt-5"
"#,
)
.await?;
persist_model_selection(
codex_home.path(),
Some("dev"),
"o4-high",
Some(ReasoningEffort::Medium),
)
.await?;
let serialized = tokio::fs::read_to_string(config_path).await?;
let parsed: ConfigToml = toml::from_str(&serialized)?;
let dev_profile = parsed
.profiles
.get("dev")
.expect("dev profile should survive updates");
assert_eq!(dev_profile.model.as_deref(), Some("o4-high"));
assert_eq!(
dev_profile.model_reasoning_effort,
Some(ReasoningEffort::Medium)
);
assert_eq!(
parsed
.profiles
.get("prod")
.and_then(|profile| profile.model.as_deref()),
Some("gpt-5"),
);
Ok(())
}
struct PrecedenceTestFixture {
cwd: TempDir,
codex_home: TempDir,
cfg: ConfigToml,
model_provider_map: HashMap<String, ModelProviderInfo>,
openai_provider: ModelProviderInfo,
openai_chat_completions_provider: ModelProviderInfo,
}
impl PrecedenceTestFixture {
fn cwd(&self) -> PathBuf {
self.cwd.path().to_path_buf()
}
fn codex_home(&self) -> PathBuf {
self.codex_home.path().to_path_buf()
}
}
fn create_test_fixture() -> std::io::Result<PrecedenceTestFixture> {
let toml = r#"
model = "o3"
approval_policy = "untrusted"
# Can be used to determine which profile to use if not specified by
# `ConfigOverrides`.
profile = "gpt3"
[model_providers.openai-chat-completions]
name = "OpenAI using Chat Completions"
base_url = "https://api.openai.com/v1"
env_key = "OPENAI_API_KEY"
wire_api = "chat"
request_max_retries = 4 # retry failed HTTP requests
stream_max_retries = 10 # retry dropped SSE streams
stream_idle_timeout_ms = 300000 # 5m idle timeout
[profiles.o3]
model = "o3"
model_provider = "openai"
approval_policy = "never"
model_reasoning_effort = "high"
model_reasoning_summary = "detailed"
[profiles.gpt3]
model = "gpt-3.5-turbo"
model_provider = "openai-chat-completions"
[profiles.zdr]
model = "o3"
model_provider = "openai"
approval_policy = "on-failure"
[profiles.gpt5]
model = "gpt-5"
model_provider = "openai"
approval_policy = "on-failure"
model_reasoning_effort = "high"
model_reasoning_summary = "detailed"
model_verbosity = "high"
"#;
let cfg: ConfigToml = toml::from_str(toml).expect("TOML deserialization should succeed");
// Use a temporary directory for the cwd so it does not contain an
// AGENTS.md file.
let cwd_temp_dir = TempDir::new().unwrap();
let cwd = cwd_temp_dir.path().to_path_buf();
// Make it look like a Git repo so it does not search for AGENTS.md in
// a parent folder, either.
std::fs::write(cwd.join(".git"), "gitdir: nowhere")?;
let codex_home_temp_dir = TempDir::new().unwrap();
let openai_chat_completions_provider = ModelProviderInfo {
name: "OpenAI using Chat Completions".to_string(),
base_url: Some("https://api.openai.com/v1".to_string()),
env_key: Some("OPENAI_API_KEY".to_string()),
wire_api: crate::WireApi::Chat,
env_key_instructions: None,
query_params: None,
http_headers: None,
env_http_headers: None,
request_max_retries: Some(4),
stream_max_retries: Some(10),
stream_idle_timeout_ms: Some(300_000),
requires_openai_auth: false,
};
let model_provider_map = {
let mut model_provider_map = built_in_model_providers();
model_provider_map.insert(
"openai-chat-completions".to_string(),
openai_chat_completions_provider.clone(),
);
model_provider_map
};
let openai_provider = model_provider_map
.get("openai")
.expect("openai provider should exist")
.clone();
Ok(PrecedenceTestFixture {
cwd: cwd_temp_dir,
codex_home: codex_home_temp_dir,
cfg,
model_provider_map,
openai_provider,
openai_chat_completions_provider,
})
}
/// Users can specify config values at multiple levels that have the
/// following precedence:
///
/// 1. custom command-line argument, e.g. `--model o3`
/// 2. as part of a profile, where the `--profile` is specified via a CLI
/// (or in the config file itself)
/// 3. as an entry in `config.toml`, e.g. `model = "o3"`
/// 4. the default value for a required field defined in code, e.g.,
/// `crate::flags::OPENAI_DEFAULT_MODEL`
///
/// Note that profiles are the recommended way to specify a group of
/// configuration options together.
#[test]
fn test_precedence_fixture_with_o3_profile() -> std::io::Result<()> {
let fixture = create_test_fixture()?;
let o3_profile_overrides = ConfigOverrides {
config_profile: Some("o3".to_string()),
cwd: Some(fixture.cwd()),
..Default::default()
};
let o3_profile_config: Config = Config::load_from_base_config_with_overrides(
fixture.cfg.clone(),
o3_profile_overrides,
fixture.codex_home(),
)?;
assert_eq!(
Config {
model: "o3".to_string(),
Review Mode (Core) (#3401) ## 📝 Review Mode -- Core This PR introduces the Core implementation for Review mode: - New op `Op::Review { prompt: String }:` spawns a child review task with isolated context, a review‑specific system prompt, and a `Config.review_model`. - `EnteredReviewMode`: emitted when the child review session starts. Every event from this point onwards reflects the review session. - `ExitedReviewMode(Option<ReviewOutputEvent>)`: emitted when the review finishes or is interrupted, with optional structured findings: ```json { "findings": [ { "title": "<≤ 80 chars, imperative>", "body": "<valid Markdown explaining *why* this is a problem; cite files/lines/functions>", "confidence_score": <float 0.0-1.0>, "priority": <int 0-3>, "code_location": { "absolute_file_path": "<file path>", "line_range": {"start": <int>, "end": <int>} } } ], "overall_correctness": "patch is correct" | "patch is incorrect", "overall_explanation": "<1-3 sentence explanation justifying the overall_correctness verdict>", "overall_confidence_score": <float 0.0-1.0> } ``` ## Questions ### Why separate out its own message history? We want the review thread to match the training of our review models as much as possible -- that means using a custom prompt, removing user instructions, and starting a clean chat history. We also want to make sure the review thread doesn't leak into the parent thread. ### Why do this as a mode, vs. sub-agents? 1. We want review to be a synchronous task, so it's fine for now to do a bespoke implementation. 2. We're still unclear about the final structure for sub-agents. We'd prefer to land this quickly and then refactor into sub-agents without rushing that implementation.
2025-09-12 16:25:10 -07:00
review_model: "gpt-5".to_string(),
model_family: find_family_for_model("o3").expect("known model slug"),
feat: show number of tokens remaining in UI (#1388) When using the OpenAI Responses API, we now record the `usage` field for a `"response.completed"` event, which includes metrics about the number of tokens consumed. We also introduce `openai_model_info.rs`, which includes current data about the most common OpenAI models available via the API (specifically `context_window` and `max_output_tokens`). If Codex does not recognize the model, you can set `model_context_window` and `model_max_output_tokens` explicitly in `config.toml`. When then introduce a new event type to `protocol.rs`, `TokenCount`, which includes the `TokenUsage` for the most recent turn. Finally, we update the TUI to record the running sum of tokens used so the percentage of available context window remaining can be reported via the placeholder text for the composer: ![Screenshot 2025-06-25 at 11 20 55 PM](https://github.com/user-attachments/assets/6fd6982f-7247-4f14-84b2-2e600cb1fd49) We could certainly get much fancier with this (such as reporting the estimated cost of the conversation), but for now, we are just trying to achieve feature parity with the TypeScript CLI. Though arguably this improves upon the TypeScript CLI, as the TypeScript CLI uses heuristics to estimate the number of tokens used rather than using the `usage` information directly: https://github.com/openai/codex/blob/296996d74e345b1b05d8c3451a06ace21c5ada96/codex-cli/src/utils/approximate-tokens-used.ts#L3-L16 Fixes https://github.com/openai/codex/issues/1242
2025-06-25 23:31:11 -07:00
model_context_window: Some(200_000),
model_max_output_tokens: Some(100_000),
model_auto_compact_token_limit: None,
model_provider_id: "openai".to_string(),
model_provider: fixture.openai_provider.clone(),
approval_policy: AskForApproval::Never,
sandbox_policy: SandboxPolicy::new_read_only_policy(),
feat: introduce support for shell_environment_policy in config.toml (#1061) To date, when handling `shell` and `local_shell` tool calls, we were spawning new processes using the environment inherited from the Codex process itself. This means that the sensitive `OPENAI_API_KEY` that Codex needs to talk to OpenAI models was made available to everything run by `shell` and `local_shell`. While there are cases where that might be useful, it does not seem like a good default. This PR introduces a complex `shell_environment_policy` config option to control the `env` used with these tool calls. It is inevitably a bit complex so that it is possible to override individual components of the policy so without having to restate the entire thing. Details are in the updated `README.md` in this PR, but here is the relevant bit that explains the individual fields of `shell_environment_policy`: | Field | Type | Default | Description | | ------------------------- | -------------------------- | ------- | ----------------------------------------------------------------------------------------------------------------------------------------------- | | `inherit` | string | `core` | Starting template for the environment:<br>`core` (`HOME`, `PATH`, `USER`, …), `all` (clone full parent env), or `none` (start empty). | | `ignore_default_excludes` | boolean | `false` | When `false`, Codex removes any var whose **name** contains `KEY`, `SECRET`, or `TOKEN` (case-insensitive) before other rules run. | | `exclude` | array&lt;string&gt; | `[]` | Case-insensitive glob patterns to drop after the default filter.<br>Examples: `"AWS_*"`, `"AZURE_*"`. | | `set` | table&lt;string,string&gt; | `{}` | Explicit key/value overrides or additions – always win over inherited values. | | `include_only` | array&lt;string&gt; | `[]` | If non-empty, a whitelist of patterns; only variables that match _one_ pattern survive the final step. (Generally used with `inherit = "all"`.) | In particular, note that the default is `inherit = "core"`, so: * if you have extra env variables that you want to inherit from the parent process, use `inherit = "all"` and then specify `include_only` * if you have extra env variables where you want to hardcode the values, the default `inherit = "core"` will work fine, but then you need to specify `set` This configuration is not battle-tested, so we will probably still have to play with it a bit. `core/src/exec_env.rs` has the critical business logic as well as unit tests. Though if nothing else, previous to this change: ``` $ cargo run --bin codex -- debug seatbelt -- printenv OPENAI_API_KEY # ...prints OPENAI_API_KEY... ``` But after this change it does not print anything (as desired). One final thing to call out about this PR is that the `configure_command!` macro we use in `core/src/exec.rs` has to do some complex logic with respect to how it builds up the `env` for the process being spawned under Landlock/seccomp. Specifically, doing `cmd.env_clear()` followed by `cmd.envs(&$env_map)` (which is arguably the most intuitive way to do it) caused the Landlock unit tests to fail because the processes spawned by the unit tests started failing in unexpected ways! If we forgo `env_clear()` in favor of updating env vars one at a time, the tests still pass. The comment in the code talks about this a bit, and while I would like to investigate this more, I need to move on for the moment, but I do plan to come back to it to fully understand what is going on. For example, this suggests that we might not be able to spawn a C program that calls `env_clear()`, which would be...weird. We may still have to fiddle with our Landlock config if that is the case.
2025-05-22 09:51:19 -07:00
shell_environment_policy: ShellEnvironmentPolicy::default(),
user_instructions: None,
notify: None,
cwd: fixture.cwd(),
mcp_servers: HashMap::new(),
model_providers: fixture.model_provider_map.clone(),
project_doc_max_bytes: PROJECT_DOC_MAX_BYTES,
codex_home: fixture.codex_home(),
feat: record messages from user in ~/.codex/history.jsonl (#939) This is a large change to support a "history" feature like you would expect in a shell like Bash. History events are recorded in `$CODEX_HOME/history.jsonl`. Because it is a JSONL file, it is straightforward to append new entries (as opposed to the TypeScript file that uses `$CODEX_HOME/history.json`, so to be valid JSON, each new entry entails rewriting the entire file). Because it is possible for there to be multiple instances of Codex CLI writing to `history.jsonl` at once, we use advisory file locking when working with `history.jsonl` in `codex-rs/core/src/message_history.rs`. Because we believe history is a sufficiently useful feature, we enable it by default. Though to provide some safety, we set the file permissions of `history.jsonl` to be `o600` so that other users on the system cannot read the user's history. We do not yet support a default list of `SENSITIVE_PATTERNS` as the TypeScript CLI does: https://github.com/openai/codex/blob/3fdf9df1335ac9501e3fb0e61715359145711e8b/codex-cli/src/utils/storage/command-history.ts#L10-L17 We are going to take a more conservative approach to this list in the Rust CLI. For example, while `/\b[A-Za-z0-9-_]{20,}\b/` might exclude sensitive information like API tokens, it would also exclude valuable information such as references to Git commits. As noted in the updated documentation, users can opt-out of history by adding the following to `config.toml`: ```toml [history] persistence = "none" ``` Because `history.jsonl` could, in theory, be quite large, we take a[n arguably overly pedantic] approach in reading history entries into memory. Specifically, we start by telling the client the current number of entries in the history file (`history_entry_count`) as well as the inode (`history_log_id`) of `history.jsonl` (see the new fields on `SessionConfiguredEvent`). The client is responsible for keeping new entries in memory to create a "local history," but if the user hits up enough times to go "past" the end of local history, then the client should use the new `GetHistoryEntryRequest` in the protocol to fetch older entries. Specifically, it should pass the `history_log_id` it was given originally and work backwards from `history_entry_count`. (It should really fetch history in batches rather than one-at-a-time, but that is something we can improve upon in subsequent PRs.) The motivation behind this crazy scheme is that it is designed to defend against: * The `history.jsonl` being truncated during the session such that the index into the history is no longer consistent with what had been read up to that point. We do not yet have logic to enforce a `max_bytes` for `history.jsonl`, but once we do, we will aspire to implement it in a way that should result in a new inode for the file on most systems. * New items from concurrent Codex CLI sessions amending to the history. Because, in absence of truncation, `history.jsonl` is an append-only log, so long as the client reads backwards from `history_entry_count`, it should always get a consistent view of history. (That said, it will not be able to read _new_ commands from concurrent sessions, but perhaps we will introduce a `/` command to reload latest history or something down the road.) Admittedly, my testing of this feature thus far has been fairly light. I expect we will find bugs and introduce enhancements/fixes going forward.
2025-05-15 16:26:23 -07:00
history: History::default(),
file_opener: UriBasedFileOpener::VsCode,
codex_linux_sandbox_exe: None,
hide_agent_reasoning: false,
show_raw_agent_reasoning: false,
model_reasoning_effort: Some(ReasoningEffort::High),
model_reasoning_summary: ReasoningSummary::Detailed,
feat(gpt5): add model_verbosity for GPT‑5 via Responses API (#2108) **Summary** - Adds `model_verbosity` config (values: low, medium, high). - Sends `text.verbosity` only for GPT‑5 family models via the Responses API. - Updates docs and adds serialization tests. **Motivation** - GPT‑5 introduces a verbosity control to steer output length/detail without pro mpt surgery. - Exposing it as a config knob keeps prompts stable and makes behavior explicit and repeatable. **Changes** - Config: - Added `Verbosity` enum (low|medium|high). - Added optional `model_verbosity` to `ConfigToml`, `Config`, and `ConfigProfi le`. - Request wiring: - Extended `ResponsesApiRequest` with optional `text` object. - Populates `text.verbosity` only when model family is `gpt-5`; omitted otherw ise. - Tests: - Verifies `text.verbosity` serializes when set and is omitted when not set. - Docs: - Added “GPT‑5 Verbosity” section in `codex-rs/README.md`. - Added `model_verbosity` section to `codex-rs/config.md`. **Usage** - In `~/.codex/config.toml`: - `model = "gpt-5"` - `model_verbosity = "low"` (or `"medium"` default, `"high"`) - CLI override example: - `codex -c model="gpt-5" -c model_verbosity="high"` **API Impact** - Requests to GPT‑5 via Responses API include: `text: { verbosity: "low|medium|h igh" }` when configured. - For legacy models or Chat Completions providers, `text` is omitted. **Backward Compatibility** - Default behavior unchanged when `model_verbosity` is not set (server default “ medium”). **Testing** - Added unit tests for serialization/omission of `text.verbosity`. - Ran `cargo fmt` and `cargo test --all-features` (all green). **Docs** - `README.md`: new “GPT‑5 Verbosity” note under Config with example. - `config.md`: new `model_verbosity` section. **Out of Scope** - No changes to temperature/top_p or other GPT‑5 parameters. - No changes to Chat Completions wiring. **Risks / Notes** - If OpenAI changes the wire shape for verbosity, we may need to update `Respons esApiRequest`. - Behavior gated to `gpt-5` model family to avoid unexpected effects elsewhere. **Checklist** - [x] Code gated to GPT‑5 family only - [x] Docs updated (`README.md`, `config.md`) - [x] Tests added and passing - [x] Formatting applied Release note: Add `model_verbosity` config to control GPT‑5 output verbosity via the Responses API (low|medium|high).
2025-08-22 17:12:10 +01:00
model_verbosity: None,
chatgpt_base_url: "https://chatgpt.com/backend-api/".to_string(),
experimental_resume: None,
base_instructions: None,
include_plan_tool: false,
include_apply_patch_tool: false,
tools_web_search_request: false,
use_experimental_streamable_shell_tool: false,
use_experimental_unified_exec_tool: false,
include_view_image_tool: true,
active_profile: Some("o3".to_string()),
disable_paste_burst: false,
},
o3_profile_config
);
Ok(())
}
#[test]
fn test_precedence_fixture_with_gpt3_profile() -> std::io::Result<()> {
let fixture = create_test_fixture()?;
let gpt3_profile_overrides = ConfigOverrides {
config_profile: Some("gpt3".to_string()),
cwd: Some(fixture.cwd()),
..Default::default()
};
let gpt3_profile_config = Config::load_from_base_config_with_overrides(
fixture.cfg.clone(),
gpt3_profile_overrides,
fixture.codex_home(),
)?;
let expected_gpt3_profile_config = Config {
model: "gpt-3.5-turbo".to_string(),
Review Mode (Core) (#3401) ## 📝 Review Mode -- Core This PR introduces the Core implementation for Review mode: - New op `Op::Review { prompt: String }:` spawns a child review task with isolated context, a review‑specific system prompt, and a `Config.review_model`. - `EnteredReviewMode`: emitted when the child review session starts. Every event from this point onwards reflects the review session. - `ExitedReviewMode(Option<ReviewOutputEvent>)`: emitted when the review finishes or is interrupted, with optional structured findings: ```json { "findings": [ { "title": "<≤ 80 chars, imperative>", "body": "<valid Markdown explaining *why* this is a problem; cite files/lines/functions>", "confidence_score": <float 0.0-1.0>, "priority": <int 0-3>, "code_location": { "absolute_file_path": "<file path>", "line_range": {"start": <int>, "end": <int>} } } ], "overall_correctness": "patch is correct" | "patch is incorrect", "overall_explanation": "<1-3 sentence explanation justifying the overall_correctness verdict>", "overall_confidence_score": <float 0.0-1.0> } ``` ## Questions ### Why separate out its own message history? We want the review thread to match the training of our review models as much as possible -- that means using a custom prompt, removing user instructions, and starting a clean chat history. We also want to make sure the review thread doesn't leak into the parent thread. ### Why do this as a mode, vs. sub-agents? 1. We want review to be a synchronous task, so it's fine for now to do a bespoke implementation. 2. We're still unclear about the final structure for sub-agents. We'd prefer to land this quickly and then refactor into sub-agents without rushing that implementation.
2025-09-12 16:25:10 -07:00
review_model: "gpt-5".to_string(),
model_family: find_family_for_model("gpt-3.5-turbo").expect("known model slug"),
feat: show number of tokens remaining in UI (#1388) When using the OpenAI Responses API, we now record the `usage` field for a `"response.completed"` event, which includes metrics about the number of tokens consumed. We also introduce `openai_model_info.rs`, which includes current data about the most common OpenAI models available via the API (specifically `context_window` and `max_output_tokens`). If Codex does not recognize the model, you can set `model_context_window` and `model_max_output_tokens` explicitly in `config.toml`. When then introduce a new event type to `protocol.rs`, `TokenCount`, which includes the `TokenUsage` for the most recent turn. Finally, we update the TUI to record the running sum of tokens used so the percentage of available context window remaining can be reported via the placeholder text for the composer: ![Screenshot 2025-06-25 at 11 20 55 PM](https://github.com/user-attachments/assets/6fd6982f-7247-4f14-84b2-2e600cb1fd49) We could certainly get much fancier with this (such as reporting the estimated cost of the conversation), but for now, we are just trying to achieve feature parity with the TypeScript CLI. Though arguably this improves upon the TypeScript CLI, as the TypeScript CLI uses heuristics to estimate the number of tokens used rather than using the `usage` information directly: https://github.com/openai/codex/blob/296996d74e345b1b05d8c3451a06ace21c5ada96/codex-cli/src/utils/approximate-tokens-used.ts#L3-L16 Fixes https://github.com/openai/codex/issues/1242
2025-06-25 23:31:11 -07:00
model_context_window: Some(16_385),
model_max_output_tokens: Some(4_096),
model_auto_compact_token_limit: None,
model_provider_id: "openai-chat-completions".to_string(),
model_provider: fixture.openai_chat_completions_provider.clone(),
approval_policy: AskForApproval::UnlessTrusted,
sandbox_policy: SandboxPolicy::new_read_only_policy(),
feat: introduce support for shell_environment_policy in config.toml (#1061) To date, when handling `shell` and `local_shell` tool calls, we were spawning new processes using the environment inherited from the Codex process itself. This means that the sensitive `OPENAI_API_KEY` that Codex needs to talk to OpenAI models was made available to everything run by `shell` and `local_shell`. While there are cases where that might be useful, it does not seem like a good default. This PR introduces a complex `shell_environment_policy` config option to control the `env` used with these tool calls. It is inevitably a bit complex so that it is possible to override individual components of the policy so without having to restate the entire thing. Details are in the updated `README.md` in this PR, but here is the relevant bit that explains the individual fields of `shell_environment_policy`: | Field | Type | Default | Description | | ------------------------- | -------------------------- | ------- | ----------------------------------------------------------------------------------------------------------------------------------------------- | | `inherit` | string | `core` | Starting template for the environment:<br>`core` (`HOME`, `PATH`, `USER`, …), `all` (clone full parent env), or `none` (start empty). | | `ignore_default_excludes` | boolean | `false` | When `false`, Codex removes any var whose **name** contains `KEY`, `SECRET`, or `TOKEN` (case-insensitive) before other rules run. | | `exclude` | array&lt;string&gt; | `[]` | Case-insensitive glob patterns to drop after the default filter.<br>Examples: `"AWS_*"`, `"AZURE_*"`. | | `set` | table&lt;string,string&gt; | `{}` | Explicit key/value overrides or additions – always win over inherited values. | | `include_only` | array&lt;string&gt; | `[]` | If non-empty, a whitelist of patterns; only variables that match _one_ pattern survive the final step. (Generally used with `inherit = "all"`.) | In particular, note that the default is `inherit = "core"`, so: * if you have extra env variables that you want to inherit from the parent process, use `inherit = "all"` and then specify `include_only` * if you have extra env variables where you want to hardcode the values, the default `inherit = "core"` will work fine, but then you need to specify `set` This configuration is not battle-tested, so we will probably still have to play with it a bit. `core/src/exec_env.rs` has the critical business logic as well as unit tests. Though if nothing else, previous to this change: ``` $ cargo run --bin codex -- debug seatbelt -- printenv OPENAI_API_KEY # ...prints OPENAI_API_KEY... ``` But after this change it does not print anything (as desired). One final thing to call out about this PR is that the `configure_command!` macro we use in `core/src/exec.rs` has to do some complex logic with respect to how it builds up the `env` for the process being spawned under Landlock/seccomp. Specifically, doing `cmd.env_clear()` followed by `cmd.envs(&$env_map)` (which is arguably the most intuitive way to do it) caused the Landlock unit tests to fail because the processes spawned by the unit tests started failing in unexpected ways! If we forgo `env_clear()` in favor of updating env vars one at a time, the tests still pass. The comment in the code talks about this a bit, and while I would like to investigate this more, I need to move on for the moment, but I do plan to come back to it to fully understand what is going on. For example, this suggests that we might not be able to spawn a C program that calls `env_clear()`, which would be...weird. We may still have to fiddle with our Landlock config if that is the case.
2025-05-22 09:51:19 -07:00
shell_environment_policy: ShellEnvironmentPolicy::default(),
user_instructions: None,
notify: None,
cwd: fixture.cwd(),
mcp_servers: HashMap::new(),
model_providers: fixture.model_provider_map.clone(),
project_doc_max_bytes: PROJECT_DOC_MAX_BYTES,
codex_home: fixture.codex_home(),
feat: record messages from user in ~/.codex/history.jsonl (#939) This is a large change to support a "history" feature like you would expect in a shell like Bash. History events are recorded in `$CODEX_HOME/history.jsonl`. Because it is a JSONL file, it is straightforward to append new entries (as opposed to the TypeScript file that uses `$CODEX_HOME/history.json`, so to be valid JSON, each new entry entails rewriting the entire file). Because it is possible for there to be multiple instances of Codex CLI writing to `history.jsonl` at once, we use advisory file locking when working with `history.jsonl` in `codex-rs/core/src/message_history.rs`. Because we believe history is a sufficiently useful feature, we enable it by default. Though to provide some safety, we set the file permissions of `history.jsonl` to be `o600` so that other users on the system cannot read the user's history. We do not yet support a default list of `SENSITIVE_PATTERNS` as the TypeScript CLI does: https://github.com/openai/codex/blob/3fdf9df1335ac9501e3fb0e61715359145711e8b/codex-cli/src/utils/storage/command-history.ts#L10-L17 We are going to take a more conservative approach to this list in the Rust CLI. For example, while `/\b[A-Za-z0-9-_]{20,}\b/` might exclude sensitive information like API tokens, it would also exclude valuable information such as references to Git commits. As noted in the updated documentation, users can opt-out of history by adding the following to `config.toml`: ```toml [history] persistence = "none" ``` Because `history.jsonl` could, in theory, be quite large, we take a[n arguably overly pedantic] approach in reading history entries into memory. Specifically, we start by telling the client the current number of entries in the history file (`history_entry_count`) as well as the inode (`history_log_id`) of `history.jsonl` (see the new fields on `SessionConfiguredEvent`). The client is responsible for keeping new entries in memory to create a "local history," but if the user hits up enough times to go "past" the end of local history, then the client should use the new `GetHistoryEntryRequest` in the protocol to fetch older entries. Specifically, it should pass the `history_log_id` it was given originally and work backwards from `history_entry_count`. (It should really fetch history in batches rather than one-at-a-time, but that is something we can improve upon in subsequent PRs.) The motivation behind this crazy scheme is that it is designed to defend against: * The `history.jsonl` being truncated during the session such that the index into the history is no longer consistent with what had been read up to that point. We do not yet have logic to enforce a `max_bytes` for `history.jsonl`, but once we do, we will aspire to implement it in a way that should result in a new inode for the file on most systems. * New items from concurrent Codex CLI sessions amending to the history. Because, in absence of truncation, `history.jsonl` is an append-only log, so long as the client reads backwards from `history_entry_count`, it should always get a consistent view of history. (That said, it will not be able to read _new_ commands from concurrent sessions, but perhaps we will introduce a `/` command to reload latest history or something down the road.) Admittedly, my testing of this feature thus far has been fairly light. I expect we will find bugs and introduce enhancements/fixes going forward.
2025-05-15 16:26:23 -07:00
history: History::default(),
file_opener: UriBasedFileOpener::VsCode,
codex_linux_sandbox_exe: None,
hide_agent_reasoning: false,
show_raw_agent_reasoning: false,
model_reasoning_effort: None,
model_reasoning_summary: ReasoningSummary::default(),
feat(gpt5): add model_verbosity for GPT‑5 via Responses API (#2108) **Summary** - Adds `model_verbosity` config (values: low, medium, high). - Sends `text.verbosity` only for GPT‑5 family models via the Responses API. - Updates docs and adds serialization tests. **Motivation** - GPT‑5 introduces a verbosity control to steer output length/detail without pro mpt surgery. - Exposing it as a config knob keeps prompts stable and makes behavior explicit and repeatable. **Changes** - Config: - Added `Verbosity` enum (low|medium|high). - Added optional `model_verbosity` to `ConfigToml`, `Config`, and `ConfigProfi le`. - Request wiring: - Extended `ResponsesApiRequest` with optional `text` object. - Populates `text.verbosity` only when model family is `gpt-5`; omitted otherw ise. - Tests: - Verifies `text.verbosity` serializes when set and is omitted when not set. - Docs: - Added “GPT‑5 Verbosity” section in `codex-rs/README.md`. - Added `model_verbosity` section to `codex-rs/config.md`. **Usage** - In `~/.codex/config.toml`: - `model = "gpt-5"` - `model_verbosity = "low"` (or `"medium"` default, `"high"`) - CLI override example: - `codex -c model="gpt-5" -c model_verbosity="high"` **API Impact** - Requests to GPT‑5 via Responses API include: `text: { verbosity: "low|medium|h igh" }` when configured. - For legacy models or Chat Completions providers, `text` is omitted. **Backward Compatibility** - Default behavior unchanged when `model_verbosity` is not set (server default “ medium”). **Testing** - Added unit tests for serialization/omission of `text.verbosity`. - Ran `cargo fmt` and `cargo test --all-features` (all green). **Docs** - `README.md`: new “GPT‑5 Verbosity” note under Config with example. - `config.md`: new `model_verbosity` section. **Out of Scope** - No changes to temperature/top_p or other GPT‑5 parameters. - No changes to Chat Completions wiring. **Risks / Notes** - If OpenAI changes the wire shape for verbosity, we may need to update `Respons esApiRequest`. - Behavior gated to `gpt-5` model family to avoid unexpected effects elsewhere. **Checklist** - [x] Code gated to GPT‑5 family only - [x] Docs updated (`README.md`, `config.md`) - [x] Tests added and passing - [x] Formatting applied Release note: Add `model_verbosity` config to control GPT‑5 output verbosity via the Responses API (low|medium|high).
2025-08-22 17:12:10 +01:00
model_verbosity: None,
chatgpt_base_url: "https://chatgpt.com/backend-api/".to_string(),
experimental_resume: None,
base_instructions: None,
include_plan_tool: false,
include_apply_patch_tool: false,
tools_web_search_request: false,
use_experimental_streamable_shell_tool: false,
use_experimental_unified_exec_tool: false,
include_view_image_tool: true,
active_profile: Some("gpt3".to_string()),
disable_paste_burst: false,
};
assert_eq!(expected_gpt3_profile_config, gpt3_profile_config);
// Verify that loading without specifying a profile in ConfigOverrides
// uses the default profile from the config file (which is "gpt3").
let default_profile_overrides = ConfigOverrides {
cwd: Some(fixture.cwd()),
..Default::default()
};
let default_profile_config = Config::load_from_base_config_with_overrides(
fixture.cfg.clone(),
default_profile_overrides,
fixture.codex_home(),
)?;
assert_eq!(expected_gpt3_profile_config, default_profile_config);
Ok(())
}
#[test]
fn test_precedence_fixture_with_zdr_profile() -> std::io::Result<()> {
let fixture = create_test_fixture()?;
let zdr_profile_overrides = ConfigOverrides {
config_profile: Some("zdr".to_string()),
cwd: Some(fixture.cwd()),
..Default::default()
};
let zdr_profile_config = Config::load_from_base_config_with_overrides(
fixture.cfg.clone(),
zdr_profile_overrides,
fixture.codex_home(),
)?;
let expected_zdr_profile_config = Config {
model: "o3".to_string(),
Review Mode (Core) (#3401) ## 📝 Review Mode -- Core This PR introduces the Core implementation for Review mode: - New op `Op::Review { prompt: String }:` spawns a child review task with isolated context, a review‑specific system prompt, and a `Config.review_model`. - `EnteredReviewMode`: emitted when the child review session starts. Every event from this point onwards reflects the review session. - `ExitedReviewMode(Option<ReviewOutputEvent>)`: emitted when the review finishes or is interrupted, with optional structured findings: ```json { "findings": [ { "title": "<≤ 80 chars, imperative>", "body": "<valid Markdown explaining *why* this is a problem; cite files/lines/functions>", "confidence_score": <float 0.0-1.0>, "priority": <int 0-3>, "code_location": { "absolute_file_path": "<file path>", "line_range": {"start": <int>, "end": <int>} } } ], "overall_correctness": "patch is correct" | "patch is incorrect", "overall_explanation": "<1-3 sentence explanation justifying the overall_correctness verdict>", "overall_confidence_score": <float 0.0-1.0> } ``` ## Questions ### Why separate out its own message history? We want the review thread to match the training of our review models as much as possible -- that means using a custom prompt, removing user instructions, and starting a clean chat history. We also want to make sure the review thread doesn't leak into the parent thread. ### Why do this as a mode, vs. sub-agents? 1. We want review to be a synchronous task, so it's fine for now to do a bespoke implementation. 2. We're still unclear about the final structure for sub-agents. We'd prefer to land this quickly and then refactor into sub-agents without rushing that implementation.
2025-09-12 16:25:10 -07:00
review_model: "gpt-5".to_string(),
model_family: find_family_for_model("o3").expect("known model slug"),
feat: show number of tokens remaining in UI (#1388) When using the OpenAI Responses API, we now record the `usage` field for a `"response.completed"` event, which includes metrics about the number of tokens consumed. We also introduce `openai_model_info.rs`, which includes current data about the most common OpenAI models available via the API (specifically `context_window` and `max_output_tokens`). If Codex does not recognize the model, you can set `model_context_window` and `model_max_output_tokens` explicitly in `config.toml`. When then introduce a new event type to `protocol.rs`, `TokenCount`, which includes the `TokenUsage` for the most recent turn. Finally, we update the TUI to record the running sum of tokens used so the percentage of available context window remaining can be reported via the placeholder text for the composer: ![Screenshot 2025-06-25 at 11 20 55 PM](https://github.com/user-attachments/assets/6fd6982f-7247-4f14-84b2-2e600cb1fd49) We could certainly get much fancier with this (such as reporting the estimated cost of the conversation), but for now, we are just trying to achieve feature parity with the TypeScript CLI. Though arguably this improves upon the TypeScript CLI, as the TypeScript CLI uses heuristics to estimate the number of tokens used rather than using the `usage` information directly: https://github.com/openai/codex/blob/296996d74e345b1b05d8c3451a06ace21c5ada96/codex-cli/src/utils/approximate-tokens-used.ts#L3-L16 Fixes https://github.com/openai/codex/issues/1242
2025-06-25 23:31:11 -07:00
model_context_window: Some(200_000),
model_max_output_tokens: Some(100_000),
model_auto_compact_token_limit: None,
model_provider_id: "openai".to_string(),
model_provider: fixture.openai_provider.clone(),
approval_policy: AskForApproval::OnFailure,
sandbox_policy: SandboxPolicy::new_read_only_policy(),
feat: introduce support for shell_environment_policy in config.toml (#1061) To date, when handling `shell` and `local_shell` tool calls, we were spawning new processes using the environment inherited from the Codex process itself. This means that the sensitive `OPENAI_API_KEY` that Codex needs to talk to OpenAI models was made available to everything run by `shell` and `local_shell`. While there are cases where that might be useful, it does not seem like a good default. This PR introduces a complex `shell_environment_policy` config option to control the `env` used with these tool calls. It is inevitably a bit complex so that it is possible to override individual components of the policy so without having to restate the entire thing. Details are in the updated `README.md` in this PR, but here is the relevant bit that explains the individual fields of `shell_environment_policy`: | Field | Type | Default | Description | | ------------------------- | -------------------------- | ------- | ----------------------------------------------------------------------------------------------------------------------------------------------- | | `inherit` | string | `core` | Starting template for the environment:<br>`core` (`HOME`, `PATH`, `USER`, …), `all` (clone full parent env), or `none` (start empty). | | `ignore_default_excludes` | boolean | `false` | When `false`, Codex removes any var whose **name** contains `KEY`, `SECRET`, or `TOKEN` (case-insensitive) before other rules run. | | `exclude` | array&lt;string&gt; | `[]` | Case-insensitive glob patterns to drop after the default filter.<br>Examples: `"AWS_*"`, `"AZURE_*"`. | | `set` | table&lt;string,string&gt; | `{}` | Explicit key/value overrides or additions – always win over inherited values. | | `include_only` | array&lt;string&gt; | `[]` | If non-empty, a whitelist of patterns; only variables that match _one_ pattern survive the final step. (Generally used with `inherit = "all"`.) | In particular, note that the default is `inherit = "core"`, so: * if you have extra env variables that you want to inherit from the parent process, use `inherit = "all"` and then specify `include_only` * if you have extra env variables where you want to hardcode the values, the default `inherit = "core"` will work fine, but then you need to specify `set` This configuration is not battle-tested, so we will probably still have to play with it a bit. `core/src/exec_env.rs` has the critical business logic as well as unit tests. Though if nothing else, previous to this change: ``` $ cargo run --bin codex -- debug seatbelt -- printenv OPENAI_API_KEY # ...prints OPENAI_API_KEY... ``` But after this change it does not print anything (as desired). One final thing to call out about this PR is that the `configure_command!` macro we use in `core/src/exec.rs` has to do some complex logic with respect to how it builds up the `env` for the process being spawned under Landlock/seccomp. Specifically, doing `cmd.env_clear()` followed by `cmd.envs(&$env_map)` (which is arguably the most intuitive way to do it) caused the Landlock unit tests to fail because the processes spawned by the unit tests started failing in unexpected ways! If we forgo `env_clear()` in favor of updating env vars one at a time, the tests still pass. The comment in the code talks about this a bit, and while I would like to investigate this more, I need to move on for the moment, but I do plan to come back to it to fully understand what is going on. For example, this suggests that we might not be able to spawn a C program that calls `env_clear()`, which would be...weird. We may still have to fiddle with our Landlock config if that is the case.
2025-05-22 09:51:19 -07:00
shell_environment_policy: ShellEnvironmentPolicy::default(),
user_instructions: None,
notify: None,
cwd: fixture.cwd(),
mcp_servers: HashMap::new(),
model_providers: fixture.model_provider_map.clone(),
project_doc_max_bytes: PROJECT_DOC_MAX_BYTES,
codex_home: fixture.codex_home(),
feat: record messages from user in ~/.codex/history.jsonl (#939) This is a large change to support a "history" feature like you would expect in a shell like Bash. History events are recorded in `$CODEX_HOME/history.jsonl`. Because it is a JSONL file, it is straightforward to append new entries (as opposed to the TypeScript file that uses `$CODEX_HOME/history.json`, so to be valid JSON, each new entry entails rewriting the entire file). Because it is possible for there to be multiple instances of Codex CLI writing to `history.jsonl` at once, we use advisory file locking when working with `history.jsonl` in `codex-rs/core/src/message_history.rs`. Because we believe history is a sufficiently useful feature, we enable it by default. Though to provide some safety, we set the file permissions of `history.jsonl` to be `o600` so that other users on the system cannot read the user's history. We do not yet support a default list of `SENSITIVE_PATTERNS` as the TypeScript CLI does: https://github.com/openai/codex/blob/3fdf9df1335ac9501e3fb0e61715359145711e8b/codex-cli/src/utils/storage/command-history.ts#L10-L17 We are going to take a more conservative approach to this list in the Rust CLI. For example, while `/\b[A-Za-z0-9-_]{20,}\b/` might exclude sensitive information like API tokens, it would also exclude valuable information such as references to Git commits. As noted in the updated documentation, users can opt-out of history by adding the following to `config.toml`: ```toml [history] persistence = "none" ``` Because `history.jsonl` could, in theory, be quite large, we take a[n arguably overly pedantic] approach in reading history entries into memory. Specifically, we start by telling the client the current number of entries in the history file (`history_entry_count`) as well as the inode (`history_log_id`) of `history.jsonl` (see the new fields on `SessionConfiguredEvent`). The client is responsible for keeping new entries in memory to create a "local history," but if the user hits up enough times to go "past" the end of local history, then the client should use the new `GetHistoryEntryRequest` in the protocol to fetch older entries. Specifically, it should pass the `history_log_id` it was given originally and work backwards from `history_entry_count`. (It should really fetch history in batches rather than one-at-a-time, but that is something we can improve upon in subsequent PRs.) The motivation behind this crazy scheme is that it is designed to defend against: * The `history.jsonl` being truncated during the session such that the index into the history is no longer consistent with what had been read up to that point. We do not yet have logic to enforce a `max_bytes` for `history.jsonl`, but once we do, we will aspire to implement it in a way that should result in a new inode for the file on most systems. * New items from concurrent Codex CLI sessions amending to the history. Because, in absence of truncation, `history.jsonl` is an append-only log, so long as the client reads backwards from `history_entry_count`, it should always get a consistent view of history. (That said, it will not be able to read _new_ commands from concurrent sessions, but perhaps we will introduce a `/` command to reload latest history or something down the road.) Admittedly, my testing of this feature thus far has been fairly light. I expect we will find bugs and introduce enhancements/fixes going forward.
2025-05-15 16:26:23 -07:00
history: History::default(),
file_opener: UriBasedFileOpener::VsCode,
codex_linux_sandbox_exe: None,
hide_agent_reasoning: false,
show_raw_agent_reasoning: false,
model_reasoning_effort: None,
model_reasoning_summary: ReasoningSummary::default(),
feat(gpt5): add model_verbosity for GPT‑5 via Responses API (#2108) **Summary** - Adds `model_verbosity` config (values: low, medium, high). - Sends `text.verbosity` only for GPT‑5 family models via the Responses API. - Updates docs and adds serialization tests. **Motivation** - GPT‑5 introduces a verbosity control to steer output length/detail without pro mpt surgery. - Exposing it as a config knob keeps prompts stable and makes behavior explicit and repeatable. **Changes** - Config: - Added `Verbosity` enum (low|medium|high). - Added optional `model_verbosity` to `ConfigToml`, `Config`, and `ConfigProfi le`. - Request wiring: - Extended `ResponsesApiRequest` with optional `text` object. - Populates `text.verbosity` only when model family is `gpt-5`; omitted otherw ise. - Tests: - Verifies `text.verbosity` serializes when set and is omitted when not set. - Docs: - Added “GPT‑5 Verbosity” section in `codex-rs/README.md`. - Added `model_verbosity` section to `codex-rs/config.md`. **Usage** - In `~/.codex/config.toml`: - `model = "gpt-5"` - `model_verbosity = "low"` (or `"medium"` default, `"high"`) - CLI override example: - `codex -c model="gpt-5" -c model_verbosity="high"` **API Impact** - Requests to GPT‑5 via Responses API include: `text: { verbosity: "low|medium|h igh" }` when configured. - For legacy models or Chat Completions providers, `text` is omitted. **Backward Compatibility** - Default behavior unchanged when `model_verbosity` is not set (server default “ medium”). **Testing** - Added unit tests for serialization/omission of `text.verbosity`. - Ran `cargo fmt` and `cargo test --all-features` (all green). **Docs** - `README.md`: new “GPT‑5 Verbosity” note under Config with example. - `config.md`: new `model_verbosity` section. **Out of Scope** - No changes to temperature/top_p or other GPT‑5 parameters. - No changes to Chat Completions wiring. **Risks / Notes** - If OpenAI changes the wire shape for verbosity, we may need to update `Respons esApiRequest`. - Behavior gated to `gpt-5` model family to avoid unexpected effects elsewhere. **Checklist** - [x] Code gated to GPT‑5 family only - [x] Docs updated (`README.md`, `config.md`) - [x] Tests added and passing - [x] Formatting applied Release note: Add `model_verbosity` config to control GPT‑5 output verbosity via the Responses API (low|medium|high).
2025-08-22 17:12:10 +01:00
model_verbosity: None,
chatgpt_base_url: "https://chatgpt.com/backend-api/".to_string(),
experimental_resume: None,
base_instructions: None,
include_plan_tool: false,
include_apply_patch_tool: false,
tools_web_search_request: false,
use_experimental_streamable_shell_tool: false,
use_experimental_unified_exec_tool: false,
include_view_image_tool: true,
active_profile: Some("zdr".to_string()),
disable_paste_burst: false,
};
assert_eq!(expected_zdr_profile_config, zdr_profile_config);
Ok(())
}
#[test]
fn test_precedence_fixture_with_gpt5_profile() -> std::io::Result<()> {
let fixture = create_test_fixture()?;
let gpt5_profile_overrides = ConfigOverrides {
config_profile: Some("gpt5".to_string()),
cwd: Some(fixture.cwd()),
..Default::default()
};
let gpt5_profile_config = Config::load_from_base_config_with_overrides(
fixture.cfg.clone(),
gpt5_profile_overrides,
fixture.codex_home(),
)?;
let expected_gpt5_profile_config = Config {
model: "gpt-5".to_string(),
Review Mode (Core) (#3401) ## 📝 Review Mode -- Core This PR introduces the Core implementation for Review mode: - New op `Op::Review { prompt: String }:` spawns a child review task with isolated context, a review‑specific system prompt, and a `Config.review_model`. - `EnteredReviewMode`: emitted when the child review session starts. Every event from this point onwards reflects the review session. - `ExitedReviewMode(Option<ReviewOutputEvent>)`: emitted when the review finishes or is interrupted, with optional structured findings: ```json { "findings": [ { "title": "<≤ 80 chars, imperative>", "body": "<valid Markdown explaining *why* this is a problem; cite files/lines/functions>", "confidence_score": <float 0.0-1.0>, "priority": <int 0-3>, "code_location": { "absolute_file_path": "<file path>", "line_range": {"start": <int>, "end": <int>} } } ], "overall_correctness": "patch is correct" | "patch is incorrect", "overall_explanation": "<1-3 sentence explanation justifying the overall_correctness verdict>", "overall_confidence_score": <float 0.0-1.0> } ``` ## Questions ### Why separate out its own message history? We want the review thread to match the training of our review models as much as possible -- that means using a custom prompt, removing user instructions, and starting a clean chat history. We also want to make sure the review thread doesn't leak into the parent thread. ### Why do this as a mode, vs. sub-agents? 1. We want review to be a synchronous task, so it's fine for now to do a bespoke implementation. 2. We're still unclear about the final structure for sub-agents. We'd prefer to land this quickly and then refactor into sub-agents without rushing that implementation.
2025-09-12 16:25:10 -07:00
review_model: "gpt-5".to_string(),
model_family: find_family_for_model("gpt-5").expect("known model slug"),
model_context_window: Some(272_000),
model_max_output_tokens: Some(128_000),
model_auto_compact_token_limit: None,
model_provider_id: "openai".to_string(),
model_provider: fixture.openai_provider.clone(),
approval_policy: AskForApproval::OnFailure,
sandbox_policy: SandboxPolicy::new_read_only_policy(),
shell_environment_policy: ShellEnvironmentPolicy::default(),
user_instructions: None,
notify: None,
cwd: fixture.cwd(),
mcp_servers: HashMap::new(),
model_providers: fixture.model_provider_map.clone(),
project_doc_max_bytes: PROJECT_DOC_MAX_BYTES,
codex_home: fixture.codex_home(),
history: History::default(),
file_opener: UriBasedFileOpener::VsCode,
codex_linux_sandbox_exe: None,
hide_agent_reasoning: false,
show_raw_agent_reasoning: false,
model_reasoning_effort: Some(ReasoningEffort::High),
model_reasoning_summary: ReasoningSummary::Detailed,
model_verbosity: Some(Verbosity::High),
chatgpt_base_url: "https://chatgpt.com/backend-api/".to_string(),
experimental_resume: None,
base_instructions: None,
include_plan_tool: false,
include_apply_patch_tool: false,
tools_web_search_request: false,
use_experimental_streamable_shell_tool: false,
use_experimental_unified_exec_tool: false,
include_view_image_tool: true,
active_profile: Some("gpt5".to_string()),
disable_paste_burst: false,
};
assert_eq!(expected_gpt5_profile_config, gpt5_profile_config);
Ok(())
}
#[test]
fn test_set_project_trusted_writes_explicit_tables() -> anyhow::Result<()> {
let project_dir = Path::new("/some/path");
let mut doc = DocumentMut::new();
set_project_trusted_inner(&mut doc, project_dir)?;
let contents = doc.to_string();
let raw_path = project_dir.to_string_lossy();
let path_str = if raw_path.contains('\\') {
format!("'{raw_path}'")
} else {
format!("\"{raw_path}\"")
};
let expected = format!(
r#"[projects.{path_str}]
trust_level = "trusted"
"#
);
assert_eq!(contents, expected);
Ok(())
}
#[test]
fn test_set_project_trusted_converts_inline_to_explicit() -> anyhow::Result<()> {
let project_dir = Path::new("/some/path");
// Seed config.toml with an inline project entry under [projects]
let raw_path = project_dir.to_string_lossy();
let path_str = if raw_path.contains('\\') {
format!("'{raw_path}'")
} else {
format!("\"{raw_path}\"")
};
// Use a quoted key so backslashes don't require escaping on Windows
let initial = format!(
r#"[projects]
{path_str} = {{ trust_level = "untrusted" }}
"#
);
let mut doc = initial.parse::<DocumentMut>()?;
// Run the function; it should convert to explicit tables and set trusted
set_project_trusted_inner(&mut doc, project_dir)?;
let contents = doc.to_string();
// Assert exact output after conversion to explicit table
let expected = format!(
r#"[projects]
[projects.{path_str}]
trust_level = "trusted"
"#
);
assert_eq!(contents, expected);
Ok(())
}
#[test]
fn test_set_project_trusted_migrates_top_level_inline_projects_preserving_entries()
-> anyhow::Result<()> {
let initial = r#"toplevel = "baz"
projects = { "/Users/mbolin/code/codex4" = { trust_level = "trusted", foo = "bar" } , "/Users/mbolin/code/codex3" = { trust_level = "trusted" } }
model = "foo""#;
let mut doc = initial.parse::<DocumentMut>()?;
// Approve a new directory
let new_project = Path::new("/Users/mbolin/code/codex2");
set_project_trusted_inner(&mut doc, new_project)?;
let contents = doc.to_string();
// Since we created the [projects] table as part of migration, it is kept implicit.
// Expect explicit per-project tables, preserving prior entries and appending the new one.
let expected = r#"toplevel = "baz"
model = "foo"
[projects."/Users/mbolin/code/codex4"]
trust_level = "trusted"
foo = "bar"
[projects."/Users/mbolin/code/codex3"]
trust_level = "trusted"
[projects."/Users/mbolin/code/codex2"]
trust_level = "trusted"
"#;
assert_eq!(contents, expected);
Ok(())
}
}