valknar/llmx
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

chore: consolidate model utils and drive-by cleanups (#476)

Browse Source 
Signed-off-by: Thibault Sottiaux <tibo@openai.com>
This commit is contained in:
Thibault Sottiaux
2025-04-21 12:33:57 -04:00
committed by GitHub
parent dc276999a9
commit 3c4f1fea9b
21 changed files with 196 additions and 205 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
codex-cli/src/approvals.ts
View File

@@ -136,8 +136,8 @@ export function canAutoApprove(
// bashCmd could be a mix of strings and operators, e.g.:
// "ls || (true && pwd)" => [ 'ls', { op: '||' }, '(', 'true', { op: '&&' }, 'pwd', ')' ]
// We try to ensure that *every* command segment is deemed safe and that
// all operators belong to an allow‑list. If so, the entire expression is
// considered auto‑approvable.
// all operators belong to an allow-list. If so, the entire expression is
// considered auto-approvable.
const shellSafe = isEntireShellExpressionSafe(bashCmd);
if (shellSafe != null) {
@@ -333,7 +333,7 @@ export function isSafeCommand(
};
case "true":
return {
reason: "No‑op (true)",
reason: "No-op (true)",
group: "Utility",
};
case "echo":
@@ -442,10 +442,10 @@ function isValidSedNArg(arg: string | undefined): boolean {
// ---------------- Helper utilities for complex shell expressions -----------------
// A conservative allow‑list of bash operators that do not, on their own, cause
// A conservative allow-list of bash operators that do not, on their own, cause
// side effects. Redirections (>, >>, <, etc.) and command substitution `$()`
// are intentionally excluded. Parentheses used for grouping are treated as
// strings by `shell‑quote`, so we do not add them here. Reference:
// strings by `shell-quote`, so we do not add them here. Reference:
// https://github.com/substack/node-shell-quote#parsecmd-opts
const SAFE_SHELL_OPERATORS: ReadonlySet<string> = new Set([
"&&", // logical AND
@@ -471,7 +471,7 @@ function isEntireShellExpressionSafe(
}
try {
// Collect command segments delimited by operators. `shell‑quote` represents
// Collect command segments delimited by operators. `shell-quote` represents
// subshell grouping parentheses as literal strings "(" and ")"; treat them
// as unsafe to keep the logic simple (since subshells could introduce
// unexpected scope changes).
@@ -539,7 +539,7 @@ function isParseEntryWithOp(
return (
typeof entry === "object" &&
entry != null &&
// Using the safe `in` operator keeps the check property‑safe even when
// Using the safe `in` operator keeps the check property-safe even when
// `entry` is a `string`.
"op" in entry &&
typeof (entry as { op?: unknown }).op === "string"

47
codex-cli/src/cli.tsx
View File

@@ -136,7 +136,7 @@ const cli = meow(
},
noProjectDoc: {
type: "boolean",
description: "Disable automatic inclusion of project‑level codex.md",
description: "Disable automatic inclusion of project-level codex.md",
},
projectDoc: {
type: "string",
@@ -202,19 +202,20 @@ complete -c codex -a '(_fish_complete_path)' -d 'file path'`,
console.log(script);
process.exit(0);
}
// Show help if requested
// For --help, show help and exit.
if (cli.flags.help) {
cli.showHelp();
}
// Handle config flag: open instructions file in editor and exit
// For --config, open custom instructions file in editor and exit.
if (cli.flags.config) {
// Ensure configuration and instructions file exist
try {
loadConfig();
loadConfig(); // Ensures the file is created if it doesn't already exit.
} catch {
// ignore errors
}
const filePath = INSTRUCTIONS_FILEPATH;
const editor =
process.env["EDITOR"] || (process.platform === "win32" ? "notepad" : "vi");
@@ -237,13 +238,13 @@ let config = loadConfig(undefined, undefined, {
const prompt = cli.input[0];
const model = cli.flags.model ?? config.model;
const imagePaths = cli.flags.image;
const provider = cli.flags.provider ?? config.provider;
const provider = cli.flags.provider ?? config.provider ?? "openai";
const apiKey = getApiKey(provider);
if (!apiKey) {
// eslint-disable-next-line no-console
console.error(
`\n${chalk.red("Missing OpenAI API key.")}\n\n` +
`\n${chalk.red(`Missing ${provider} API key.`)}\n\n` +
`Set the environment variable ${chalk.bold("OPENAI_API_KEY")} ` +
`and re-run this command.\n` +
`You can create a key here: ${chalk.bold(
@@ -262,13 +263,11 @@ config = {
provider,
};
// Check for updates after loading config
// This is important because we write state file in the config dir
// Check for updates after loading config. This is important because we write state file in
// the config dir.
await checkForUpdates().catch();
// ---------------------------------------------------------------------------
// --flex-mode validation (only allowed for o3 and o4-mini)
// ---------------------------------------------------------------------------
// For --flex-mode, validate and exit if incorrect.
if (cli.flags.flexMode) {
const allowedFlexModels = new Set(["o3", "o4-mini"]);
if (!allowedFlexModels.has(config.model)) {
@@ -282,13 +281,13 @@ if (cli.flags.flexMode) {
}
if (
!(await isModelSupportedForResponses(config.model)) &&
!(await isModelSupportedForResponses(provider, config.model)) &&
(!provider || provider.toLowerCase() === "openai")
) {
// eslint-disable-next-line no-console
console.error(
`The model "${config.model}" does not appear in the list of models ` +
`available to your account. Double‑check the spelling (use\n` +
`available to your account. Double-check the spelling (use\n` +
` openai models list\n` +
`to see the full list) or choose another model with the --model flag.`,
);
@@ -297,6 +296,7 @@ if (
let rollout: AppRollout | undefined;
// For --view, optionally load an existing rollout from disk, display it and exit.
if (cli.flags.view) {
const viewPath = cli.flags.view;
const absolutePath = path.isAbsolute(viewPath)
@@ -312,7 +312,7 @@ if (cli.flags.view) {
}
}
// If we are running in --fullcontext mode, do that and exit.
// For --fullcontext, run the separate cli entrypoint and exit.
if (fullContextMode) {
await runSinglePass({
originalPrompt: prompt,
@@ -328,11 +328,8 @@ const additionalWritableRoots: ReadonlyArray<string> = (
cli.flags.writableRoot ?? []
).map((p) => path.resolve(p));
// If we are running in --quiet mode, do that and exit.
const quietMode = Boolean(cli.flags.quiet);
const fullStdout = Boolean(cli.flags.fullStdout);
if (quietMode) {
// For --quiet, run the cli without user interactions and exit.
if (cli.flags.quiet) {
process.env["CODEX_QUIET_MODE"] = "1";
if (!prompt || prompt.trim() === "") {
// eslint-disable-next-line no-console
@@ -389,7 +386,7 @@ const instance = render(
imagePaths={imagePaths}
approvalPolicy={approvalPolicy}
additionalWritableRoots={additionalWritableRoots}
fullStdout={fullStdout}
fullStdout={Boolean(cli.flags.fullStdout)}
/>,
{
patchConsole: process.env["DEBUG"] ? false : true,
@@ -501,13 +498,13 @@ process.on("SIGQUIT", exit);
process.on("SIGTERM", exit);
// ---------------------------------------------------------------------------
// Fallback for Ctrl‑C when stdin is in raw‑mode
// Fallback for Ctrl-C when stdin is in raw-mode
// ---------------------------------------------------------------------------
if (process.stdin.isTTY) {
// Ensure we do not leave the terminal in raw mode if the user presses
// Ctrl‑C while some other component has focus and Ink is intercepting
// input. Node does *not* emit a SIGINT in raw‑mode, so we listen for the
// Ctrl-C while some other component has focus and Ink is intercepting
// input. Node does *not* emit a SIGINT in raw-mode, so we listen for the
// corresponding byte (0x03) ourselves and trigger a graceful shutdown.
const onRawData = (data: Buffer | string): void => {
const str = Buffer.isBuffer(data) ? data.toString("utf8") : data;
@@ -518,6 +515,6 @@ if (process.stdin.isTTY) {
process.stdin.on("data", onRawData);
}
// Ensure terminal clean‑up always runs, even when other code calls
// Ensure terminal clean-up always runs, even when other code calls
// `process.exit()` directly.
process.once("exit", onExit);

2
codex-cli/src/components/chat/multiline-editor.tsx
View File

@@ -14,7 +14,7 @@ import React, { useRef, useState } from "react";
* The real `process.stdin` object exposed by Node.js inherits these methods
* from `Socket`, but the lightweight stub used in tests only extends
* `EventEmitter`. Ink calls the two methods when enabling/disabling raw
* mode, so make them harmless no‑ops when they're absent to avoid runtime
* mode, so make them harmless no-ops when they're absent to avoid runtime
* failures during unit tests.
* ----------------------------------------------------------------------- */

113
codex-cli/src/components/chat/terminal-chat-utils.ts
View File

@@ -1,113 +0,0 @@
import type { ResponseItem } from "openai/resources/responses/responses.mjs";
import { approximateTokensUsed } from "../../utils/approximate-tokens-used.js";
/**
* Type‑guard that narrows a {@link ResponseItem} to one that represents a
* user‑authored message. The OpenAI SDK represents both input *and* output
* messages with a discriminated union where:
* • `type` is the string literal "message" and
* • `role` is one of "user" | "assistant" | "system" | "developer".
*
* For the purposes of de‑duplication we only care about *user* messages so we
* detect those here in a single, reusable helper.
*/
function isUserMessage(
item: ResponseItem,
): item is ResponseItem & { type: "message"; role: "user"; content: unknown } {
return item.type === "message" && (item as { role?: string }).role === "user";
}
/**
* Returns the maximum context length (in tokens) for a given model.
* These numbers are best‑effort guesses and provide a basis for UI percentages.
*/
export function maxTokensForModel(model: string): number {
const lower = model.toLowerCase();
if (lower.includes("32k")) {
return 32000;
}
if (lower.includes("16k")) {
return 16000;
}
if (lower.includes("8k")) {
return 8000;
}
if (lower.includes("4k")) {
return 4000;
}
// Default to 128k for newer long‑context models
return 128000;
}
/**
* Calculates the percentage of tokens remaining in context for a model.
*/
export function calculateContextPercentRemaining(
items: Array<ResponseItem>,
model: string,
): number {
const used = approximateTokensUsed(items);
const max = maxTokensForModel(model);
const remaining = Math.max(0, max - used);
return (remaining / max) * 100;
}
/**
* Deduplicate the stream of {@link ResponseItem}s before they are persisted in
* component state.
*
* Historically we used the (optional) {@code id} field returned by the
* OpenAI streaming API as the primary key: the first occurrence of any given
* {@code id} “won” and subsequent duplicates were dropped. In practice this
* proved brittle because locally‑generated user messages don’t include an
* {@code id}. The result was that if a user quickly pressed <Enter> twice the
* exact same message would appear twice in the transcript.
*
* The new rules are therefore:
* 1. If a {@link ResponseItem} has an {@code id} keep only the *first*
* occurrence of that {@code id} (this retains the previous behaviour for
* assistant / tool messages).
* 2. Additionally, collapse *consecutive* user messages with identical
* content. Two messages are considered identical when their serialized
* {@code content} array matches exactly. We purposefully restrict this
* to **adjacent** duplicates so that legitimately repeated questions at
* a later point in the conversation are still shown.
*/
export function uniqueById(items: Array<ResponseItem>): Array<ResponseItem> {
const seenIds = new Set<string>();
const deduped: Array<ResponseItem> = [];
for (const item of items) {
// ──────────────────────────────────────────────────────────────────
// Rule #1 – de‑duplicate by id when present
// ──────────────────────────────────────────────────────────────────
if (typeof item.id === "string" && item.id.length > 0) {
if (seenIds.has(item.id)) {
continue; // skip duplicates
}
seenIds.add(item.id);
}
// ──────────────────────────────────────────────────────────────────
// Rule #2 – collapse consecutive identical user messages
// ──────────────────────────────────────────────────────────────────
if (isUserMessage(item) && deduped.length > 0) {
const prev = deduped[deduped.length - 1]!;
if (
isUserMessage(prev) &&
// Note: the `content` field is an array of message parts. Performing
// a deep compare is over‑kill here; serialising to JSON is sufficient
// (and fast for the tiny payloads involved).
JSON.stringify(prev.content) === JSON.stringify(item.content)
) {
continue; // skip duplicate user message
}
}
deduped.push(item);
}
return deduped;
}

57
codex-cli/src/components/chat/terminal-chat.tsx
View File

@@ -6,10 +6,6 @@ import type { ResponseItem } from "openai/resources/responses/responses.mjs";
import TerminalChatInput from "./terminal-chat-input.js";
import { TerminalChatToolCallCommand } from "./terminal-chat-tool-call-command.js";
import {
calculateContextPercentRemaining,
uniqueById,
} from "./terminal-chat-utils.js";
import TerminalMessageHistory from "./terminal-message-history.js";
import { formatCommandForDisplay } from "../../format-command.js";
import { useConfirmation } from "../../hooks/use-confirmation.js";
@@ -22,7 +18,11 @@ import { extractAppliedPatches as _extractAppliedPatches } from "../../utils/ext
import { getGitDiff } from "../../utils/get-diff.js";
import { createInputItem } from "../../utils/input-utils.js";
import { log } from "../../utils/logger/log.js";
import { getAvailableModels } from "../../utils/model-utils.js";
import {
getAvailableModels,
calculateContextPercentRemaining,
uniqueById,
} from "../../utils/model-utils.js";
import { CLI_VERSION } from "../../utils/session.js";
import { shortCwd } from "../../utils/short-path.js";
import { saveRollout } from "../../utils/storage/save-rollout.js";
@@ -106,11 +106,8 @@ async function generateCommandExplanation(
} catch (error) {
log(`Error generating command explanation: ${error}`);
// Improved error handling with more specific error information
let errorMessage = "Unable to generate explanation due to an error.";
if (error instanceof Error) {
// Include specific error message for better debugging
errorMessage = `Unable to generate explanation: ${error.message}`;
// If it's an API error, check for more specific information
@@ -141,18 +138,17 @@ export default function TerminalChat({
additionalWritableRoots,
fullStdout,
}: Props): React.ReactElement {
// Desktop notification setting
const notify = config.notify;
const [model, setModel] = useState<string>(config.model);
const [provider, setProvider] = useState<string>(config.provider || "openai");
const [lastResponseId, setLastResponseId] = useState<string | null>(null);
const [items, setItems] = useState<Array<ResponseItem>>([]);
const [loading, setLoading] = useState<boolean>(false);
// Allow switching approval modes at runtime via an overlay.
const [approvalPolicy, setApprovalPolicy] = useState<ApprovalPolicy>(
initialApprovalPolicy,
);
const [thinkingSeconds, setThinkingSeconds] = useState(0);
const handleCompact = async () => {
setLoading(true);
try {
@@ -185,6 +181,7 @@ export default function TerminalChat({
setLoading(false);
}
};
const {
requestConfirmation,
confirmationPrompt,
@@ -215,13 +212,13 @@ export default function TerminalChat({
// DEBUG: log every render w/ key bits of state
// ────────────────────────────────────────────────────────────────
log(
`render – agent? ${Boolean(agentRef.current)} loading=${loading} items=${
`render - agent? ${Boolean(agentRef.current)} loading=${loading} items=${
items.length
}`,
);
useEffect(() => {
// Skip recreating the agent if awaiting a decision on a pending confirmation
// Skip recreating the agent if awaiting a decision on a pending confirmation.
if (confirmationPrompt != null) {
log("skip AgentLoop recreation due to pending confirmationPrompt");
return;
@@ -234,7 +231,7 @@ export default function TerminalChat({
)} approvalPolicy=${approvalPolicy}`,
);
// Tear down any existing loop before creating a new one
// Tear down any existing loop before creating a new one.
agentRef.current?.terminate();
const sessionId = crypto.randomUUID();
@@ -267,11 +264,9 @@ export default function TerminalChat({
<TerminalChatToolCallCommand commandForDisplay={commandForDisplay} />,
);
// If the user wants an explanation, generate one and ask again
// If the user wants an explanation, generate one and ask again.
if (review === ReviewDecision.EXPLAIN) {
log(`Generating explanation for command: ${commandForDisplay}`);
// Generate an explanation using the same model
const explanation = await generateCommandExplanation(
command,
model,
@@ -279,7 +274,7 @@ export default function TerminalChat({
);
log(`Generated explanation: ${explanation}`);
// Ask for confirmation again, but with the explanation
// Ask for confirmation again, but with the explanation.
const confirmResult = await requestConfirmation(
<TerminalChatToolCallCommand
commandForDisplay={commandForDisplay}
@@ -287,11 +282,11 @@ export default function TerminalChat({
/>,
);
// Update the decision based on the second confirmation
// Update the decision based on the second confirmation.
review = confirmResult.decision;
customDenyMessage = confirmResult.customDenyMessage;
// Return the final decision with the explanation
// Return the final decision with the explanation.
return { review, customDenyMessage, applyPatch, explanation };
}
@@ -299,7 +294,7 @@ export default function TerminalChat({
},
});
// force a render so JSX below can "see" the freshly created agent
// Force a render so JSX below can "see" the freshly created agent.
forceUpdate();
log(`AgentLoop created: ${inspect(agentRef.current, { depth: 1 })}`);
@@ -320,7 +315,7 @@ export default function TerminalChat({
additionalWritableRoots,
]);
// whenever loading starts/stops, reset or start a timer — but pause the
// Whenever loading starts/stops, reset or start a timer — but pause the
// timer while a confirmation overlay is displayed so we don't trigger a
// re‑render every second during apply_patch reviews.
useEffect(() => {
@@ -345,14 +340,15 @@ export default function TerminalChat({
};
}, [loading, confirmationPrompt]);
// Notify desktop with a preview when an assistant response arrives
// Notify desktop with a preview when an assistant response arrives.
const prevLoadingRef = useRef<boolean>(false);
useEffect(() => {
// Only notify when notifications are enabled
// Only notify when notifications are enabled.
if (!notify) {
prevLoadingRef.current = loading;
return;
}
if (
prevLoadingRef.current &&
!loading &&
@@ -389,7 +385,7 @@ export default function TerminalChat({
prevLoadingRef.current = loading;
}, [notify, loading, confirmationPrompt, items, PWD]);
// Let's also track whenever the ref becomes available
// Let's also track whenever the ref becomes available.
const agent = agentRef.current;
useEffect(() => {
log(`agentRef.current is now ${Boolean(agent)}`);
@@ -412,7 +408,7 @@ export default function TerminalChat({
const inputItems = [
await createInputItem(initialPrompt || "", initialImagePaths || []),
];
// Clear them to prevent subsequent runs
// Clear them to prevent subsequent runs.
setInitialPrompt("");
setInitialImagePaths([]);
agent?.run(inputItems);
@@ -447,7 +443,7 @@ export default function TerminalChat({
// eslint-disable-next-line react-hooks/exhaustive-deps
}, []);
// Just render every item in order, no grouping/collapse
// Just render every item in order, no grouping/collapse.
const lastMessageBatch = items.map((item) => ({ item }));
const groupCounts: Record<string, number> = {};
const userMsgCount = items.filter(
@@ -626,10 +622,10 @@ export default function TerminalChat({
agent?.cancel();
setLoading(false);
// Select default model for the new provider
// Select default model for the new provider.
const defaultModel = model;
// Save provider to config
// Save provider to config.
const updatedConfig = {
...config,
provider: newProvider,
@@ -669,13 +665,12 @@ export default function TerminalChat({
<ApprovalModeOverlay
currentMode={approvalPolicy}
onSelect={(newMode) => {
// update approval policy without cancelling an in-progress session
// Update approval policy without cancelling an in-progress session.
if (newMode === approvalPolicy) {
return;
}
// update state
setApprovalPolicy(newMode as ApprovalPolicy);
// update existing AgentLoop instance
if (agentRef.current) {
(
agentRef.current as unknown as {

4
codex-cli/src/components/chat/terminal-header.tsx
View File

@@ -34,9 +34,9 @@ const TerminalHeader: React.FC<TerminalHeaderProps> = ({
{terminalRows < 10 ? (
// Compact header for small terminal windows
<Text>
● Codex v{version} – {PWD} – {model} ({provider}) –{" "}
● Codex v{version} - {PWD} - {model} ({provider}) -{" "}
<Text color={colorsByPolicy[approvalPolicy]}>{approvalPolicy}</Text>
{flexModeEnabled ? " – flex-mode" : ""}
{flexModeEnabled ? " - flex-mode" : ""}
</Text>
) : (
<>

4
codex-cli/src/components/singlepass-cli-app.tsx
View File

@@ -399,8 +399,8 @@ export function SinglePassApp({
});
const openai = new OpenAI({
apiKey: getApiKey(config.provider),
baseURL: getBaseUrl(config.provider),
apiKey: getApiKey(config.provider ?? "openai"),
baseURL: getBaseUrl(config.provider ?? "openai"),
timeout: OPENAI_TIMEOUT_MS,
});
const chatResp = await openai.beta.chat.completions.parse({

1
codex-cli/src/hooks/use-confirmation.ts
View File

@@ -1,4 +1,3 @@
// use-confirmation.ts
import type { ReviewDecision } from "../utils/agent/review";
import type React from "react";

4
codex-cli/src/text-buffer.ts
View File

@@ -423,7 +423,7 @@ export default class TextBuffer {
/** Delete the word to the *left* of the caret, mirroring common
* Ctrl/Alt+Backspace behaviour in editors & terminals. Both the adjacent
* whitespace *and* the word characters immediately preceding the caret are
* removed. If the caret is already at column‑0 this becomes a no‑op. */
* removed. If the caret is already at column‑0 this becomes a no-op. */
deleteWordLeft(): void {
dbg("deleteWordLeft", { beforeCursor: this.getCursor() });
@@ -710,7 +710,7 @@ export default class TextBuffer {
}
endSelection(): void {
// no‑op for now, kept for API symmetry
// no-op for now, kept for API symmetry
// we rely on anchor + current cursor to compute selection
}

4
codex-cli/src/utils/agent/agent-loop.ts
View File

@@ -744,7 +744,7 @@ export class AgentLoop {
for await (const event of stream as AsyncIterable<ResponseEvent>) {
log(`AgentLoop.run(): response event ${event.type}`);
// process and surface each item (no‑op until we can depend on streaming events)
// process and surface each item (no-op until we can depend on streaming events)
if (event.type === "response.output_item.done") {
const item = event.item;
// 1) if it's a reasoning item, annotate it
@@ -936,7 +936,7 @@ export class AgentLoop {
],
});
} catch {
/* no‑op – emitting the error message is best‑effort */
/* no-op – emitting the error message is best‑effort */
}
this.onLoading(false);
return;

2
codex-cli/src/utils/agent/handle-exec-command.ts
View File

@@ -144,7 +144,7 @@ export async function handleExecCommand(
abortSignal,
);
// If the operation was aborted in the meantime, propagate the cancellation
// upward by returning an empty (no‑op) result so that the agent loop will
// upward by returning an empty (no-op) result so that the agent loop will
// exit cleanly without emitting spurious output.
if (abortSignal?.aborted) {
return {

4
codex-cli/src/utils/config.ts
View File

@@ -41,7 +41,7 @@ export function setApiKey(apiKey: string): void {
OPENAI_API_KEY = apiKey;
}
export function getBaseUrl(provider: string = "openai"): string | undefined {
export function getBaseUrl(provider: string): string | undefined {
const providerInfo = providers[provider.toLowerCase()];
if (providerInfo) {
return providerInfo.baseURL;
@@ -49,7 +49,7 @@ export function getBaseUrl(provider: string = "openai"): string | undefined {
return undefined;
}
export function getApiKey(provider: string = "openai"): string | undefined {
export function getApiKey(provider: string): string | undefined {
const providerInfo = providers[provider.toLowerCase()];
if (providerInfo) {
if (providerInfo.name === "Ollama") {

124
codex-cli/src/utils/model-utils.ts
View File

@@ -1,3 +1,6 @@
import type { ResponseItem } from "openai/resources/responses/responses.mjs";
import { approximateTokensUsed } from "./approximate-tokens-used.js";
import { getBaseUrl, getApiKey } from "./config";
import OpenAI from "openai";
@@ -11,9 +14,8 @@ export const RECOMMENDED_MODELS: Array<string> = ["o4-mini", "o3"];
* enters interactive mode. The request is made exactly once during the
* lifetime of the process and the results are cached for subsequent calls.
*/
async function fetchModels(provider: string): Promise<Array<string>> {
// If the user has not configured an API key we cannot hit the network.
// If the user has not configured an API key we cannot retrieve the models.
if (!getApiKey(provider)) {
throw new Error("No API key configured for provider: " + provider);
}
@@ -26,7 +28,7 @@ async function fetchModels(provider: string): Promise<Array<string>> {
for await (const model of list as AsyncIterable<{ id?: string }>) {
if (model && typeof model.id === "string") {
let modelStr = model.id;
// fix for gemini
// Fix for gemini.
if (modelStr.startsWith("models/")) {
modelStr = modelStr.replace("models/", "");
}
@@ -40,6 +42,7 @@ async function fetchModels(provider: string): Promise<Array<string>> {
}
}
/** Returns the list of models available for the provided key / credentials. */
export async function getAvailableModels(
provider: string,
): Promise<Array<string>> {
@@ -47,11 +50,11 @@ export async function getAvailableModels(
}
/**
* Verify that the provided model identifier is present in the set returned by
* {@link getAvailableModels}. The list of models is fetched from the OpenAI
* `/models` endpoint the first time it is required and then cached in‑process.
* Verifies that the provided model identifier is present in the set returned by
* {@link getAvailableModels}.
*/
export async function isModelSupportedForResponses(
provider: string,
model: string | undefined | null,
): Promise<boolean> {
if (
@@ -64,7 +67,7 @@ export async function isModelSupportedForResponses(
try {
const models = await Promise.race<Array<string>>([
getAvailableModels("openai"),
getAvailableModels(provider),
new Promise<Array<string>>((resolve) =>
setTimeout(() => resolve([]), MODEL_LIST_TIMEOUT_MS),
),
@@ -82,3 +85,110 @@ export async function isModelSupportedForResponses(
return true;
}
}
/** Returns the maximum context length (in tokens) for a given model. */
function maxTokensForModel(model: string): number {
// TODO: These numbers are best‑effort guesses and provide a basis for UI percentages. They
// should be provider & model specific instead of being wild guesses.
const lower = model.toLowerCase();
if (lower.includes("32k")) {
return 32000;
}
if (lower.includes("16k")) {
return 16000;
}
if (lower.includes("8k")) {
return 8000;
}
if (lower.includes("4k")) {
return 4000;
}
return 128000; // Default to 128k for any other model.
}
/** Calculates the percentage of tokens remaining in context for a model. */
export function calculateContextPercentRemaining(
items: Array<ResponseItem>,
model: string,
): number {
const used = approximateTokensUsed(items);
const max = maxTokensForModel(model);
const remaining = Math.max(0, max - used);
return (remaining / max) * 100;
}
/**
* Type‑guard that narrows a {@link ResponseItem} to one that represents a
* user‑authored message. The OpenAI SDK represents both input *and* output
* messages with a discriminated union where:
* • `type` is the string literal "message" and
* • `role` is one of "user" | "assistant" | "system" | "developer".
*
* For the purposes of de‑duplication we only care about *user* messages so we
* detect those here in a single, reusable helper.
*/
function isUserMessage(
item: ResponseItem,
): item is ResponseItem & { type: "message"; role: "user"; content: unknown } {
return item.type === "message" && (item as { role?: string }).role === "user";
}
/**
* Deduplicate the stream of {@link ResponseItem}s before they are persisted in
* component state.
*
* Historically we used the (optional) {@code id} field returned by the
* OpenAI streaming API as the primary key: the first occurrence of any given
* {@code id} “won” and subsequent duplicates were dropped. In practice this
* proved brittle because locally‑generated user messages don’t include an
* {@code id}. The result was that if a user quickly pressed <Enter> twice the
* exact same message would appear twice in the transcript.
*
* The new rules are therefore:
* 1. If a {@link ResponseItem} has an {@code id} keep only the *first*
* occurrence of that {@code id} (this retains the previous behaviour for
* assistant / tool messages).
* 2. Additionally, collapse *consecutive* user messages with identical
* content. Two messages are considered identical when their serialized
* {@code content} array matches exactly. We purposefully restrict this
* to **adjacent** duplicates so that legitimately repeated questions at
* a later point in the conversation are still shown.
*/
export function uniqueById(items: Array<ResponseItem>): Array<ResponseItem> {
const seenIds = new Set<string>();
const deduped: Array<ResponseItem> = [];
for (const item of items) {
// ──────────────────────────────────────────────────────────────────
// Rule #1 – de‑duplicate by id when present
// ──────────────────────────────────────────────────────────────────
if (typeof item.id === "string" && item.id.length > 0) {
if (seenIds.has(item.id)) {
continue; // skip duplicates
}
seenIds.add(item.id);
}
// ──────────────────────────────────────────────────────────────────
// Rule #2 – collapse consecutive identical user messages
// ──────────────────────────────────────────────────────────────────
if (isUserMessage(item) && deduped.length > 0) {
const prev = deduped[deduped.length - 1]!;
if (
isUserMessage(prev) &&
// Note: the `content` field is an array of message parts. Performing
// a deep compare is over‑kill here; serialising to JSON is sufficient
// (and fast for the tiny payloads involved).
JSON.stringify(prev.content) === JSON.stringify(item.content)
) {
continue; // skip duplicate user message
}
}
deduped.push(item);
}
return deduped;
}

0
codex-cli/src/utils/terminal-chat-utils.ts Normal file
View File

2
codex-cli/tests/agent-project-doc.test.ts
View File

@@ -50,7 +50,7 @@ vi.mock("openai", () => {
// The AgentLoop pulls these helpers in order to decide whether a command can
// be auto‑approved. None of that matters for this test, so we stub the module
// with minimal no‑op implementations.
// with minimal no-op implementations.
vi.mock("../src/approvals.js", () => {
return {
__esModule: true,

2
codex-cli/tests/approvals.test.ts
View File

@@ -79,7 +79,7 @@ describe("canAutoApprove()", () => {
test("true command is considered safe", () => {
expect(check(["true"])).toEqual({
type: "auto-approve",
reason: "No‑op (true)",
reason: "No-op (true)",
group: "Utility",
runInSandbox: false,
});

2
codex-cli/tests/config.test.tsx
View File

@@ -26,7 +26,7 @@ vi.mock("fs", async () => {
memfs[path] = data;
},
mkdirSync: () => {
// no‑op in in‑memory store
// no-op in in‑memory store
},
rmSync: (path: string) => {
// recursively delete any key under this prefix

9
codex-cli/tests/model-utils-network-error.test.ts
View File

@@ -44,7 +44,7 @@ describe("model-utils – offline resilience", () => {
"../src/utils/model-utils.js"
);
const supported = await isModelSupportedForResponses("o4-mini");
const supported = await isModelSupportedForResponses("openai", "o4-mini");
expect(supported).toBe(true);
});
@@ -63,8 +63,11 @@ describe("model-utils – offline resilience", () => {
"../src/utils/model-utils.js"
);
// Should resolve true despite the network failure
const supported = await isModelSupportedForResponses("some-model");
// Should resolve true despite the network failure.
const supported = await isModelSupportedForResponses(
"openai",
"some-model",
);
expect(supported).toBe(true);
});
});

2
codex-cli/tests/multiline-history-behavior.test.tsx
View File

@@ -57,7 +57,7 @@ async function type(
await flush();
}
/** Build a set of no‑op callbacks so <TerminalChatInput> renders with minimal
/** Build a set of no-op callbacks so <TerminalChatInput> renders with minimal
* scaffolding.
*/
function stubProps(): any {

2
codex-cli/tests/text-buffer.test.ts
View File

@@ -127,7 +127,7 @@ describe("TextBuffer – basic editing parity with Rust suite", () => {
expect(buf.getCursor()).toEqual([0, 2]); // after 'b'
});
it("is a no‑op at the very beginning of the buffer", () => {
it("is a no-op at the very beginning of the buffer", () => {
const buf = new TextBuffer("ab");
buf.backspace(); // caret starts at (0,0)

2
codex-cli/tests/typeahead-scroll.test.tsx
View File

@@ -26,7 +26,7 @@ vi.mock("../src/components/select-input/select-input.js", () => {
// Ink's <TextInput> toggles raw‑mode which calls .ref() / .unref() on stdin.
// The test environment's mock streams don't implement those methods, so we
// polyfill them to no‑ops on the prototype *before* the component tree mounts.
// polyfill them to no-ops on the prototype *before* the component tree mounts.
import { EventEmitter } from "node:events";
if (!(EventEmitter.prototype as any).ref) {
(EventEmitter.prototype as any).ref = () => {};