fix: exclude sysprompt etc from context left % (#2446)

- Prevents the % left indicator from immediately decrementing to ~97%.
- Tested by prompting "hi" and noting it only decremented to 99%. And by
adding a bunch of debug logs and observing numbers.

codex-rs/protocol/src/protocol.rs

@@ -508,6 +508,33 @@ impl TokenUsage {
         self.total_tokens
             .saturating_sub(self.reasoning_output_tokens.unwrap_or(0))
     }
+
+    /// Estimate the remaining user-controllable percentage of the model's context window.
+    ///
+    /// `context_window` is the total size of the model's context window.
+    /// `baseline_used_tokens` should capture tokens that are always present in
+    /// the context (e.g., system prompt and fixed tool instructions) so that
+    /// the percentage reflects the portion the user can influence.
+    ///
+    /// This normalizes both the numerator and denominator by subtracting the
+    /// baseline, so immediately after the first prompt the UI shows 100% left
+    /// and trends toward 0% as the user fills the effective window.
+    pub fn percent_of_context_window_remaining(
+        &self,
+        context_window: u64,
+        baseline_used_tokens: u64,
+    ) -> u8 {
+        if context_window <= baseline_used_tokens {
+            return 0;
+        }
+
+        let effective_window = context_window - baseline_used_tokens;
+        let used = self
+            .tokens_in_context_window()
+            .saturating_sub(baseline_used_tokens);
+        let remaining = effective_window.saturating_sub(used);
+        ((remaining as f32 / effective_window as f32) * 100.0).clamp(0.0, 100.0) as u8
+    }
 }
 
 #[derive(Debug, Clone, Deserialize, Serialize)]

codex-rs/tui/src/bottom_pane/chat_composer.rs

@@ -45,6 +45,15 @@ struct TokenUsageInfo {
     total_token_usage: TokenUsage,
     last_token_usage: TokenUsage,
     model_context_window: Option<u64>,
+    /// Baseline token count present in the context before the user's first
+    /// message content is considered. This is used to normalize the
+    /// "context left" percentage so it reflects the portion the user can
+    /// influence rather than fixed prompt overhead (system prompt, tool
+    /// instructions, etc.).
+    ///
+    /// Preferred source is `cached_input_tokens` from the first turn (when
+    /// available), otherwise we fall back to 0.
+    initial_prompt_tokens: u64,
 }
 
 pub(crate) struct ChatComposer {
@@ -134,10 +143,17 @@ impl ChatComposer {
         last_token_usage: TokenUsage,
         model_context_window: Option<u64>,
     ) {
+        let initial_prompt_tokens = self
+            .token_usage_info
+            .as_ref()
+            .map(|info| info.initial_prompt_tokens)
+            .unwrap_or_else(|| last_token_usage.cached_input_tokens.unwrap_or(0));
+
         self.token_usage_info = Some(TokenUsageInfo {
             total_token_usage,
             last_token_usage,
             model_context_window,
+            initial_prompt_tokens,
         });
     }
 
@@ -673,11 +689,10 @@ impl WidgetRef for &ChatComposer {
                     let last_token_usage = &token_usage_info.last_token_usage;
                     if let Some(context_window) = token_usage_info.model_context_window {
                         let percent_remaining: u8 = if context_window > 0 {
-                            let percent = 100.0
-                                - (last_token_usage.tokens_in_context_window() as f32
-                                    / context_window as f32
-                                    * 100.0);
-                            percent.clamp(0.0, 100.0) as u8
+                            last_token_usage.percent_of_context_window_remaining(
+                                context_window,
+                                token_usage_info.initial_prompt_tokens,
+                            )
                         } else {
                             100
                         };