Expose invocation sources to extension tools (#40382)

## What changed

- Add `ToolCallSource` to extension `ToolCall`s and propagate direct or Code Mode
  invocation metadata, including the runtime cell and nested tool-call IDs.
- Add `ToolCall::response_byte_budget` so extension tools can bound direct-call
  responses by the host truncation policy while allowing Code Mode responses up
  to the tool's own limit.
- Share the skills extension's response-size limit between its `list` and `read`
  tools.

## Testing

- Verify that Code Mode source metadata is preserved when core constructs an
  extension tool call.

GitOrigin-RevId: b10aef01be30d35b67bee3af24a1a4175c9ee112

codex-rs/ext/skills/src/tools/list.rs

@@ -15,6 +15,7 @@ use crate::render::truncate_utf8_to_bytes;
 use crate::warnings::bounded_warnings;
 
 use super::MAX_HANDLE_BYTES;
+use super::MAX_SKILL_RESPONSE_BYTES;
 use super::SkillToolAuthority;
 use super::SkillToolAuthoritySelector;
 use super::SkillToolContext;
@@ -29,7 +30,6 @@ use super::skill_tool_name;
 
 const TOOL_NAME: &str = "list";
 const MAX_SKILLS_PER_PAGE: usize = 20;
-const MAX_LIST_RESPONSE_BYTES: usize = 512 * 1024;
 const OVERSIZED_ENTRY_WARNING: &str =
     "Some skills were omitted because their metadata is too large.";
 
@@ -114,7 +114,7 @@ impl ToolExecutor<ToolCall> for ListTool {
                     warnings: warnings.clone(),
                     next_cursor: (end < skills.len()).then(|| pagination_cursor(&skills, end)),
                 };
-                if serialized_len(&response)? <= MAX_LIST_RESPONSE_BYTES {
+                if serialized_len(&response)? <= MAX_SKILL_RESPONSE_BYTES {
                     return skill_json_output(&response, args.authority);
                 }
                 if end.saturating_sub(start) > 1 {
@@ -137,7 +137,7 @@ fn single_entry_response_is_bounded(skill: &ListedSkill) -> bool {
         warnings: Vec::new(),
         next_cursor: Some(pagination_cursor(skill, usize::MAX)),
     })
-    .is_ok_and(|size| size <= MAX_LIST_RESPONSE_BYTES)
+    .is_ok_and(|size| size <= MAX_SKILL_RESPONSE_BYTES)
 }
 
 fn listed_skill(entry: SkillCatalogEntry) -> Option<ListedSkill> {

codex-rs/ext/skills/src/tools/mod.rs

@@ -50,6 +50,7 @@ mod schema;
 
 const SKILLS_NAMESPACE: &str = "skills";
 const MAX_HANDLE_BYTES: usize = 2_048;
+const MAX_SKILL_RESPONSE_BYTES: usize = 512 * 1024;
 
 pub(crate) fn skill_tools(
     providers: SkillProviders,

codex-rs/ext/skills/src/tools/read.rs

@@ -14,6 +14,7 @@ use crate::provider::SkillReadRequest;
 use crate::render::build_alias_plan;
 
 use super::MAX_HANDLE_BYTES;
+use super::MAX_SKILL_RESPONSE_BYTES;
 use super::SkillToolAuthority;
 use super::SkillToolContext;
 use super::pagination_cursor;
@@ -26,7 +27,6 @@ use super::skill_tool_name;
 use super::validate_handle;
 
 const TOOL_NAME: &str = "read";
-const MAX_READ_RESPONSE_BYTES: usize = 512 * 1024;
 
 #[derive(Deserialize, JsonSchema)]
 struct ReadArgs {
@@ -237,7 +237,7 @@ fn page_response(
         next_cursor,
     };
     let complete = response(contents.len(), None);
-    if serialized_len(&complete)? <= MAX_READ_RESPONSE_BYTES {
+    if serialized_len(&complete)? <= MAX_SKILL_RESPONSE_BYTES {
         return Ok(complete);
     }
 
@@ -248,7 +248,7 @@ fn page_response(
             end -= 1;
         }
         let candidate = response(end, Some(pagination_cursor(contents, end)));
-        if serialized_len(&candidate)? <= MAX_READ_RESPONSE_BYTES {
+        if serialized_len(&candidate)? <= MAX_SKILL_RESPONSE_BYTES {
             return Ok(candidate);
         }
     }

codex-rs/ext/skills/tests/skills_extension.rs

@@ -24,6 +24,7 @@ use codex_extension_api::SkillInvocationInput;
 use codex_extension_api::SkillInvocationKind;
 use codex_extension_api::ThreadStartInput;
 use codex_extension_api::ToolCall;
+use codex_extension_api::ToolCallSource;
 use codex_extension_api::ToolPayload;
 use codex_extension_api::TurnInputContext;
 use codex_extension_api::WorldStateContributionInput;
@@ -1604,6 +1605,7 @@ async fn skills_list_only_returns_model_visible_bounded_metadata() -> TestResult
             model: "gpt-test".to_string(),
             codex_turn_metadata: None,
             truncation_policy: TruncationPolicy::Bytes(1_024),
+            source: ToolCallSource::Direct,
             conversation_history: ConversationHistory::default(),
             turn_item_emitter: Arc::new(NoopTurnItemEmitter),
             environments: Vec::new(),