Capture new Claude Code hook fields (effort, duration_ms, is_interrupt)

[Abhijeet Prasad (AbhiPrasad)]

What was missing

Claude Code hooks now send more input fields; the same set showed up in claude-agent-sdk 0.2.166 (braintrustdata/braintrust-sdk-python#973). The Claude Code translator reads payload fields directly instead of going through an allowlist, so last_assistant_message, error, and prompt_id were already being used. Three fields were still dropped:

Hook field	Hooks	Before
effort ({"level": "high"})	all	ignored
duration_ms	PostToolUse / PostToolUseFailure	used only to work out start_ms when there was no PreToolUse, never recorded
is_interrupt	PostToolUseFailure	ignored

The fix

Effort → turn metadata. effort is decoded in HookContext and tracked in observe_session_details together with permission_mode. Claude Code doesn't send effort on UserPromptSubmit; it shows up on later hooks (Stop, PreToolUse, PostToolUse). So reading it only when a turn opens isn't enough. Each turn tracks the level it has recorded, and when a different level arrives it is merged into the active turn:

UserPromptSubmit                   ──►  Turn 1 opens (effort unknown)
PreToolUse {effort:{level:"high"}} ──►  merge Turn 1  metadata.effort = "high"
Stop                               ──►  Turn 1 closes
/effort max
UserPromptSubmit                   ──►  Turn 2 opens  metadata.effort = "high" (carried)
Stop {effort:{level:"max"}}        ──►  merge Turn 2  metadata.effort = "max"

fn record_turn_effort(&mut self, ops: &mut Vec<SpanOp>) {
    let Some(turn) = self.turn.as_mut() else { return };
    if self.effort.is_none() || turn.effort == self.effort {
        return;
    }
    turn.effort = self.effort.clone();
    ops.push(SpanOp::Merge(SpanRow {
        span_id: turn.id.clone(),
        metadata: Some(json!({ "effort": turn.effort })),
        // ...
    }));
}

Tool timing / interrupts → tool metadata. finish_tool reads duration_ms once. That one value sets both native_duration_ms (the key the Cursor translator already uses) and the fallback start time, so the span timing and the metadata can't disagree:

let duration = event.payload.get("duration_ms").and_then(Value::as_i64);
if let Some(duration) = duration {
    metadata["native_duration_ms"] = json!(duration);
}
if let Some(interrupted) = event.payload.get("is_interrupt").and_then(Value::as_bool) {
    metadata["is_interrupt"] = json!(interrupted);
}
// ...
start_ms: Some(event.ts_ms.saturating_sub(duration.unwrap_or(0))),

Span timing doesn't change when PreToolUse exists: the span still runs from PreToolUse to PostToolUse.

prompt_id comment. The old comment said native hooks never send prompt_id. They now send it after the first user input, though older Claude Code versions leave it out. The matching code already handled both cases, so only the comment changed.

Test cleanup

The Claude translator tests went from 33 to 29 by merging overlapping cases: the three root-output tests, the two SubagentStart-validation tests, and the success and failure tool-outcome tests. Hand-built Envelope literals were replaced with the existing claude_event helper.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-19T23:33:27.653106Z	e6bf913	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "Codex (@codex) review" or "Codex (@codex) security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 552e777d85

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "Codex (@codex) review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "Codex (@codex) address that feedback".

[chatgpt-codex-connector]

Update the active turn when effort first appears

Claude does not consistently include effort on UserPromptSubmit: for example, subagent-compact/events.ndjson line 2 omits it and the corresponding Stop on line 4 reports high. Because open_turn snapshots self.effort only when the turn opens, this assignment merely caches the late value for the next turn and never adds it to the turn that produced the hook. Consequently, the first turn loses known effort metadata, and after an effort change a later turn can retain the previous level until another turn opens; merge newly observed effort into the currently active turn as well.

Useful? React with 👍 / 👎.

[Abhijeet Prasad (AbhiPrasad)]

Good catch, fixed in 8e34e04. The recorded fixtures confirm that UserPromptSubmit never carries effort; it only arrives on later hooks. Each turn now tracks the level it has recorded, and a newly observed level is merged into the active turn (record_turn_effort). The real-fixture test now asserts that every turn gets xhigh.