Use executed tool input on Claude tool spans

[Abhijeet Prasad (AbhiPrasad)]

Fixes #156.

The bug

Claude Code PreToolUse hooks can rewrite a proposed tool call by returning updatedInput. The terminal hook (PostToolUse, PostToolUseFailure, or PermissionDenied) reports the input the call actually ran with, or was denied with, in tool_input.

The Claude translator creates the tool span from PreToolUse.tool_input. When the terminal hook arrived, finish_tool merged in output, approval, error and metadata, but never input or name:

PreToolUse   tool_input: {"command": "rm -rf build"}  ──► Insert  name:  "Terminal: rm -rf build"
                                                               input: {"command": "rm -rf build"}
     │  hook returns updatedInput
     ▼
PostToolUse  tool_input: {"command": "ls build"}      ──► Merge   output, approval, metadata
                                                               (input and name untouched)

Resulting span: input says `rm -rf build`, output is from `ls build`

The span's metadata was rebuilt from the terminal input, so the span could also disagree with itself: its input/name showed the proposed call while its input-derived metadata reflected the one that ran.

The fix

PendingTool now remembers the proposed input. When the terminal hook reports a different tool_input, the merge replaces the span's input and recomputes its name:

let rewritten = raw_tool_input(&event.payload).is_some() && input != pending.input;
ops.push(SpanOp::Merge(SpanRow {
    name: if rewritten { tool_span_name(&tool_name, &input) } else { String::new() },
    input: rewritten.then_some(input),
    // ...
}));

• Unchanged input is not resent. Most calls are never rewritten, and Write/Edit inputs can carry whole files, so the merge only includes input when it actually changed. An empty name leaves the existing span name as is in the sink.
• A missing tool_input keeps the proposed input. raw_tool_input() tells "the hook reported no input" apart from an empty one, so the {} fallback in tool_input() can't overwrite the original.
• Denied calls follow the same rule. The permission check evaluates the rewritten call, so PermissionDenied.tool_input is what was actually denied.
• Replays stay no-ops. A duplicate terminal hook for an already-finished call is still ignored, so it cannot bring back the proposed input.

raw_tool_input() is now the only place that looks up the tool_input/input keys; tool_input() is built on top of it.

The test helper reduce() now applies input and name from merges (it silently dropped them before, which is how this went unnoticed). The tool-span tests also share one tool_span() lookup helper, and two overlapping hook-outcome tests are merged into one.

Known limitation

Braintrust deep-merges object fields on merge. If a hook removes a key from the input, the old key stays on the span; changed values are replaced correctly. Fixing that needs _merge_paths support in the sink, which the SDK's SpanLog builder doesn't expose yet.

🤖 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 2026-10-19T23:33:20.298216Z	2527fd7	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.