fix(actions): recover the action array from one-shot agent responses - #920
Conversation
Detect Actions has returned zero actions for every repo added since 2026-07-07. The agent explores the project and answers correctly, but the response never survives parsing. Two things had to line up: - `BasicMessageWriter` — the writer behind `run_acp_prompt` — appended `[Tool: …]` / `[Result: …]` markers to the same buffer it collects agent text in, so `get_text()` returned a transcript rather than an answer. Every caller of that buffer (action detection, run-phase autodetect, badge short names) parses it as JSON. - `extract_json_array` sliced from the first `[` to the last `]`. Once the transcript led with `[Tool: Terminal]`, that slice covered the whole marker-riddled transcript and never parsed. Stop recording tool activity in `BasicMessageWriter`, and replace the first-bracket-to-last-bracket slice with a backward scan: candidate `[` positions are tried from the end of the response, each prefix-parsed into `Vec<SuggestedAction>`. That takes the agent's final array, ignores trailing text such as a closing code fence, and skips arrays of some other shape. Verified against the 8 KB response logged for `block/berd`, which the old extractor rejected and the new one parses into all 32 actions. Detection results are not backfilled: contexts already marked detected need the manual Detect Actions button pressed once. Signed-off-by: Matt Toohey <contact@matttoohey.com>
…rses Review dae123fa flagged a diagnosability regression in the backward-scan extractor: a near-miss array — valid JSON, but with e.g. one malformed actionType among 32 otherwise-valid actions — was silently skipped, and the caller only saw the generic "could not find valid JSON array" error. The old first-to-last-bracket slice at least surfaced the serde error. Now, when no candidate deserializes into Vec<SuggestedAction>, the scan remembers the typed serde error from the closest candidate that parses as a JSON array of objects, and reports that instead of the generic message. Arrays of non-objects (file lists, exit codes) still don't count, so their unhelpful type errors can't mask a real near-miss earlier in the response. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: Matt Toohey <contact@matttoohey.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1da25ef403
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| let candidate = &response[start..]; | ||
| let mut stream = | ||
| serde_json::Deserializer::from_str(candidate).into_iter::<Vec<SuggestedAction>>(); | ||
| match stream.next() { | ||
| Some(Ok(actions)) => return Ok(actions), |
There was a problem hiding this comment.
When an otherwise valid action contains a literal JSON-looking array in a string field, such as "command": "echo []", this raw reverse scan starts at that inner [. Because the streaming deserializer accepts [] without checking the trailing quote and remaining object text, it returns an empty Vec<SuggestedAction> immediately and silently discards the enclosing valid action array. The previous whole-response parse handled this correctly; candidate discovery needs to be string-aware or otherwise verify that a candidate is an actual top-level array before returning it.
Useful? React with 👍 / 👎.
1 participant
Summary
Detect Actions has returned zero actions for every repo added since 2026-07-07. The agent explores the project and answers correctly, but the response never survives parsing. Two things had to line up:
BasicMessageWriter— the writer behindrun_acp_prompt— appended[Tool: …]/[Result: …]markers to the same buffer it collects agent text in, soget_text()returned a transcript rather than an answer. Every caller of that buffer (action detection, run-phase autodetect, badge short names) parses it as JSON.extract_json_arraysliced from the first[to the last]. Once the transcript led with[Tool: Terminal], that slice covered the whole marker-riddled transcript and never parsed.Changes
acp-client:BasicMessageWriterno longer records tool calls or tool results, sorun_acp_promptreturns the agent's text only. Documented on the type, onrun_acp_prompt, and in the crate README.builderbot-actions: replaced the first-bracket-to-last-bracket slice with a backward scan. Candidate[positions are tried from the end of the response, each prefix-parsed intoVec<SuggestedAction>. That takes the agent's final array, ignores trailing text such as a closing code fence, and skips arrays of some other shape.actionTypeamong 32 valid actions) stays diagnosable. Arrays of non-objects (file lists, exit codes) don't count as near misses, so their unhelpful type errors can't mask a real one earlier in the response.Verification
Verified against the 8 KB response logged for
block/berd, which the old extractor rejected and the new one parses into all 32 actions. Added unit tests covering prose-wrapped arrays, bare arrays, fenced arrays after tool markers, last-array-wins, non-action arrays, near-miss shape errors, and the writer dropping tool activity.Note
Detection results are not backfilled: contexts already marked detected need the manual Detect Actions button pressed once.
🤖 Generated with Claude Code