feat(brain): add code findings persistence and enhance explore agent reports #27
Conversation
humanbeeng
commented
Jan 13, 2026
humanbeeng
commented
- Add FindingsPersister interface for caching/deduplicating explore results
- Implement findings cache with similarity matching for query deduplication
- Update Explore() to accept issueID for findings persistence
- Enhance explore agent system prompt for comprehensive reports:
- Strategy: encourage thorough exploration over efficiency
- Token budget awareness (~60k tokens for medium thoroughness)
- Output format: numbered sections, tables, code snippets
- Wire FindingsPersister into orchestrator
- Remove outdated token consumption specs
…reports - Add FindingsPersister interface for caching/deduplicating explore results - Implement findings cache with similarity matching for query deduplication - Update Explore() to accept issueID for findings persistence - Enhance explore agent system prompt for comprehensive reports: - Strategy: encourage thorough exploration over efficiency - Token budget awareness (~60k tokens for medium thoroughness) - Output format: numbered sections, tables, code snippets - Wire FindingsPersister into orchestrator - Remove outdated token consumption specs
![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: 1d2cad0603
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 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 address that feedback".
| issue.CodeFindings[i] = finding | ||
| _, err = p.issues.Upsert(ctx, issue) | ||
| return err |
There was a problem hiding this comment.
AddFinding reads the issue earlier in the same function and then writes the whole record back with Upsert. Because executeExploresParallel spawns multiple explore goroutines, two explores for the same issue can race: both read the same CodeFindings slice, append/replace in memory, and then the last Upsert overwrites the other’s changes. In that scenario one of the newly generated findings is silently dropped, which defeats caching/deduplication. Consider serializing writes per issue or adding an atomic append/merge method in the store (or an optimistic version check) to avoid last-write-wins loss.
Useful? React with 👍 / 👎.
