oad-debug

name: oad/debug description: OpenAgentd workflow for investigating bugs, regressions, sessions, and runtime issues.

Debug the reported issue.

Workflow:

1. Triage the report

   • Extract the symptom, expected behavior, reproduction steps, affected surface (backend/frontend/desktop/agent/provider), session id, workspace, model, logs, and timing clues.
   • If the report is ambiguous, inspect available evidence first; ask only when a missing decision blocks safe progress.

2. Choose the fastest evidence path

   • Live session / agent behavior: prefer manual/ helpers before ad-hoc probing:

     uv run python -m manual.health
     uv run python -m manual.backend_log
     uv run python -m manual.backend_log --contains drop_partial_tool_call_bad_json
     uv run python -m manual.team_sessions --id <SESSION_ID>
     uv run python -m manual.team_history <SESSION_ID>
     uv run python -m manual.team_timeline <SESSION_ID> --full
     uv run python -m manual.team_sse "message" --session <SESSION_ID>
     uv run python -m manual.queued_injection
     

   • Backend/API issue: hit the smallest route/service path, inspect logs, then add/adjust pytest coverage.
   • Frontend issue: inspect relevant components/hooks/stores, run focused bun checks/tests, and use existing UI state patterns.
   • Desktop/CLI/provider issue: inspect the specific command/provider path, environment assumptions, and existing smoke scripts under manual/ or manual/try_providers/.
   • See manual/AGENTS.md for the full manual script catalogue.

3. Reproduce narrowly

   • Recreate the smallest scenario that demonstrates the bug.
   • Match the user's mode/workspace/model/message sequence when relevant.
   • Capture durable evidence: raw HTTP response, persisted history, SSE events, logs, failing test output, or UI state.

4. Diagnose from code and evidence

   • Search for existing patterns before editing.
   • Identify the boundary that failed: route validation, persistence, queueing, stream emission, agent loop, hook, tool, provider, frontend store, or renderer.
   • Preserve unrelated work; do not reset or overwrite changes you did not make.

5. Fix surgically

   • Make the smallest change that addresses the proven root cause.
   • Add or update focused regression coverage at the closest layer.
   • Update related docs only when behavior, API contract, or operator workflow changed.

6. Verify and report

   • Re-run the reproduction and the focused tests/checks for touched areas.
   • If feasible, run the repository's standard lint/type/test commands for the changed surface.
   • Report root cause, changed files, checks run with results, and any remaining risk or unverified area.