Change cleanup logic from "check and maybe kill" to "kill unconditionally".
The previous approach allowed agents to read stale output from lingering
external agent processes, causing context pollution and reuse of old data.
New behavior:
- Kill ALL background shells as the FIRST action
- Do NOT read output from stale shells (BashOutput forbidden)
- Do NOT check what processes are doing - just terminate them
- Only proceed to Step 1 after all shells are dead
Adds support for delegating adversarial code reviews to external CLI agents
(Codex, Gemini, or Claude) when available. This provides independent, unbiased
code reviews from a different AI model.
Changes:
- Add invoke-bash and set-var tags to workflow.xml execution engine
- Add external_review_agents configuration to install-config.yaml
- Rewrite code-review workflow to detect and invoke external agents
- Cache agent detection in config.yaml to avoid repeated CLI checks
- Add fallback to built-in review if external agents unavailable/fail
- Update checklist to reflect new external agent workflow
External agent invocation:
- Codex: codex exec --full-auto "prompt"
- Gemini: gemini -p "prompt" --yolo
- Claude: claude -p "prompt" --dangerously-skip-permissions
- Fix checklist to only accept 'review' status (not 'ready-for-review')
- Include MEDIUM issues in done/in-progress status determination
- Initialize and track fixed_count/action_count variables for summary
- Add sprint-status.yaml sync when story status changes
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude <noreply@anthropic.com>
Scott Jennings
Alex Verkhovsky
Brian Madison