CC Iterator — Autonomous Development Loop

Drive OpenLinkOS project forward by managing CC (Claude Code) background tasks.

Core Loop (Pull-based)

1. •Check — process action:list for active CC sessions
2. •Harvest — If CC completed: verify with pnpm test, git push, close issue
3. •Next — gh issue list --state open → has open issue → start CC
4. •Wait — No open issues → HEARTBEAT_OK (never auto-create issues)

Issues come from discussions, user feedback, or bug reports. Patrol executes, it does not decide.

Starting CC

exec pty:true background:true workdir:/data/code/github.com/openlinkos/agent command:"IS_SANDBOX=1 claude --dangerously-skip-permissions -p 'PROMPT' --output-format text"

• •No timeout
• •Prompt ends with: When done: gh issue close N && openclaw system event --text "Done: SUMMARY" --mode now
• •Return immediately after starting, never poll-wait

CC Prompt Template

1. Task context: implementing X, Issue #N
2. Current state: N tests passing
3. Specific files and responsibilities
4. Requirements: tests, typecheck, build, no any types
5. Closing: git commit + gh issue close + openclaw system event

Failure Detection

• •Runtime < 2 min + minimal output = CC didn't start (API issue) → retry
• •Exit code 143 (SIGTERM) = killed, check if timed out
• •No output after 25+ min = likely hung → kill and retry
• •Check git status to confirm no code changes before retrying
• •If CC partially completed: check git log, continue from breakpoint

Lessons Learned

• •CC self-check won't catch CI environment differences (local has dist cache, CI doesn't)
• •Self-check prompt should include: verify CI workflow step ordering
• •Simple fixes (e.g. reordering CI steps) can be done manually, no need to spawn CC

Review Cadence

• •Every 3-4 development issues, create a batch review issue
• •Review checks: type safety, error handling, resource cleanup, API consistency

Project Context

• •Repo: /data/code/github.com/openlinkos/agent
• •Devlog: memory/openlinkos-devlog.md
• •GitHub: https://github.com/openlinkos/agent