DarkFactory
/

Rule 14 — Harness-agnostic agent runtime and conversational lifecycle

Requirement

An autonomous AI agent runs containerized in GitHub Actions (docker/Dockerfile.agent). It is harness-agnostic: no pipeline code knows which coding-agent CLI is executing.

  • Harness registry: .github/scripts/harnesses.py declares each CLI — Antigravity (agy), Claude Code (claude), OpenAI Codex (codex), Kimi (kimi), Grok (grok), Cursor (cursor-agent), and opencode (opencode) — as a binary, an argv template, the credentials it accepts, and its quota pools. Adding a harness is a data change; changing one is a configuration change.
  • No hardcoded invocation: every field is overridable at runtime through the AGENT_HARNESS_CONFIG repository variable, and the order through AGENT_HARNESS_CHAIN, so an upstream flag rename never requires a code change or a container rebuild.
  • Graceful degradation: harnesses whose binary is absent from PATH, or whose credentials are unset, are skipped rather than failed. An image carrying four of seven CLIs is a working image with a shorter fallback chain.
  • Resilience: exhaustion moves to the next account, then the next pool, then the next harness. Only when every account of every pool of every harness is spent does the agent checkpoint and block. Authentication uses repository secrets only; the account being run is named in the log, its credential never is.
  • Conversational lifecycle: incoming issues are auto-classified, the agent answers feedback and executes adjustments on Request issues and pull requests, bots are ignored to prevent self-reply loops, and on plan approval the agent opens a bot-authored Draft PR and runs a review loop bounded by MAX_REVIEW_ITERATIONS. On quota exhaustion the agent saves a checkpoint, moves the item to Blocked, comments the resume instructions, and exits cleanly.

Rationale

Outputs must be classified by their real outcome, not by exit code, and no provider's short-term quota may idly sleep while an unused account still has capacity.

Enforcement

  • docker/Dockerfile.agent image; .github/scripts/harnesses.py registry.
  • Runner/quota and review-cap tests in tests/test_agent_runner.py.
  • PRD section 7 states runtime and continuity outcomes.

Exceptions

None.

Change control

Owner is rotation (outcome classification, cross-provider handoff, repository-variable quota state, scheduled resume). Provider, model, and account details are manifest declarations, not rule text. Checkpoint filenames and concrete mechanics are operational details, not product requirements.