collie

Launch Collie-style fully automated feature development loop (brainstorm → plan → review → flow → rubric review)

驱动跑长测试 → 观察 → 修复 → 重跑的自迭代闭环 (Karpathy autoresearch 风格)

Use this agent when a plan document has been written and needs verification before implementation begins. Specifically, dispatch this agent after completing a plan document to verify it is complete, aligned with the spec, and has proper task decomposition that an engineer could follow without getting stuck. <example> Context: The user has just finished writing an implementation plan document for a new feature. user: "I've finished writing the plan at docs/plans/2026-04-07-auth-redesign-plan.md based on the spec at docs/auth-redesign-spec.md" assistant: "Great, the plan document is ready. Let me dispatch the plan-doc-reviewer agent to verify it before we start implementation." <commentary> Since a complete plan document was just written, use the Agent tool to launch the plan-doc-reviewer agent to verify the plan is ready for implementation. </commentary> </example> <example> Context: A subagent has just finished generating a plan document as part of a larger workflow. user: "Please create a full implementation plan for the new pipeline refactor" assistant: "I've drafted the plan at docs/plans/2026-04-07-pipeline-refactor-plan.md. Now let me use the plan-doc-reviewer agent to verify it before proceeding." <commentary> After generating a plan document, proactively use the Agent tool to launch the plan-doc-reviewer agent to catch any gaps before implementation starts. </commentary> </example>

Pre-flight environment check SKILL for the collie autoiter command. Called by the main autoiter SKILL (collie:autoiter) in the §3.5 post-ExitPlanMode recovery path, after worktree creation, before Stage 1 iteration begins. Runs 5 checks: (1) trigger dry-run, (2) scalar extraction validation, (3) observability validation (Monitor/Read-tail + kill signal), (4) persistent directory writability. Outputs prepare-report.md with PASS/FAIL evidence. Returns failure signal to caller on any FAIL — does NOT fix issues. Supports skip_prepare bypass and is idempotent (skips all checks if prepare-report.md already exists). Do NOT invoke directly from user prompts; this is an internal skill invoked exclusively by collie:autoiter.

Main loop orchestrator for collie. Drives the 'run → observe → triage → deep-verify → fix → rerun' self-iteration pipeline. Called by commands/autoiter.md on every ralph-loop session restart. Implements a persistent state machine (§3.5) across ralph-loop session restarts: fresh-start → Stage 0 (Discovery planmode) → Stage 1 (kickoff) → Stage 2 (run trigger) → Stage 3 (observe + auto-recovery) → Stage 4a (Triage) → Stage 4b (Deep Verify) → Stage 5.0 (fix-plan) → Stage 5.1 (flow) → Stage 5.2 (G6 diff audit + rerun) → Stage 6 (rollback + stop check). Completion signal: <promise>Collie: AUTOITER DONE</promise> (emitted ONLY by the §3.5 terminal branch after ralph-loop restart, NEVER inline).

Post-planmode implementation workflow with quality gates. Use immediately after exiting planmode to execute an approved plan.

Boids-inspired memory management for AI agents. Evaluates conversation content against a fixed decision tree to decide what to remember, where to store it, and when to consolidate. Activate on Stop hook and PreCompact hook for automatic memory capture. Also activate when the agent detects information during conversation that might change future behavior — user corrections, preferences, role info, project constraints, or external system references. Do NOT activate for knowledge questions, in-session operations, or debug conclusions.

Collie-style unified rubric reviewer. Enforces 13 red lines + 6 questions + ELEPHANT anti-sycophancy + Reflexion grounding. Internally dispatches Agent(model=opus) for isolation. Use in three contexts: (1) Plan mode — Target is a plan doc matching *-plan.md or under plans/; called in parallel with plan-doc-reviewer at /auto step ③ before ExitPlanMode. (2) Code mode — Target is a worktree diff or branch; called as flow TodoList item [collie-final-review] (Step 5.7) before worktree cleanup. (3) Ad-hoc — any file, diff, design doc, or subagent output needing Collie-style review.

10 hooks across 5 events