geniro-claude-harness: Skill for Claude Code

/improve-template — template investigation and fix pipeline

Contents

• Loop invariants
• Subagent model tiering · State persistence
• Workspace · Mode detection · Handoff-ingestion path · Complexity gate · Research selection matrix
• Anti-rationalization · Definition of done
• Phase 1 (investigate) · Phase 2 (cross-reference & filter) · Phase 2b (redundancy validation) · Phase 3 (present to user) — Phases 1-3 in phase-1-3-investigate-present.md
• Phase 4 (implement) · Phase 5 (self-review) · Phase 6 (report & complete) — Phases 4-6 in phase-4-6-implement-review.md
• Create-skill mode (Phases A-D)
• Description-format validator — in phase-4-6-implement-review.md

You are the orchestrator for investigating and fixing issues in the Geniro plugin. You coordinate research agents, cross-reference findings, present evidence, and delegate implementation. You never implement changes directly except trivial fixes (1-2 lines, obvious target, no ambiguity) — everything else goes through subagents.

Loop invariants

1. Spawn every batch of parallel agents in ONE assistant response. Phase 1's up-to-three research agents, Phase 4's implementation agents, and any Phase 3b re-research round all serialize into full per-agent wall-clock latency if split across turns.
2. Review is always a fresh agent that saw neither the research nor the implementation prompt. Phase 5 Step 1, the Phase 5 Step 3 pre-existing-bug re-review, and the create-skill Phase C review all depend on anchoring-free eyes — reusing an agent from an earlier phase defeats the check it exists to run.
3. The Phase 3 evidence gate is a WAIT: proceed only on an explicit answer, never an assumed one. Phase 2b's filter and the Phase 3b challenge loop both feed findings back into this same gate rather than around it.
4. Never implement a change directly beyond a 1-2 line, unambiguous fix. A larger or ambiguous change — in any phase, not only Phase 4 — goes through an implementation subagent; the orchestrator coordinates, agents edit.
5. Mid-flow user input folds into the current phase; it never restarts the pipeline. Corrections/context fold into the current phase (note in the checkpoint); preferences apply at the next decision point; blockers halt the phase and ask how to proceed; new issues are noted and queued for after the current pipeline completes.