harness-forge

Create or upgrade a self-improving Claude Code harness in the current project — self-healing failure ledger, per-file lesson memory, regression evals, small-model elevation agents, deterministic safety hooks. Modes: (no args) auto-detect new vs upgrade; 'doctor' = validate only. Evidence-gated: scan, propose plan, confirm, apply, verify.

Self-maintenance pass over the .claude/ harness itself: validate every mechanical invariant, detect drift and dead weight, and propose upgrades. Run periodically, after any harness change, or via /loop.

Run one iteration of this project's maintenance loop: finish in-flight work, convert failures into lessons, checkpoint hygiene, audit the harness, or start the next pending task — whichever is highest-value right now. Also the remedy when capability-gate.sh blocks a source edit because maintenance is overdue.

Remove dead code, unused dependencies, commented-out blocks, and orphaned files once a checkpoint is reached — for long-term maintainability. Evidence-gated and reversible: nothing is removed without proof it is unreferenced, and build+tests must stay green after every batch. Run after a milestone/feature completes, not mid-implementation.

Convert this session's failures and user corrections into durable per-file lessons in .claude/memory/lessons/, following the evidence-backed write policy (retrieve-similar-first, delta updates, weighted vote lifecycle, hook promotion). Run at the end of any session that contained a correction or a repeated failure.

Holistic, single-pass senior-engineer review of the target project's actual code (not the harness) across security, performance, code quality, and production-readiness. Use for 'is this ready to ship', a full code review, a security audit, or a performance pass.

Structured debugging flow — check issues-solved/ first, reproduce, fix the root cause, verify end-to-end, and capture a new issues-solved entry if it was non-trivial. Use when a bug is reported or a test fails unexpectedly.

Test-driven implementation flow for one TASKS.md item — load, red, green, refactor, verify, archive — driving the six-status vocabulary. Use when picking up any pending task.

Measure and defend the always-loaded context budget (CLAUDE.md + unscoped rules). A self-improving harness accretes rules and lessons over time; without this regulator it silently bloats until instruction-following degrades. Run periodically, after adding rules, or when CLAUDE.md grows.

Session-resumption summary — reads git state, TASKS.md, and status.json to produce a terse Goal/Done/In-flight/Next/Watch-out summary. Use at the start of a session after time away, or pass 'handoff' to write one before ending a session.

Confirmed commit/push/PR pipeline — stage, commit, push, open a PR. Never runs unattended; every destructive or outward-facing step is confirmed first.