Runs an iterative multi-phase quality loop that refines code changes until convergence.
npx skills add https://github.com/camercu/dotfiles --skill harden-camercu
Please help me install this Agent Skill. Skill: harden Source: https://github.com/camercu/dotfiles/tree/main/common/.config/agents/skills/harden Command: npx skills add https://github.com/camercu/dotfiles --skill harden-camercu
.harden/LEDGER.md plus a findings file, so runs survive /clear, crashes, and session limits. - Fresh-agent evaluative passes: Delegates judgment passes (review, simplify, architecture) to fresh agents so the reviewer is never the author, with a spawn contract that forces findings to disk as they are found. - Use Case: After implementing a new public API, invoke the loop to dogfood the surface as a real consumer, gate fix decisions with the user, land TDD slices with per-slice commits, and run independent review passes until convergence. ## Quick Start Ask the AI to harden the current change with a full quality pass and keep going until a complete round comes back clean.High-intent search queries and answers about installing and using this skill.
Invoke the harden skill on the change and it sequences dogfooding, decision gates, TDD slices, and post-implementation passes (review, soundness, simplify, architecture, test-health, docs). The loop repeats until a full round finds nothing significant, then confirms the original friction is gone.
Evaluative passes run in fresh agents that receive only the commit range, verification matrix, and authority-doc locations, never the author's rationale. The implementing context relays the independent verdict verbatim, so the reviewer is never primed by its own work.
Yes. Run state is checkpointed to an on-disk ledger at .harden/LEDGER.md and findings are appended to a findings file as they occur. After a /clear or crash, re-invoking resumes at the recorded position instead of restarting.
Repo-wide scope is decomposed into a prioritized slice queue ranked by consumer-surface breadth, correctness risk, churn, and staleness. Each slice runs the full loop independently, with checkpoints between slices and cross-slice findings re-queued.
Skip or scale down for changes with no consumer surface, such as internal test tweaks, doc edits, or lockfile updates. Passes acting on consumer surface are marked N/A in the ledger rather than run, and machinery is right-sized to blast radius.
abzhawRun verification commands and require explicit confirmation before claiming completion.
lithiumtoastStage and commit a single logical change after three-gate manual approval.
raddueOrchestrate the full software development lifecycle with design, planning, TDD, review, and quality gates.
ConsiliencyWrites an implementable plan for one bounded code change with verification and handoff.