tyabu12

Run one Mac-side model evaluation battery cycle — drive a candidate GGUF through the 6-cell scenario battery on pastura-harness with real local-LLM inference, score each transcript against the scenario-refine rubric, and append the local eval digest. Use when…

Consume agent-ready GitHub issues into Draft PRs — one overnight queue run. Fetches open agent-ready issues oldest-first, implements each on an agent/issue-<N> branch, tests, passes a mandatory code review, opens a Draft PR, and appends the run digest. Use…

Run one scenario-factory cycle — generate 3 fresh scenario YAMLs, execute each on pastura-harness with real local-LLM inference, judge the transcripts in-session, and append the local digest. Use when the user asks to run the scenario factory, run a factory…

Run one scenario-refine cycle — evaluate a rotating slice of the EXISTING shipped scenario inventory (bundled presets + gallery) on pastura-harness with real local-LLM inference, judge each against a category-aware rubric with regression detection,…

Orchestrate feature implementation from plan to PR — worktree isolation, TDD, review, and PR creation.

Prune the prose this branch added — compress or delete self-evident, redundant, or duplicated comments and documentation, behind verification gates that stop a deletion from breaking something load-bearing. Use when asked to simplify or compress docs or…

Run one code-health-audit cycle — sweep a code layer (Engine/LLM) with category-weighted read-only finders, Vet each finding, and write a ranked digest (files no issues, schedules nothing). Use when the user asks to run code-health-audit, audit code health,…

Detect mechanically-verifiable documentation/ADR drift and act on it — open a docs-fix Draft PR for fixes whose value is uniquely determined by an authoritative source, or file an issue (with confidence + counter-evidence) for drift that needs human judgment.…

Fan out a batch of independent, well-specified tasks (backlog items, issue numbers) to parallel worktree-isolated implementer subagents, each ending at a draft PR plus a critic review. For a single large feature, use the project's own /orchestrate instead…

Scaffold or upgrade a project-owned /orchestrate skill (plan → issue → worktree → implement → review → PR) in the current repo's .claude/skills/. Use when a project needs an orchestration workflow, or to upgrade one generated earlier.

Triage per-user memory: promote durable entries into a rules file (global or project .claude/rules/) and retire SHIPPED trackers. Use for a memory triage round.

Multi-perspective, bias-resistant risk review of a working-tree diff or a plan/design decision: pre-mortem risk axes fanned out to parallel `critic` subagents, then synthesized. Read-only by instruction. Use for change-impact and design risk.

Monthly evidence-driven retrospective of this kit's skills: gather friction logs, invocation traces, PR outcomes and docs drift, then propose SKILL.md fixes as a draft PR — only where evidence exists. No evidence → no change.

Extract and format a work log from conversation history under ~/.claude/projects/. Argument is the number of days to look back (default 7).

Draft a new Architecture Decision Record into the repo's ADR directory (new files only — not for amending an existing ADR). Use when asked to write an ADR, record an architecture decision, or document a design trade-off.