Git-native preparation & convergence skill — four routines: init/prepare/converge/prune. prepare is local-first: work stays on local prepare/* by default; remote push is a cross-machine opt-in (push-prepare-branches: yes); prepare freshness-rebases in-progress branches onto default (schema 8; opt-out rebase-prepare-onto-default), and converge offers the refresh. Human takeover is recognized by containment; the redundant prepare/<slug> is surfaced for cleanup, never auto-deleted. converge accept defaults to a handoff-only rename (prepare/<slug> → accepted-prefix/<slug>), with a human-marked maintenance exception; prune is a human-present, read-only-until-approved sweep of unused local + remote branches. Committed .tsugu/ is a three-part layer (policy.md + context.md + evidence/) with no infrastructure file; evidence/ is per-ref and emptied at landing, and no merge driver auto-resolves .tsugu/ any more, so the agent resolves such a conflict and declines one it cannot reconcile; accepted-prefixes default feature/* bugfix/* chore/*; state derived from refs, no intake layer; never auto-merges; invokes no user-installed skill by default; 015 adds a standing POST-HANDOFF cleanup block in context.md + an agent-md routing pointer; 017 adds a ## Blindspots section (prepare surfaces unknown unknowns, material + grounded) + a converge verify-findings reminder (no schema bump)
1 skill · 4 commands
Natural language constraints → deterministic test artifacts. Write constraints in structured Markdown, generate unit tests / PBT / mutation tests, enforce with deterministic tools.
3 skills
Assisted review loop — local reviewers answer blind and in parallel, each returns its own convergence verdict (loop-until-dry, facilitator owns K); a behaviour claim about a runtime prompt is settled by five independent simulation runs, not by a string comparison; a settled finding keeps the command that settled it, and the loop re-runs every kept command at the start of each round and before each commit; a finding no command settled keeps nothing; roster enrolled per host via /review-loop:init (routine or on-request reviewers, a private-content rule for each); the author-stop shape is published at A0 and then held (author pass, one batched message per round that has a T2/T3 finding, the hand-off, plus named exceptions that re-publish it), and a T3 that comes back after a fix is read as a drift signal: that round's message re-confirms the direction of the whole change with the author, as well as asking for the item's own fix; never merges autonomously
1 skill · 2 commands
FP-thinking review and refactor rules for pre-RSC React projects
1 skill · 1 command
Describe multi-step agent workflows using an Arrow-style DSL (>>>, ***, &&&, |||, ?, loop, \, let...in, (), ;) with epistemic operator conventions and structural validation
1 skill
Explore an unfamiliar codebase together with the user, one step at a time, through its data flow and its types
1 skill
Socratic dialogue for reflecting on human-AI stewardship
1 skill