These skills supply mechanics within the root-owned lifecycle, durable learning, and user-intent discovery. They do not own nested orchestration stages.
k-ai-kb
Automatic hooks retrieve and stage relevant capsules; the root owns admission and one final verified learning batch. Preserve relevance/workspace filters and admitted-ID deduplication. Do not add a recall or scribe invocation per skill, correction or worker. When delegation is forbidden, the same search-first, evidence-backed write/readback mechanics run inline without another model.
k-proof
| Field | Value |
|---|
| Use when | an explicit proof-ledger/receipt request, auditable security/auth/data-migration/destructive effect, or named handoff/resume consumer needs a durable freeform receipt |
| Source | exact_k-proof |
| CLI | ,proof stores proof state outside worktrees under $AGENT_PROOF_HOME, $XDG_STATE_HOME, or ~/.local/state; reports require a finalized ledger with an intact seal |
| Boundary | runtime/UI/external checks, multi-file scope, failed commands, and late completion challenges use inline evidence unless one of the receipt triggers above independently applies |
k-proof is available in two ways. Explicit receipt requests route through the skill frontmatter and SKILL.md. Non-review/non-build iteration gets the same narrow receipt gate from the always-on SOP and the shared verification prefix that perturn_recall.py re-injects after material context growth or a compaction. The ledger is a durable receipt, not verification itself: evidence collection remains mandatory and inline by default. When a receipt trigger applies, choose the topic and criteria before ledger-bound evidence collection, finalize the ledger, and only then generate a report.
| Field | Value |
|---|
| Routing | model-invoked only after that explicit user intent; ordinary task size, complexity, or instructions to continue never trigger it |
k-interview-me
| Field | Value |
|---|
| Use when | reverse-interviewing the user until intent is fully clear |
| Source | exact_k-interview-me |
| Routing | manual |
k-spec
| Field | Value |
|---|
| Use when | developing an idea, feature request, or bug into a compact packet with planned final acceptance checks |
| Source | exact_k-spec |
Fork-closing consults a domain overlay's planning fork checklist when the verified target repo has one. Forks that cannot close locally (external sign-off, another team's decision) go in the packet's External dependencies section — owner, blocked criteria, recommended default — instead of blocking assembly; consumers must not start blocked criteria. Plan checks without running a red-check ceremony merely to approve the packet. The packet also carries an impact map (affected callers/consumers, invariants, co-edit set) from the SOP §3.1 shared assessment; none requires the light-path proof.
k-build
| Field | Value |
|---|
| Use when | implementing an approved spec packet through production and one integrated final verification |
| Source | exact_k-build |
| Routing | manual |
An already-authorized target needs no duplicate approval gate. Strong research settles material questions; implementation-band workers produce substantial settled edits; deterministic tools execute known mechanical operations. The root integrates artifacts and owns final checks and strong review/refutation. Workers do not run private QA. Every co-edit-set member named in the impact map is updated in the same change or recorded as unaffected with evidence, and the final criteria verification checks that. Commits, pushes and publication retain their separate authority requirements.
k-converge
| Field | Value |
|---|
| Use when | a claim or changeset should be re-attacked until a round comes back dry (mutation probes + fresh refuters) |
| Source | exact_k-converge |
| Routing | manual (disable-model-invocation: true); build/review/light-review do not invoke it automatically |
Declare the exit (a full round with zero changes to code, tests, or published text and no unresolved findings or mutation verdicts) and the correctness-only filter (vacuous test, production bug, false published claim; everything else refused, not deferred) before round 1. Each round pins a baseline snapshot, mutates every behavioral change before arguing, fans out fresh k-agent-adversarial-verifier refuters on distinct dimensions, re-verifies their findings, applies in-scope fixes through the implement lane, and reruns the covering probes plus required regression checks. A changed round repeats; a dry round stops. Authorship never gates entry: on someone else's branch, the mutation and refutation steps still run against a disposable worktree and fixes come back as proposals. Close with the honest residue — what the loop never covered.
The shared workflow-handoff.md reference preserves the caller's frozen scope, criteria, receipts, approval and unresolved decisions. Existing read-only, ownership, compatibility, commit and publication gates remain binding; the handoff grants no extra edit, commit, push, or publication permission. No worker owns convergence or another model invocation. This is the only unbounded loop in the setup — an ordinary review fix pass is one round and must not be looped to imitate it.
k-text-tournament
| Field | Value |
|---|
| Use when | the user explicitly requests comparison of materially different prose alternatives against a stated rubric |
| Source | exact_k-text-tournament |
| Routing | manual (disable-model-invocation: true); ordinary prose edits do not trigger it |
| Boundary | no evaluator spawn, two-order judging, worker SELF_CHECK block, or repeated tournament per edit |
Produce useful alternatives and explain their tradeoffs while preserving instruction safety and factual fidelity. In a larger task, alternatives belong to Understand/Produce and final assessment belongs to the existing Verify stage. Comparison stays inline unless a substantial independent production assignment warrants isolation; it never creates another verification workflow.
k-improve-local
| Field | Value |
|---|
| Use when | proposing one evidence-backed improvement to local changes |
| Source | exact_k-improve-local |
| Routing | manual |
k-improve-branch
| Field | Value |
|---|
| Use when | proposing one evidence-backed improvement to the current branch, PR, or issue goal |
| Source | exact_k-improve-branch |
| Routing | manual |
k-improve-targeted
| Field | Value |
|---|
| Use when | proposing one evidence-backed improvement to a targeted codebase area |
| Source | exact_k-improve-targeted |
| Routing | manual |
k-improve-codebase
| Field | Value |
|---|
| Use when | proposing one evidence-backed improvement to the whole codebase |
| Source | exact_k-improve-codebase |
| Routing | manual |