Opinionated multi-model delegation for very large repos.
Use when a completed change needs a readiness check and PR-body draft: require fresh cross-family review, per-lane verification, applicable UAT and unchanged frozen paths; prepare an engineering summary and branch handoff only, without push, PR creation, merge, deploy or publication.
Delegates: omh:reviewer/omh-verification-gate
Contract: 1
Python 3.11+ for bundled read-only routing; explicit project model choices and genuinely bound executor/reviewer channels. Optional evidence assessment requires the pinned OMH peer on Hermes, Node 18+, Python 3.11+ and verified loaded provenance and reviewer bindings.
Install: npx skills add thunderock/thunderkit -s tk-ship -g
Thunderkit owns this local preparation gate. Inspect existing evidence, report whether the exact change is ready, and print a branch handoff and PR-body draft. Preparation is not delivery authorization, and a native assessor is not a replacement for Thunderkit's completion gates.
Model class: cheapest selected executor, as assigned to preparation by tk-router. Choose only within classes.executors using known cost/availability, not a hardcoded model or an invented price ranking. The optional evidence assessor separately uses classes.reviewers. Read this skill's model roster, catalog and config schema; validate with the bundled model helper. Preserve all choices, array order, literal reviewers "all", the family minimum and frozen paths. Legacy normalization is a preview, not a write.
Every operation here is model-bearing, including owned/off/fallback assembly. Before work, prove the actual executor channel's effective host/provider/wire-model and supported effort match its selected catalog member; prove reviewer bindings separately for any assessment. Valid configuration, a model name in a prompt or a skill load does not bind the current root. Represent selected plural members without silently narrowing the set. Never change global config, credentials, effort or fallback chains, or assume a running session changes after a config edit. Missing selected channels block work rather than using an arbitrary current model.
Follow delegation.md and dependencies.json. Set SKILL_ROOT to the directory containing this loaded skill and PROJECT_ROOT to the actual project/worktree being prepared. Resolve resources only from this skill's own scripts/ and references/; missing assets block, with no borrowed checkout or presumed sibling copy. Config and capability inputs must resolve within the explicit project root, without symlink escapes. When considering a native component, CAPABILITIES is a current project-contained snapshot of actual loaded descriptors, provenance, tools and effective bindings, not secrets.
python3 "$SKILL_ROOT/scripts/tk-resolve.py" \
--skill tk-ship --operation prepare --project-root "$PROJECT_ROOT" \
--config "$PROJECT_ROOT/.thunderkit/config.json" --capabilities "$CAPABILITIES" --jsonprepare is the only operation and the default. With delegation off or no enabled target, omit --capabilities and perform no native discovery, loading, routing, doctor or installation. Configuration and the owned executor binding are still required.
Only omh:reviewer/omh-verification-gate is eligible, on Hermes in component mode with tool:skill and model-binding:reviewers. Verify oh-my-hermes@2.0.3, its pinned source, bundle root containing manifest.json, canonical name verification-gate, loaded entrypoint and every provenance-map SHA-256, including the shared skills/guide/omh-routing/references/skill-common-rail.md companion. The bundle root is not the skills directory or HERMES_HOME. Missing/quarantined files, self-reported hashes or a same-named foreign skill do not qualify. The local registry supplies the trusted identities.
On delegate, use only the verified categorized selector through an actual supported, selected-reviewer-bound host channel. The internal address is not a slash command. Give the assessor the immutable target, supplied evidence and a bounded read-only question; it returns findings only. It cannot edit, rerun a workflow, change gates or deliver. Prefer an already proven nonmutating binding; do not call omh_delegate_route from this read-only preparation. If that boundary cannot be enforced, do not invoke the component. OpenCode and Codex have no eligible OMO preparation target: never substitute a native delivery workflow.
Keep the resolver's fixed decision record immutable: schema, operation, reason, target, requested/effective bindings, null pre-invocation observation, runtime home and evidence paths. Exit 0 means only that routing was computed, not work executed or readiness proved. Blocked or malformed input stops dispatch. Later invocation failures and findings get separate records; do not rewrite delegate to claim native completion or erase a failure with an owned route.
Bind all checks to one target: actual project/worktree and branch, base/head commit and tree, exact diff digest including in-scope staged/unstaged/untracked bytes, approved scope and relevant config/plan/evidence artifact identities. Name excluded local changes. Missing identity is unverified, not a guessed hash. Recheck these identities immediately before reporting readiness.
REVIEW.md and underlying independent evidence cover every lane; each lane is complete, with no unresolved blocker or major finding or assessment. A done label or report's existence alone is insufficient; retain minor findings and risks.
unchanged review_families_min catalog families, including one different from each lane's author. Require every explicit reviewer; "all" considers all catalog candidates and preserves unavailable optional candidates. Multiple harnesses or same-family variants do not add families. single-family-review, missing author identity or quota-lost required review means not ready, never a reduced-confidence pass or a lower minimum.
on the exact target: command/argv, cwd, exit status and sanitized output/counts. Missing, failed, skipped required or stale checks block. Inspect supplied evidence here; missing execution returns to its owning stage, not an invented pass or an automatic test/fix loop.
CLI/API/visual surface. Applicable UAT requires current actual observations in UAT.md, with no gaps or unresolved failures. Preserve an explicit not-applicable decision and its reason/scope/target identity; do not turn it into a claimed executed pass. An optional stage that never ran is not evidence that required UAT is unnecessary. Missing applicability or required UAT blocks; a prior not-applicable decision is stale if the surface/scope changes.
against config.json.frozen_paths, including additions, deletions and renames. A changed frozen path blocks; do not unfreeze it, exclude it from the diff or edit config to pass.
changes invalidate dependent review, verification and UAT. A newer timestamp or a native PASS on a narrower claim cannot refresh them. Missing identities block readiness.
HOLD/BLOCK prevents readiness until the named issue is resolved with current evidence. Unknown/incomplete native outcomes remain blocked/unverified. Native PASS adds evidence only; it cannot replace any gate above. An optional assessor never invoked is recorded as not used, not as a passed assessment or a missing required UAT waiver.
Any failed, missing or ambiguous gate means not ready. Name the exact gap, affected target and owning correction/check; retain useful evidence without certifying readiness. Check that tk-review, tk-verify-work, tk-router or any other requested sibling is actually available before handoff; absent siblings are prerequisites, not assumed paths or implicit installs.
Assemble, don't re-derive: goal + non-goals from SPEC.md; decisions from CONTEXT.md/ DECISIONS.md; lanes + verification from PLAN.md/REVIEW.md; risks from PLAN.md; UAT evidence from UAT.md. These are inputs, not public citations. Trace each output claim to underlying engineering facts: actual commits/diffs, code behavior, tests, verification commands and observed results. If a claim lacks that support, omit or qualify it; do not invent coverage.
Write normal engineering prose: purpose and scope, implementation choices and trade-offs, tests and their real results, compatibility/migration impact, remaining risks and limitations. The public draft contains no .thunderkit or planning-artifact paths/names, internal receipts, stage/lane bookkeeping, model-routing history or process narration. Do not disguise internal filenames as aliases or encoded citations. Keep internal traceability in project context, separate from the PR body; this does not change the product's committed-context convention.
owned/disabled, no enabled target or an unsupported host uses the same preparationprocedure, but only through a genuinely selected-executor-bound channel. A computed owned route cannot waive model readiness or the completion gates.
owned fallback, not an undeclared peer or model substitution. Preserve the resolver reason. If owned binding, explicit selections or required evidence cannot be honored, stop and record a separate blocked outcome. Give operator guidance, never install, log in or repair global configuration automatically.
partial output, mark blocked/unknown, and inspect that same session before any retry or fallback. If its termination/outcome is unprovable, remain blocked; never duplicate work. A captured resume ID does not prove that the session is currently runnable.
Return two clearly separated outputs:
current branch/base/head/tree/diff identities, in-scope and excluded local changes; each owned gate's result and named gaps; actual review-family coverage, every lane's verification and UAT applicability (including explicit not-applicable reasons). Preserve the unchanged resolver record and separate invocation outcome, requested/effective/observed model and family, package/version/selector, actual native artifact path/SHA-256 and genuine session ID using the common delegated-run fields. Unknown facts remain null/unverified. Keep native artifacts at their real paths, without moving, rewriting or mirroring native state.
gates pass. On failure, label any partial draft not ready and list blockers separately, never as a hidden warning beneath a readiness claim. No public artifact/process references.
Do not include credentials in either output. Readiness applies only to the recorded identity, not future edits. Print the proposed handoff; do not create or alter branches/commits, apply fixes, start missing stages or issue remote delivery commands as part of this skill.
No push, PR creation, merge (including automatic/local merge), deploy or publish commands. Never call OMO --ship/--make-pr, a deployment workflow or a release publisher. A ready summary, native PASS or request to run tk-ship grants none of that authority. Delivery needs a separate explicit user instruction outside this skill; stop at the local preparation result even when every gate passes.