[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7D2TQX8HD9Z730DRQVQJCTS":3,"skill-same:01M7D2TQX8HD9Z730DRQVQJCTS":54},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":32,"commitSha":33,"skillDirectory":34,"forksCount":35,"repositoryHtmlUrl":36,"repositoryHomepageUrl":6,"ownerHtmlUrl":37,"ownerBlogUrl":38,"ownerDescription":6,"updatedAt":39,"install":40,"resources":53},"01M7D2TQX8HD9Z730DRQVQJCTS","super-verify",null,"Turns a completion claim into receipts: for each acceptance criterion, the command or probe that would prove it, run fresh, with exit status, output digest, revision and environment identity recorded. Use when a specific implemented behaviour has to be shown to work. Not for judging whether code is well written, not for fixing what fails, and not for ratifying an account of a green run.","Pibomeister\u002Fagent-kit","Pibomeister","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F11633791?v=4","01M7D2NBHEKARMHN9MPV2J80XD","Eduardo Picazo","01M7D2NBJ3Z33SGSYCZ0KBD5DE",0,"MIT","https:\u002F\u002Fgithub.com\u002FPibomeister\u002Fagent-kit\u002Fblob\u002Fmain\u002F.cursor\u002Fskills\u002Fsuper-verify\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: super-verify\ndescription: \"Turns a completion claim into receipts: for each acceptance criterion, the command or probe that would prove it, run fresh, with exit status, output digest, revision and environment identity recorded. Use when a specific implemented behaviour has to be shown to work. Not for judging whether code is well written, not for fixing what fails, and not for ratifying an account of a green run.\"\nlicense: MIT\nmetadata:\n  ak_catalog_id: super-verify\n  ak:\n    mode: autonomous\n    autonomy_unenforceable:\n      - \"artifact-write is storage only: the host does not compute or check the artifact hash, so envelope hash binding is this package's own work.\"\n      - kb-write is not provided by the host; the knowledgebase adapter supplies it and refuses rather than falling back to a repository path when no knowledgebase is configured, which is why it does not cap this row while that adapter is attached (ruling `fail-closed-adapter-lifts-ceiling`).\n      - environment identity is not attested by the host. This package records what it observed of the environment and never reports an identity it could not read.\n      - independent-context and trusted-evidence are not provided by the host. The attached runner supplies them fail-closed; without it this skill is guided and receipts are host-unattested.\n---\n\n## When to use\n\nUse when a caller claims a specific behaviour works and the claim can be named as an acceptance\ncriterion: the export includes archived rows; a 400 is no longer retried; the migration runs twice\nwithout changing the second result.\n\nUse when a ticket's acceptance criteria need evidence before the work can be reported done, and the\nevidence has to be re-checkable by someone who was not in the run.\n\nUse when the code has moved since the last receipts were taken and a claim that was true of the old\nrevision has to be established again for the new one.\n\nUse when a repair has landed and what follows it is new verification rather than a re-reading of the\nreceipts taken before it.\n\n## Not for\n\nNot for an opinion about whether code is well written, well factored or well named. That is a review\nlane's judgement, and this skill has no instrument for it: a passing command says the behaviour\nholds, not that the code deserves to.\n\nNot for producing the fix when a check fails. A failing check is a reproduction, and a reproduction\nenters through `diagnose`, which states its hypothesis before any edit and emits either a verified\nbounded patch under a grant or a work packet — never a patch this lane wrote on the way past (ruling\n`diagnose-patch-or-packet-never-both`).\n\nNot for verifying a claim nobody stated. \"Check that everything still works\" names no criterion, so\nnothing would settle it and any result would be an impression.\n\nNot for ratifying someone's account of a run. A description of a green test suite is the thing a\nreceipt exists to replace, and re-reporting it with this skill's name on it launders an assertion\ninto evidence.\n\n## Authority\n\nAuthority: `model`. A controller or a parent skill starts it when a claim needs evidence; no slash\ncommand exposes it and no human act is required to start it.\n\nNo grant covers delegation, because no phase operation exposes this skill\n([`policies\u002Finvocation.yaml`](..\u002F..\u002Freferences\u002Fshared\u002Fpolicies\u002Finvocation.yaml)). The implementer may start the skill, but may not fill its verifier seat\nor author gate-eligible receipts. The runner assigns a distinct [`verifier`](..\u002F..\u002Freferences\u002Fshared\u002Froles\u002Fverifier\u002FROLE.md) seat, and closure depends\non that seat's receipts plus the applicable policy rule, never on the caller's confidence. A seat\nthat cannot be filled independently is `unavailable` and is never backfilled by the implementer,\nauthor, spec approver or another seated role (ruling `missing-supervisor-never-implementer`).\n\n## Inputs\n\nThe claim, resolved to named acceptance criteria ([`schemas\u002Fticket.schema.json`](..\u002F..\u002Freferences\u002Fshared\u002Fschemas\u002Fticket.schema.json) `acceptance_criteria`,\nby AC id). Absent, or phrased so that no command could settle it: stop and report `needs-input` with\nthe criterion it would need.\n\nThe repository at the revision being verified, readable (`repository-read`). Unreadable, or no\nrevision named: stop and report `failed`. A receipt whose revision is null is not a receipt.\n\nThe project's content-addressed verification recipe and surface declarations. With no matching\nrecipe, discover a candidate from project files, but do not substitute tests for a runnable surface;\npublish a recipe only after its complete build, launch, readiness, drive and cleanup path succeeds.\n\nOne runner-assigned verifier context (`independent-context`) and trusted evidence storage for the\nautonomous form. A fresh host context without attestation remains valid in guided mode and records\n`host-unattested`; the gate refuses it once the build gate records an implementer seat or the run\nholds any bypass or delegated grant record, for the rest of that run. Under the unattended policy\n(ADR-0014) it is a recorded note instead, whose boundary is the live merge-gate read (an exception\nto ruling `closure-requires-independent-verification`), and the gate is recorded with its evidence\ncommand, as [the unattended-lifecycle reference](..\u002F..\u002Freferences\u002Fshared\u002Freferences\u002Funattended-lifecycle\u002FREFERENCE.md) says.\n\nThe environment the checks run in, identified ([`schemas\u002Fverification.schema.json`](..\u002F..\u002Freferences\u002Fshared\u002Fschemas\u002Fverification.schema.json)). Test execution\nbelongs in the appropriate isolated environment, and repository test code does not receive production\ncredentials by default.\n\nOptionally, receipts from an earlier run. They are history: a receipt is evidence for the revision it\nnames and for no other, and one taken before the code moved is read as invalidated rather than as a\nhead start. The code moved when either the revision or the working-tree diff hash differs from the\nreceipt's (`common#\u002F$defs\u002Frevision_ref`); an uncommitted edit on the same revision is a move.\n\n## Workflow\n\n1. Resolve the claim into the acceptance criteria it is made of, by id. A claim that resolves to no\n   criterion stops here with `needs-input`.\n2. Resolve the matching project-declared recipe by id and content hash. If none exists, discover the\n   setup, build, launch, readiness, drive and cleanup steps from project files. Record a candidate\n   only after the whole path works; never write a host-specific recipe into the application tree.\n   Load [verification evidence](..\u002F..\u002Freferences\u002Fshared\u002Freferences\u002Fverification-evidence\u002FREFERENCE.md) before matching\n   a surface to an evidence kind.\n3. Seat [`roles\u002Fverifier\u002FROLE.md`](..\u002F..\u002Freferences\u002Fshared\u002Froles\u002Fverifier\u002FROLE.md) outside the implementer lineage. Give it the claim, criteria,\n   revision, recipe, permitted commands, environment and required evidence kinds—never the\n   implementer's narrative, claimed result or reviewer verdict. An unfillable seat is `unavailable`.\n4. Run each command fresh and complete, at the revision under verification. A partial run, a cached\n   result and a previous run's output are not this run's evidence.\n   Where a criterion declares a runnable surface, build and launch the application, wait for the\n   declared readiness signal, and drive the declared path or request. Tests, type checks and builds\n   may support the receipt; none substitutes for the running-app evidence kind the criterion names.\n5. Read the whole output: exit status first, then the counts the output reports. A suite that reports\n   failures while exiting zero is read by its output, and the disagreement is recorded as\n   `exit_disagreement` with `verdict_from: output` and an `output_reports` quote.\n6. Record one receipt per check with the command as an argument vector or the probe with its target,\n   the exit status, the digest of the relevant output, the revision, the environment identity, and\n   the criteria it supports. For an opened run the receipt also names the ticket verification id it\n   executed as `check`, lists the output log under `artifacts` with that digest, and carries the\n   ticket ref `{id, hash}` whose hash is `artifactHash` of the ticket (`common#\u002F$defs\u002Fhash`: sha256\n   over its canonical JSON without `approvals`), the value `open` wrote as the ticket's hash in the\n   run record.\n   Each receipt also records the recipe id and hash, `evidence_kind`, every evidence artifact and\n   digest, and the verifier seat id plus its runner attestation or `host-unattested` state.\n7. Set the outcome of each check to what happened: it ran and confirmed, it ran and refuted, it did\n   not run, it does not apply to this change, or it ran and settled nothing. Each outcome that is not\n   a pass carries a reason.\n8. Assemble the matrix: every criterion on one axis, every receipt on the other, and no criterion\n   left without a cell.\n9. Never rewrite a recorded receipt. To correct one, record a new passed or failed receipt for the\n   same check, covering every criterion it supports, whose `supersedes` names its `{id, hash}`.\n10. Claim exactly what the receipts support, and publish the matrix and the receipts through the\n    knowledgebase adapter's `publishArtifact` operation with a run-artifact placement (under a\n    knowledgebase coverage limit, skip it and name them as not published). Return the\n    receipts and matrix exactly as artifacts. Do not add a prose approval over them; each criterion's\n    outcome is the verdict. Publish a successful newly discovered recipe as a content-addressed\n    `verification-recipe` artifact for later runs.\n11. When every criterion is confirmed, record the gate with every receipt file recorded at this\n    head: `node \u003Cthis skill's directory>\u002F..\u002F..\u002Fbin\u002Fak-gate.mjs record --gate verify --receipt \u003Cfile>`\n    (repeat `--receipt` for each receipt). The bundle's gate copies each receipt and every\n    artifact it lists into the run store and writes references in the v2 phase record; it does not turn those\n    references into a verdict. It skips, with a note, a receipt bound to another revision and\n    refuses one with a listed artifact it cannot find, and the pre-ship check refuses one whose ticket ref\n    is not the run record's `artifactHash` value from step 6. Re-recording a corrected receipt under\n    the same id replaces the earlier one at this head, except a failed receipt: correct it by step 9.\n    The record names this revision and diff hash, so any later edit makes it stale and super-ship\n    sends you back here.\n    Run it from the project checkout; the run defaults to the branch's opened-run pointer (or the\n    branch-named v1 run when none was opened) and records default to the repository's git directory.\n    A binding's brief supplies `--dir` for the run opened with it. A run that was never opened,\n    whether branch-named or supplied by a binding, has no task record, so retain its compatible\n    marker-only call without `--receipt`; it is history, not proof. A record on a run that\n    `ship-preflight` has closed is refused; the task needs a new `open`.\n\n## Hard gates\n\nGate: no completion claim without fresh evidence for the revision claimed. An agent's description of\na green run is not a receipt — a receipt carries the command or probe, exit status, output digest,\nrevision and environment identity — and a patch that changed after a receipt was taken does not\ninherit it (ruling `closure-requires-independent-verification`).\n\nGate: a runnable frontend or backend criterion is never passed by tests, type checks or a build alone.\nIt needs the project-declared running-surface evidence kind, or an explicit non-pass outcome.\n\nGate: an implementer-authored or host-unattested receipt cannot satisfy autonomous closure; under the\nunattended policy a host-unattested one counts, labelled worker-attested, whose boundary is the live\nmerge-gate read (an exception to ruling `closure-requires-independent-verification`), and an\nimplementer-authored one still never does. A missing\nindependent seat is `unavailable`, never self-verification with a note.\n\nGate: a classifier, a confidence score and a reviewer's judgement are advisory and are recorded as\nsuch. None of them turns missing proof into a pass, and a criterion with no check is reported as\nhaving none.\n\nGate: a check is never skipped, an assertion never weakened, a threshold never lowered and coverage\nnever removed in order to obtain a passing receipt. Each of those is a separate decision, recorded on\nthe receipt that follows it, and a required product-code change is not a repair this lane absorbs:\nwhat returns from it is a bounded patch with new verification, not the old receipt read again (ruling\n`ci-repair-restricts-purpose-not-permission`).\n\nGate: a receipt names the head it verified, and that head is never conflated with a review run's\ncomparison base or its reviewed head. Where architecture, requirements, the comparison base or the\naffected surface has changed materially, the affected receipts are invalidated and a new baseline is\ndeliberately established — a new scope with its own first pass, not another delta loop (ruling\n`delta-baseline-reset-not-third-loop`).\n\nGate: an outcome that is not a pass is recorded with its reason and never omitted. A criterion\nmissing from the matrix reads as covered to everyone downstream, which is the one thing a missing\nresult must never do.\n\nGate: the run reports what it found. A run whose every check failed has completed, and reporting that\nis the result; a run that reports nothing because the news was bad has failed at the only job it has.\n\n| The thought | Why it is wrong | Do this instead |\n|---|---|---|\n| \"The caller already ran the tests and says they pass — re-running them wastes a few minutes.\" | A description of a green run is the assertion a receipt exists to replace (ruling `closure-requires-independent-verification`), and the caller's run was at a revision this one cannot name. | Identify the command, run it fresh at the revision under verification, and record exit status, digest, revision and environment. |\n| \"The last receipt was green and the change since then only touched a comment.\" | Whether the change was material is exactly what the receipt cannot tell you, and \"only a comment\" is a judgement made by the party who wants to skip the run. | Treat the moved revision as invalidating, re-run, and record the invalidation against the old receipt rather than deleting it. |\n| \"The suite exits zero, so it passed, even though the summary line lists two failures.\" | The exit status and the output disagree, and reading only the convenient half is how a green result is manufactured from a red run. | Record both, set the outcome from what the output shows, and add `exit_disagreement` with the output's failure report. |\n| \"There is no sensible way to check this one, so I will leave it out of the matrix.\" | A criterion absent from the matrix is indistinguishable from a criterion that passed, and the reader has no way to learn otherwise. | Record it as not applicable or inconclusive with the reason, so the gap is visible where the evidence would have been. |\n| \"This test is flaky — re-running until it goes green is how everyone handles it.\" | A pass selected from repeated attempts is a statement about the sampling, not about the behaviour, and the failing runs are the evidence being discarded. | Record every run as its own receipt, report the instability as the finding it is, and route the flake to diagnosis rather than to a retry. |\n| \"The check only fails because the assertion is too strict; loosening it is the obvious fix.\" | Changing what a check asserts changes the claim, and a receipt taken afterwards proves the weaker claim while reading as though it proved the original (ruling `ci-repair-restricts-purpose-not-permission`). | Leave the assertion alone, report the failure, and let the separate decision that would weaken it be taken and recorded where decisions are. |\n\n## Outputs\n\nOne receipt per check ([`schemas\u002Fverification.schema.json`](..\u002F..\u002Freferences\u002Fshared\u002Fschemas\u002Fverification.schema.json)), each naming the criteria it supports and\nbound to the revision it describes. Each receipt names its recipe digest, evidence kind, evidence\nartifacts and verifier seat. A receipt that supports nothing is not part of the matrix and is not\nevidence for anything.\n\nThe acceptance-to-evidence matrix: every acceptance criterion with its outcome and the receipt that\nproduced it, including the criteria that were not run, do not apply, or settled nothing. It never\npromotes `host-unattested` evidence into an independent result.\n\nA successful newly discovered recipe, published as a content-addressed `verification-recipe`\nartifact for later runs.\n\nInvalidation records against receipts this run supersedes, naming what changed. Nothing is rewritten\nin place.\n\nAll of it is published through the knowledgebase adapter's `publishArtifact` operation with a\nrun-artifact placement ([`adapters\u002Fknowledgebase\u002FCONTRACT.md`](..\u002F..\u002Freferences\u002Fshared\u002Fadapters\u002Fknowledgebase\u002FCONTRACT.md)); the commands are in\n[the knowledgebase-binding reference pack](..\u002F..\u002Freferences\u002Fshared\u002Freferences\u002Fknowledgebase-binding\u002FREFERENCE.md).\nThis skill names no repository path for project-derived content.\n\n## Side effects\n\n`process-exec`, `scratch-write`, `artifact-write`, `kb-publish`.\n\n`kb-publish` is a remote effect: the idempotency key is derived from the content hash of the receipt\nset per [`adapters\u002Frunner-contract\u002FCONTRACT.md`](..\u002F..\u002Freferences\u002Fshared\u002Fadapters\u002Frunner-contract\u002FCONTRACT.md) §5, and the returned record reference is read back\nbefore the run reports. Republishing an unchanged set after an interruption is a no-op success rather\nthan a second record.\n\nNo `workspace-write`, no `local-commit`: verification does not change what it is verifying. Whatever\na check writes lives in scratch and is cleaned up, and a check that can only pass by editing the tree\nis reported, not accommodated.\n\n## Stop conditions\n\n`complete`: every acceptance criterion has an outcome and a reason where that outcome is not a pass,\nand the receipts are published (or, under a knowledgebase coverage limit, named as not published).\nA run whose checks all failed is complete; the failures are the result.\n\n`needs-input`: the claim resolves to no acceptance criterion, or no revision was named. Returns the\ncriterion or the revision it would need, and no partial matrix presented as a verdict.\n\nUnder the unattended policy, a ticket that declares no acceptance criteria and no verification at\nall is a coverage limit, never `needs-input`: record it once naming its cause and verify with the\nrepository checks actually run, the project's declared test command among them when it declares one\n(its verification recipes, [`schemas\u002Fproject.schema.json`](..\u002F..\u002Freferences\u002Fshared\u002Fschemas\u002Fproject.schema.json)). The independent verifier seat records\neach on the `verify` gate with its real command, exit status and output digest, declaring itself\nwith `--author-kind`, `--host` and `--seat-id`; a verify must always declare an independent seat, so\nthe build's implementer seat, or no seat, is refused. The gate\ncounts those checks and nothing else, and records the limit on its ship decision and in the ship\ndisclosure. A ticket that declares criteria or verification the gate cannot read is judged like any\nother. Outside the policy it is `unavailable`.\n\n`failed`: the repository is unreadable at the named revision, the environment cannot be identified,\nor a configured knowledgebase refuses the write. The receipts are returned unpublished with the\nrefusal rather than written to a path in the repository.\n\nA knowledgebase coverage limit — `.cursor\u002Fbin\u002Fak kb check` exit 3 (kb.unbound, kb.unregistered), a publish exit 3 (kb.read-only), kb.in-repo-missing, or a\nmissing or unknown `.cursor\u002Fbin\u002Fak kb` command — never stops this phase under any policy: reads are unavailable,\neach skipped publish is named in the result as not published, and only exit 1 stops.\n\n`cancelled`: the caller withdrew the claim mid-run. Receipts already taken are kept and published as\nwhat they are — evidence about the checks that ran — because a completed check is a fact whether or\nnot anyone still wants it.\n\n## Limits\n\nThis skill declares no numeric cap of its own. The runner's elapsed-time budget is the only bound it\nenforces, and a budget the runner did not supply is not enforced and not guessed\n([`policies\u002Flimits.yaml`](..\u002F..\u002Freferences\u002Fshared\u002Fpolicies\u002Flimits.yaml)).\n\nReceipts per criterion: at least one, always (gate). A criterion may carry several, and none of them\nreplaces another.\n\nRe-runs: every run is its own receipt (gate). A second run does not overwrite the first, and a pass\nselected out of repeated attempts is reported as instability rather than as a pass.\n\nCommands: the project's own, discovered (gate). This skill runs no command it chose by convention\nrather than by discovery, and a project with no discoverable check for a criterion produces a\nrecorded outcome, not an invented command.\n","## When to use\n\nUse when a caller claims a specific behaviour works and the claim can be named as an acceptance\ncriterion: the export includes archived rows; a 400 is no longer retried; the migration runs twice\nwithout changing the second result.\n\nUse when a ticket's acceptance criteria need evidence before the work can be reported done, and the\nevidence has to be re-checkable by someone who was not in the run.\n\nUse when the code has moved since the last receipts were taken and a claim that was true of the old\nrevision has to be established again for the new one.\n\nUse when a repair has landed and what follows it is new verification rather than a re-reading of the\nreceipts taken before it.\n\n## Not for\n\nNot for an opinion about whether code is well written, well factored or well named. That is a review\nlane's judgement, and this skill has no instrument for it: a passing command says the behaviour\nholds, not that the code deserves to.\n\nNot for producing the fix when a check fails. A failing check is a reproduction, and a reproduction\nenters through `diagnose`, which states its hypothesis before any edit and emits either a verified\nbounded patch under a grant or a work packet — never a patch this lane wrote on the way past (ruling\n`diagnose-patch-or-packet-never-both`).\n\nNot for verifying a claim nobody stated. \"Check that everything still works\" names no criterion, so\nnothing would settle it and any result would be an impression.\n\nNot for ratifying someone's account of a run. A description of a green test suite is the thing a\nreceipt exists to replace, and re-reporting it with this skill's name on it launders an assertion\ninto evidence.\n\n## Authority\n\nAuthority: `model`. A controller or a parent skill starts it when a claim needs evidence; no slash\ncommand exposes it and no human act is required to start it.\n\nNo grant covers delegation, because no phase operation exposes this skill\n([`policies\u002Finvocation.yaml`](..\u002F..\u002Freferences\u002Fshared\u002Fpolicies\u002Finvocation.yaml)). The implementer may start the skill, but may not fill its verifier seat\nor author gate-eligible receipts. The runner assigns a distinct [`verifier`](..\u002F..\u002Freferences\u002Fshared\u002Froles\u002Fverifier\u002FROLE.md) seat, and closure depends\non that seat's receipts plus the applicable policy rule, never on the caller's confidence. A seat\nthat cannot be filled independently is `unavailable` and is never backfilled by the implementer,\nauthor, spec approver or another seated role (ruling `missing-supervisor-never-implementer`).\n\n## Inputs\n\nThe claim, resolved to named acceptance criteria ([`schemas\u002Fticket.schema.json`](..\u002F..\u002Freferences\u002Fshared\u002Fschemas\u002Fticket.schema.json) `acceptance_criteria`,\nby AC id). Absent, or phrased so that no command could settle it: stop and report `needs-input` with\nthe criterion it would need.\n\nThe repository at the revision being verified, readable (`repository-read`). Unreadable, or no\nrevision named: stop and report `failed`. A receipt whose revision is null is not a receipt.\n\nThe project's content-addressed verification recipe and surface declarations. With no matching\nrecipe, discover a candidate from project files, but do not substitute tests for a runnable surface;\npublish a recipe only after its complete build, launch, readiness, drive and cleanup path succeeds.\n\nOne runner-assigned verifier context (`independent-context`) and trusted evidence storage for the\nautonomous form. A fresh host context without attestation remains valid in guided mode and records\n`host-unattested`; the gate refuses it once the build gate records an implementer seat or the run\nholds any bypass or delegated grant record, for the rest of that run. Under the unattended policy\n(ADR-0014) it is a recorded note instead, whose boundary is the live merge-gate read (an exception\nto ruling `closure-requires-independent-verification`), and the gate is recorded with its evidence\ncommand, as [the unattended-lifecycle reference](..\u002F..\u002Freferences\u002Fshared\u002Freferences\u002Funattended-lifecycle\u002FREFERENCE.md) says.\n\nThe environment the checks run in, identified ([`schemas\u002Fverification.schema.json`](..\u002F..\u002Freferences\u002Fshared\u002Fschemas\u002Fverification.schema.json)). Test execution\nbelongs in the appropriate isolated environment, and repository test code does not receive production\ncredentials by default.\n\nOptionally, receipts from an earlier run. They are history: a receipt is evidence for the revision it\nnames and for no other, and one taken before the code moved is read as invalidated rather than as a\nhead start. The code moved when either the revision or the working-tree diff hash differs from the\nreceipt's (`common#\u002F$defs\u002Frevision_ref`); an uncommitted edit on the same revision is a move.\n\n## Workflow\n\n1. Resolve the claim into the acceptance criteria it is made of, by id. A claim that resolves to no\n   criterion stops here with `needs-input`.\n2. Resolve the matching project-declared recipe by id and content hash. If none exists, discover the\n   setup, build, launch, readiness, drive and cleanup steps from project files. Record a candidate\n   only after the whole path works; never write a host-specific recipe into the application tree.\n   Load [verification evidence](..\u002F..\u002Freferences\u002Fshared\u002Freferences\u002Fverification-evidence\u002FREFERENCE.md) before matching\n   a surface to an evidence kind.\n3. Seat [`roles\u002Fverifier\u002FROLE.md`](..\u002F..\u002Freferences\u002Fshared\u002Froles\u002Fverifier\u002FROLE.md) outside the implementer lineage. Give it the claim, criteria,\n   revision, recipe, permitted commands, environment and required evidence kinds—never the\n   implementer's narrative, claimed result or reviewer verdict. An unfillable seat is `unavailable`.\n4. Run each command fresh and complete, at the revision under verification. A partial run, a cached\n   result and a previous run's output are not this run's evidence.\n   Where a criterion declares a runnable surface, build and launch the application, wait for the\n   declared readiness signal, and drive the declared path or request. Tests, type checks and builds\n   may support the receipt; none substitutes for the running-app evidence kind the criterion names.\n5. Read the whole output: exit status first, then the counts the output reports. A suite that reports\n   failures while exiting zero is read by its output, and the disagreement is recorded as\n   `exit_disagreement` with `verdict_from: output` and an `output_reports` quote.\n6. Record one receipt per check with the command as an argument vector or the probe with its target,\n   the exit status, the digest of the relevant output, the revision, the environment identity, and\n   the criteria it supports. For an opened run the receipt also names the ticket verification id it\n   executed as `check`, lists the output log under `artifacts` with that digest, and carries the\n   ticket ref `{id, hash}` whose hash is `artifactHash` of the ticket (`common#\u002F$defs\u002Fhash`: sha256\n   over its canonical JSON without `approvals`), the value `open` wrote as the ticket's hash in the\n   run record.\n   Each receipt also records the recipe id and hash, `evidence_kind`, every evidence artifact and\n   digest, and the verifier seat id plus its runner attestation or `host-unattested` state.\n7. Set the outcome of each check to what happened: it ran and confirmed, it ran and refuted, it did\n   not run, it does not apply to this change, or it ran and settled nothing. Each outcome that is not\n   a pass carries a reason.\n8. Assemble the matrix: every criterion on one axis, every receipt on the other, and no criterion\n   left without a cell.\n9. Never rewrite a recorded receipt. To correct one, record a new passed or failed receipt for the\n   same check, covering every criterion it supports, whose `supersedes` names its `{id, hash}`.\n10. Claim exactly what the receipts support, and publish the matrix and the receipts through the\n    knowledgebase adapter's `publishArtifact` operation with a run-artifact placement (under a\n    knowledgebase coverage limit, skip it and name them as not published). Return the\n    receipts and matrix exactly as artifacts. Do not add a prose approval over them; each criterion's\n    outcome is the verdict. Publish a successful newly discovered recipe as a content-addressed\n    `verification-recipe` artifact for later runs.\n11. When every criterion is confirmed, record the gate with every receipt file recorded at this\n    head: `node \u003Cthis skill's directory>\u002F..\u002F..\u002Fbin\u002Fak-gate.mjs record --gate verify --receipt \u003Cfile>`\n    (repeat `--receipt` for each receipt). The bundle's gate copies each receipt and every\n    artifact it lists into the run store and writes references in the v2 phase record; it does not turn those\n    references into a verdict. It skips, with a note, a receipt bound to another revision and\n    refuses one with a listed artifact it cannot find, and the pre-ship check refuses one whose ticket ref\n    is not the run record's `artifactHash` value from step 6. Re-recording a corrected receipt under\n    the same id replaces the earlier one at this head, except a failed receipt: correct it by step 9.\n    The record names this revision and diff hash, so any later edit makes it stale and super-ship\n    sends you back here.\n    Run it from the project checkout; the run defaults to the branch's opened-run pointer (or the\n    branch-named v1 run when none was opened) and records default to the repository's git directory.\n    A binding's brief supplies `--dir` for the run opened with it. A run that was never opened,\n    whether branch-named or supplied by a binding, has no task record, so retain its compatible\n    marker-only call without `--receipt`; it is history, not proof. A record on a run that\n    `ship-preflight` has closed is refused; the task needs a new `open`.\n\n## Hard gates\n\nGate: no completion claim without fresh evidence for the revision claimed. An agent's description of\na green run is not a receipt — a receipt carries the command or probe, exit status, output digest,\nrevision and environment identity — and a patch that changed after a receipt was taken does not\ninherit it (ruling `closure-requires-independent-verification`).\n\nGate: a runnable frontend or backend criterion is never passed by tests, type checks or a build alone.\nIt needs the project-declared running-surface evidence kind, or an explicit non-pass outcome.\n\nGate: an implementer-authored or host-unattested receipt cannot satisfy autonomous closure; under the\nunattended policy a host-unattested one counts, labelled worker-attested, whose boundary is the live\nmerge-gate read (an exception to ruling `closure-requires-independent-verification`), and an\nimplementer-authored one still never does. A missing\nindependent seat is `unavailable`, never self-verification with a note.\n\nGate: a classifier, a confidence score and a reviewer's judgement are advisory and are recorded as\nsuch. None of them turns missing proof into a pass, and a criterion with no check is reported as\nhaving none.\n\nGate: a check is never skipped, an assertion never weakened, a threshold never lowered and coverage\nnever removed in order to obtain a passing receipt. Each of those is a separate decision, recorded on\nthe receipt that follows it, and a required product-code change is not a repair this lane absorbs:\nwhat returns from it is a bounded patch with new verification, not the old receipt read again (ruling\n`ci-repair-restricts-purpose-not-permission`).\n\nGate: a receipt names the head it verified, and that head is never conflated with a review run's\ncomparison base or its reviewed head. Where architecture, requirements, the comparison base or the\naffected surface has changed materially, the affected receipts are invalidated and a new baseline is\ndeliberately established — a new scope with its own first pass, not another delta loop (ruling\n`delta-baseline-reset-not-third-loop`).\n\nGate: an outcome that is not a pass is recorded with its reason and never omitted. A criterion\nmissing from the matrix reads as covered to everyone downstream, which is the one thing a missing\nresult must never do.\n\nGate: the run reports what it found. A run whose every check failed has completed, and reporting that\nis the result; a run that reports nothing because the news was bad has failed at the only job it has.\n\n| The thought | Why it is wrong | Do this instead |\n|---|---|---|\n| \"The caller already ran the tests and says they pass — re-running them wastes a few minutes.\" | A description of a green run is the assertion a receipt exists to replace (ruling `closure-requires-independent-verification`), and the caller's run was at a revision this one cannot name. | Identify the command, run it fresh at the revision under verification, and record exit status, digest, revision and environment. |\n| \"The last receipt was green and the change since then only touched a comment.\" | Whether the change was material is exactly what the receipt cannot tell you, and \"only a comment\" is a judgement made by the party who wants to skip the run. | Treat the moved revision as invalidating, re-run, and record the invalidation against the old receipt rather than deleting it. |\n| \"The suite exits zero, so it passed, even though the summary line lists two failures.\" | The exit status and the output disagree, and reading only the convenient half is how a green result is manufactured from a red run. | Record both, set the outcome from what the output shows, and add `exit_disagreement` with the output's failure report. |\n| \"There is no sensible way to check this one, so I will leave it out of the matrix.\" | A criterion absent from the matrix is indistinguishable from a criterion that passed, and the reader has no way to learn otherwise. | Record it as not applicable or inconclusive with the reason, so the gap is visible where the evidence would have been. |\n| \"This test is flaky — re-running until it goes green is how everyone handles it.\" | A pass selected from repeated attempts is a statement about the sampling, not about the behaviour, and the failing runs are the evidence being discarded. | Record every run as its own receipt, report the instability as the finding it is, and route the flake to diagnosis rather than to a retry. |\n| \"The check only fails because the assertion is too strict; loosening it is the obvious fix.\" | Changing what a check asserts changes the claim, and a receipt taken afterwards proves the weaker claim while reading as though it proved the original (ruling `ci-repair-restricts-purpose-not-permission`). | Leave the assertion alone, report the failure, and let the separate decision that would weaken it be taken and recorded where decisions are. |\n\n## Outputs\n\nOne receipt per check ([`schemas\u002Fverification.schema.json`](..\u002F..\u002Freferences\u002Fshared\u002Fschemas\u002Fverification.schema.json)), each naming the criteria it supports and\nbound to the revision it describes. Each receipt names its recipe digest, evidence kind, evidence\nartifacts and verifier seat. A receipt that supports nothing is not part of the matrix and is not\nevidence for anything.\n\nThe acceptance-to-evidence matrix: every acceptance criterion with its outcome and the receipt that\nproduced it, including the criteria that were not run, do not apply, or settled nothing. It never\npromotes `host-unattested` evidence into an independent result.\n\nA successful newly discovered recipe, published as a content-addressed `verification-recipe`\nartifact for later runs.\n\nInvalidation records against receipts this run supersedes, naming what changed. Nothing is rewritten\nin place.\n\nAll of it is published through the knowledgebase adapter's `publishArtifact` operation with a\nrun-artifact placement ([`adapters\u002Fknowledgebase\u002FCONTRACT.md`](..\u002F..\u002Freferences\u002Fshared\u002Fadapters\u002Fknowledgebase\u002FCONTRACT.md)); the commands are in\n[the knowledgebase-binding reference pack](..\u002F..\u002Freferences\u002Fshared\u002Freferences\u002Fknowledgebase-binding\u002FREFERENCE.md).\nThis skill names no repository path for project-derived content.\n\n## Side effects\n\n`process-exec`, `scratch-write`, `artifact-write`, `kb-publish`.\n\n`kb-publish` is a remote effect: the idempotency key is derived from the content hash of the receipt\nset per [`adapters\u002Frunner-contract\u002FCONTRACT.md`](..\u002F..\u002Freferences\u002Fshared\u002Fadapters\u002Frunner-contract\u002FCONTRACT.md) §5, and the returned record reference is read back\nbefore the run reports. Republishing an unchanged set after an interruption is a no-op success rather\nthan a second record.\n\nNo `workspace-write`, no `local-commit`: verification does not change what it is verifying. Whatever\na check writes lives in scratch and is cleaned up, and a check that can only pass by editing the tree\nis reported, not accommodated.\n\n## Stop conditions\n\n`complete`: every acceptance criterion has an outcome and a reason where that outcome is not a pass,\nand the receipts are published (or, under a knowledgebase coverage limit, named as not published).\nA run whose checks all failed is complete; the failures are the result.\n\n`needs-input`: the claim resolves to no acceptance criterion, or no revision was named. Returns the\ncriterion or the revision it would need, and no partial matrix presented as a verdict.\n\nUnder the unattended policy, a ticket that declares no acceptance criteria and no verification at\nall is a coverage limit, never `needs-input`: record it once naming its cause and verify with the\nrepository checks actually run, the project's declared test command among them when it declares one\n(its verification recipes, [`schemas\u002Fproject.schema.json`](..\u002F..\u002Freferences\u002Fshared\u002Fschemas\u002Fproject.schema.json)). The independent verifier seat records\neach on the `verify` gate with its real command, exit status and output digest, declaring itself\nwith `--author-kind`, `--host` and `--seat-id`; a verify must always declare an independent seat, so\nthe build's implementer seat, or no seat, is refused. The gate\ncounts those checks and nothing else, and records the limit on its ship decision and in the ship\ndisclosure. A ticket that declares criteria or verification the gate cannot read is judged like any\nother. Outside the policy it is `unavailable`.\n\n`failed`: the repository is unreadable at the named revision, the environment cannot be identified,\nor a configured knowledgebase refuses the write. The receipts are returned unpublished with the\nrefusal rather than written to a path in the repository.\n\nA knowledgebase coverage limit — `.cursor\u002Fbin\u002Fak kb check` exit 3 (kb.unbound, kb.unregistered), a publish exit 3 (kb.read-only), kb.in-repo-missing, or a\nmissing or unknown `.cursor\u002Fbin\u002Fak kb` command — never stops this phase under any policy: reads are unavailable,\neach skipped publish is named in the result as not published, and only exit 1 stops.\n\n`cancelled`: the caller withdrew the claim mid-run. Receipts already taken are kept and published as\nwhat they are — evidence about the checks that ran — because a completed check is a fact whether or\nnot anyone still wants it.\n\n## Limits\n\nThis skill declares no numeric cap of its own. The runner's elapsed-time budget is the only bound it\nenforces, and a budget the runner did not supply is not enforced and not guessed\n([`policies\u002Flimits.yaml`](..\u002F..\u002Freferences\u002Fshared\u002Fpolicies\u002Flimits.yaml)).\n\nReceipts per criterion: at least one, always (gate). A criterion may carry several, and none of them\nreplaces another.\n\nRe-runs: every run is its own receipt (gate). A second run does not overwrite the first, and a pass\nselected out of repeated attempts is reported as instability rather than as a pass.\n\nCommands: the project's own, discovered (gate). This skill runs no command it chose by convention\nrather than by discovery, and a project with no discoverable check for a criterion produces a\nrecorded outcome, not an invented command.\n",{"name":5,"license":15,"metadata":24,"description":7},{"ak":25,"ak_catalog_id":5},{"mode":26,"autonomy_unenforceable":27},"autonomous",[28,29,30,31],"artifact-write is storage only: the host does not compute or check the artifact hash, so envelope hash binding is this package's own work.","kb-write is not provided by the host; the knowledgebase adapter supplies it and refuses rather than falling back to a repository path when no knowledgebase is configured, which is why it does not cap this row while that adapter is attached (ruling `fail-closed-adapter-lifts-ceiling`).","environment identity is not attested by the host. This package records what it observed of the environment and never reports an identity it could not read.","independent-context and trusted-evidence are not provided by the host. The attached runner supplies them fail-closed; without it this skill is guided and receipts are host-unattested.","main","e0232821c4abbcd4338d6018c20fb70f1b74336b",".cursor\u002Fskills\u002Fsuper-verify",3,"https:\u002F\u002Fgithub.com\u002FPibomeister\u002Fagent-kit","https:\u002F\u002Fgithub.com\u002FPibomeister","www.eduardopicazoenriquez.com","2026-11-15 00:56:58",{"source":8,"skillName":5,"commands":41,"prompt":51,"zipAvailable":52},[42,45,48],{"tool":43,"command":44},"npx","npx skills add Pibomeister\u002Fagent-kit --skill super-verify",{"tool":46,"command":47},"bunx","bunx skills add Pibomeister\u002Fagent-kit --skill super-verify",{"tool":49,"command":50},"pnpm","pnpm dlx skills add Pibomeister\u002Fagent-kit --skill super-verify","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FPibomeister\u002Fagent-kit\u002Fblob\u002Fmain\u002F.cursor\u002Fskills\u002Fsuper-verify\u002FSKILL.md",true,[],[55,62,67,72,77,82],{"skillNo":56,"name":57,"nameZh":6,"description":58,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":59,"categories":60,"keywords":61,"publishedAt":19},"01M7D3JZP2026RGY6XX1MC3NQQ","alpha","Human-started command: it runs only when the human's message begins with `\u002Fak:alpha`. On any other request do not load or follow it; tell the human to type that command. Runs the alpha workflow when a human asks for it.","https:\u002F\u002Fgithub.com\u002FPibomeister\u002Fagent-kit\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fvalid\u002Fminimal\u002Fskills\u002Falpha\u002FSKILL.md",[],[],{"skillNo":63,"name":57,"nameZh":6,"description":58,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":64,"categories":65,"keywords":66,"publishedAt":19},"01M7D3JA593Q3RBYJXKVZ37Y2K","https:\u002F\u002Fgithub.com\u002FPibomeister\u002Fagent-kit\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Finvalid\u002F25-article-tier-outside-vocabulary\u002Fskills\u002Falpha\u002FSKILL.md",[],[],{"skillNo":68,"name":57,"nameZh":6,"description":58,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":69,"categories":70,"keywords":71,"publishedAt":19},"01M7D3HQDDAZTAPGWCPGDPNVBQ","https:\u002F\u002Fgithub.com\u002FPibomeister\u002Fagent-kit\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Finvalid\u002F24-article-without-triggers\u002Fskills\u002Falpha\u002FSKILL.md",[],[],{"skillNo":73,"name":57,"nameZh":6,"description":58,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":74,"categories":75,"keywords":76,"publishedAt":19},"01M7D3HK074TK3ZWK1ZZNQVPB0","https:\u002F\u002Fgithub.com\u002FPibomeister\u002Fagent-kit\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Finvalid\u002F23-enforced-by-without-predicate\u002Fskills\u002Falpha\u002FSKILL.md",[],[],{"skillNo":78,"name":57,"nameZh":6,"description":58,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":79,"categories":80,"keywords":81,"publishedAt":19},"01M7D3HGHQM5C5TBSXBCE2WFHW","https:\u002F\u002Fgithub.com\u002FPibomeister\u002Fagent-kit\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Finvalid\u002F22-article-without-source-span\u002Fskills\u002Falpha\u002FSKILL.md",[],[],{"skillNo":83,"name":57,"nameZh":6,"description":58,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":84,"categories":85,"keywords":86,"publishedAt":19},"01M7D3HDV6HMK2ZHRM0P368KNT","https:\u002F\u002Fgithub.com\u002FPibomeister\u002Fagent-kit\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Finvalid\u002F21-fragment-redeclares-id\u002Fskills\u002Falpha\u002FSKILL.md",[],[]]