[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7D3CE8C3EQFW7CP5HW56CF2":3,"skill-same:01M7D3CE8C3EQFW7CP5HW56CF2":53},{"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":25,"commitSha":26,"skillDirectory":27,"forksCount":28,"repositoryHtmlUrl":29,"repositoryHomepageUrl":6,"ownerHtmlUrl":30,"ownerBlogUrl":31,"ownerDescription":6,"updatedAt":32,"install":33,"resources":46},"01M7D3CE8C3EQFW7CP5HW56CF2","super-build",null,"Executes one approved implementation ticket in a worktree the ticket owns, test first, then has the result checked on two independent axes before the ticket is reported done. Use when asked to build, implement or execute a ticket (\"build ticket T2\", \"implement AK-214\"), including as a worker's task-local implementer, and the ticket names its acceptance criteria and the verification that shows each one met. Not for a decision ticket, an unapproved draft, a finding too vague to specify, or a quick fix with no ticket behind it.","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\u002Fskills\u002Fsuper-build\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: super-build\ndescription: >-\n  Executes one approved implementation ticket in a worktree the ticket owns, test first, then has the\n  result checked on two independent axes before the ticket is reported done. Use when asked to build,\n  implement or execute a ticket (\"build ticket T2\", \"implement AK-214\"), including as a worker's\n  task-local implementer, and the ticket names its acceptance criteria and the verification that\n  shows each one met. Not for a decision ticket, an unapproved draft, a finding too vague to specify,\n  or a quick fix with no ticket behind it.\nlicense: MIT\nmetadata:\n  ak_catalog_id: super-build\n---\n\n## When to use\n\nUse when a ticket exists, is approved, is typed `implementation`, and names its acceptance criteria\nand the verification that will show each one met. The ticket is the unit of work: one ticket, one\nworktree, one implementer, one pass of checks, one commit.\n\nUse when a run has reached its build phase and the ticket at the head of the ready set has no\nunfinished dependency — including the dependencies a dependency edge does not express.\n\nUse when a finding from a review lane or a diagnostic packet has already been sharpened into such a\nticket. What decides whether it could be is the finding's own grading, not how actionable its prose\nreads.\n\n## Not for\n\nNot for a ticket typed `decision`. A decision ticket asks which of two things to do; executing one of\nthem answers the question by doing it and calls the answer an implementation. The type field settles\nthis (`schemas\u002Fticket.schema.json`), and implementation-shaped prose in the body does not reopen it.\n\nNot for an unapproved or draft ticket. Approval binds to the ticket's artifact hash, so a ticket\nedited after approval is unapproved again and re-enters the step that approves it.\n\nNot for a finding that is not yet a ticket. A finding graded `spec_quality: smell`, or carrying\n`difficulty: null`, names a doubt rather than a change; sharpening it is a separate act by a lane\nthat holds that authority, and narrowing it here to something a fixer can act on invents the\nspecification the grading says is missing.\n\nNot for a bug report with a reproduction and no diagnosis behind it. A bug enters through `diagnose`,\nand what reaches this lane is the diagnosed result. A packet that already carried a verified bounded\npatch is finished work, and re-implementing it here is the same patch written twice (ruling\n`diagnose-patch-or-packet-never-both`).\n\nNot for \"just fix this quickly\" with no ticket at all. The ticket is what carries the acceptance\ncriteria, the allowed changes and the write ownership, and a run that improvises them has nothing to\ncheck the result against.\n\n## Authority\n\nAuthority: `model`. A controller or a parent skill starts it when an approved ticket is ready; no\nslash command 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` records model-invoked skills as exposing none by construction. What\nauthority governs here is the ticket rather than the caller: the approval the ticket carries, and the\ncharter checkpoint the run took before its build phase opened. A skill that cannot see either stops\nrather than treating its own invocation as the approval. Under the unattended policy (ADR-0014) that approval is a recorded unattended advance, a ticket whose class makes a human the author is built by the agent with its true class recorded (an exception to ruling `delegation-class-is-authority-not-finding`), and each gate is recorded with its evidence command, as [the unattended-lifecycle reference](..\u002F..\u002Freferences\u002Funattended-lifecycle\u002FREFERENCE.md) says. An unscored ticket (the delegation reference's coverage limit) is built the same way with no class recorded and its implementer seat still declared.\n\n## Inputs\n\nOne approved ticket typed `implementation` (`schemas\u002Fticket.schema.json`), with its\n`acceptance_criteria`, `verification`, `allowed_changes`, `read_dependencies`, `write_ownership` and\n`integration_owner`. Absent, unapproved, or typed `decision`: stop and report `needs-input` naming\nthe field that decided it.\n\nA writable checkout the run owns, and a worktree this ticket alone holds (`repository-write`,\n`isolated-worktree`, `vcs-local`). Unavailable, or the branch is already checked out elsewhere: stop\nand report `failed`.\n\nTwo independent contexts for the check seats (`independent-context`). A seat that cannot be filled\nindependently is `unavailable`, and unavailability blocks the ticket; it is never backfilled by the\nimplementer, by the author, by whoever approved the spec, or by the other seat (ruling\n`missing-supervisor-never-implementer`).\n\nThe source finding, where the ticket came from one (`schemas\u002Ffinding.schema.json`). A finding\narriving from a peer review lane graded `autofix_class: safe_auto` is remapped to `gated_auto` on\nintake and is never dropped for its class: the restriction is on emission by a seat, never on the\nvocabulary (ruling `safe-auto-restricted-per-seat`).\n\nOptionally a ticket budget from the runner. Absent: the cap is not enforced and not guessed, and the\nabsence is recorded (`policies\u002Flimits.yaml`).\n\n## Workflow\n\n1. Read the ticket and refuse what this lane does not execute — `decision` type, missing approval, or\n   a source finding graded `smell` or carrying `difficulty: null`. Record the refusal with the field\n   that decided it. Do not re-grade the finding to make it executable.\n2. Decide parallelism by inspection, not by the absence of a dependency edge. For each pair of\n   tickets with no edge between them, read the declared write ownership, generated artifacts,\n   migration sequence and interfaces, and read the files themselves where a declaration is silent.\n   Contention that survives inspection serializes exactly the contending tickets; the rest of the\n   layer still dispatches together.\n3. Record the base commit, then open the worktree this ticket owns\n   (`protocols\u002Fworktree-ownership\u002FPROTOCOL.md`). The worktree handle is bound to this ticket and is\n   retired when the ticket integrates, never retasked.\n   On the standalone path, open the task-bound run next, from inside that worktree with the\n   ticket's branch checked out and before implementation:\n   `node \u003Cthis skill's directory>\u002F..\u002F..\u002Fbin\u002Fak-gate.mjs open --ticket \u003Cticket-file>`. A worker that\n   holds a Firstmate binding opens the run with the command its brief gives, which appends\n   `--binding \u003Cbinding file> --dir \u003Cevidence store>`, so the run records the binding and the gate\n   finds its delegated grants. The run binds\n   to the branch checked out where it runs, and later gate commands on that branch resolve the\n   pointer it writes. A run never closed by `ship-preflight` stays the branch's default until a\n   new `open`, and the gate does not tell an earlier task's unclosed run from this one, so every\n   new task opens a new run, even on a branch that already has one.\n4. Dispatch one implementer (`roles\u002Fimplementer\u002FROLE.md`) with the ticket as its single source of\n   requirements. It spawns no implementers of its own, and no second implementer runs against this\n   worktree. Hand each implementer [the work-trail protocol](..\u002F..\u002Fprotocols\u002Fwork-trail\u002FPROTOCOL.md)\n   and the location of its own stretch, prepared in the runner's scratch space as that protocol\n   says, and no earlier stretch or pointer to one (ruling `work-trail-is-account-not-evidence`).\n5. Work the ticket test first (`protocols\u002Ftdd\u002FPROTOCOL.md`): a failing test that names the behaviour,\n   then the change that passes it. Where the behaviour is genuinely not testable, record the\n   alternative verification plan in the ticket rather than proceeding with neither. Load\n   [the engineering-principles reference pack](..\u002F..\u002Freferences\u002Fengineering-principles\u002FREFERENCE.md) for the standards the change is held to.\n6. Handle the implementer's report by what it says. Concerns about correctness or scope are addressed\n   before any check runs; missing context is supplied and the same implementer re-dispatched; a\n   blocker is assessed, not returned to the implementer as another attempt at the same wall.\n7. Record what was noticed and not touched — file and line, the observation, why it is out of scope.\n   Each entry is a candidate ticket, never an edit inside this one.\n8. Hand the diff to both check seats at once, as a file, with the ticket's constraints copied\n   verbatim from the ticket. Neither prompt carries the implementer's account of the work or any\n   expectation of what the seat will find.\n9. Keep the axes separate: `reviewer-spec` checks whether the ticket's obligations were met,\n   `reviewer-standards` checks the result against the project's own designated rules, rule by rule.\n   Findings are not merged or re-ranked across the two; one axis passing does not cover the other.\n10. Fix and re-check, at most five rounds per ticket. Rounds one to three resume the implementer that\n    already holds the context; rounds four and five dispatch a fresh implementer told how many\n    attempts preceded it and that it owns the task from here, and told nothing of their framing. Each\n    re-check is scoped per finding — addressed or not addressed — plus the fix diff read for new\n    breakage. A resumed implementer continues its own stretch; a fresh one is seated as step 4 says.\n11. At the round cap, adjudicate every finding still open, one at a time: park it with a written\n    ruling naming why the code stands and what it costs if that is wrong, or rule it load-bearing and\n    stop. A silent discard is forbidden.\n12. Load [verification evidence](..\u002F..\u002Freferences\u002Fverification-evidence\u002FREFERENCE.md), run the\n    ticket's named checks as implementer self-checks and record each receipt with\n    `created_by.role: implementer`. They are self-check evidence: once the build gate records the\n    implementer seat (step 13) they never satisfy the verify gate, and without it they count only\n    for checks no recipe binds. Hand the claim, criteria, frozen revision, project\n    recipe and permitted commands to `super-verify`; its verifier seat reruns them and produces\n    gate-eligible receipts. Commit the work on the ticket's own branch, then publish the receipts and\n    ticket result through the knowledgebase adapter's `publishArtifact` operation.\n13. When both check seats pass, record the gate: `node \u003Cthis skill's directory>\u002F..\u002F..\u002Fbin\u002Fak-gate.mjs record --gate build-checks`\n    (the bundle's `bin\u002F`, two directories above this skill). super-ship refuses to ship without it.\n    When a runner assigned the implementer seat, declare it by appending\n    `--class \u003Cdelegation class> --author-kind \u003Chuman|agent> --host \u003Chost adapter id> --seat-id \u003Cimplementer seat>`;\n    the seat is recorded on this gate, not on a receipt, and stays declared for commits descended\n    from the recorded one. After an amend or rebase, pass the flags again. An unscored ticket has no class to pass: declare the seat with `--author-kind \u003Chuman|agent> --host \u003Chost adapter id> --seat-id \u003Cimplementer seat>` and no `--class`, never a derived one or a seeded default, so a receipt that seat wrote is still refused.\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.\n14. A committed ticket then runs the trail, steps 14 to 16, a resumed run by the work-trail\n    protocol's resume rule; steps 12 and 13 never wait on it (ruling\n    `work-trail-is-account-not-evidence`). One stopped before its commit gets no audit or published\n    stretch, as the report says; a configured knowledgebase (exit 1) failing a receipt publish stops the build short of the\n    trail, while under a knowledgebase coverage limit the receipts are named as not published and the build goes on. Each prepared stretch is handled and sent as the work-trail protocol's build steps say;\n    none it holds back is sent. A refused or failed trail publish is never repaired or resent and is\n    not itself a build gate; the protocol's stop conditions describe shared-file failures that\n    can block later required receipts. The trail writes no gate record; `complete` does not depend on it.\n15. Seat [the trail auditor](..\u002F..\u002Froles\u002Ftrail-auditor\u002FROLE.md) once, even for an empty or missing\n    stretch, in a context that filled no other seat on the ticket and is not the build's own; hand\n    it its packet, each unpublished stretch named so, from its location. The audit is unavailable\n    with no such context, a binding not permitting the role, a packet it cannot be handed, a failed\n    seat, or a returned record not naming kind `trail-audit` and the id it was handed: withhold the\n    audit record, not the stretches, block nothing and let no seat fill in, the implementer included\n    (ruling `work-trail-is-account-not-evidence`). Otherwise publish the audit exactly as returned.\n16. Report every written ruling and out-of-scope observation last, the only existing act waiting on\n    the trail, with the audit's reference, result and items or why it is unavailable or unpublished,\n    and each missing or unpublished stretch, an unpublished one with its reason. Repeat as given an\n    audit notice that something read holds a secret, credential, token or personal data: where it\n    is, never the value. For a published stretch or audit, add that it cannot be corrected and the\n    value is exposed, for its owner to rotate, as `super-ship` does for a secret in history. The\n    attention list and that notice go only in the report and audit record, never in a receipt, check\n    result, finding, gate record or checkpoint card (ruling `work-trail-is-account-not-evidence`).\n\n## Hard gates\n\nGate: a ticket typed `decision` is never executed as implementation work, and never re-typed here to\nmake it executable. It returns to the lane that resolves decisions.\n\nGate: a finding graded `smell`, or carrying `difficulty: null`, is refused as an implementation\nticket (`schemas\u002Ffinding.schema.json`); the shared rule for applying a finding at all is\n`protocols\u002Fapply-findings\u002FPROTOCOL.md`. Difficulty without a solution class is not a specification,\nand a re-grading performed by the lane that wants to execute it is the refusal being routed around.\n\nGate: a receipt created by the implementer, change author, spec approver or recipe author is\nself-check evidence only. It never fills or substitutes for the verifier seat, even when its command\nand output are identical to the later independent run (ruling\n`missing-supervisor-never-implementer`).\n\nGate: neither check seat emits `autofix_class: safe_auto`. At review time a code edit has no single\nmechanically correct answer, so a seat's classification is a proposal and applying it is the caller's\ndecision under its own authorization; `gated_auto`, `manual` and `advisory` are what a seat here may\nemit (ruling `safe-auto-restricted-per-seat`).\n\nGate: the implementer never sits on the panel that checks its work, and no seat approves a patch it\nproduced. Two seats that are the same context in sequence are one seat (ruling\n`missing-supervisor-never-implementer`).\n\nGate: only independent verification evidence, plus a policy rule saying that evidence suffices for\nthat finding, closes a finding. A description of a green run is not a receipt — a receipt carries the\ncommand or probe, exit status, output digest, revision and environment identity — and a patch that\nchanged after a receipt was taken does not inherit it (ruling\n`closure-requires-independent-verification`).\n\nGate: writes stay inside the ticket's `allowed_changes` and its declared ownership. A write outside\nevery worker's declared exclusive set aborts the wave, and a change no worker accounts for may be the\nuser's: it is preserved and reconciled, never discarded to get the tree clean.\n\nGate: a stubbed test, a skipped test, an unimplemented branch or a marker left in place of the work\nis a blocker, not progress. The ticket is reported blocked with what is missing.\n\nGate: the round cap ends the loop, not the adjudication. Reaching it with findings open is a\n`cap-reached` result carrying each one and its ruling.\n\nGate: the wall (ruling `work-trail-is-account-not-evidence`). No seat whose output is a finding, a\nreceipt, a verdict or a signoff input is shown the trail, the trail audit or the attention list.\nNothing from them reaches a check seat or the verification hand-off, quoted, summarized or as a\npointer, and the build's context carries nothing it saw of them into a ruling it writes, a synthesis\nor verdict it sets, or a packet it hands a seat the wall covers, for the same change.\n\n| The thought | Why it is wrong | Do this instead |\n|---|---|---|\n| \"The implementer just fixed it and is sitting right there — having it confirm the fix is faster than dispatching a seat.\" | A seat that cannot be filled independently is unavailable, and unavailability blocks; it is never backfilled by the implementer (ruling `missing-supervisor-never-implementer`). The context that produced the change is the one context that cannot tell you whether it is right. | Mark the seat `unavailable`, stop with `needs-input` naming which seat and why, and report the ticket as not done. |\n| \"The implementer's report says the tests passed, so the criterion is met.\" | An agent's description of a green run is not a receipt (ruling `closure-requires-independent-verification`). The report is the author's account of its own work, and it is the thing the receipt exists to replace. | Run the ticket's named verification, record command, exit status, output digest, revision and environment, and bind each receipt to the criteria it supports. |\n| \"The finding is only a smell, but I can see what it means — I will write the ticket from it.\" | Difficulty without a solution class is not a specification, and reading a change out of a smell is inventing the one the grading says is missing. The invented specification is then checked against itself. | Refuse it as an implementation ticket, and return it for sharpening, diagnosis or escalation with the field that decided it. |\n| \"Round five ended with two small findings open; they are minor and the ticket is otherwise done.\" | A silent discard is exactly what the round cap exists to prevent, and the findings least likely to be written down are the ones a later reader most needs. | Park each one with a written ruling naming what it costs if wrong, or rule it load-bearing and stop; return `cap-reached` with both attached. |\n| \"These two tickets have no dependency edge, so they can run in parallel.\" | An edge records a declared dependency, not shared write surface. A shared generated artifact, a migration sequence or one exported interface makes two edgeless tickets contending writers, and the DAG cannot see it. | Read the ownership declarations and the files, serialize exactly the contending pair, and dispatch the rest of the layer together. |\n| \"The seat should know I already considered the null case, so I will say so in the prompt.\" | A prompt that says what to expect tells the seat where to stop looking, and the axis returns the author's confidence in the author's own words. | Hand over the diff and the ticket's verbatim constraints, and nothing about what the previous attempt believed. |\n\n## Outputs\n\nOne commit per ticket, on the branch that ticket owns, recorded against the ticket by id. The commit\nis the patch link the lane's result carries; nothing is merged here.\n\nVerification receipts (`schemas\u002Fverification.schema.json`), one per acceptance criterion the ticket\nnames, each naming the criteria it supports and bound to the revision it describes.\n\nThe per-ticket check result: the findings each axis raised (`schemas\u002Ffinding.schema.json`), their\ndisposition, and every ruling written at the round cap. A per-ticket check is a lighter instrument\nthan a review panel — it emits no `review` artifact, and a later panel never reads it as a lane it no\nlonger has to cover.\n\nAll of it is published through the knowledgebase adapter's `publishArtifact` operation with a\nrun-artifact placement (`adapters\u002Fknowledgebase\u002FCONTRACT.md`); the commands are in\n[the knowledgebase-binding reference pack](..\u002F..\u002Freferences\u002Fknowledgebase-binding\u002FREFERENCE.md).\nThis skill names no repository path for project-derived content. Each implementer seat's stretch\n(`schemas\u002Ftrail.schema.json`) and the trail audit (`schemas\u002Ftrail-audit.schema.json`) go through\n`publishArtifact` too, as an account, never evidence (ruling `work-trail-is-account-not-evidence`).\n\n## Side effects\n\n`workspace-write`, `branch-create`, `local-commit`, `process-exec`, `artifact-write`, `kb-publish`,\n`scratch-write`.\n\n`kb-publish` is a remote effect: the idempotency key is derived from the content hash of the\npublished set per `adapters\u002Frunner-contract\u002FCONTRACT.md` §5, and the returned record reference is\nread back before the ticket reports. Republishing unchanged receipts, ticket results and check\nresults after an interruption is a no-op success. Trail and trail-audit records instead follow the\nattempt-mark and resume rules in [the work trail protocol](..\u002F..\u002Fprotocols\u002Fwork-trail\u002FPROTOCOL.md).\n\nNo `remote-push`, no `pr-open`: a commit on an owned branch is where this lane ends.\n\n## Stop conditions\n\n`complete`: the ticket is committed, every acceptance criterion has a receipt bound to the committed\nrevision, and every finding either closed on that evidence or parked with a written ruling.\n\n`cap-reached`: the fifth fix round ended with findings still open. Returns the cap, the commit as it\nstands, and every open finding with its adjudication. It is not a pass and the ticket is not done.\n\n`needs-input`: no approved ticket, a ticket typed `decision`, a source finding too vague to specify,\nor a check seat that cannot be filled independently. Returns what is missing and the lane that owns\nit, never a narrowed version of the work.\n\n`failed`: no writable worktree, the branch is checked out elsewhere, or a configured knowledgebase refuses a\nreceipt, ticket result or check result. The worktree and the ledger are left intact for the resume. A knowledgebase coverage limit — `ak kb check` exit 3 (kb.unbound, kb.unregistered), a publish exit 3 (kb.read-only), kb.in-repo-missing, or a missing or unknown `ak kb` command — never stops this phase under any policy: reads are unavailable, each skipped publish is named in the result as not published, and only exit 1 stops.\n\n`cancelled`: the caller withdrew the ticket mid-run. The branch, the worktree and the recorded\nrulings are kept rather than cleaned up, because the next run reads them instead of re-deriving them.\n\n## Limits\n\nFix rounds: 5 per ticket (gate). Rounds one to three resume the same implementer; four and five\ndispatch a fresh one. This is the per-ticket implementer-and-check loop, and it is not\n`policies\u002Flimits.yaml`'s `fix_cycles`, which bounds the fix-and-verify cycles that follow a review\nrun's first pass and does not name this skill. Two caps, two scopes, and neither relaxes the other.\n\nImplementers per worktree: 1 (gate). Parallel dispatch is across tickets, never inside one.\n\nParallel tickets: a bounded batch rather than every independent ticket at once (guidance). A unit too\nsmall to outweigh its own dispatch runs inline or batched with related units; the runner's ticket\nbudget is the only cap this skill enforces on how many start.\n\nChanged-line target and test-pyramid shape: guidance, not a gate; configurable per project, neither\nvalidated nor enforced, not on its own grounds for a finding, an exception recorded rather than an artificial\nsplit or a meaningless test. [The reference loaded at step 5](..\u002F..\u002Freferences\u002Fengineering-principles\u002FREFERENCE.md)\nstates each (ruling `numeric-heuristics-are-guidance`). Reads of the calling agent: [read-economy](..\u002F..\u002Freferences\u002Fread-economy\u002FREFERENCE.md).\n","## When to use\n\nUse when a ticket exists, is approved, is typed `implementation`, and names its acceptance criteria\nand the verification that will show each one met. The ticket is the unit of work: one ticket, one\nworktree, one implementer, one pass of checks, one commit.\n\nUse when a run has reached its build phase and the ticket at the head of the ready set has no\nunfinished dependency — including the dependencies a dependency edge does not express.\n\nUse when a finding from a review lane or a diagnostic packet has already been sharpened into such a\nticket. What decides whether it could be is the finding's own grading, not how actionable its prose\nreads.\n\n## Not for\n\nNot for a ticket typed `decision`. A decision ticket asks which of two things to do; executing one of\nthem answers the question by doing it and calls the answer an implementation. The type field settles\nthis (`schemas\u002Fticket.schema.json`), and implementation-shaped prose in the body does not reopen it.\n\nNot for an unapproved or draft ticket. Approval binds to the ticket's artifact hash, so a ticket\nedited after approval is unapproved again and re-enters the step that approves it.\n\nNot for a finding that is not yet a ticket. A finding graded `spec_quality: smell`, or carrying\n`difficulty: null`, names a doubt rather than a change; sharpening it is a separate act by a lane\nthat holds that authority, and narrowing it here to something a fixer can act on invents the\nspecification the grading says is missing.\n\nNot for a bug report with a reproduction and no diagnosis behind it. A bug enters through `diagnose`,\nand what reaches this lane is the diagnosed result. A packet that already carried a verified bounded\npatch is finished work, and re-implementing it here is the same patch written twice (ruling\n`diagnose-patch-or-packet-never-both`).\n\nNot for \"just fix this quickly\" with no ticket at all. The ticket is what carries the acceptance\ncriteria, the allowed changes and the write ownership, and a run that improvises them has nothing to\ncheck the result against.\n\n## Authority\n\nAuthority: `model`. A controller or a parent skill starts it when an approved ticket is ready; no\nslash command 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` records model-invoked skills as exposing none by construction. What\nauthority governs here is the ticket rather than the caller: the approval the ticket carries, and the\ncharter checkpoint the run took before its build phase opened. A skill that cannot see either stops\nrather than treating its own invocation as the approval. Under the unattended policy (ADR-0014) that approval is a recorded unattended advance, a ticket whose class makes a human the author is built by the agent with its true class recorded (an exception to ruling `delegation-class-is-authority-not-finding`), and each gate is recorded with its evidence command, as [the unattended-lifecycle reference](..\u002F..\u002Freferences\u002Funattended-lifecycle\u002FREFERENCE.md) says. An unscored ticket (the delegation reference's coverage limit) is built the same way with no class recorded and its implementer seat still declared.\n\n## Inputs\n\nOne approved ticket typed `implementation` (`schemas\u002Fticket.schema.json`), with its\n`acceptance_criteria`, `verification`, `allowed_changes`, `read_dependencies`, `write_ownership` and\n`integration_owner`. Absent, unapproved, or typed `decision`: stop and report `needs-input` naming\nthe field that decided it.\n\nA writable checkout the run owns, and a worktree this ticket alone holds (`repository-write`,\n`isolated-worktree`, `vcs-local`). Unavailable, or the branch is already checked out elsewhere: stop\nand report `failed`.\n\nTwo independent contexts for the check seats (`independent-context`). A seat that cannot be filled\nindependently is `unavailable`, and unavailability blocks the ticket; it is never backfilled by the\nimplementer, by the author, by whoever approved the spec, or by the other seat (ruling\n`missing-supervisor-never-implementer`).\n\nThe source finding, where the ticket came from one (`schemas\u002Ffinding.schema.json`). A finding\narriving from a peer review lane graded `autofix_class: safe_auto` is remapped to `gated_auto` on\nintake and is never dropped for its class: the restriction is on emission by a seat, never on the\nvocabulary (ruling `safe-auto-restricted-per-seat`).\n\nOptionally a ticket budget from the runner. Absent: the cap is not enforced and not guessed, and the\nabsence is recorded (`policies\u002Flimits.yaml`).\n\n## Workflow\n\n1. Read the ticket and refuse what this lane does not execute — `decision` type, missing approval, or\n   a source finding graded `smell` or carrying `difficulty: null`. Record the refusal with the field\n   that decided it. Do not re-grade the finding to make it executable.\n2. Decide parallelism by inspection, not by the absence of a dependency edge. For each pair of\n   tickets with no edge between them, read the declared write ownership, generated artifacts,\n   migration sequence and interfaces, and read the files themselves where a declaration is silent.\n   Contention that survives inspection serializes exactly the contending tickets; the rest of the\n   layer still dispatches together.\n3. Record the base commit, then open the worktree this ticket owns\n   (`protocols\u002Fworktree-ownership\u002FPROTOCOL.md`). The worktree handle is bound to this ticket and is\n   retired when the ticket integrates, never retasked.\n   On the standalone path, open the task-bound run next, from inside that worktree with the\n   ticket's branch checked out and before implementation:\n   `node \u003Cthis skill's directory>\u002F..\u002F..\u002Fbin\u002Fak-gate.mjs open --ticket \u003Cticket-file>`. A worker that\n   holds a Firstmate binding opens the run with the command its brief gives, which appends\n   `--binding \u003Cbinding file> --dir \u003Cevidence store>`, so the run records the binding and the gate\n   finds its delegated grants. The run binds\n   to the branch checked out where it runs, and later gate commands on that branch resolve the\n   pointer it writes. A run never closed by `ship-preflight` stays the branch's default until a\n   new `open`, and the gate does not tell an earlier task's unclosed run from this one, so every\n   new task opens a new run, even on a branch that already has one.\n4. Dispatch one implementer (`roles\u002Fimplementer\u002FROLE.md`) with the ticket as its single source of\n   requirements. It spawns no implementers of its own, and no second implementer runs against this\n   worktree. Hand each implementer [the work-trail protocol](..\u002F..\u002Fprotocols\u002Fwork-trail\u002FPROTOCOL.md)\n   and the location of its own stretch, prepared in the runner's scratch space as that protocol\n   says, and no earlier stretch or pointer to one (ruling `work-trail-is-account-not-evidence`).\n5. Work the ticket test first (`protocols\u002Ftdd\u002FPROTOCOL.md`): a failing test that names the behaviour,\n   then the change that passes it. Where the behaviour is genuinely not testable, record the\n   alternative verification plan in the ticket rather than proceeding with neither. Load\n   [the engineering-principles reference pack](..\u002F..\u002Freferences\u002Fengineering-principles\u002FREFERENCE.md) for the standards the change is held to.\n6. Handle the implementer's report by what it says. Concerns about correctness or scope are addressed\n   before any check runs; missing context is supplied and the same implementer re-dispatched; a\n   blocker is assessed, not returned to the implementer as another attempt at the same wall.\n7. Record what was noticed and not touched — file and line, the observation, why it is out of scope.\n   Each entry is a candidate ticket, never an edit inside this one.\n8. Hand the diff to both check seats at once, as a file, with the ticket's constraints copied\n   verbatim from the ticket. Neither prompt carries the implementer's account of the work or any\n   expectation of what the seat will find.\n9. Keep the axes separate: `reviewer-spec` checks whether the ticket's obligations were met,\n   `reviewer-standards` checks the result against the project's own designated rules, rule by rule.\n   Findings are not merged or re-ranked across the two; one axis passing does not cover the other.\n10. Fix and re-check, at most five rounds per ticket. Rounds one to three resume the implementer that\n    already holds the context; rounds four and five dispatch a fresh implementer told how many\n    attempts preceded it and that it owns the task from here, and told nothing of their framing. Each\n    re-check is scoped per finding — addressed or not addressed — plus the fix diff read for new\n    breakage. A resumed implementer continues its own stretch; a fresh one is seated as step 4 says.\n11. At the round cap, adjudicate every finding still open, one at a time: park it with a written\n    ruling naming why the code stands and what it costs if that is wrong, or rule it load-bearing and\n    stop. A silent discard is forbidden.\n12. Load [verification evidence](..\u002F..\u002Freferences\u002Fverification-evidence\u002FREFERENCE.md), run the\n    ticket's named checks as implementer self-checks and record each receipt with\n    `created_by.role: implementer`. They are self-check evidence: once the build gate records the\n    implementer seat (step 13) they never satisfy the verify gate, and without it they count only\n    for checks no recipe binds. Hand the claim, criteria, frozen revision, project\n    recipe and permitted commands to `super-verify`; its verifier seat reruns them and produces\n    gate-eligible receipts. Commit the work on the ticket's own branch, then publish the receipts and\n    ticket result through the knowledgebase adapter's `publishArtifact` operation.\n13. When both check seats pass, record the gate: `node \u003Cthis skill's directory>\u002F..\u002F..\u002Fbin\u002Fak-gate.mjs record --gate build-checks`\n    (the bundle's `bin\u002F`, two directories above this skill). super-ship refuses to ship without it.\n    When a runner assigned the implementer seat, declare it by appending\n    `--class \u003Cdelegation class> --author-kind \u003Chuman|agent> --host \u003Chost adapter id> --seat-id \u003Cimplementer seat>`;\n    the seat is recorded on this gate, not on a receipt, and stays declared for commits descended\n    from the recorded one. After an amend or rebase, pass the flags again. An unscored ticket has no class to pass: declare the seat with `--author-kind \u003Chuman|agent> --host \u003Chost adapter id> --seat-id \u003Cimplementer seat>` and no `--class`, never a derived one or a seeded default, so a receipt that seat wrote is still refused.\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.\n14. A committed ticket then runs the trail, steps 14 to 16, a resumed run by the work-trail\n    protocol's resume rule; steps 12 and 13 never wait on it (ruling\n    `work-trail-is-account-not-evidence`). One stopped before its commit gets no audit or published\n    stretch, as the report says; a configured knowledgebase (exit 1) failing a receipt publish stops the build short of the\n    trail, while under a knowledgebase coverage limit the receipts are named as not published and the build goes on. Each prepared stretch is handled and sent as the work-trail protocol's build steps say;\n    none it holds back is sent. A refused or failed trail publish is never repaired or resent and is\n    not itself a build gate; the protocol's stop conditions describe shared-file failures that\n    can block later required receipts. The trail writes no gate record; `complete` does not depend on it.\n15. Seat [the trail auditor](..\u002F..\u002Froles\u002Ftrail-auditor\u002FROLE.md) once, even for an empty or missing\n    stretch, in a context that filled no other seat on the ticket and is not the build's own; hand\n    it its packet, each unpublished stretch named so, from its location. The audit is unavailable\n    with no such context, a binding not permitting the role, a packet it cannot be handed, a failed\n    seat, or a returned record not naming kind `trail-audit` and the id it was handed: withhold the\n    audit record, not the stretches, block nothing and let no seat fill in, the implementer included\n    (ruling `work-trail-is-account-not-evidence`). Otherwise publish the audit exactly as returned.\n16. Report every written ruling and out-of-scope observation last, the only existing act waiting on\n    the trail, with the audit's reference, result and items or why it is unavailable or unpublished,\n    and each missing or unpublished stretch, an unpublished one with its reason. Repeat as given an\n    audit notice that something read holds a secret, credential, token or personal data: where it\n    is, never the value. For a published stretch or audit, add that it cannot be corrected and the\n    value is exposed, for its owner to rotate, as `super-ship` does for a secret in history. The\n    attention list and that notice go only in the report and audit record, never in a receipt, check\n    result, finding, gate record or checkpoint card (ruling `work-trail-is-account-not-evidence`).\n\n## Hard gates\n\nGate: a ticket typed `decision` is never executed as implementation work, and never re-typed here to\nmake it executable. It returns to the lane that resolves decisions.\n\nGate: a finding graded `smell`, or carrying `difficulty: null`, is refused as an implementation\nticket (`schemas\u002Ffinding.schema.json`); the shared rule for applying a finding at all is\n`protocols\u002Fapply-findings\u002FPROTOCOL.md`. Difficulty without a solution class is not a specification,\nand a re-grading performed by the lane that wants to execute it is the refusal being routed around.\n\nGate: a receipt created by the implementer, change author, spec approver or recipe author is\nself-check evidence only. It never fills or substitutes for the verifier seat, even when its command\nand output are identical to the later independent run (ruling\n`missing-supervisor-never-implementer`).\n\nGate: neither check seat emits `autofix_class: safe_auto`. At review time a code edit has no single\nmechanically correct answer, so a seat's classification is a proposal and applying it is the caller's\ndecision under its own authorization; `gated_auto`, `manual` and `advisory` are what a seat here may\nemit (ruling `safe-auto-restricted-per-seat`).\n\nGate: the implementer never sits on the panel that checks its work, and no seat approves a patch it\nproduced. Two seats that are the same context in sequence are one seat (ruling\n`missing-supervisor-never-implementer`).\n\nGate: only independent verification evidence, plus a policy rule saying that evidence suffices for\nthat finding, closes a finding. A description of a green run is not a receipt — a receipt carries the\ncommand or probe, exit status, output digest, revision and environment identity — and a patch that\nchanged after a receipt was taken does not inherit it (ruling\n`closure-requires-independent-verification`).\n\nGate: writes stay inside the ticket's `allowed_changes` and its declared ownership. A write outside\nevery worker's declared exclusive set aborts the wave, and a change no worker accounts for may be the\nuser's: it is preserved and reconciled, never discarded to get the tree clean.\n\nGate: a stubbed test, a skipped test, an unimplemented branch or a marker left in place of the work\nis a blocker, not progress. The ticket is reported blocked with what is missing.\n\nGate: the round cap ends the loop, not the adjudication. Reaching it with findings open is a\n`cap-reached` result carrying each one and its ruling.\n\nGate: the wall (ruling `work-trail-is-account-not-evidence`). No seat whose output is a finding, a\nreceipt, a verdict or a signoff input is shown the trail, the trail audit or the attention list.\nNothing from them reaches a check seat or the verification hand-off, quoted, summarized or as a\npointer, and the build's context carries nothing it saw of them into a ruling it writes, a synthesis\nor verdict it sets, or a packet it hands a seat the wall covers, for the same change.\n\n| The thought | Why it is wrong | Do this instead |\n|---|---|---|\n| \"The implementer just fixed it and is sitting right there — having it confirm the fix is faster than dispatching a seat.\" | A seat that cannot be filled independently is unavailable, and unavailability blocks; it is never backfilled by the implementer (ruling `missing-supervisor-never-implementer`). The context that produced the change is the one context that cannot tell you whether it is right. | Mark the seat `unavailable`, stop with `needs-input` naming which seat and why, and report the ticket as not done. |\n| \"The implementer's report says the tests passed, so the criterion is met.\" | An agent's description of a green run is not a receipt (ruling `closure-requires-independent-verification`). The report is the author's account of its own work, and it is the thing the receipt exists to replace. | Run the ticket's named verification, record command, exit status, output digest, revision and environment, and bind each receipt to the criteria it supports. |\n| \"The finding is only a smell, but I can see what it means — I will write the ticket from it.\" | Difficulty without a solution class is not a specification, and reading a change out of a smell is inventing the one the grading says is missing. The invented specification is then checked against itself. | Refuse it as an implementation ticket, and return it for sharpening, diagnosis or escalation with the field that decided it. |\n| \"Round five ended with two small findings open; they are minor and the ticket is otherwise done.\" | A silent discard is exactly what the round cap exists to prevent, and the findings least likely to be written down are the ones a later reader most needs. | Park each one with a written ruling naming what it costs if wrong, or rule it load-bearing and stop; return `cap-reached` with both attached. |\n| \"These two tickets have no dependency edge, so they can run in parallel.\" | An edge records a declared dependency, not shared write surface. A shared generated artifact, a migration sequence or one exported interface makes two edgeless tickets contending writers, and the DAG cannot see it. | Read the ownership declarations and the files, serialize exactly the contending pair, and dispatch the rest of the layer together. |\n| \"The seat should know I already considered the null case, so I will say so in the prompt.\" | A prompt that says what to expect tells the seat where to stop looking, and the axis returns the author's confidence in the author's own words. | Hand over the diff and the ticket's verbatim constraints, and nothing about what the previous attempt believed. |\n\n## Outputs\n\nOne commit per ticket, on the branch that ticket owns, recorded against the ticket by id. The commit\nis the patch link the lane's result carries; nothing is merged here.\n\nVerification receipts (`schemas\u002Fverification.schema.json`), one per acceptance criterion the ticket\nnames, each naming the criteria it supports and bound to the revision it describes.\n\nThe per-ticket check result: the findings each axis raised (`schemas\u002Ffinding.schema.json`), their\ndisposition, and every ruling written at the round cap. A per-ticket check is a lighter instrument\nthan a review panel — it emits no `review` artifact, and a later panel never reads it as a lane it no\nlonger has to cover.\n\nAll of it is published through the knowledgebase adapter's `publishArtifact` operation with a\nrun-artifact placement (`adapters\u002Fknowledgebase\u002FCONTRACT.md`); the commands are in\n[the knowledgebase-binding reference pack](..\u002F..\u002Freferences\u002Fknowledgebase-binding\u002FREFERENCE.md).\nThis skill names no repository path for project-derived content. Each implementer seat's stretch\n(`schemas\u002Ftrail.schema.json`) and the trail audit (`schemas\u002Ftrail-audit.schema.json`) go through\n`publishArtifact` too, as an account, never evidence (ruling `work-trail-is-account-not-evidence`).\n\n## Side effects\n\n`workspace-write`, `branch-create`, `local-commit`, `process-exec`, `artifact-write`, `kb-publish`,\n`scratch-write`.\n\n`kb-publish` is a remote effect: the idempotency key is derived from the content hash of the\npublished set per `adapters\u002Frunner-contract\u002FCONTRACT.md` §5, and the returned record reference is\nread back before the ticket reports. Republishing unchanged receipts, ticket results and check\nresults after an interruption is a no-op success. Trail and trail-audit records instead follow the\nattempt-mark and resume rules in [the work trail protocol](..\u002F..\u002Fprotocols\u002Fwork-trail\u002FPROTOCOL.md).\n\nNo `remote-push`, no `pr-open`: a commit on an owned branch is where this lane ends.\n\n## Stop conditions\n\n`complete`: the ticket is committed, every acceptance criterion has a receipt bound to the committed\nrevision, and every finding either closed on that evidence or parked with a written ruling.\n\n`cap-reached`: the fifth fix round ended with findings still open. Returns the cap, the commit as it\nstands, and every open finding with its adjudication. It is not a pass and the ticket is not done.\n\n`needs-input`: no approved ticket, a ticket typed `decision`, a source finding too vague to specify,\nor a check seat that cannot be filled independently. Returns what is missing and the lane that owns\nit, never a narrowed version of the work.\n\n`failed`: no writable worktree, the branch is checked out elsewhere, or a configured knowledgebase refuses a\nreceipt, ticket result or check result. The worktree and the ledger are left intact for the resume. A knowledgebase coverage limit — `ak kb check` exit 3 (kb.unbound, kb.unregistered), a publish exit 3 (kb.read-only), kb.in-repo-missing, or a missing or unknown `ak kb` command — never stops this phase under any policy: reads are unavailable, each skipped publish is named in the result as not published, and only exit 1 stops.\n\n`cancelled`: the caller withdrew the ticket mid-run. The branch, the worktree and the recorded\nrulings are kept rather than cleaned up, because the next run reads them instead of re-deriving them.\n\n## Limits\n\nFix rounds: 5 per ticket (gate). Rounds one to three resume the same implementer; four and five\ndispatch a fresh one. This is the per-ticket implementer-and-check loop, and it is not\n`policies\u002Flimits.yaml`'s `fix_cycles`, which bounds the fix-and-verify cycles that follow a review\nrun's first pass and does not name this skill. Two caps, two scopes, and neither relaxes the other.\n\nImplementers per worktree: 1 (gate). Parallel dispatch is across tickets, never inside one.\n\nParallel tickets: a bounded batch rather than every independent ticket at once (guidance). A unit too\nsmall to outweigh its own dispatch runs inline or batched with related units; the runner's ticket\nbudget is the only cap this skill enforces on how many start.\n\nChanged-line target and test-pyramid shape: guidance, not a gate; configurable per project, neither\nvalidated nor enforced, not on its own grounds for a finding, an exception recorded rather than an artificial\nsplit or a meaningless test. [The reference loaded at step 5](..\u002F..\u002Freferences\u002Fengineering-principles\u002FREFERENCE.md)\nstates each (ruling `numeric-heuristics-are-guidance`). Reads of the calling agent: [read-economy](..\u002F..\u002Freferences\u002Fread-economy\u002FREFERENCE.md).\n",{"name":5,"license":15,"metadata":24,"description":7},{"ak_catalog_id":5},"main","e0232821c4abbcd4338d6018c20fb70f1b74336b","skills\u002Fsuper-build",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":34,"prompt":44,"zipAvailable":45},[35,38,41],{"tool":36,"command":37},"npx","npx skills add Pibomeister\u002Fagent-kit --skill super-build",{"tool":39,"command":40},"bunx","bunx skills add Pibomeister\u002Fagent-kit --skill super-build",{"tool":42,"command":43},"pnpm","pnpm dlx skills add Pibomeister\u002Fagent-kit --skill super-build","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FPibomeister\u002Fagent-kit\u002Fblob\u002Fmain\u002Fskills\u002Fsuper-build\u002FSKILL.md",true,[47],{"path":48,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":51,"sourceHtmlUrl":52},"skill.yaml","REFERENCE","参考文档",23966,"https:\u002F\u002Fgithub.com\u002FPibomeister\u002Fagent-kit\u002Fblob\u002Fmain\u002Fskills\u002Fsuper-build\u002Fskill.yaml",[54,61,66,71,76,81],{"skillNo":55,"name":56,"nameZh":6,"description":57,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":58,"categories":59,"keywords":60,"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":62,"name":56,"nameZh":6,"description":57,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":63,"categories":64,"keywords":65,"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":67,"name":56,"nameZh":6,"description":57,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":68,"categories":69,"keywords":70,"publishedAt":19},"01M7D3HQDDAZTAPGWCPGDPNVBQ","https:\u002F\u002Fgithub.com\u002FPibomeister\u002Fagent-kit\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Finvalid\u002F24-article-without-triggers\u002Fskills\u002Falpha\u002FSKILL.md",[],[],{"skillNo":72,"name":56,"nameZh":6,"description":57,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":73,"categories":74,"keywords":75,"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":77,"name":56,"nameZh":6,"description":57,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"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":82,"name":56,"nameZh":6,"description":57,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":83,"categories":84,"keywords":85,"publishedAt":19},"01M7D3HDV6HMK2ZHRM0P368KNT","https:\u002F\u002Fgithub.com\u002FPibomeister\u002Fagent-kit\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Finvalid\u002F21-fragment-redeclares-id\u002Fskills\u002Falpha\u002FSKILL.md",[],[]]