[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M79A10WWWN33Y2G4JFQ0ES9X":3,"skill-same:01M79A10WWWN33Y2G4JFQ0ES9X":48},{"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":30,"ownerHtmlUrl":31,"ownerBlogUrl":6,"ownerDescription":32,"updatedAt":33,"install":34,"resources":47},"01M79A10WWWN33Y2G4JFQ0ES9X","arbeitsplan-run",null,"Executes a compiled workflow as a redundant swarm — N candidates over the same scope in their own git worktrees, judged blind, exactly one landed. Use after arbeitsplan-compile, or when the user asks to run the workflow, start the swarm, or build several versions and pick one. Halts and surfaces rather than retrying when candidates fail; convergence comes from widening, never from repeating. Writes only inside candidate worktrees until the landing step.","Anselmoo\u002Fwerkstoff","Anselmoo","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F13209783?v=4","01M799YTB3D7RZ4P3B8CSR2X85","Anselm Hahn","01M799YTBRAM6719D9V0NRGB4R",1,"MIT","https:\u002F\u002Fgithub.com\u002FAnselmoo\u002Fwerkstoff\u002Fblob\u002Fmain\u002Fplugins\u002Farbeitsplan\u002Fskills\u002Farbeitsplan-run\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: arbeitsplan-run\ndescription: Executes a compiled workflow as a redundant swarm — N candidates over the same scope in their own git worktrees, judged blind, exactly one landed. Use after arbeitsplan-compile, or when the user asks to run the workflow, start the swarm, or build several versions and pick one. Halts and surfaces rather than retrying when candidates fail; convergence comes from widening, never from repeating. Writes only inside candidate worktrees until the landing step.\nargument-hint: \"[runId]\"\n---\n\n# Run the swarm\n\nN candidates build the **same scope** in isolation. Exactly one lands. The rest are deleted.\n\nThat is what makes a merge conflict impossible here: this is the same worktree isolation\n`superpowers` uses, inverted. There, N worktrees do *different* work and must all be\nreconciled — which is why `subagent-driven-development` forbids parallel implementers inside\none worktree \"(conflicts)\" and pushes the parallelism up to where the conflicts reappear.\nHere, N worktrees do *the same* work and N−1 are discarded. Nothing is ever merged.\n\n## Steps\n\n1. **Read `workflow.json`.** Read the file; do not work from a summary of it. Refuse a\n   `schemaVersion` other than `\"2\"` rather than guessing at it. Then branch on\n   `backend.kind` — it is a decision the spec already made, never yours to re-make:\n   `workflow` → follow [The workflow backend](#the-workflow-backend) below; `matrix` →\n   `arbeitsplan-matrix`; `in-session` → the steps here.\n\n   A phase with `mode: \"plan\"` never runs under the lock: `worktree_pool.py open` refuses it,\n   because plan mode plus an open lock leaves no legal write. Close, run that phase in plan\n   mode, then open the next one.\n\n2. **Open the run scope lock** before any dispatch:\n\n   ```bash\n   python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Fworktree_pool.py\" open --spec analysis\u002Farbeitsplan\u002F\u003CrunId>\u002Fworkflow.json --phase build\n   ```\n\n   This writes `analysis\u002Farbeitsplan\u002Frun_scope.json`, which arms the guard. It is a\n   **per-dispatch lock, not repo-level state** — that distinction is why this guard does not\n   sweep up every other plugin's edits the way one in this repository once did.\n\n3. **Create one worktree per candidate**, one angle each, from `workflow.json`'s `angles`.\n\n4. **Dispatch the batch in ONE message.** Multiple dispatch calls in one response run in\n   parallel; one per response is sequential. Each `arbeitsplan:candidate-builder` gets its\n   angle, its worktree, the scope, and the acceptance list — and **never** another\n   candidate's angle or output.\n\n   These are in-session dispatches, so the guard sees each one. The workflow backend is a\n   different run, compiled for it — never switch to it mid-run.\n\n   **Record each builder's result** as it returns — the JSON it printed, saved to a file —\n   with its own worktree:\n\n   ```bash\n   python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Frecord_event.py\" candidate --run \u003CrunId> --phase build --result c2.json --tree .arbeitsplan\u002F\u003CrunId>\u002Fc2\n   ```\n\n   This writes `candidates\u002Fc2.json`, the file `reconcile.py` (step 8) and `land_candidate.py`\n   (step 9) read; nothing else writes it on this backend. `--tree` replaces the builder's\n   self-reported `diff` and `filesTouched` with what its worktree actually holds, and notes\n   whether they differed — builders paraphrase long diffs, and a paraphrased diff does not\n   apply. It refuses a phase that was never opened, a result missing any field of\n   `candidate-builder`'s output shape, and a candidate already recorded, writing nothing.\n\n5. **Apply the breaker, per batch, never cumulatively.**\n\n   - Exclude every `measured: false` candidate from the denominator. It is not a rejection;\n     it is a candidate that was never fairly tried.\n   - `measured == 0` → **acquisition problem.** The environment failed, not the contract. Fix\n     the environment and re-run. Do not re-dispatch into a broken environment.\n   - `accepted * 3 \u003C measured * 2` → **contract problem.** Halt and say so: *the correct\n     response is a better contract, not more agents.*\n\n6. **Referee blind.** One `arbeitsplan:candidate-referee` per surviving candidate, each given\n   the criteria and that candidate's diff only — never the builder's rationale, never another\n   candidate. Landing is an **allowlist**: only `accepted`.\n\n   Record the batch — every verdict the referees returned, as one JSON list — under the referee\n   phase you opened:\n\n   ```bash\n   python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Frecord_event.py\" referee --run \u003CrunId> --phase \u003CrefereePhase> --verdict verdicts.json\n   ```\n\n   This writes `referee\u002F\u003Cid>.json` (what `land_candidate.py` reads) **and** the\n   `referee\u002F\u003Cphase>\u002F\u003Cid>.json` twin that `rounds.py` rebuilds rounds from. Without it an\n   in-session run has no rounds, and step 7's `rounds.py decide` can never route a shared hole\n   or stop a moving residual. It refuses a phase that was never opened, a malformed verdict,\n   and a (phase, candidate) already recorded — and writes nothing when it refuses.\n\n7. **Select by rule**, not by preference: most criteria met, then fewest files touched, then\n   candidate id. If no candidate is `accepted`, **halt and surface**. All N failing the same\n   way is a statement about the contract — do not just widen and re-dispatch. Run\n   `scripts\u002Frounds.py record --run \u003CrunId>` (from the root holding `analysis\u002Farbeitsplan\u002F`) to\n   rebuild this run's rounds, then `scripts\u002Frounds.py decide --rounds \u003Cfile> --spec\n   workflow.json` (see [Rounds: a shared hole or a moving residual](#rounds-a-shared-hole-or-a-moving-residual-78-93)\n   below) to learn which of three things this halt is: `ROUTE CONTINUE` (nothing to say yet),\n   `ROUTE SYNTHESIZE criterion=\u003Cid>` (a **sharedHole**: relaunch at the single-writer phase with\n   `carry.sharedHole` set, rather than building N more candidates against the same hole), or\n   `ROUTE HALT moving-residual` (stop outright: the last several rounds each \"advanced\" on a\n   *different* blocker, which is not convergence).\n\n8. **Measure the winner against its own tree, before landing** (#76). A builder's self-reported\n   exit codes are never trusted on their own:\n\n   ```bash\n   python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Freconcile.py\" --run \u003CrunId> --run-checks --candidate c2 --tree .arbeitsplan\u002F\u003CrunId>\u002Fc2\n   ```\n\n   This RUNS every checked acceptance criterion now, with `cwd` set to `c2`'s OWN worktree, and\n   records each as an `execute_tool` event in `run.jsonl`, carrying `detail.candidate == c2`. Any\n   disagreement with what `c2` itself reported — in either direction — prints `CONTRADICTION` and\n   exits 1. `land_candidate.py --apply` (next step) **refuses** a candidate lacking this\n   measurement for any checked criterion, or whose latest measurement contradicts its report;\n   another candidate's measurement never unlocks it. An honestly-reported failure a referee\n   already accepted is not blocked here — the gate is \"unmeasured or contradicted\", nothing more.\n\n9. **Land exactly one diff.**\n\n   ```bash\n   python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Fworktree_pool.py\" open --spec ... --phase land\n   python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Fland_candidate.py\" --run \u003CrunId> --candidate c2 --apply\n   ```\n\n   Optional gated synthesis, only if the spec opted in: one further writer, given the winner\n   plus **explicitly named** elements to borrow. It must beat the plain winner on at least one\n   declared criterion, or the plain winner lands unchanged.\n\n10. **Delete the losers** — worktrees and branches — then record the phase as closed and close\n    the lock:\n\n    ```bash\n    python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Frecord_event.py\" phase --run \u003CrunId> --phase land --status closed\n    python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Fworktree_pool.py\" close\n    ```\n\n    `close` **refuses** a phase that recorded no terminal event. On a halt, close with\n    `--halt \"\u003Cthe specific reason>\"`: a halt is an event in `run.jsonl`, never an absence.\n\n11. **End the run in the record** -- landed or halted, always last:\n\n    ```bash\n    python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Frecord_event.py\" finish --run \u003CrunId>\n    ```\n\n    A landed run gets `complete.json` (refused while any phase is still open); a halted one\n    gets `FAILED-\u003Cstamp>.json` carrying the halt's reason; anything else is refused, because\n    stopping is not an ending. This marker is what `sweep_artifacts.py` waits for -- a run\n    without it is kept forever, whatever state its worktrees are in.\n\n    Deletion itself goes through `worktree_pool.py destroy` (or `sweep_artifacts.py` for a whole\n    finished run), which never just discards a loser's uncommitted state (#80): a DIRTY worktree\n    is committed and preserved on `kept\u002F\u003CrunId>-\u003Ccid>` before its worktree and throwaway\n    `arbeitsplan\u002F\u003CrunId>\u002F\u003Ccid>` branch are removed; a clean worktree gets no `kept\u002F` branch.\n\n## Referee-owned artifacts (#77)\n\nIf `workflow.json` declares `refereeOwned`, its `referee-fixture` phase runs first, exactly\nlike any other phase (open the lock, dispatch its `agentType`, record the output, close). Once\nit has written those paths, baseline them **once**:\n\n```bash\npython3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Freferee_owned.py\" record --run \u003CrunId>\n```\n\nA second `record` for the same run is refused: the baseline is taken at creation, never\nre-taken to accommodate a later change. `worktree_pool.py open` already narrows a fan-out\nphase's lock to exclude every `refereeOwned` path, so the guard denies a candidate's Edit\u002FWrite\nbefore it lands — but that covers only the tool calls the guard's matcher sees. Re-verify\nbefore landing (step 9), as the belt to that guard's suspenders:\n\n```bash\npython3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Freferee_owned.py\" verify --run \u003CrunId>\n```\n\n`land_candidate.py` independently refuses (citing `refereeOwned` by name) any candidate diff\nthat touches one of these paths, so the same rule is checked three ways: at the lock, at\nlanding, and by content.\n\n## Rounds: a shared hole or a moving residual (#78, #93)\n\nA no-accept halt is arithmetic (`accepted * den \u003C measured * num`) and never looks at WHY.\n`scripts\u002Frounds.py` is the one place that does, and both rules below read the SAME derived\nrounds — one round per referee phase, rebuilt from `referee\u002F\u003Cphase>\u002F\u003Cid>.json` and the\nadjudicator's recorded `round: {outcome, blocking}` — so a fix to one is a fix to both. Those\nfiles come from `record_event.py workflow` (the workflow backend) or `record_event.py referee`\n(in-session, step 6); a batch recorded any other way is invisible here:\n\n```bash\npython3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Frounds.py\" record --run \u003CrunId> > \u002Ftmp\u002Frounds.json\npython3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Frounds.py\" decide --rounds \u002Ftmp\u002Frounds.json --spec analysis\u002Farbeitsplan\u002F\u003CrunId>\u002Fworkflow.json\n```\n\n- **`ROUTE SYNTHESIZE criterion=\u003Cid>` — a `sharedHole` (#78).** The LATEST round accepted no\n  one and >= 2 rejected candidates share one unmet criterion — a structural hole every\n  candidate hit, not noise. Relaunch `workflows\u002Frun.js` with `startAt` at the single-writer\n  phase and `carry.sharedHole = {criterion, candidates: [{candidateId, diff}]}` (the run.js\n  halt's own `rejectionsByCriterion`, paired with each candidate's diff, builds this). The\n  synthesizer then works from every rejected diff toward that one criterion, instead of N fresh\n  candidates re-discovering the same hole from scratch.\n- **`ROUTE HALT moving-residual` (#93).** A per-batch breaker and a `sharedHole` both look at\n  ONE round; a run that \"advances, not closes\" every round, each time naming a *different*\n  blocking condition, passes both forever. This fires when the last N **judged** rounds\n  (`judge.outcome != \"none\"`; an unjudged round neither breaks nor counts) are all `\"advanced\"`\n  with non-null, pairwise-**distinct** `judge.blocking` ids — N genuinely different obstacles in\n  a row, not real convergence. Stop outright: **do not** re-compile and **do not** widen again.\n  `N` is `--max-advancing-rounds`, else the spec's `roundBreaker.maxAdvancingRounds`, else 3.\n  When both rules would fire on the same rounds, the `moving-residual` HALT wins.\n- **`ROUTE CONTINUE`** — neither condition holds; proceed as normal.\n\n## Stacked fan-outs across waves (#79)\n\nA later wave's builders should sometimes start from an **earlier wave's refereed winner**,\nnot from HEAD — that is what a `fanout-redundant` phase's `base: \"\u003CphaseId>\"` declares. Once a\nphase's candidate is selected (step 7 above) and would normally just land (step 9), promote it\ninstead if a later phase names it as `base`:\n\n```bash\npython3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Fworktree_pool.py\" promote --run \u003CrunId> --phase build-w1 --candidate c2\n```\n\nThis commits everything sitting in that candidate's worktree — untracked files included — and\npoints `arbeitsplan\u002F\u003CrunId>\u002Fbase\u002Fbuild-w1` at the new commit. The next wave's `create` then\nreads it automatically:\n\n```bash\npython3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Fworktree_pool.py\" create --spec analysis\u002Farbeitsplan\u002F\u003CrunId>\u002Fworkflow.json --phase build-w2 --count 2\n```\n\nIf `build-w2` declares `base: \"build-w1\"`, every worktree this creates starts from the promoted\nbranch instead of HEAD; if `build-w1` was never promoted, `create` **refuses**, naming the\nmissing branch, and creates nothing. A `--phase` whose phase carries no `base` (or `create` with\nno `--phase` at all) behaves exactly as before: from HEAD.\n\nBase branches deliberately **survive** an ordinary `worktree_pool.py destroy --run \u003CrunId>` — a\nstill-pending later wave may need to stack on one. Only pass `--bases` once the whole run is\nactually done with them:\n\n```bash\npython3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Fworktree_pool.py\" destroy --run \u003CrunId> --bases\n```\n\n`compile_spec.py --strict` is the check that a stacked run actually declared every `base` it\nneeded: a `fanout-redundant` phase whose transitive `requires` reach another `fanout-redundant`\nphase with no `base` chain reaching it back prints `WARNING ... [AP-SIBLING-INVISIBLE] ...`\nnaming both phases. A plain compile still writes past a warning (exit 0); `--strict` treats one\nas a rejection (exit 1) — run compilation with `--strict` before dispatching a stacked run.\n\n## The workflow backend\n\nWhen `backend.kind` is `\"workflow\"`, the whole phase graph runs inside `workflows\u002Frun.js`,\nand this session does only the things the Workflow tool cannot: read files, run plan-mode\nphases, measure the accepted candidate with `reconcile.py --run-checks` (#76), and land.\n\n1. **Launch it with the spec as data.** The Workflow tool has no filesystem, so pass the\n   parsed `workflow.json` verbatim: `args: {spec}` (plus `startAt` and `carry` on a resume).\n   No run-scope lock is open while it runs — nothing inside writes the shared tree.\n1b. **Arm the script allowlist when the spec has a `script` phase (#107).** `compile_spec.py\n   --write` already wrote `analysis\u002Farbeitsplan\u002F\u003CrunId>\u002Fscripts.json` from the spec; point the\n   guard at it before launching:\n\n   ```bash\n   printf '{\"runId\": \"%s\"}\\n' \u003CrunId> > analysis\u002Farbeitsplan\u002Fscripts_armed.json\n   ```\n\n   The guard judges every Bash call `arbeitsplan:script-runner` makes — lock or no lock — and\n   allows exactly one declared command per dispatch. Without an armed allowlist the runner is\n   denied everything, which is the safe direction. For `{placeholder}` slots, pass their values\n   as `carry.scriptArgs.\u003CphaseId>`; a value outside `[A-Za-z0-9._\u002F,=:@+-]` halts the phase\n   before dispatch.\n2. **Persist what it returns, before reading it.** Every return carries `events`:\n\n   ```bash\n   python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Frecord_event.py\" workflow --run \u003CrunId> --result result.json\n   ```\n\n   A halt is an event. Record it even — especially — when the run aborted.\n3. **On `pending_plan_node`, run that phase here, in plan mode**, by dispatching its\n   `agentType` with the carried data. Record its output, add it to `carry` under the phase id\n   (`carry.contract.acceptance` replaces the compiled acceptance for later phases), and\n   relaunch with `startAt: \u003CresumeWith>`. Loop state lives here, never in the script.\n   That same command writes `candidates\u002F\u003Cid>.json` and `referee\u002F\u003Cid>.json` for every\n   candidate the return carries, once each, never overwriting.\n3b. **Record a plan-mode phase's output** the same way:\n   `record_event.py phase-output --run \u003CrunId> --phase \u003Cid> --output \u003Cfile>`. An adjudicator's\n   `verdict: \"land\"` is what gives a borrowed synthesis its referee record; without it,\n   `land_candidate.py` refuses the synthesis like any unjudged candidate.\n3c. **On a NO CANDIDATE ACCEPTED halt, `rounds.py` decides the next move, not this session.**\n   The halt's `rejectionsByCriterion` is already grouped by criterion; run `scripts\u002Frounds.py\n   record --run \u003CrunId>` then `scripts\u002Frounds.py decide --rounds \u003Cfile> --spec workflow.json`\n   (see [Rounds: a shared hole or a moving residual](#rounds-a-shared-hole-or-a-moving-residual-78-93)\n   above — same rules, same `sharedHole` and `moving-residual` outcomes, whichever backend ran\n   the halted batch). A `SYNTHESIZE` route relaunches `run.js` with `startAt` at the\n   single-writer phase and `carry.sharedHole` set, from the same halt's `rejectionsByCriterion`\n   and `candidates`; a `moving-residual` HALT stops the run outright.\n4. **On completion, measure then land in-session.** `worktree_pool.py open --phase \u003Clast>`, then\n\n   ```bash\n   python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Freconcile.py\" --run \u003CrunId> --run-checks --candidate \u003Cid> --tree .arbeitsplan\u002F\u003CrunId>\u002F\u003Cid>\n   ```\n\n   to measure every checked criterion against the accepted candidate's own tree (#76), then\n   `land_candidate.py`, where the hook, the `writeScope` check, and this measurement gate all\n   run. The workflow never lands. `land_candidate.py` writes `landed.json`, with `divergedFrom`\n   if what landed differs from the recorded candidate. Then end the run with\n   `record_event.py finish --run \u003CrunId>` -- after a halt too -- exactly as step 11 above.\n\n## Rules\n\n- **Never re-dispatch an identical prompt.** The hook denies it, and the denial is correct:\n  past the cap those rounds do not converge, they only cost. Widen instead — a new angle\n  nobody tried — or stop.\n- **Never let a builder write to the shared tree.** Only the landing step touches it.\n- **Never treat an unmeasured candidate as a failure.** It has no rate, only missing data.\n- **Never raise the budget mid-run.** That is a decision to re-compile.\n- **Delegation nests at most 3 deep, and never in a cycle** — both enforced by the hook, see\n  `references\u002Fdelegation.md`. A dispatch to arbeitsplan's own agents is fan-out and is not\n  counted; a dispatch to another plugin is.\n- **A denial from another plugin's guard is that plugin doing its job.** Report it; never\n  reach for its escape hatch.\n\n## Output format\n\n```\narbeitsplan run — ap-2026-09-12-a3f1, phase build\n  c1  middleware-layer       measured   checks 3\u002F3   diff 2 files\n  c2  decorator-per-route    measured   checks 3\u002F3   diff 2 files\n  c3  reverse-proxy-config   UNMEASURED no nginx toolchain in the worktree\n\n  breaker  accepted 2 \u002F measured 2  (c3 excluded from the denominator) — pass\n\nphase referee (blind)\n  c1  rejected   a3 not met: requirements.txt gains 'slowapi'\n  c2  accepted   3\u002F3, evidence at src\u002Fapi\u002Fsearch.py:41\n\nphase land\n  winner   c2 (3 criteria met, 2 files touched)\n  applied  src\u002Fapi\u002Fsearch.py, tests\u002Ftest_ratelimit.py\n  deleted  c1, c3 worktrees and branches — nothing merged\n  marker   .takt\u002Fap-2026-09-12-a3f1\u002Flanded\n\ndispatches 7\u002F9 budget.\n```\n\nA halt is an output, not a crash:\n\n```\narbeitsplan run — ap-2026-09-12-a3f1 HALTED at phase referee\n  accepted 0 \u002F measured 3 — every candidate failed criterion a1 the same way.\n  All three read the limit from a per-process dict, so a1 (\"429 across workers\")\n  cannot be met inside the declared writeScope at all.\n  This is a statement about the contract, not about the candidates.\n  The correct response is a better contract, not more agents: re-compile with a\n  scope that admits shared state, or drop a1.\n  Nothing was landed. Worktrees kept at .arbeitsplan\u002Fap-2026-09-12-a3f1\u002F for reading.\n```\n\n## Resources\n\n- `references\u002Fcandidate-contract.md` — what builders and referees exchange, and why the\n  referee is starved.\n- `references\u002Fdelegation.md` — read before dispatching another plugin: the registry every\n  plugin declares, the append-only ledger, and the depth cap plus cycle detection that bound\n  how far a delegation may nest.\n- `references\u002Fpatterns.md` — the breaker's thresholds and the reason they are per-batch.\n- `workflows\u002Frun.js` — the workflow backend: executes `spec.phases`, halts before each\n  plan-mode phase, counts the dispatch budget in code, and returns span-shaped `events`.\n- `references\u002Fbackend-selection.md` — why a run was compiled for the backend it names.\n","# Run the swarm\n\nN candidates build the **same scope** in isolation. Exactly one lands. The rest are deleted.\n\nThat is what makes a merge conflict impossible here: this is the same worktree isolation\n`superpowers` uses, inverted. There, N worktrees do *different* work and must all be\nreconciled — which is why `subagent-driven-development` forbids parallel implementers inside\none worktree \"(conflicts)\" and pushes the parallelism up to where the conflicts reappear.\nHere, N worktrees do *the same* work and N−1 are discarded. Nothing is ever merged.\n\n## Steps\n\n1. **Read `workflow.json`.** Read the file; do not work from a summary of it. Refuse a\n   `schemaVersion` other than `\"2\"` rather than guessing at it. Then branch on\n   `backend.kind` — it is a decision the spec already made, never yours to re-make:\n   `workflow` → follow [The workflow backend](#the-workflow-backend) below; `matrix` →\n   `arbeitsplan-matrix`; `in-session` → the steps here.\n\n   A phase with `mode: \"plan\"` never runs under the lock: `worktree_pool.py open` refuses it,\n   because plan mode plus an open lock leaves no legal write. Close, run that phase in plan\n   mode, then open the next one.\n\n2. **Open the run scope lock** before any dispatch:\n\n   ```bash\n   python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Fworktree_pool.py\" open --spec analysis\u002Farbeitsplan\u002F\u003CrunId>\u002Fworkflow.json --phase build\n   ```\n\n   This writes `analysis\u002Farbeitsplan\u002Frun_scope.json`, which arms the guard. It is a\n   **per-dispatch lock, not repo-level state** — that distinction is why this guard does not\n   sweep up every other plugin's edits the way one in this repository once did.\n\n3. **Create one worktree per candidate**, one angle each, from `workflow.json`'s `angles`.\n\n4. **Dispatch the batch in ONE message.** Multiple dispatch calls in one response run in\n   parallel; one per response is sequential. Each `arbeitsplan:candidate-builder` gets its\n   angle, its worktree, the scope, and the acceptance list — and **never** another\n   candidate's angle or output.\n\n   These are in-session dispatches, so the guard sees each one. The workflow backend is a\n   different run, compiled for it — never switch to it mid-run.\n\n   **Record each builder's result** as it returns — the JSON it printed, saved to a file —\n   with its own worktree:\n\n   ```bash\n   python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Frecord_event.py\" candidate --run \u003CrunId> --phase build --result c2.json --tree .arbeitsplan\u002F\u003CrunId>\u002Fc2\n   ```\n\n   This writes `candidates\u002Fc2.json`, the file `reconcile.py` (step 8) and `land_candidate.py`\n   (step 9) read; nothing else writes it on this backend. `--tree` replaces the builder's\n   self-reported `diff` and `filesTouched` with what its worktree actually holds, and notes\n   whether they differed — builders paraphrase long diffs, and a paraphrased diff does not\n   apply. It refuses a phase that was never opened, a result missing any field of\n   `candidate-builder`'s output shape, and a candidate already recorded, writing nothing.\n\n5. **Apply the breaker, per batch, never cumulatively.**\n\n   - Exclude every `measured: false` candidate from the denominator. It is not a rejection;\n     it is a candidate that was never fairly tried.\n   - `measured == 0` → **acquisition problem.** The environment failed, not the contract. Fix\n     the environment and re-run. Do not re-dispatch into a broken environment.\n   - `accepted * 3 \u003C measured * 2` → **contract problem.** Halt and say so: *the correct\n     response is a better contract, not more agents.*\n\n6. **Referee blind.** One `arbeitsplan:candidate-referee` per surviving candidate, each given\n   the criteria and that candidate's diff only — never the builder's rationale, never another\n   candidate. Landing is an **allowlist**: only `accepted`.\n\n   Record the batch — every verdict the referees returned, as one JSON list — under the referee\n   phase you opened:\n\n   ```bash\n   python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Frecord_event.py\" referee --run \u003CrunId> --phase \u003CrefereePhase> --verdict verdicts.json\n   ```\n\n   This writes `referee\u002F\u003Cid>.json` (what `land_candidate.py` reads) **and** the\n   `referee\u002F\u003Cphase>\u002F\u003Cid>.json` twin that `rounds.py` rebuilds rounds from. Without it an\n   in-session run has no rounds, and step 7's `rounds.py decide` can never route a shared hole\n   or stop a moving residual. It refuses a phase that was never opened, a malformed verdict,\n   and a (phase, candidate) already recorded — and writes nothing when it refuses.\n\n7. **Select by rule**, not by preference: most criteria met, then fewest files touched, then\n   candidate id. If no candidate is `accepted`, **halt and surface**. All N failing the same\n   way is a statement about the contract — do not just widen and re-dispatch. Run\n   `scripts\u002Frounds.py record --run \u003CrunId>` (from the root holding `analysis\u002Farbeitsplan\u002F`) to\n   rebuild this run's rounds, then `scripts\u002Frounds.py decide --rounds \u003Cfile> --spec\n   workflow.json` (see [Rounds: a shared hole or a moving residual](#rounds-a-shared-hole-or-a-moving-residual-78-93)\n   below) to learn which of three things this halt is: `ROUTE CONTINUE` (nothing to say yet),\n   `ROUTE SYNTHESIZE criterion=\u003Cid>` (a **sharedHole**: relaunch at the single-writer phase with\n   `carry.sharedHole` set, rather than building N more candidates against the same hole), or\n   `ROUTE HALT moving-residual` (stop outright: the last several rounds each \"advanced\" on a\n   *different* blocker, which is not convergence).\n\n8. **Measure the winner against its own tree, before landing** (#76). A builder's self-reported\n   exit codes are never trusted on their own:\n\n   ```bash\n   python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Freconcile.py\" --run \u003CrunId> --run-checks --candidate c2 --tree .arbeitsplan\u002F\u003CrunId>\u002Fc2\n   ```\n\n   This RUNS every checked acceptance criterion now, with `cwd` set to `c2`'s OWN worktree, and\n   records each as an `execute_tool` event in `run.jsonl`, carrying `detail.candidate == c2`. Any\n   disagreement with what `c2` itself reported — in either direction — prints `CONTRADICTION` and\n   exits 1. `land_candidate.py --apply` (next step) **refuses** a candidate lacking this\n   measurement for any checked criterion, or whose latest measurement contradicts its report;\n   another candidate's measurement never unlocks it. An honestly-reported failure a referee\n   already accepted is not blocked here — the gate is \"unmeasured or contradicted\", nothing more.\n\n9. **Land exactly one diff.**\n\n   ```bash\n   python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Fworktree_pool.py\" open --spec ... --phase land\n   python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Fland_candidate.py\" --run \u003CrunId> --candidate c2 --apply\n   ```\n\n   Optional gated synthesis, only if the spec opted in: one further writer, given the winner\n   plus **explicitly named** elements to borrow. It must beat the plain winner on at least one\n   declared criterion, or the plain winner lands unchanged.\n\n10. **Delete the losers** — worktrees and branches — then record the phase as closed and close\n    the lock:\n\n    ```bash\n    python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Frecord_event.py\" phase --run \u003CrunId> --phase land --status closed\n    python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Fworktree_pool.py\" close\n    ```\n\n    `close` **refuses** a phase that recorded no terminal event. On a halt, close with\n    `--halt \"\u003Cthe specific reason>\"`: a halt is an event in `run.jsonl`, never an absence.\n\n11. **End the run in the record** -- landed or halted, always last:\n\n    ```bash\n    python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Frecord_event.py\" finish --run \u003CrunId>\n    ```\n\n    A landed run gets `complete.json` (refused while any phase is still open); a halted one\n    gets `FAILED-\u003Cstamp>.json` carrying the halt's reason; anything else is refused, because\n    stopping is not an ending. This marker is what `sweep_artifacts.py` waits for -- a run\n    without it is kept forever, whatever state its worktrees are in.\n\n    Deletion itself goes through `worktree_pool.py destroy` (or `sweep_artifacts.py` for a whole\n    finished run), which never just discards a loser's uncommitted state (#80): a DIRTY worktree\n    is committed and preserved on `kept\u002F\u003CrunId>-\u003Ccid>` before its worktree and throwaway\n    `arbeitsplan\u002F\u003CrunId>\u002F\u003Ccid>` branch are removed; a clean worktree gets no `kept\u002F` branch.\n\n## Referee-owned artifacts (#77)\n\nIf `workflow.json` declares `refereeOwned`, its `referee-fixture` phase runs first, exactly\nlike any other phase (open the lock, dispatch its `agentType`, record the output, close). Once\nit has written those paths, baseline them **once**:\n\n```bash\npython3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Freferee_owned.py\" record --run \u003CrunId>\n```\n\nA second `record` for the same run is refused: the baseline is taken at creation, never\nre-taken to accommodate a later change. `worktree_pool.py open` already narrows a fan-out\nphase's lock to exclude every `refereeOwned` path, so the guard denies a candidate's Edit\u002FWrite\nbefore it lands — but that covers only the tool calls the guard's matcher sees. Re-verify\nbefore landing (step 9), as the belt to that guard's suspenders:\n\n```bash\npython3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Freferee_owned.py\" verify --run \u003CrunId>\n```\n\n`land_candidate.py` independently refuses (citing `refereeOwned` by name) any candidate diff\nthat touches one of these paths, so the same rule is checked three ways: at the lock, at\nlanding, and by content.\n\n## Rounds: a shared hole or a moving residual (#78, #93)\n\nA no-accept halt is arithmetic (`accepted * den \u003C measured * num`) and never looks at WHY.\n`scripts\u002Frounds.py` is the one place that does, and both rules below read the SAME derived\nrounds — one round per referee phase, rebuilt from `referee\u002F\u003Cphase>\u002F\u003Cid>.json` and the\nadjudicator's recorded `round: {outcome, blocking}` — so a fix to one is a fix to both. Those\nfiles come from `record_event.py workflow` (the workflow backend) or `record_event.py referee`\n(in-session, step 6); a batch recorded any other way is invisible here:\n\n```bash\npython3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Frounds.py\" record --run \u003CrunId> > \u002Ftmp\u002Frounds.json\npython3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Frounds.py\" decide --rounds \u002Ftmp\u002Frounds.json --spec analysis\u002Farbeitsplan\u002F\u003CrunId>\u002Fworkflow.json\n```\n\n- **`ROUTE SYNTHESIZE criterion=\u003Cid>` — a `sharedHole` (#78).** The LATEST round accepted no\n  one and >= 2 rejected candidates share one unmet criterion — a structural hole every\n  candidate hit, not noise. Relaunch `workflows\u002Frun.js` with `startAt` at the single-writer\n  phase and `carry.sharedHole = {criterion, candidates: [{candidateId, diff}]}` (the run.js\n  halt's own `rejectionsByCriterion`, paired with each candidate's diff, builds this). The\n  synthesizer then works from every rejected diff toward that one criterion, instead of N fresh\n  candidates re-discovering the same hole from scratch.\n- **`ROUTE HALT moving-residual` (#93).** A per-batch breaker and a `sharedHole` both look at\n  ONE round; a run that \"advances, not closes\" every round, each time naming a *different*\n  blocking condition, passes both forever. This fires when the last N **judged** rounds\n  (`judge.outcome != \"none\"`; an unjudged round neither breaks nor counts) are all `\"advanced\"`\n  with non-null, pairwise-**distinct** `judge.blocking` ids — N genuinely different obstacles in\n  a row, not real convergence. Stop outright: **do not** re-compile and **do not** widen again.\n  `N` is `--max-advancing-rounds`, else the spec's `roundBreaker.maxAdvancingRounds`, else 3.\n  When both rules would fire on the same rounds, the `moving-residual` HALT wins.\n- **`ROUTE CONTINUE`** — neither condition holds; proceed as normal.\n\n## Stacked fan-outs across waves (#79)\n\nA later wave's builders should sometimes start from an **earlier wave's refereed winner**,\nnot from HEAD — that is what a `fanout-redundant` phase's `base: \"\u003CphaseId>\"` declares. Once a\nphase's candidate is selected (step 7 above) and would normally just land (step 9), promote it\ninstead if a later phase names it as `base`:\n\n```bash\npython3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Fworktree_pool.py\" promote --run \u003CrunId> --phase build-w1 --candidate c2\n```\n\nThis commits everything sitting in that candidate's worktree — untracked files included — and\npoints `arbeitsplan\u002F\u003CrunId>\u002Fbase\u002Fbuild-w1` at the new commit. The next wave's `create` then\nreads it automatically:\n\n```bash\npython3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Fworktree_pool.py\" create --spec analysis\u002Farbeitsplan\u002F\u003CrunId>\u002Fworkflow.json --phase build-w2 --count 2\n```\n\nIf `build-w2` declares `base: \"build-w1\"`, every worktree this creates starts from the promoted\nbranch instead of HEAD; if `build-w1` was never promoted, `create` **refuses**, naming the\nmissing branch, and creates nothing. A `--phase` whose phase carries no `base` (or `create` with\nno `--phase` at all) behaves exactly as before: from HEAD.\n\nBase branches deliberately **survive** an ordinary `worktree_pool.py destroy --run \u003CrunId>` — a\nstill-pending later wave may need to stack on one. Only pass `--bases` once the whole run is\nactually done with them:\n\n```bash\npython3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Fworktree_pool.py\" destroy --run \u003CrunId> --bases\n```\n\n`compile_spec.py --strict` is the check that a stacked run actually declared every `base` it\nneeded: a `fanout-redundant` phase whose transitive `requires` reach another `fanout-redundant`\nphase with no `base` chain reaching it back prints `WARNING ... [AP-SIBLING-INVISIBLE] ...`\nnaming both phases. A plain compile still writes past a warning (exit 0); `--strict` treats one\nas a rejection (exit 1) — run compilation with `--strict` before dispatching a stacked run.\n\n## The workflow backend\n\nWhen `backend.kind` is `\"workflow\"`, the whole phase graph runs inside `workflows\u002Frun.js`,\nand this session does only the things the Workflow tool cannot: read files, run plan-mode\nphases, measure the accepted candidate with `reconcile.py --run-checks` (#76), and land.\n\n1. **Launch it with the spec as data.** The Workflow tool has no filesystem, so pass the\n   parsed `workflow.json` verbatim: `args: {spec}` (plus `startAt` and `carry` on a resume).\n   No run-scope lock is open while it runs — nothing inside writes the shared tree.\n1b. **Arm the script allowlist when the spec has a `script` phase (#107).** `compile_spec.py\n   --write` already wrote `analysis\u002Farbeitsplan\u002F\u003CrunId>\u002Fscripts.json` from the spec; point the\n   guard at it before launching:\n\n   ```bash\n   printf '{\"runId\": \"%s\"}\\n' \u003CrunId> > analysis\u002Farbeitsplan\u002Fscripts_armed.json\n   ```\n\n   The guard judges every Bash call `arbeitsplan:script-runner` makes — lock or no lock — and\n   allows exactly one declared command per dispatch. Without an armed allowlist the runner is\n   denied everything, which is the safe direction. For `{placeholder}` slots, pass their values\n   as `carry.scriptArgs.\u003CphaseId>`; a value outside `[A-Za-z0-9._\u002F,=:@+-]` halts the phase\n   before dispatch.\n2. **Persist what it returns, before reading it.** Every return carries `events`:\n\n   ```bash\n   python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Frecord_event.py\" workflow --run \u003CrunId> --result result.json\n   ```\n\n   A halt is an event. Record it even — especially — when the run aborted.\n3. **On `pending_plan_node`, run that phase here, in plan mode**, by dispatching its\n   `agentType` with the carried data. Record its output, add it to `carry` under the phase id\n   (`carry.contract.acceptance` replaces the compiled acceptance for later phases), and\n   relaunch with `startAt: \u003CresumeWith>`. Loop state lives here, never in the script.\n   That same command writes `candidates\u002F\u003Cid>.json` and `referee\u002F\u003Cid>.json` for every\n   candidate the return carries, once each, never overwriting.\n3b. **Record a plan-mode phase's output** the same way:\n   `record_event.py phase-output --run \u003CrunId> --phase \u003Cid> --output \u003Cfile>`. An adjudicator's\n   `verdict: \"land\"` is what gives a borrowed synthesis its referee record; without it,\n   `land_candidate.py` refuses the synthesis like any unjudged candidate.\n3c. **On a NO CANDIDATE ACCEPTED halt, `rounds.py` decides the next move, not this session.**\n   The halt's `rejectionsByCriterion` is already grouped by criterion; run `scripts\u002Frounds.py\n   record --run \u003CrunId>` then `scripts\u002Frounds.py decide --rounds \u003Cfile> --spec workflow.json`\n   (see [Rounds: a shared hole or a moving residual](#rounds-a-shared-hole-or-a-moving-residual-78-93)\n   above — same rules, same `sharedHole` and `moving-residual` outcomes, whichever backend ran\n   the halted batch). A `SYNTHESIZE` route relaunches `run.js` with `startAt` at the\n   single-writer phase and `carry.sharedHole` set, from the same halt's `rejectionsByCriterion`\n   and `candidates`; a `moving-residual` HALT stops the run outright.\n4. **On completion, measure then land in-session.** `worktree_pool.py open --phase \u003Clast>`, then\n\n   ```bash\n   python3 -B \"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Freconcile.py\" --run \u003CrunId> --run-checks --candidate \u003Cid> --tree .arbeitsplan\u002F\u003CrunId>\u002F\u003Cid>\n   ```\n\n   to measure every checked criterion against the accepted candidate's own tree (#76), then\n   `land_candidate.py`, where the hook, the `writeScope` check, and this measurement gate all\n   run. The workflow never lands. `land_candidate.py` writes `landed.json`, with `divergedFrom`\n   if what landed differs from the recorded candidate. Then end the run with\n   `record_event.py finish --run \u003CrunId>` -- after a halt too -- exactly as step 11 above.\n\n## Rules\n\n- **Never re-dispatch an identical prompt.** The hook denies it, and the denial is correct:\n  past the cap those rounds do not converge, they only cost. Widen instead — a new angle\n  nobody tried — or stop.\n- **Never let a builder write to the shared tree.** Only the landing step touches it.\n- **Never treat an unmeasured candidate as a failure.** It has no rate, only missing data.\n- **Never raise the budget mid-run.** That is a decision to re-compile.\n- **Delegation nests at most 3 deep, and never in a cycle** — both enforced by the hook, see\n  `references\u002Fdelegation.md`. A dispatch to arbeitsplan's own agents is fan-out and is not\n  counted; a dispatch to another plugin is.\n- **A denial from another plugin's guard is that plugin doing its job.** Report it; never\n  reach for its escape hatch.\n\n## Output format\n\n```\narbeitsplan run — ap-2026-09-12-a3f1, phase build\n  c1  middleware-layer       measured   checks 3\u002F3   diff 2 files\n  c2  decorator-per-route    measured   checks 3\u002F3   diff 2 files\n  c3  reverse-proxy-config   UNMEASURED no nginx toolchain in the worktree\n\n  breaker  accepted 2 \u002F measured 2  (c3 excluded from the denominator) — pass\n\nphase referee (blind)\n  c1  rejected   a3 not met: requirements.txt gains 'slowapi'\n  c2  accepted   3\u002F3, evidence at src\u002Fapi\u002Fsearch.py:41\n\nphase land\n  winner   c2 (3 criteria met, 2 files touched)\n  applied  src\u002Fapi\u002Fsearch.py, tests\u002Ftest_ratelimit.py\n  deleted  c1, c3 worktrees and branches — nothing merged\n  marker   .takt\u002Fap-2026-09-12-a3f1\u002Flanded\n\ndispatches 7\u002F9 budget.\n```\n\nA halt is an output, not a crash:\n\n```\narbeitsplan run — ap-2026-09-12-a3f1 HALTED at phase referee\n  accepted 0 \u002F measured 3 — every candidate failed criterion a1 the same way.\n  All three read the limit from a per-process dict, so a1 (\"429 across workers\")\n  cannot be met inside the declared writeScope at all.\n  This is a statement about the contract, not about the candidates.\n  The correct response is a better contract, not more agents: re-compile with a\n  scope that admits shared state, or drop a1.\n  Nothing was landed. Worktrees kept at .arbeitsplan\u002Fap-2026-09-12-a3f1\u002F for reading.\n```\n\n## Resources\n\n- `references\u002Fcandidate-contract.md` — what builders and referees exchange, and why the\n  referee is starved.\n- `references\u002Fdelegation.md` — read before dispatching another plugin: the registry every\n  plugin declares, the append-only ledger, and the depth cap plus cycle detection that bound\n  how far a delegation may nest.\n- `references\u002Fpatterns.md` — the breaker's thresholds and the reason they are per-batch.\n- `workflows\u002Frun.js` — the workflow backend: executes `spec.phases`, halts before each\n  plan-mode phase, counts the dispatch budget in code, and returns span-shaped `events`.\n- `references\u002Fbackend-selection.md` — why a run was compiled for the backend it names.\n",{"name":5,"description":7,"argument-hint":24},"[runId]","main","6158c1e4e112c06af2c7ac01b13b2ed7214c0fcf","plugins\u002Farbeitsplan\u002Fskills\u002Farbeitsplan-run",0,"https:\u002F\u002Fgithub.com\u002FAnselmoo\u002Fwerkstoff","https:\u002F\u002Fanselmoo.github.io\u002Fwerkstoff\u002F","https:\u002F\u002Fgithub.com\u002FAnselmoo","Likes to Develop","2026-11-02 04:20:59",{"source":8,"skillName":5,"commands":35,"prompt":45,"zipAvailable":46},[36,39,42],{"tool":37,"command":38},"npx","npx skills add Anselmoo\u002Fwerkstoff --skill arbeitsplan-run",{"tool":40,"command":41},"bunx","bunx skills add Anselmoo\u002Fwerkstoff --skill arbeitsplan-run",{"tool":43,"command":44},"pnpm","pnpm dlx skills add Anselmoo\u002Fwerkstoff --skill arbeitsplan-run","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FAnselmoo\u002Fwerkstoff\u002Fblob\u002Fmain\u002Fplugins\u002Farbeitsplan\u002Fskills\u002Farbeitsplan-run\u002FSKILL.md",true,[],[49,56,63,68,73,78],{"skillNo":50,"name":51,"nameZh":6,"description":52,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":53,"categories":54,"keywords":55,"publishedAt":19},"01M79AJTRBNBX03WWKYJ8ENM45","toy-check","A skill that checks things in the repository and reports what it finds.","https:\u002F\u002Fgithub.com\u002FAnselmoo\u002Fwerkstoff\u002Fblob\u002Fmain\u002Ftest\u002Fworkflows\u002Ffixtures\u002Ftoy-plugin\u002Fskills\u002Ftoy-check\u002FSKILL.md",[],[],{"skillNo":57,"name":58,"nameZh":6,"description":59,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":60,"categories":61,"keywords":62,"publishedAt":19},"01M79AJS292F3HTDRXWZVRSN5B","demo-check","Reports which generated files changed since the last generation run. Use when the user asks what drifted. Read-only; never edits a file.","https:\u002F\u002Fgithub.com\u002FAnselmoo\u002Fwerkstoff\u002Fblob\u002Fmain\u002Ftest\u002Fplugins\u002Ffixtures\u002Fnacharbeit-uncompleted-run\u002Fplugins\u002Fdemo\u002Fskills\u002Fdemo-check\u002FSKILL.md",[],[],{"skillNo":64,"name":58,"nameZh":6,"description":59,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":65,"categories":66,"keywords":67,"publishedAt":19},"01M79AJP4TFQXKT1AAK1WC1JYK","https:\u002F\u002Fgithub.com\u002FAnselmoo\u002Fwerkstoff\u002Fblob\u002Fmain\u002Ftest\u002Fplugins\u002Ffixtures\u002Fnacharbeit-stale-lock\u002Fplugins\u002Fdemo\u002Fskills\u002Fdemo-check\u002FSKILL.md",[],[],{"skillNo":69,"name":58,"nameZh":6,"description":59,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":70,"categories":71,"keywords":72,"publishedAt":19},"01M79AJM6Y50WP3G2BR4H0M60Z","https:\u002F\u002Fgithub.com\u002FAnselmoo\u002Fwerkstoff\u002Fblob\u002Fmain\u002Ftest\u002Fplugins\u002Ffixtures\u002Fnacharbeit-preflight-mixed\u002Fplugins\u002Fdemo\u002Fskills\u002Fdemo-check\u002FSKILL.md",[],[],{"skillNo":74,"name":58,"nameZh":6,"description":59,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":75,"categories":76,"keywords":77,"publishedAt":19},"01M79AJGR6DMT8KCCK071V78XM","https:\u002F\u002Fgithub.com\u002FAnselmoo\u002Fwerkstoff\u002Fblob\u002Fmain\u002Ftest\u002Fplugins\u002Ffixtures\u002Fnacharbeit-hooks-shape\u002Fplugins\u002Fdemo\u002Fskills\u002Fdemo-check\u002FSKILL.md",[],[],{"skillNo":79,"name":80,"nameZh":6,"description":81,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":82,"categories":83,"keywords":84,"publishedAt":19},"01M79AJBJ4CHC4PZVWKERWF3HF","zirkel-verify-assumptions","Checks exactly ONE named assumption against real evidence in at most 3 steps and reports one outcome: confidence raised (>=90), reading changed, or still unresolved. Use to resolve a single load-bearing uncertainty — typically a blocking entry surfaced by zirkel-clarify-scope: \"verify this one assumption\", \"is it actually true that X\", \"check this before we rely on it\". For multiple uncertainties, invoke once per uncertainty — the step budget is never shared.","https:\u002F\u002Fgithub.com\u002FAnselmoo\u002Fwerkstoff\u002Fblob\u002Fmain\u002Fplugins\u002Fzirkel\u002Fskills\u002Fzirkel-verify-assumptions\u002FSKILL.md",[],[]]