This page records what this marketplace took from, or rejected in, the pstack/ collection of
cursor/plugins (MIT). Provenance lives here and in plugin
CHANGELOGs, never in skill bodies, where it is noise to the agent. A citation an agent uses while it
works is not a provenance record and stays where it is. The Matt Pocock record is
mattpocock-skills.md; the Conflict column below checks each row against
the skills that record lists.
Last audited upstream state: cursor/plugins@bf0c7cd80abdbf7fe47b634978a7d1b4071f71a7 under pstack/ (pstack 0.15.42)
Between 88d47514b8 (pstack 0.15.29) and ae5650febf, 23 upstream commits touched pstack/. They
added statechart and ketchup-video, rewrote interrogate and deleted its reference files, cut
autopilot-stack to a variant of autopilot-full, and rewrote orchestrate and deleted its
scripts/orch state CLI. Each changed row says what was re-checked at ae5650febf. Between
ae5650febf and ec56937961 no pstack/ path changed (the upstream compare lists only the
marketplace manifest, README.md and third_party/ files). Rows marked Decided 2026-11-12 record
the dispositions of map decision
#15661, made at
ec56937961. Between ec56937961 and bf0c7cd80a one upstream commit touched pstack/
(cursor/plugins#1236): it cut why to one paragraph
and deleted all its reference files, cut how to one subagent and deleted both its prompts, replaced
every other skill's call to how with an inline trace, changed the default models and the
setup-pstack migration of old defaults, reworded sharpen's last sentence, and fixed
statechart's Mermaid renderer with new tests. Each row that commit touched says what was re-checked
at bf0c7cd80a. The statechart and benny changes touch no row; benny only drops how from
its dependency list, and statechart's renderer and tests were listed, not read, because its
adoption (#15648) reads the
skill at its then-current head. A row whose latest stamp is older covers paths that did not change
since that stamp's commit, so that check stands at this pin.
Recheck trigger: a change, between the pinned commit and upstream HEAD, to a path a row we took
or rejected something from links, a unit removed or added under the scope: re-audit the affected
rows. Row links point at whole skill directories where a verdict rests on more than SKILL.md, so
the trigger reaches references/, playbooks and scripts/. scripts/check-upstream-drift.sh
computes the changed paths.
Adaptation posture. Every entry is a reauthor, not a fork and not a vendored copy. We keep the
substance where it earns its place, fit it to this marketplace's conventions, and write the prose
ourselves. The playbooks plugin's boris pack is the deliberate contrast: it keeps an upstream copy
so drift can be detected against it.
Upstream targets Cursor, so three kinds of machinery are dropped from every row without a note:
hard-coded model identifiers and model-role tables, Cursor's MCP discovery, and Cursor transcript
paths under ~/.cursor/. Cursor-only product surfaces (its PR forge, cloud agents, review bot and
runtime UI) are dropped the same way.
Decision cells give our decision and the outcome of the re-check at the pin. pending adoption:
names the skill a decided change lands in and links the issue that tracks it; the row is updated when
it lands. A conflict means the
upstream unit and a Pocock-derived skill of ours prescribe incompatible behavior for the same
situation; overlap alone is not a conflict. A pr-pipeline: note records where a PR-workflow unit
disagrees with this repository's PR pipeline; nothing there
is adopted by this record.
| Upstream (links at the pin) | Ours | Relation | Decision | Conflict with a Pocock-derived skill |
|---|---|---|---|---|
why |
discovery:trace-intent and its evidence categories |
Derived | We take: intent is recovered from records outside the code; one investigation per evidence category; every run ends with a coverage map that also records the categories where nothing turned up; only two skip reasons are valid; direct evidence, inference, competing hypotheses and gaps are reported apart; the five tier names, renamed the intent-evidence tier because the axis measures distance from a stated intent, not certainty; the Preserve/Change/Avoid/Risk handoff. We add a per-citation source-reliability note. We reject the seven-category roster (four categories have no surface here), the vendor-named source playbooks, the separate synthesizer agent, and code-shape inference (below). Re-checked at 23e4138: holds. Upstream thinned SKILL.md and moved its output contract into a reference file; the reference files changed punctuation only. Taken later: when the target only matters on a failure path (a branch for one error code, a delay or cap with no derivation nearby, a check the type already excludes), the tracker search adds incident and follow-up tickets from the month before the introducing commit, and the long-form search adds postmortems, queried by error text or constant value as well as by symbol (plugins/discovery/skills/trace-intent/context/evidence-categories.md:20-30). Re-checked at 118e73b0e1: holds. Upstream cut SKILL.md, epistemics.md, the investigator and synthesizer prompts and the source playbooks to shorter restatements and changed the synthesizer's default model; every rule we take remains, and the investigator prompt still bars reading intent from code behavior or style. Re-checked at 88d47514b8: holds; the only change is a line deferring each subagent's reasoning level and speed to poteto-mode's model rule, machinery we drop. Re-checked at bf0c7cd80a: changed. Upstream cut SKILL.md to one paragraph and deleted every reference file: the category roster and its MCP discovery, the investigator and synthesizer prompts, epistemics.md and the source playbooks are gone. Already have: every rule we took is now ours alone, stated in discovery:trace-intent (tiers plugins/discovery/skills/trace-intent/SKILL.md:96-109, categories and skips :136-171, Sources consulted :184-186). The new paragraph answers from the record, not the code, cites what the authors wrote, marks inferences and says what the record does not answer, which our tiers and Sources consulted line already require; it now agrees with our departure (see below). Reject: reading the commits and pull requests first and other sources only when that trail is thin. Our skill decides no source is empty in advance, because an empty category costs one search and a missed design document costs a wrong answer (plugins/discovery/skills/trace-intent/SKILL.md:169-171), and it investigates only three categories here (:138-145). |
none |
recall |
session-flow:recall; see below |
Adopted (new skill) | The earlier omission is reversed: session-flow:recall rebuilds a topic's history from this repository's past transcripts, condensed per session by subagents, adds reverted commits, closed unmerged pull requests and open tracker items, and states the scope it searched, never reporting a capped scan as complete (plugins/session-flow/skills/recall/SKILL.md:20-21,73,89,110). It reads the transcript_scope setting it shares with session-flow:find-handoff, the narrowest layer winning (plugins/session-flow/skills/recall/SKILL.md:37-52). Still rejected: the closed status-tag set (the worktree status vocabulary already owns one) and the cut-detail rule (session-flow:recommend-skills owns it). Re-checked at 23e4138: holds; upstream removed its trigger phrases and trimmed rationale sentences, neither of which we took. Re-checked at 88d47514b8: holds; the only change is the subagent model-rule line, machinery we drop. Decided 2026-11-12: the sweep of incidents and user reports is taken inside session-flow:recall itself, not by calling discovery:trace-intent, which writes a file and tiers evidence while recall writes nothing (plugins/session-flow/skills/recall/SKILL.md:144). Recall's own tracker step extends to open and closed items with a "Still reported" list, plus a search of tracked documents for postmortems (pending adoption: session-flow:recall, #15839). Re-checked at bf0c7cd80a: changed. Already have: the shared-record sweep no longer hands off to why's investigators; it spawns one investigator per source itself and asks what is current, what was tried and did not hold, and what users still report. That is the 2026-11-12 decision above, so upstream now matches it and #15839 is unchanged. |
none |
teach |
Two rules in education:teach's lesson contract; education:explain and its starting-rung setting; see below |
Absorbed (skill omitted) | We take the incremental diagram sequence for multi-part systems and the rule that an inventory of identifiers does not count as a lesson. Re-checked at 23e4138: changed (corrections and reversals only); both absorbed rules hold. Upstream removed its trigger phrases and nine sentences, several of them restating rules we absorbed, and split long sentences; neither absorbed rule was removed. The record's claim that we have no how was wrong: discovery:explore and discovery:trace-intent are that pair. Taken later, all in education:explain: for a subsystem of the current codebase it invokes /discovery:explore to describe the code's behavior and /discovery:trace-intent for the reasons behind it, when both are available, and keeps trace-intent's hedged wording (plugins/education/skills/explain/SKILL.md:45-53); its register rule admits no reader quiz, no role tag in front of a sentence and no remark on the explanation's own tempo (:128-134), which reverses the earlier rejection in explain only; and explain_starting_rung (plain by default, or peer) sets where it starts (plugins/education/reference/config.md:36). Still declined: the image-generation branch (no image-generation route in our visualization skills, so not adopted) and the voice paragraph (owned by docs-hygiene:write-for-humans). Re-checked at bf0c7cd80a: holds; upstream dropped the advice to keep why narrow because its full sweep is slow, which follows from why's cut (see its row). Neither absorbed rule changed. |
education:teach: upstream forbids quizzes and asking the learner to restate; ours coaches with questions and quizzes. decided: ours, because the anti-pacing rule lands only in the one-shot education:explain and teach stays a multi-session tutor. Starting altitude: the explain_starting_rung setting in education. |
tdd |
testing:write's decline list and debugging:debug phase 5; implementation:implement's fix steps; see below |
Absorbed (skill omitted) | We take the impracticality triggers, preferring to add no test rather than a weak one, and checking that a red test goes red because of the bug itself. The skill stays omitted: it runs only on explicit request. Re-checked at 23e4138: changed: upstream dropped its nearby-validation step, its warning against skipping the regression step, and its line against weak-signal tests; our requirement to name the trigger and the substitute check is now ours alone and stays. The rejection of the skip warning is moot. Taken later: debugging:debug puts a redacted failing loop run from before the fix and a passing one from after it into the user report and the PR description (plugins/debugging/skills/debug/SKILL.md:269), which reverses the earlier marginal verdict; and implementation:implement's fix steps commit the reported case alone, with each sibling in its own later commit and test (plugins/implementation/skills/implement/context/bugfix.md:13-14). |
testing:write: upstream runs test-first only on request or with a cheap target; ours is test-first under the project's declared cadence. decided: ours, because the cadence is already the project's setting and the impracticality branch covers upstream's skip case. |
reflect |
One routing rule in running-retro and retro, and the retro_lenses setting; see below |
Absorbed (skill omitted) | We take routing an accepted new-skill candidate to /playbooks:skill-authoring, gated on /skill-quality:check, in both retro skills. Still rejected: the skill (the retro skills own the axis), the structural-enforcement step (harness-config:audit-automation-gaps owns it), auto-filed backlog items, the missed-trigger finding class (discipline:use-your-skills owns it), and per-lens vendor pins. Re-checked at 23e4138: holds; upstream now fires only on explicit request, which supports the omission. The three-lens rejection is reversed and taken: retro_lenses in session-flow (1 by default, or 3) runs session mode's analysis as one pass or as three lens subagents side by side, merged into one list before scoring (plugins/session-flow/skills/retro/context/session.md:128-166, plugins/session-flow/reference/config.md:17). session-flow:retro also keeps a finding only when it is durable, specific and decision-changing (plugins/session-flow/skills/retro/context/session.md:226-233) and tries a placement fix before writing a new rule (plugins/session-flow/skills/retro/context/session.md:318-322). Re-checked at 118e73b0e1: holds. Upstream merged its steps into numbered sections, shortened the three lens prompts and changed the lens models; the structural-enforcement step and the automatic backlog filing we reject remain. Re-checked at 88d47514b8: holds; the only change is the subagent model-rule line, machinery we drop. Re-checked at ae5650febf: holds; the one addition sends approved skill edits to a writable copy instead of an installed plugin's cache, on Cursor's paths. Nothing taken: both retro skills hand a skill candidate to /playbooks:skill-authoring, and a search of both for plugins/cache finds nothing. Re-checked at bf0c7cd80a: holds; the tooling lens's default model and one example in the synthesizer's drop list changed, neither of which we took. |
none |
arena |
No skill. Four ideas in architecture:improve Design-It-Twice, prototype's shared discipline and naming:name-it-better tournament |
Absorbed (skill omitted; see below) | We take: a rejected-alternatives field in the Design-It-Twice result; a graft ledger that names what a hybrid took from each design and what it left, with the reason; reading the spread of orthogonal candidates, where only convergence or diverging assumptions carry signal; and fixing the rubric before generation in naming's tournament. We reject the runner skill (no second consumer) and a hidden rubric (naming's rubric is the project's public standard). Taken later: competing attempts as a setting in implementation:implement-dispatch "Competing attempts" (plugins/implementation/skills/implement-dispatch/SKILL.md:230-264). For a phase planning:plan marks multi-shape (plugins/planning/skills/plan/SKILL.md "Step 4.5: Execution-Shape Analysis", the Attempts cell; line hint :295), each attempt gets its own local worktree, its own stated constraint and its own phase-verifier; the plan writes the selection rule before dispatch; the chosen attempt lands by fast-forward. The settings are competing_attempts (off, suggest default, auto) and competing_attempt_count (2 to 5, default 3). We still reject a runner skill. Landed: a fifth idea in the same Design-It-Twice file, a comparison note the parent writes in the dispatch turn, before any return arrives, naming which of the three comparison axes counts most for this candidate and what result would settle the pick, and kept out of every brief (plugins/architecture/skills/improve/research/deepening/interface-design.md:51-55). It is the same fixed-before-generation move as naming's rubric, applied to design comparison; the hidden-rubric rejection stands for naming, whose rubric is public. Re-checked at 23e4138: holds; the changes touch rationale and model selection, neither of which we took. Re-checked at 118e73b0e1: holds. Upstream folded the phase list into one line, cut its default runners to two models and moved the graft and rejection record into the closing synthesis note; the four absorbed ideas remain. Re-checked at 88d47514b8: holds; the runner step gained the subagent model-rule line and a family fallback for cursor-grok-* and composer-* slugs, both model machinery we drop. Re-checked at ae5650febf: holds; the changes are runner and judge fallbacks by model family, including running with no judge, which is model machinery we drop. The four absorbed ideas are unchanged. Re-checked at bf0c7cd80a: holds; the default runners and judge pool changed and the judge rule now groups model families into vendors, all model machinery we drop. |
none |
technical-writing |
docs-hygiene:write-for-humans |
Derived, re-postured | We take the four-layer model (mode, address, load, ambiguity), the Diátaxis mode picker, the three rules above the layers, the rhythm-and-specificity section, the layer rule sets and a review checklist. Re-postured: these apply as a named, replaceable default only after a search for the consuming project's own guide finds none, because a house style shipped as the default conflicts in some repository (docs/plugin-philosophy.md). We reject the commit and PR scope (other skills own their shape), tab indentation (a repository's linter decides), upstream's worked example (it names a file that does not exist in a consumer's tree), and editing another plugin's tell catalog (a private, attribution-bound surface). Re-checked at 23e4138: holds. Upstream removed its review checklist, its source stamps and two prompts, so ours now carry them alone; its catalog instruction became a proposed diff, which still targets a private surface. Upstream's voice guidance now lives here, not in unslop. Own words (2026-10-04): the skill body and reference/sentence-rules.md were restated from the rules alone, in our own structure (a mode table by voice, opening, body and exclusions; a table of checks for rhythm; address rules grouped by the part of the page they govern; load rules in one table ordered from paragraph to sentence to verb to word; the first-read goal and the instruction to drop a rule that makes a sentence worse moved into the after-writing checks) and with our own examples (a made-up backup tool throughout the sentence rules, a webhook receiver in the tutorial row, a worker-reload passage with its fixes listed in sentence order as the worked example). A 5-word scan of those two files against e43c7ee26e, 23e4138 and 73f8be4873 finds one shared run at all three: "like a machine wrote it", in the description's trigger phrase "this doc reads like a machine wrote it", a user's wording kept so the skill fires on it. Against 73f8be4873 only, the title The Global English Style Guide, which both texts cite as a source, is also shared. The skill names no upstream; this row is its only credit. Re-checked at 118e73b0e1: holds; upstream shortened sentences without changing a rule. |
none |
blast-radius |
review:quality-gate downstream and its downstream_probe setting |
Partial (scope adopted, mechanics rejected) | We take the outward scope the diff-scoped review lane lacked: a caller list alone is not the deliverable; the surfaces a grep cannot see (library source and pinned version, serialization, timing and lifecycle, flag reach, cross-language readers); confirmed risks reported apart from cleared ones; an empty search counted as an answer; a handoff to the cheapest test. We reject the five-rung proof ladder (it would be a ninth evidence ladder here; its unverified-claim floor is kept under playbooks:fable-5's verification chapter), a single safety fact as the report's spine (demoted to a first probe), a third severity axis, and routing to arena. Taken later: downstream_probe in review, where run (the default) writes the safety-fact probe in a temporary directory outside the tree and runs it, report states it without running it, and report from any layer wins (plugins/review/reference/config.md:12). Re-checked at 23e4138: holds; upstream removed emphasis lines only. Re-checked at 118e73b0e1: holds; one sentence changed in the step that routes to arena, which we reject. Re-checked at bf0c7cd80a: holds; two steps that pointed at why now state its rule inline (pull the PR and commits; cite a real file:line, and an empty search is still an answer). The scope we took is unchanged. |
none |
unslop |
ai-slop:audit's tell catalog, section General-prose additions and reference/rewrite-guide.md |
Absorbed (skill omitted) | We take seven patterns the Wikipedia-derived inventory lacked (three script rules, four rubric tells), the plain-word trio into the AI-vocabulary default (density-gated), and a fix-time guide with plain rewrites, a substitution guardrail and a closing self-audit. Every other upstream pattern maps to an existing entry, recorded as catalogued, not covered. We reject name-dropping for general prose, the general form of generic conclusions, a script rule for abstract metaphor nouns (rubric only, after calibration), and an always-apply posture (the audit stays triggered and read-only by default). Re-checked at 23e4138: changed: upstream removed its voice section, its self-audit step and five rules, so our self-audit and voice passes are now ours alone, and the voice pass cites no upstream (credit stays on this page). We reject the voice list's planted-imperfection item: it changes how text scores, not how it reads. The name-dropping rejection is moot. Own words (2026-10-04): the rewrite guide's plain-speech, replacement and voice sections and the catalog's examples were restated in our own structure and examples, against e43c7ee26e and 73f8be4873 (the last commit with the voice list); the shared 5-word runs left are quoted tell phrases the detector matches and upstream pattern names in the catalog's overlap map. The README, catalog and guide link neither upstream nor this page (credit lives here only), and no plugin file names this skill: the catalog section holding these tells is "General-prose additions", the detector's tests for its three script rules are named for those rules, and the catalog's model-era record runs its absence check against the Wikipedia page only. Adopted (2026-10-04): ai-slop:audit gains an over-compression tell, with the compress_articles setting in docs-hygiene:compress (articles kept by default), and a rubric tell for a figure of speech standing in for a fact, both written in our own structure and examples against e43c7ee26e. This is the first re-check of this row at a pinned commit. Re-checked at 118e73b0e1: holds. Upstream dropped "Must always apply." from its description, so it now matches our rejection of an always-apply posture; it folded its process steps into one sentence and moved each abstract metaphor noun's plain replacement beside the noun, with the same replacements. |
none |
show-me-your-work |
Four rules in implement-dispatch's deviation log; the formula-injection guard in install_state.py; see below |
Absorbed (skill omitted) | We take into DEVIATIONS.md: append and supersede, never edit; evidence as a re-runnable pointer; an explicit outcome or unverified; one entry per decision. We take the spreadsheet formula guard into audit-install-state's CSV writer, stricter than upstream (it quotes, and treats a leading tab, CR or LF as a formula lead). We reject the skill and a shared decision-log convention (one real consumer), the per-reply Attention section, and a second log-versus-transcript self-audit. Re-checked at 23e4138: changed: upstream's end-of-run audit now supersedes wrong rows, which agrees with ours; the formula guard is unchanged. The cross-model trail-review rejection is reversed and taken: implementation:implement-dispatch runs a fresh-context review of the trail at the end of every autonomous run, by default, and never in an interactive run (plugins/implementation/skills/implement-dispatch/SKILL.md:228-235). A resumed implementation:implement-dispatch run's first DEVIATIONS.md entry is a run-boundary entry naming the session and the earlier entries it did not write, and it writes a new one when another session appended in between (plugins/implementation/skills/implement-dispatch/SKILL.md:223, type list :224). Re-checked at 88d47514b8: holds; the only change is the subagent model-rule line in the cross-model review step, machinery we drop. Corrected 2026-11-12: this row said upstream dropped the one-decision-per-entry rule; it did not, its log is still one row per decision (pstack/skills/show-me-your-work/SKILL.md:13), so that rule is shared, not ours alone (plugins/implementation/skills/implement-dispatch/SKILL.md:222). The implement-dispatch citations above are re-pointed to the lines that now hold them. |
none |
create-verification-skill, maintain-verification-skill, principle-prove-it-works |
testing:map-features and testing:refresh-feature-map (retired; app-control:map in the two rows below replaces them), verification:confirm, testing:run-e2e |
Adopted (two new skills, since retired; merged rules) | We take the feature map, its health check and the upkeep pass as two opt-in, user-invoked skills built on the bundled launch-recipe generator, with product gaps routed to bugs:write and a skip when nothing changed since the last clean pass. Verification rules merge into verification:confirm and testing:run-e2e. Disagreements become settings: an end-to-end driver and reuse of a running instance in testing, a proof level in verification. We reject a verbatim generator and its writes under a Cursor-only path; neither is adopted. testing:map-features writes the feature map once, opt-in, to a repository location set by the feature_map_dir key of docs/conventions/testing.yaml, refusing unsafe locations; testing:run-e2e reads the map, checks it before the first drive, drives every entry point and reports one line each (plugins/testing/skills/run-e2e/SKILL.md:220-236); and verification:confirm holds its verdict below confirmed when an entry point is missing from that report (plugins/verification/skills/confirm/SKILL.md:166). testing:refresh-feature-map keeps an existing map current and ends clean, changed or blocked; its edits stay in the map directory; it skips when nothing changed since the last clean pass, keeping that state under the plugin data directory; an unattended pass refuses a map driven by chrome; a behavior the app lost goes to bugs:write or the report, never into the map; and a changed pass uses one fixed branch and one pull request. The upkeep routine class row, now control-skill-upkeep, is at plugins/autonomy/reference/routines.md:247 (status join: proven recurring manual pattern) and its Class parameters bullet at plugins/autonomy/reference/routines.md:343. Re-checked at 23e4138: holds; the two skills are unchanged since the old pin, and principle-prove-it-works is unchanged since the comparison at 12d587dfb2. Re-checked at 118e73b0e1: changed in create-verification-skill only: a launch from no stored data has to run schema setup and fixture loading and end in a state someone can work in, and a process that answers a health check is not ready yet. Decided 2026-11-12: taken, in app-control, which replaced testing:map-features (#15798): a launch starts from empty data with migrations and seed, in a per-run data directory where the app has one (plugins/app-control/reference/control-skill.md:118-147), and Ready shows the instance is usable (:162-175); the row below has the detail. maintain-verification-skill and principle-prove-it-works are unchanged. |
none |
create-verification-skill, principle-prove-it-works |
app-control:map create and the control-skill contract |
Adopted (new skill) | Read at 88d47514b8. We take: the repository is interviewed before the person (plugins/app-control/skills/map/SKILL.md:55-84); the generated skill covers launch, a read-only health check, drive, evidence and cleanup, with seeded features; and one end-to-end proof that launches from empty state, drives one feature, captures, cleans up and confirms the evidence survives (:140-155, detailed in plugins/app-control/skills/map/reference/prove.md:18-80), with any step a stand-in performed that the real target will not marked unverified (plugins/app-control/skills/map/reference/prove.md:82-84). A process that answers a health check is not yet usable (plugins/app-control/skills/map/SKILL.md:276-277). Taken with #15798, from create-verification-skill/SKILL.md:19 and :27: the hub's ## Launch records the launch model, a Data: line for the per-run data directory with migrations and seed, a Side by side: line for whether two instances can run at once, and for a short-lived app a Session: line for build once and one isolated session per drive (plugins/app-control/reference/control-skill.md:118-147); Ready checks the port belongs to this run's instance and the credentials are valid (:162-175); /app-control:map check reports a hub missing the first two lines (launch-model), and refresh treats that as doc drift. Every value stays the consumer's. We add what the consumer report behind this skill asked for: discovery over fixed rows, side effects and shared systems included (plugins/app-control/skills/map/reference/discovery.md); questions only for what discovery left open, through /planning:interview when installed; an unattended run that is a dry run and writes nothing; layered config and safety floors that ship safe and that the generated driver CLI enforces (plugins/app-control/reference/control-skill.md); and a check that no process the run started is still alive. We reject the Cursor-only output path and a fixed launch or doctor section set: the contract's hub sections govern. It replaces testing:map-features, credited in the row above and now retired. Decided 2026-11-12 (map decision #15659): app-control is aligned further with these two skills, each in its own work item whose pull request updates this row in detail: drive steps state their result and each feature ends with a proof step, preconditions and per-step evidence, and a journey never weakened to pass, launch, isolation and ready meaning usable from empty data, which settles the readiness candidate in the row above, and driving conventions and control-tool traits. Landed from the first of these, #15797, read at ec56937961: upstream's drive section pairs each user action with an exact command and an observable result and ends with a proof bullet (create-verification-skill/references/feature-map-example/README.md, create-note.md). We take it as a contract rule: every Drive step names the action, its command or handle and, after Result:, what a user should observe, and the steps end in a **Proof.** step whose Result: is the end state and whose Captures: names the captures that show it (plugins/app-control/reference/control-skill.md:333-351). The result stays on its step, with no separate proof section (:348-349). /app-control:map check reports a step with no Result: as drive-result (plugins/app-control/skills/map/scripts/check-map.sh:270, tested at plugins/app-control/skills/map/scripts/check-map.test.sh:222-233); create writes the results and proof step (plugins/app-control/skills/map/SKILL.md:94-96), and refresh treats a missing result, or a missing or stale proof step, as doc drift, and a stated result the step did not show as doc drift or a product regression by what the source says (:227-232). We add: /app-control:drive records each step's stated result beside what it observed and never compares the two (plugins/app-control/skills/drive/SKILL.md:239-245), and /testing:run-e2e starts each entry point's expected outcome from those results and the proof step, with the change's requirement on top (plugins/testing/skills/run-e2e/SKILL.md:156-163). This repository's two control skills, harness-ops:control-claude-code and .claude/skills/control-claude-code-plugins, state a result on each of their Drive steps and end each step list in a proof step; most of the second's features hand off to a driving tool's feature and have no steps of their own. |
none |
principle-prove-it-works |
verification:confirm Stage 2, "A failed check: suspect the observation first" |
Absorbed (one rule) | Read at ae5650febf. We take the rule from upstream's "Check the real thing, not a proxy" list: when verification fails, suspect the observation method before the system. It first landed as a Gotcha under the row above and now sits where the skill reads a failed check. It is stated once, in verification:confirm, which owns the verdict on a failed check, and not in debugging:debug (placement decided on #15776) (plugins/verification/skills/confirm/SKILL.md:195). We add: examples of a wrong observation (a stale build or artifact, a cached page, the wrong port or instance, a log read before the write landed, a stub left in place); a second observation by another route before a failure counts against the change, with NOT VERIFIED when no second route can run; and an Observation check: report line naming the first suspect and what ruled it in or out, in the outcome, fix and refactor report templates. |
none |
maintain-verification-skill |
app-control:map refresh |
Adopted (new action) | Read at 88d47514b8. We take: the upkeep pass ends clean, changed or blocked and edits only the control skill's own files (plugins/app-control/skills/map/reference/refresh.md:19-40); a read-only source wave of one subagent per feature file (:85-127); a live pass over every feature even when the source looks clean, under three invariants: a health check (Ready) before the first drive and after any failed one, cleanup of what each drive started, and evidence checked after cleanup (:129-150); and triage into doc drift (corrected and shipped), driver gap (fixed, re-driven live before it ships) and product regression (reported, kept out of the pull request) (:152-161), summarized in plugins/app-control/skills/map/SKILL.md:167-194. We add: waves of at most six subagents, a fixed value in the skill body; the skip when nothing changed since the last clean pass, one fixed control-skill-upkeep/<skill slug> branch and one pull request carried over from testing:refresh-feature-map; a refusal of an attended-only control skill when unattended; no fix that loosens a safety floor; and conversion of a legacy testing feature map, or flat feature files, in the same upkeep pull request (plugins/app-control/skills/map/reference/refresh.md:214-240). We reject asking which skill to refresh when several exist on an unattended pass (each is refreshed on its own branch) and a reset-or-relaunch step outside drive. It replaces testing:refresh-feature-map, credited in the row above and now retired. Decided 2026-11-12 (map decision #15659): the refresh is brought in line with her upkeep pass in its own work item, whose pull request updates this row in detail (pending adoption: app-control:map refresh, #15801). |
none |
principle-explain-the-number |
performance:verify section 5 |
Partial | New at 23e4138. No standalone principle skill. We take, as a required step before the verdict: the report names what limits the measured number (measured in a profiled run kept out of the report and traced to a code path) and at least one other thing the number could reflect, with the evidence that rules it out, in new Limiter: and Ruled out: report lines; a report that cannot fill them is not MET. Not taken: the eval-result variant (trials, models, scenario), which stays with evals:methodology; performance:snapshot is unchanged, since its host qualification and interleaved pairs already supply the spread. Re-checked at 118e73b0e1: holds; upstream merged its why paragraph into the three pattern bullets, which keep their rules. Re-checked at ae5650febf: changed: a new first pattern, applied before the run, says up front which outcome would alter the choice and how much variation is expected anyway, and stops when spread exceeds the effect. Not taken as a new step, consistent with the benchmark-checklist row's rejection of its pre-run claim sentence: performance:goal fixes the targets and floor before any work (plugins/performance/skills/goal/SKILL.md:84), and performance:climb writes its keep rule, with the host's measured noise as the gain floor, before the first attempt (plugins/performance/skills/climb/SKILL.md:47-49). The Limiter: and Ruled out: lines are unchanged. |
none |
benchmark-checklist |
performance:climb, performance:verify section 5 |
Partial | New at 23e4138. performance:climb takes the harness sanity counts (errors and finished work on every run). performance:verify takes the limiter question and the limits arithmetic (a saving cannot exceed the changed code's share of the run), and lists the tuning, error, skipped-work and repeatability questions as the alternatives its Ruled out: line excludes, each pointing at the check that already supplies the evidence. Not taken: the end-to-end relevance question (the goal's Correlation:, Event: and Unit: lines already cover it), the one-run ballpark mode (verify has no ballpark), the pre-run claim sentence and load checks (performance:snapshot qualifies the host), and the one-number PR body (the PR pipeline owns PR bodies). Its short alternation run count is rejected as a floor: performance:snapshot sets a higher interleaved-pair floor. Both takes have landed: climb refuses a frozen harness that does not print a failure tally and a finished-work tally (plugins/performance/skills/climb/SKILL.md:31-34), and verify carries the Limiter: and Ruled out: lines (plugins/performance/skills/verify/SKILL.md:113-128). Re-checked at e43c7ee26e: changed in one cross-reference only, which now points at perf-issue's ordered list instead of its strategy families (see that row); no question we took or declined changed. Re-checked at 118e73b0e1: holds; upstream shortened every question without changing one, and its notes on perf-issue and hillclimb became one line each. |
none |
autopilot-full with swarm |
session-flow:orchestrate, implementation:implement-dispatch, work-items:work-loop, review:fanout's fix pass, verification:confirm, multi-agent:assess, multi-agent:route |
Absorbed (no runner skill) | We keep our split between the agent that authors and the lane that merges and ship no runner skill. From swarm, orchestration gains a declared selection rule for a race, measurement briefs that name exact SHAs and the sampling method, and a missing field that earns one retry and then reads as a gap, never a pass. From autopilot-full, the fanout fix gains a red test for every site a proven finding touches, and dispatch and the work loop gain an operator hold that stops worker writes. We decline an owner merging its own PR and force-pushed rebases. Landed: the swarm take is in session-flow:orchestrate: imperative 2 requires a measuring worker's brief to name the exact commits and the sampling method and its return to repeat both, so a measurement without them is not evidence (plugins/session-flow/skills/orchestrate/SKILL.md:71-73); the Discipline line enters an item missing a required field as a gap, never a pass, unless one consolidated-scope respawn supplies it (:150-152). The autopilot-full take is in review:fanout's fix pass: a failing test first, covering every site of the defect inside the finding's location (plugins/review/skills/fanout/context/fix-pass-mode.md:142-149). Re-checked at 23e4138: changed: the scheduling change matches our loop lanes; the subagent-continuation change (fresh subagents by default, from 0.15.6) is now answered by the worker_continuation setting in session-flow:orchestrate, default resume, with respawn giving each new unit a fresh worker (plugins/session-flow/skills/orchestrate/SKILL.md:177-197), so the earlier conflict note is settled as a setting. The three paths credited here did not change between 23e4138 and e43c7ee26e. A merge-time rebase shortcut is a new candidate, not taken; owner to decide in source-control. Taken later: swarm's live lanes became the live_workers setting in verification:confirm, at both levels with the repository value winning, resolved beside proof_level (plugins/verification/skills/confirm/SKILL.md:112-129); Stage 2 step 1 splits the live drive by feature-map entry point, needs a Workspace environment entry for more than one local worker, names each worker's sessions apart, merges every worker's report for the missing-entry-point check and reports measured tokens and wall time per worker (:165-171); testing:run-e2e starts its own instance beside an answering app (plugins/testing/skills/run-e2e/SKILL.md:36-57) and prefixes session names with the workspace part (:194-200). The dispatch hold is in implementation:implement-dispatch: an operator hold stops every running dispatched agent with TaskStop and dispatches nothing more; SendMessage is not the hold, and no message goes to a stopped worker (plugins/implementation/skills/implement-dispatch/SKILL.md:148-158). Re-checked at 88d47514b8: changed. cursor/plugins#634 cut swarm's opening summary, its worker-count step and its result-table line; the declared race rule, the commit-and-method brief and the one-respawn-then-gap rule we took remain. autopilot-full was cut to principles: the red test across every site with the same defect and the zero-writes hold on the operator's stop remain, and the owner still merges its own PR after --force-with-lease rebases, both still declined. New, not taken: the full swarm runs one time, once the code is ready, and a rebase, a merge-conflict fix or a small fix gets one verifier alone; a merge-time judgment of whether trunk's changes can affect the patch, in place of the fixed merge-tree and changed-path checks; a hold reported to the root rather than a third rebase past merge prep; a defect trunk already shows recorded against trunk without gating the PR; and ending a tick that changed something by catching the operator up per ketchup (see its row). Re-checked at ae5650febf: changed. swarm renumbered a step and reworded its report line; the three rules we took stand. In autopilot-full the sentence behind our fanout take, a red test across every site, is gone, so that take is now ours alone (plugins/review/skills/fanout/context/fix-pass-mode.md:142-149). The zero-writes hold remains, and the owner still merges its own PR, which stays declined. New candidate, not taken: its findings rule now holds back a PR only for a defect someone can show, as a red run on the head commit or documentation this PR made false, and parks the rest as follow-ups, which widens the trunk-defect item above; owner to decide in review. Also not taken: no verifier for a wording-only change. Decided 2026-11-12: the root layer's fresh countersign on a change that relaxes a check (pstack/skills/poteto-mode/playbooks/autopilot-full.md:20) is taken in a different form, a review criterion rather than an owner stop. Our owner approval covers trust paths and security controls only; in an ordinary pull request a raised timeout, a lowered threshold or a new allowlist entry passes with a note (docs/conventions/pr-pipeline/README.md:2769-2777). The criterion lives in the review plugin, judged by a verifier wherever a loosened check appears; what counts as a gate, and whether it fails or notes, are settings (pending adoption: review:quality-gate criteria, #15832, after its interview). Re-checked at bf0c7cd80a: holds; swarm changed only its default worker model, machinery we drop, and autopilot-full did not change. |
none. pr-pipeline: an owner merging its own PR sits outside the one merge authority. Settings: pr_open_state in source-control; worker continuation in implementation. |
automate-me |
session-flow:audit-sessions, corrections:codify |
Partial | We decline the output: personal style belongs in user scope (user CLAUDE.md, user rules or an output style), never in a skill or shared repository files, and never landed through a PR. Taken: /corrections:codify preferences reads the local transcript store through /session-flow:audit-sessions excerpts-only, not workspace transcripts (plugins/corrections/skills/codify/actions/preferences.md:30-44), reads only sessions since the target's last edit (plugins/corrections/skills/codify/actions/preferences.md:20-27), keeps a preference only when distinct sessions reach recurrence.min_occurrences, default 2 (plugins/corrections/skills/codify/actions/preferences.md:45-51), asks the user to confirm each (plugins/corrections/skills/codify/actions/preferences.md:52-56), and proposes each one for a user-scope file (plugins/corrections/skills/codify/actions/preferences.md:10-19,57-61), with no separate mode skill. Codify places a personal-style lesson at user scope (plugins/corrections/skills/codify/SKILL.md:104-105). Re-checked at 23e4138: holds; unchanged since the comparison. Re-checked at 88d47514b8: holds; the only change is the subagent model-rule line, machinery we drop. |
none |
opening-a-pr |
source-control:pull-request and its settings |
Options | We take a comment-residue pass at prep, a body that does not recite the diff, and a narrow-PR flag. Options, not declines: the briefing body shape as a selectable preset (pr_body_required_sections in source-control; a repository keeps its own contract), opening ready as pr_open_state (default ready: create opens a draft and runs ready on it; a consumer opts into draft), and stacks as the native-stacks option babysit_stacked_prs. The briefing body preset tells the writer to leave out a run-by-run account of review or CI lanes and to link a longer record instead of quoting it (plugins/source-control/skills/pull-request/reference/create.md:288), and prep asks whether the branch holds parts that do not need each other and proposes one pull request per part, as advice with a fallback when no one can answer (plugins/source-control/skills/pull-request/reference/prep.md:23). pr_open_state (default ready) and babysit_stacked_prs (off by default) exist in source-control (plugins/source-control/reference/config-resolution.md:157, plugins/source-control/.claude-plugin/plugin.json:133). Re-checked at 23e4138: changed: the revised briefing is the candidate default preset. Re-checked at 118e73b0e1: holds; the briefing body keeps the same six sections and limits in shorter form, and the rule that a commit body does not restate its subject moved to the commits paragraph. Re-checked at 88d47514b8: holds; one sentence now points a stacked child at shipping's post-merge move onto trunk, a stack-landing mechanic the babysit_stacked_prs option already covers. Re-checked at ae5650febf: holds; one sentence drops interrogate from what a PR-opening subagent runs and drops the stack-ready report. We took neither. |
none. pr-pipeline: a PR reaches ready only through ready's gates, and this repository sets pr_open_state: ready; both settings exist, and it keeps the stack default (no stack merges). |
shipping |
source-control:pull-request merge |
Options | We keep merge-forward freshness and the merge gate's own readiness and recheck steps. Options: an independent, patch-bound verdict from a non-author agent, opt-in in the merge gate (pending adoption: source-control:pull-request, #6202); stack landing through babysit_stacked_prs. Re-checked at 23e4138: holds. Re-checked at 88d47514b8: changed: upstream cut the playbook from nine steps to six, each with its reason. The non-author verdict and the patch-id carry-over remain. New: after a merge, the child PR is retargeted to trunk, its patch rebased on what the merge produced and pushed with --force-with-lease; a verified PR stacked over an unverified one may rebase straight on trunk and merge by itself once verified again, when its branch neither holds nor needs the unverified code; and a merge is read back, because a queue that has only accepted the PR can still make the merge command exit cleanly. Our merge step already tells queued from merged (plugins/source-control/skills/pull-request/reference/merge.md:68,77), so nothing new is taken, and the options are unchanged. Decided 2026-11-12: the verdict carry-over by git patch-id (pstack/skills/poteto-mode/playbooks/shipping.md:7) is declined. A patch-id match ignores the base, so code no verifier saw could land on an old verdict, and a verdict here is bound to the commit it judged (docs/conventions/pr-pipeline/README.md:844, :1183-1186). Revisit only with a count of re-verifications on heads changed solely by an update-branch merge. |
none. pr-pipeline: landing from a working session is a merge path outside the one merge authority; rebasing or retargeting the bottom PR conflicts with the lanes' fast-forward-only writes; the after-merge recheck belongs to the post-merge lane; the patch-id carry-over conflicts with the pipeline's rule that a verdict is bound to the commit it judged and is not reused for any other commit. |
babysit with bugbot-triage |
source-control:babysit-prs, source-control:pull-request monitor |
Options | We take one fresh build per flake with no second retry, red-first fixes for behavior findings, and an optional consumer-owned review-bot triage rubric (review_triage_rubric) with a never-dismiss floor. We keep the standing loop lane and oldest-first order. When a finding is about behavior, monitor has the fix start from a test that fails on the current head for the finding's reason (plugins/source-control/skills/pull-request/reference/monitor.md:510). Re-checked at 23e4138: holds. Re-checked at 118e73b0e1: holds. Upstream removed two passages from the playbook, the rule that a watcher re-arm never authorizes a merge and the closing reply shape; we took neither. bugbot-triage is unchanged. Re-checked at ae5650febf: holds; one sentence moves the rebase under autopilot-stack from the root to the owner, which touches nothing we took. bugbot-triage is unchanged. |
none. pr-pipeline: once a PR is ready, the pipeline's respond lanes do this work; leaning toward dismissal on later passes differs from the pipeline's per-finding validator with a loop-cap escalation. |
autopilot-stack |
source-control stacked PRs |
Options | Operator landing matches our human-only default rung. Options: ready-on-open (pr_open_state), a per-link independent verdict, opt-in (pending adoption: source-control:pull-request, #6202), and the base-branch chain as babysit_stacked_prs. Re-checked at 23e4138: holds; the hourly tick and the push after each verifiable unit touch nothing we took. Re-checked at 88d47514b8: holds. Upstream now states only where it differs from autopilot-full: the root alone rebases or retargets a link, an item that depends on another branches from that item's branch, and once landing starts nobody writes to the chain but the post-merge retarget shipping allows. Operator landing still matches our human-only default rung, and the options are unchanged. Re-checked at ae5650febf: changed: cut to autopilot-full with no merge. A PR stacks only on one it builds on, and each owner rebases its PR onto its parent instead of the root writing the chain. Operator landing still matches our human-only default rung, and the options are unchanged; the pr-pipeline note names the owner's rebase now. |
none. pr-pipeline: an owner's rebase onto its parent and lease push conflict with the lanes' per-PR write queue and their no-stack-rebase rule; operator landing matches the pipeline's manual stack landing. |
autonomous-run |
implementation:implement, planning:draft-goal-condition |
Partial | Taken: implementation:implement reverts a block that leaves its target check where it was, back to the last green commit, before it tries the next hypothesis (plugins/implementation/skills/implement/SKILL.md:116). planning:draft-goal-condition Step 2 never relaxes a condition so a stalled run can finish; flat progress calls for a different approach, and a run that cannot finish surfaces the dead end with its evidence (plugins/planning/skills/draft-goal-condition/SKILL.md:64). The rest already matches ours. Re-checked at 23e4138: holds. Re-checked at 88d47514b8: holds; the only change is the subagent model-rule line, machinery we drop. Re-checked at ae5650febf: holds; one qualifier limits mid-run discoveries to those that support the goal, which touches neither take. |
none |
worktree-cleanup with worktree-audit.sh |
source-control:worktree cleanup |
Partial | Taken: cleanup takes a free-space reading before removal and reports it before, after and the difference (plugins/source-control/skills/worktree/context/cleanup.md:61,329), and a macOS reference measures Xcode build output and unavailable simulators and deletes each item only on a yes for that item (plugins/source-control/skills/worktree/context/macos-caches.md:1-7). Ours stays stricter elsewhere: a session claim lock, a dry run and confirmation for bulk removal. Re-checked at 23e4138: holds. Re-checked at 118e73b0e1: changed: worktree-audit.sh now checks its MERGED column against a -b ref, else origin/HEAD, then origin/main, then origin/master, where it used origin/main alone. Ours already resolves that order and takes --base (plugins/source-control/scripts/landed-work.sh:79-81), so nothing new is taken; the free-space reading we took is unchanged. Re-checked at 88d47514b8: holds. worktree-audit.sh reads its PR column through the Origin forge's CLI when that CLI can read the repository, else gh, and a draft PR now holds a worktree as an open one does; the playbook gained only the subagent model-rule line. The free-space reading we took is unchanged. Re-checked at ae5650febf: holds; worktree-audit.sh only chooses GNU or BSD flags for stat and date, and the playbook is unchanged. The free-space reading we took stands. |
none |
pause-safely |
session-flow:handoff |
Options | Taken as an option: with wip_commit in session-flow (default off), session-flow:handoff makes one local commit of tracked changes when the user asks to pause, and never pushes it (plugins/session-flow/reference/config.md:16). The rest matches session-flow:handoff. Re-checked at 23e4138: holds. |
none. pr-pipeline: pushing such a commit to a ready PR would start the lanes on a broken head; decided: ours, because the commit stays local. |
poteto-mode, poteto-agent |
session-flow:orchestrate, session-flow:workflow |
Partial | We take skip-with-reason checklists into orchestration; the rest is covered by playbooks:fable-5 and the discipline plugin. Re-checked at 23e4138: changed: the subagent-continuation default changed upstream; it is now a worker_continuation setting in session-flow:orchestrate (default resume, respawn for a fresh worker per unit; see the swarm row). Re-checked at 118e73b0e1: changed: the chat that receives /poteto-mode now runs the playbook and owns the task list and every spawn, and poteto-agent is only a delegate that reads the absolute skill path its parent passes. That matches session-flow:orchestrate, which keeps dispatch in the session; the skip-with-reason checklist we took is unchanged. Re-checked at 88d47514b8: holds. The skill's name is now poteto-mode, its mode and reminder frontmatter are gone, its principle index carries the restated summaries of cursor/plugins#633 (see those rows), and a new "Match the model to the job" rule moves each spawn's reasoning level and speed for the job from the role's model line. That rule is model-role machinery we drop; choosing model and effort per role is multi-agent:route's. poteto-agent now searches the Cursor plugin folder for the skill when its parent passes no path. Dispatch still stays in the chat that received /poteto-mode, and the skip-with-reason checklist is unchanged. Re-checked at ae5650febf: changed: a new trigger has the agent restate the goal and the smallest expected change in its first message, each subagent brief, each PR and each final report. New candidate, not taken: that restatement, beside the in-your-own-words row; owner to decide in session-flow. Its other changes, background resumes and a longer rejected-slug fallback, are model and harness machinery we drop. The skip-with-reason checklist is unchanged, and poteto-agent is unchanged. Decided 2026-11-12: of the restatement's four places, three are already owned: each subagent brief (session-flow:orchestrate, implementation:implement-dispatch), each PR (the body's Why section), and the first message (playbooks:fable-5's problem-framing restatement, plugins/playbooks/skills/fable-5/context/problem-framing.md:7-13, and work-items:work's "I read this item as" line). A goal line in every final report is not proposed. The one gap, the smallest-change clause, is taken into the playbook's restatement sentence (plugins/playbooks/skills/fable-5/context/problem-framing.md:11); see the in-your-own-words row. Re-checked at bf0c7cd80a: holds. Upstream removed the trigger that sent a nontrivial change or an "are we sure?" to how, changed the default model per role, and added a notice when the model rule predates the new defaults; all of it is model-role routing, which is multi-agent:route's. The skip-with-reason checklist and the goal restatement are unchanged, and poteto-agent did not change. |
none |
hillclimb |
performance:climb, performance:goal, performance:snapshot |
Adopted (new skill) | Landed: performance:climb runs the keep-or-revert loop over a harness performance:snapshot froze, with a keep rule written before the first attempt and a min_attempts floor from performance:goal, so an early lucky result cannot end the run (plugins/performance/skills/climb/SKILL.md:28-34 and :43-47). At 23e4138 the harness step began deferring to benchmark-checklist; see that row. Re-checked at e43c7ee26e: changed: the step that forms each hypothesis now orders candidates by perf-issue's fixed list, taking that list's order and not its stopping rule. Not taken: climb chooses the next hypothesis from evidence, preferring the smaller change when two are expected to move the counter about equally, a mechanism no logged attempt has tried, and performance:target's ranking after three reverts in a row (plugins/performance/skills/climb/SKILL.md:120-129). A fixed order would spend attempts on mechanisms the profile never pointed at, the same reason the perf-issue row gives. Re-checked at 88d47514b8: holds; the only change is the subagent model-rule line, machinery we drop. Re-checked at bf0c7cd80a: holds; the grounding step traces the target inline instead of running how, and the delegate's default model changed. The keep rule and the attempt floor we took are unchanged. |
none |
perf-issue |
performance:target, performance:climb |
Partial | Landed: the strategy-family index sits above the latency catalog as a table from what a profile shows to a family, its counter and its catalog rows (plugins/performance/reference/techniques.md:467-490), and performance:target names a candidate's mechanism only through a matching row (plugins/performance/skills/target/SKILL.md:68-71). Re-checked at e43c7ee26e: changed: upstream dropped its eight families, each gated on a trace signal, for seven shorter rules tried in a fixed order, cheapest first, stopping at the first that meets the target. Not taken: our table stays keyed to evidence, so a family with no matching row in the profile is not tried (plugins/performance/reference/techniques.md:469-471); each of the seven rules corresponds to a row the table already has, and the hedged-request row, which the new list does not name, stays. The stopping rule is not taken either: climb continues past a met target until the goal's min_attempts (plugins/performance/skills/climb/SKILL.md:106-112). Re-checked at 88d47514b8: holds; the only change is the subagent model-rule line, machinery we drop. Re-checked at bf0c7cd80a: holds; step 2 traces the code inline instead of running how, and the delegate's default model changed. The seven rules and their stopping rule are unchanged. |
none |
multi-phase-plan with check-plan.mjs |
planning:plan, source-control:pull-request ready, verification:confirm's proof level |
Partial | planning:plan skips the plan when one sentence covers the change; invoked proactively it says so and writes none, and on an explicit request it says so and asks; the reviewer gate covers any plan it presents (plugins/planning/skills/plan/SKILL.md "Purpose", the one-sentence skip, and "Gates", the Reviewer gate; line hints :34-37, :66); the template's optional Review: evidence value is in plugins/planning/skills/plan/context/plan-template.md "Phase Review tags (optional per phase)" (line hint :415-421). The performance gate and the proof level landed as one setting, proof_level in verification (path by default, live, strict): under strict, verification:confirm wants unit, live and performance evidence for each plan phase, so planning:plan has no setting of its own for it (plugins/verification/reference/config.md:11,22-25). pull-request ready reads a plan phase's Review: value, gathers evidence when it asks for some, and skips a review-concern tag (plugins/source-control/skills/pull-request/reference/ready-for-review.md:62-74). Re-checked at 23e4138: holds; the hourly tick and removed goal touch nothing we took. Re-checked at 118e73b0e1: changed in the script only: check-plan.mjs now finds fenced code by CommonMark fence rules (backtick or tilde fences, a closer at least as long as its opener), tested by a new check-plan.test.ts. We took no plan-checker script, so nothing here moves; the playbook is unchanged. Re-checked at 88d47514b8: changed in the playbook: the full verification swarm now happens one time, when the code is ready, and repeats only when the change adds behavior, is redesigned or grows by a large diff, while a rebase, a merge-conflict fix or a small fix is checked by one verifier on the resulting head; the explore step gained the subagent model-rule line. proof_level sets what evidence each plan phase needs, not how often a swarm reruns, so nothing we took moves. Re-checked at ae5650febf: holds; the playbook now picks its execution playbook by whether the run may merge and judges findings by that playbook's blocking rule (see the autopilot-full row), and check-plan.mjs changed how it matches inline code. We took no plan-checker script, and proof_level is unaffected. Re-checked at bf0c7cd80a: holds; the swarm workers default model changed and the reading list no longer names how. Nothing we took moves. |
none |
typescript-best-practices |
planning:design TypeScript reference, review:code-reviewer |
Partial | A TypeScript reference loaded on detection, not a separate skill. Landed in review:code-reviewer: an item for casts that overrule the type checker (a double cast through unknown or any, a non-null assertion on a value that can be null) and for debug output left in shipped code, raised only where no project linter or compiler setting already covers it and graded SUGGESTION (plugins/review/agents/code-reviewer.md:79,83). Re-checked at 23e4138: changed: the boundary example in the reference file changed; whether our reference follows is the owner's call (planning:design). The path did not change between 23e4138 and e43c7ee26e. Re-checked at 88d47514b8: changed: the Branded types row now requires the brand to be established by a type guard that checks the value (input is AgentId) and forbids as for it, and the reference's parseAgentId example narrows through isAgentId and returns the input with no cast. A new candidate, not taken: our TypeScript reference keeps as as the one routine cast, right after the check in a parse function (plugins/planning/skills/design/reference/type-discipline/typescript.md:80-82); owner to decide in planning:design. The reviewer's cast item is unchanged. Decided 2026-11-12: her TypeScript rules are not taken; a generic skill does not ship one language's rule table. Taken is the mechanism behind them: the type-discipline contract asks each reader to say where its language idioms live (plugins/implementation/reference/type-discipline.md:8) and implementation:implement names none, so it will read a consumer-declared idioms file for the language it detects, with no shipped fallback (pending adoption: implementation:implement, #15835). The as candidate above stays open. |
none |
architect |
planning:design, implementation:implement, architecture:improve |
Partial | No new skill. Taken: the scrap tells, as the design-signals table implementation:implement checks before filing friction as a plan problem (plugins/implementation/skills/implement/SKILL.md:153). A scaffold_stubs setting, off by default, with which planning:design adds a stub scaffold section listing each new file and its signatures before the handoff gate (plugins/planning/skills/design/SKILL.md "Handoff gate", Stub files (scaffold_stubs); line hint :325-342). In planning:design: the two-candidate step, as design's own step rather than a pointer, and Phase 3 opening from two or three caller sites and deriving the types from them (plugins/planning/skills/design/SKILL.md "Phase 2: Design Threads", Two-candidate step, and "Phase 3: Type Modeling", Call sites before types; line hints :147-157, :207-210). Re-checked at e43c7ee26e: changed: the candidate screen now assumes whoever changes the code later is an agent that reads few files and imitates what is nearby, and the red-flags reference gained four flags: state with more than one writer, two supported routes to the same result, module internals reachable from outside, and lists kept in step by hand. Not taken as a new screen, because each already has an owner here: the duplicate-writer and content-coupling types in coupling:reduce (plugins/coupling/skills/reduce/reference/coupling-model.md:36-50), its second-list hotspot (same file, :109-111), the one-way rule of discipline:reuse-or-replace (plugins/discipline/skills/reuse-or-replace/SKILL.md:14-16), and the architecture-test rung for a forbidden import (plugins/review/context/enforcement-ladder.md:44-46). The imitating-agent assumption is already the stated reason for the ladder's canonical-helper rung and its rule that the most-copied weak pattern is fixed first (plugins/review/context/enforcement-ladder.md:38-43,104-114). planning:design's two-candidate step restates none of the four flags; those owners keep them. Separately landed: the temporal-decomposition entry of the red-flags reference, in the architecture:principles vocabulary as the rule that a run order is not a reason to split modules, which group by the knowledge they keep to themselves (plugins/architecture/reference/vocabulary.md:94-98). That entry is identical at 23e4138 and at e43c7ee26e; only the four flags above changed. Re-checked at 118e73b0e1: holds; upstream shortened the overview and the implement phase, cut its default runners to two models, and kept the scrap phase and its tells. Re-checked at bf0c7cd80a: holds; Phase A and the scrap list trace the code inline instead of running how, and the default runners changed. The scrap tells and the two-candidate rule are unchanged. |
none |
figure-it-out |
planning:plan's phase order, implementation:implement, implementation:implement-dispatch |
Partial | No skill. Taken: riskiest-first as one value of the phase_order setting in planning (plugins/planning/reference/config.md "Keys", the phase_order row and its riskiest-first value; line hints :43, :68), and revert-not-patch as implementation:implement's rule for a block that leaves its check where it was (plugins/implementation/skills/implement/SKILL.md:116). Gate repair in implementation:implement-dispatch: a gamed gate is rejected and its criterion restated so the same shortcut fails, and a mistaken gate is repaired in a separate change (plugins/implementation/skills/implement-dispatch/SKILL.md:369). Re-checked at 23e4138: holds. Re-checked at 118e73b0e1: holds; the description no longer names show-me-your-work, and the loop and revert rules are unchanged. Re-checked at 88d47514b8: holds; the only change is the subagent model-rule line, machinery we drop. |
none |
how, investigation |
discovery:explore and its explore_output setting |
Options | No how skill. Taken: explore_output in discovery (auto by default, change-prep, explain) chooses the shape of the final reply, and explain walks through how the scope works with a path:line cite at each step (plugins/discovery/reference/config.md:34). Re-checked at 23e4138: holds. Re-checked at 118e73b0e1: holds. Upstream renumbered its steps and cut the explorer prompt's step list; the walkthrough output we took does not depend on the cut text, and investigation is unchanged. Re-checked at 88d47514b8: holds; the only change to how is the subagent model-rule line, and investigation is unchanged. Re-checked at bf0c7cd80a: changed. how lost its simple-or-complex split, its parallel explorers and both prompt files: one read-only subagent now explores and explains in a single pass from a one-paragraph brief, and its named output sections are gone. investigation traces the code inline and keeps the Overview to Gotchas sections. Already have: discovery:explore runs in one fresh-context subagent, and the explore_output walkthrough we took does not depend on the removed text. |
none |
interrogate |
review:fanout, review:code-reviewer, review:quality-gate with code-metrics:audit size |
Partial | Landed in review:fanout: a short intent paragraph the orchestrator writes in its own words and passes, labeled as the author's claim, to every surface it dispatches, as a description to test and never evidence (plugins/review/skills/fanout/SKILL.md:95-99), and a ## Dismissed report section listing each demotion with its reason (:145-149). Landed in review:code-reviewer: the reviewer rubric's repeat-run and dual-path item, flagging an old and a new path for one behavior both left callable with nothing saying which callers use which (plugins/review/agents/code-reviewer.md:71). Review's fanout_diversity setting, values lens (default, one reviewer agent per concern) and model (one review prompt per distinct fan-out model tier) (plugins/review/skills/fanout/SKILL.md:64-76, plugins/review/workflows/fanout-sweep.js:32, plugins/review/reference/config.md:13, plugins/review/.claude-plugin/plugin.json:37). The same rubric line holds the second rubric item, state-changing code that may run a second time after a retry or a resume (plugins/review/agents/code-reviewer.md:71), and the large-file flag is the reviewer's size-crossing note: an advisory finding for a file that went from below the consumer's size.file_lines reference at the base to at or above it at the head, counted through code-metrics:audit size (plugins/review/agents/code-reviewer.md, "Before reviewing" step 1, "Size crossings"). Re-checked at 23e4138: holds; unchanged at e43c7ee26e. Re-checked at 118e73b0e1: holds. Upstream shortened its scope, intent and synthesis steps and cut its default reviewers from three models to two; the intent paragraph, the Dismissed bucket, the agreement section and the rubric's idempotency item remain. Re-checked at 88d47514b8 (cursor/plugins#634): changed: the lead-judgment reference lost its opening, which said why the lead filters rather than aggregates, and its Verdict Calibration section, which capped the Act On list at about five items and called the Dismissed list a trust mechanism; SKILL.md gained the subagent model-rule line and a family fallback for cursor-grok-* and composer-* slugs. The Dismissed section stays in the skill's report shape, so our ## Dismissed take stands and its reason is now ours alone; we never took the five-item cap. Re-checked at ae5650febf: changed. The skill was rewritten: adversaries on the configured models must prove each finding with a run, the lead's four buckets are gone, and all four reference files were deleted. Our intent paragraph still matches its framing step. ## Dismissed (plugins/review/skills/fanout/SKILL.md:145) is now ours alone beside upstream's per-adversary dropped count; so are the repeat-run and dual-path rubric item (plugins/review/agents/code-reviewer.md:71) and the size-crossing note ("Size crossings"). fanout_diversity stands as a setting. New candidate, not taken: keeping only findings a rerun proves; owner to decide in review. Decided 2026-11-12: taken, reshaped. A plain filter would break the fan-out's recall contract and its rule that it runs no builds or tests (plugins/review/skills/fanout/SKILL.md:93-97, :200). Instead, the rule for running reviewed code moves out of /bugs:scan's verification gate into one contract both cite, and a setting on the fan-out's normalization step labels or requires run-proven findings (off by default), with unproven findings still listed as dismissed (pending adoption: review:fanout, #15833, after its interview). |
none |
no-comments, comment-sicko |
code-tidying:remove-comments, code-metrics:audit-suppressions |
Partial | Taken: code-tidying:remove-comments keeps a suppression's justification beside its waiver (dissolve-comments/SKILL.md:102, the file the line cites in this row refer to) and proposes a regression test or a lint rule for a comment that states a constraint the code could enforce (:168-172); the run's mode settles a doubtful comment, kept under the cautious modes and kept under aggressive and strip only when it is a listed survivor kind (:130-139); and code-metrics:audit-suppressions is the suppression inventory. Re-checked at 23e4138: holds. Re-checked at 118e73b0e1: holds; the agent's id is now comment-sicko, and no-comments spawns it by that id. Re-checked at 88d47514b8: holds; the only change to no-comments is the subagent model-rule line, and comment-sicko is unchanged. Decided 2026-11-12: fixing the flagged code inside the same run is declined; comment removal stays comment-only. Taken instead is the hand-off: code-tidying:remove-comments writes its reshape flags and root-cause rows to a findings file in every mode, so /review:fanout fix, /implementation:implement or /architecture:improve can act on them (pending adoption: code-tidying:remove-comments, #15841). Decided 2026-11-14: her policy is the shipped default posture, necessary. Our floor differs from comment-sicko in three places: a suppression of a correctness rule is kept and listed as root-cause work, a TODO(#issue) marker is kept, and a comment paired with a regression test is kept. New comments are stopped by a lint on the lines a change adds (comment_lint in code-tidying, off by default) and by the review rule, not by a comment count. Re-checked at bf0c7cd80a: holds; no-comments and comment-sicko now trace a thin IMPORTANT or do not remove symbol inline instead of running /how (/why stays). Our posture and floor are unchanged. |
none |
setup-pstack |
multi-agent:route |
Partial | We keep alias tiers and per-plugin configuration; an effort budget lives in multi-agent configuration. Re-checked at 23e4138: holds. Re-checked at 118e73b0e1: holds; upstream changed its budget labels and model defaults, which we do not take. Re-checked at 88d47514b8: holds; a new last step offers once to generate a formal verification skill through create-formal-verification-skill, which is not adopted (see below); the alias tiers and per-plugin configuration we keep are unchanged. Re-checked at ae5650febf: holds; one sentence on how the judge pool is ordered, model-role machinery we drop. Re-checked at bf0c7cd80a: holds. New defaults, budget steps measured from each role's default, retired how and why role lines, and a # defaults: 2 marker so a rerun replaces old defaults a full-file rule had pinned while keeping the user's choices. Already have: /multi-agent:setup apply writes only the keys the user asks to change and keeps the rest (plugins/multi-agent/skills/setup/SKILL.md:60-62), so a bundled default never goes stale in a consumer layer. |
none |
principle-attack-the-premise, principle-fix-root-causes |
debugging:debug, code-tidying:remove-comments audit |
Partial | Taken in debugging:debug: after a second failed fix, reduce what the failed fixes share to one claim a test can check, running the per-actor tally first when the symptom is uneven (plugins/debugging/skills/debug/SKILL.md:383); a failure that began after a restart with no code change ranks what the old process left on disk first in the hypothesis table (:200); and a shipped workaround needs a tracking link or a removal condition in its comment, which code-tidying:remove-comments audit reports as unjustified-workaround when both are missing (:369). Re-checked at 23e4138: holds. Re-checked at 88d47514b8 (cursor/plugins#633, landed by 99743e9c15): changed; both principles are restated as reasoning. principle-attack-the-premise dropped its census of which actors hold an imbalance and its rotate-or-randomize remedies; it now asks for one check whose result would differ were the premise wrong, judged from the pattern of where failures fall as well as how big they are. That is the one checkable claim debugging:debug already takes (:383); running the per-actor tally before that test is now ours alone, kept as one way to read the distribution. principle-fix-root-causes now defines the root cause as the earliest point where behavior departs from what was intended, keeps suspecting persisted state after a restart (:200) and fixing the code rather than a paragraph-long comment (:369), and replaces its flat ban on guards with a test: handling a state is the fix when the system may legitimately be in it, and a guard masks the defect when that state is one the system must never reach. Taken in debugging:debug under the principle placement of #7589: its "Does not fix the symptom" line now states that test. Also taken there: after a deep dig, Phase 6 proposes the cheap always-on signal that would have shown the deciding fact at first look, gated by the always_on_signal key of docs/conventions/debugging.yaml, and searches for the fixed defect's siblings. Decided 2026-11-12: a hard stop on a doubtful premise is declined; debugging:debug keeps the premise check advisory on purpose, with its reproduction loop as the only hard gate (plugins/debugging/skills/debug/SKILL.md:386). Taken: the test-fix loop gains the same check, so /testing:diagnose runs the premise check after its second failed fix, reaching it by slash reference to /debugging:debug with the sentence as fallback (pending adoption: testing:diagnose, #15836). |
none |
principle-boundary-discipline, principle-type-system-discipline, principle-model-the-domain |
implementation:implement, planning:design's type-discipline reference, review:code-reviewer, architecture:improve |
Partial | Taken: implementation:implement narrows values where they enter the system and relies on those types afterwards (plugins/implementation/skills/implement/SKILL.md:52); planning:design loads a type-discipline reference; feature mode builds a structure that names every case before it adds another branch arm, and reports that structure as a follow-up under a minimal posture (plugins/implementation/skills/implement/context/feature.md:12). The reviewer questions are in review:code-reviewer: a side-effect item asking whether I/O, clock, randomness or shared state reached for partway through logic can move to the caller and leave the logic pure (plugins/review/agents/code-reviewer.md:77), the cast and debug-output item (:79, see the typescript-best-practices row), and a second trigger on the Repeated Switches smell, a boolean added only to stay in step with an existing one, with one enum or state value as the remedy (plugins/review/reference/design-smells.md "Baseline", Repeated Switches); in architecture:improve's scan briefing, a special-case-in-a-general-module signal (plugins/architecture/skills/improve/research/deepening/scan-briefing.md:43). Re-checked at 23e4138: holds; the three paths did not change between 23e4138 and e43c7ee26e. Re-checked at 118e73b0e1: holds. principle-type-system-discipline folded its tests into its patterns and principle-model-the-domain its why paragraph into its opening; principle-boundary-discipline is unchanged. Re-checked at 88d47514b8 (cursor/plugins#634, at 99743e9c15): changed in principle-type-system-discipline only. It cut five bullets (illegal states unrepresentable, types as constructions, not lying to the type system, exhaustive matching, strengthening a type only where partiality appears) and dropped "refuse to lie to the compiler" from its description, keeping the branding, parse-at-the-boundary and derive-from-schema bullets. The matching rows of our type-discipline reference (sum types for exclusive states, an encoding that cannot hold an invalid value, tightening only where a function would fail, exhaustive branching, no unproven cast: docs/contracts/type-discipline/README.md:19-23, generated into plugins/planning/skills/design/reference/type-discipline.md) carry their own reasons there, so nothing taken is lost; upstream's typescript-best-practices still states the TypeScript forms. principle-boundary-discipline and principle-model-the-domain are unchanged. Decided 2026-11-12: principle-type-system-discipline's branding rule, for two values that share a primitive, is taken as one language-neutral row in the type-discipline contract, which today states it only in its TypeScript file; the row takes no side on the open as versus type-guard question in the typescript-best-practices row (pending adoption: docs/contracts/type-discipline/README.md, #15840). |
none |
principle-build-the-lever |
discipline:script-the-deterministic-work and its lever_scope setting, implementation:implement-dispatch |
Options | Taken: the lever check, which implementation:implement runs before a multi-site block and implementation:implement-dispatch before a fan-out, has the caller edit the first site by hand, run the lever on a clean copy of that site and diff the two before any other site (plugins/discipline/skills/script-the-deterministic-work/SKILL.md:175-178); lever_scope (deterministic by default, or non-trivial) decides which changes get a lever (plugins/discipline/reference/config.md:44). Re-checked at 23e4138: holds. Re-checked at 118e73b0e1: holds; upstream dropped its Why, Pattern and Balance labels and kept the text. |
none |
principle-encode-lessons-in-structure |
enforcement-ladder convention, corrections:codify |
Partial | Taken: the ladder's canonical-helper rung (docs/contracts/enforcement-ladder/README.md:35), and corrections:codify's rung pick, which sends a recurring mistake a check could assert to the strongest ladder rung that can decide it (plugins/corrections/skills/codify/SKILL.md:72-92). Re-checked at 23e4138: holds. Re-checked at 118e73b0e1: changed: upstream cut the skill to its judgment fallback and corollary, removing the strongest-mechanism order the canonical-helper rung came from and its feedback-loop and anti-pattern lists. Both takes stand; the rung is now ours alone. |
none |
principle-exhaust-the-design-space, principle-foundational-thinking, principle-redesign-from-first-principles |
planning:design, planning:plan, implementation:implement |
Partial | Taken: phase_order in planning (plugins/planning/reference/config.md "Keys", the phase_order row; line hint :43) and integration_posture in implementation (plugins/implementation/reference/config.md:15). In planning:design: a day-one question asked before type modeling when a requirement is added to existing code (plugins/planning/skills/design/SKILL.md "Phase 1: Problem Space Decomposition"; line hint :134); a two-candidate step for a thread with nothing to copy, several workable approaches and a hard-to-reverse decision, with criteria written first, at least two candidates that differ in structure, and rejected ones recorded, written as design's own step rather than a pointer into architecture:improve ("Phase 2: Design Threads", Two-candidate step; line hint :147-157); a shared-state thread for any type more than one concurrent actor touches (Shared-state in the same phase's thread list; :172-175); and types chosen for their most frequent read and change paths ("Phase 3: Type Modeling", Operations set structure; :211-213). Re-checked at 23e4138: holds. Re-checked at 118e73b0e1: changed: principle-exhaust-the-design-space now states its rule, two or three competing prototypes before a novel decision, in its description only, and its body keeps one line on product decisions; principle-redesign-from-first-principles shortened its description. Nothing we took depends on the cut text, and principle-foundational-thinking is unchanged. |
none |
principle-experience-first |
planning:prd, planning:design, testing:run-e2e |
Partial | The product-value questions went to planning:prd rather than design: it asks once where the easier build would leave users worse off and which side wins, asks what each option is for while keeping options a project holds configurable on purpose, and ties success metrics to the repeated task users come back for (plugins/planning/skills/prd/SKILL.md "Step 4. Drive frontier-rounds Q&A", the Goals, Non-goals and Success metrics rows, and "Step 5. Persist the PRD", required sections 3 and 5; line hints :141-145, :175-177). planning:design Phase 3 writes at least one call site from the importing package's side for a library or API (plugins/planning/skills/design/SKILL.md "Phase 3: Type Modeling", Call sites before types; line hint :207-210). Re-checked at 23e4138: holds. testing:run-e2e captures error, empty and feedback states, not only the path where everything succeeds, and lists any state it could not reach as not driven (plugins/testing/skills/run-e2e/context/e2e.md:158). Decided 2026-11-12: the preset tie-break, delight over implementation convenience, and "ship less, ship better" (pstack/skills/principle-experience-first/SKILL.md:9,12) are declined. planning:prd asks which shortcut would cost users something and leaves the choice to the user (plugins/planning/skills/prd/SKILL.md:143); a preset answer conflicts with keeping that choice the project's, and an eval guards against cutting options only to ship less (plugins/planning/skills/prd/evals/evals.json:96). No work item. |
none |
principle-laziness-protocol, principle-subtract-before-you-add, principle-migrate-callers-then-delete-legacy-apis, refactoring |
implementation:implement, architecture:improve, verification:confirm |
Partial | Taken: cut before polish, as feature mode's rule that what a finished slice showed it does not need comes out before the rest is refined (plugins/implementation/skills/implement/context/feature.md:14). From principle-laziness-protocol: a long-trace signal in architecture:improve's scan briefing, for a request that passes through file after file or layer after layer with each adding little, judged against the rest of the codebase with no fixed count (plugins/architecture/skills/improve/research/deepening/scan-briefing.md:42). A compatibility setting, refactor_compat in implementation:implement (same-wave by default migrates every caller and retires the old shape in one wave unless outside consumers depend on it; deprecate keeps an adapter with a removal condition for each consumer), resolved in refactor mode (plugins/implementation/skills/implement/context/refactor.md:25, resolution at :27, rule at :29; option named at plugins/implementation/skills/implement/SKILL.md:82; key at plugins/implementation/reference/config.md:16, plugins/implementation/schemas/implementation.schema.json:35, plugins/implementation/.claude-plugin/plugin.json:72), and staging by consumer kind in fable-5 (plugins/playbooks/skills/fable-5/context/planning.md:89). Refactor mode deletes first. It lists stale imports and config keys, one-caller wrappers and unreachable branches through /code-tidying:audit-dead-code when available, else a grep, and commits the deletions first (plugins/implementation/skills/implement/context/refactor.md:18). A typecheck or lint is not a pin (:16). A move must remove a branch, an invalid state or an unclear name, never only add indirection (:19). Old-contract tests are retargeted only for a plan-declared contract change (:21). The reshape is kept only where reader load fell, else reverted (:23). A deprecate adapter carries a removal condition, a time box and a follow-up item (rule :29; refactor_compat heading :25, Resolution :27). The equivalence check is an output diff or replayed baseline for a large reshape (plugins/verification/skills/confirm/context/refactor.md:39). Re-checked at 23e4138: holds; unchanged at e43c7ee26e. Re-checked at 118e73b0e1: changed: principle-laziness-protocol dropped its deletion, flat-call-hierarchy (more than three files), threading and small-leak bullets, so the long-trace signal is now ours alone, still with no fixed count; principle-subtract-before-you-add dropped its why paragraph and kept cut-before-polish. principle-migrate-callers-then-delete-legacy-apis and refactoring are unchanged. Re-checked at 88d47514b8 (cursor/plugins#634, at 99743e9c15): changed: principle-migrate-callers-then-delete-legacy-apis cut its rule bullets (take an inventory of callers, move them all over and remove the legacy API together; adapters only as a time-boxed exception) and its dual-path reason, keeping the description, the test-update bullet and the conditions for applying it. The same-wave default and the deprecate adapter's removal condition and time box in refactor_compat (plugins/implementation/skills/implement/context/refactor.md:25-29) no longer have a rule body behind them in the principle; upstream's description and the refactoring playbook still state same-wave migration, and the playbook gained only the subagent model-rule line. The test-update bullet still matches our old-contract test rule (:21). Re-checked at bf0c7cd80a: holds; refactoring traces the subsystem inline instead of running how and its delegate's default model changed. It still states same-wave migration, and the three principles did not change. |
none |
principle-make-operations-idempotent |
planning:devils-advocate, review:code-reviewer |
Partial | planning:devils-advocate Round 4's Re-execution row covers a second start, a stop between writes and a reconciliation step with examples, and runs on any plan with a step that writes state and can run again (plugins/planning/skills/devils-advocate/SKILL.md:207,214). The reviewer item is in review:code-reviewer: which flags state-changing code that may run again (a retry loop, a resumed job, a redelivered message) with no reason the second run is harmless or no condition that ends the repeats, graded SUGGESTION (plugins/review/agents/code-reviewer.md:71,83). Re-checked at 23e4138: holds; unchanged at e43c7ee26e. Re-checked at 88d47514b8 (cursor/plugins#633, landed by 99743e9c15): holds. The four pattern bullets (convergent startup, content-based cleanup, PID-based stale locks, idempotent scheduling) became one idea (read the current state, work out what each step still has to change, and apply only that) and one rule: anything a run leaves behind must be recognizable as stale by evidence that holds wherever the previous run stopped. That is the reconciliation step the Re-execution row already asks for, with the temp-path rename and done-marker examples; the three-question test and the reviewer item are unchanged. |
none |
principle-minimize-reader-load, principle-separate-before-serializing-shared-state |
architecture:improve, coupling:reduce |
Partial | In architecture:improve's scan briefing, a wide-state-scope signal ordering scope from function inputs to globals, plus two reader questions about the value at the center of a candidate (what produces it, and which code may change it), recorded as evidence for the confidence rating and not a third heuristic (plugins/architecture/skills/improve/research/deepening/scan-briefing.md:41,72-77); in coupling:reduce, the first rung of the shared-state remediation, which has each writer own its own target before any lock is added, and does not apply when every writer must agree on one value (plugins/coupling/skills/reduce/reference/remediations.md:38-47). Re-checked at 23e4138: holds; the two paths did not change between 23e4138 and e43c7ee26e. Re-checked at 118e73b0e1: holds; principle-minimize-reader-load folded its two axes and why paragraph into one paragraph and kept both reader questions, and principle-separate-before-serializing-shared-state is unchanged. |
none |
principle-never-block-on-the-human, principle-outcome-oriented-execution, principle-sequence-verifiable-units |
playbooks:fable-5, planning:plan, implementation:implement |
Options | Taken: described_problem in playbooks (report by default, or fix) decides whether playbooks:fable-5 fixes a problem the user only described or reports it (plugins/playbooks/reference/config.md:25); and a phase can declare a planned breakage span (the build, the tests or the lint left red until a named later phase), inside which implementation:implement allows a red commit only for the declared failures (plugins/implementation/skills/implement/SKILL.md:104). implementation:implement Step 2 applies a ready lever or hand edit one unit at a time under a per_unit_check setting, pilot by default or every (plugins/implementation/skills/implement/SKILL.md:109). With no approved plan, implementation:implement Step 0 resolves an integration_posture setting (by-kind default, day-one, minimal) by kind of change, and planning:plan Step 2 reads it from the repository file and writes a day-one redesign as plan work items (plugins/implementation/skills/implement/SKILL.md:70, plugins/planning/skills/plan/SKILL.md:234). In implementation:implement: a data-boundary line in Step 0 (plugins/implementation/skills/implement/SKILL.md:52), a base sync before the first check (:93), a revert rule for a block that leaves the failing check unchanged, to the start of a declared planned-breakage span inside one (:116), the default delivery orders (:130), design signals (:153) and a smaller-first redesign on major divergence (:172), and a completion reply listing choices made and open decisions (:262); in the feature steps, delete-first, a domain structure before a growing conditional, remove before refining and a contract-change sweep (plugins/implementation/skills/implement/context/feature.md:9,12,14,26); in the fix steps, the focused fix first with siblings in later commits and red-before, green-after evidence (plugins/implementation/skills/implement/context/bugfix.md:13-14). Re-checked at 23e4138: holds. Re-checked at 118e73b0e1: holds; principle-never-block-on-the-human flattened its pattern and boundaries into one list with the same items, and the other two are unchanged. |
none |
principle-test-behavior-not-implementation |
testing:write, testing:audit, testing:test-value |
Partial | Re-checked at 23e4138: holds. Taken: testing:write's per-cycle checklist ends with the returns-nothing check (plugins/testing/skills/write/context/write.md:88). testing:audit reports not.toThrow(), toBeGreaterThan(0) and a > 0 comparison asserted true under rule-weak-oracle (plugins/testing/skills/audit/adapters/js-jest.yaml:63-74), and fixture against fixture under rule-constant-restatement (plugins/testing/skills/audit/scripts/cant-fail-scan.awk:710-782). Both stay report-only under --strict (plugins/testing/skills/audit/SKILL.md:63). *.test-d.ts is never scanned, and a relation across rows is a cant-fail-ok reason (plugins/testing/skills/audit/SKILL.md:155-164). Rejected: a gate under --strict. Re-checked at 118e73b0e1: holds. Upstream shortened its five shapes into one list and dropped toBeInstanceOf, toHaveLength(0) and not.toBe(wrongValue) from the examples; keeping relation checks across rows and *.test-d.ts files remains. Re-checked at 88d47514b8 (cursor/plugins#633, landed by 99743e9c15): changed. The returns-undefined check and the list of failing shapes became two imagined edits: break the behavior, and the test must fail; change only the internals, and it must still pass. The contract, not the assertion's shape, decides what counts as behavior. Our returns-nothing hollow-out run is now ours alone (plugins/testing/skills/write/context/write.md:88), and the second edit is the resistance-to-refactoring pillar testing:write already scores (:95). An absence asserted beside a case where the effect happens, and invariants across tables or at compile time, remain upstream, matching our relation-check and *.test-d.ts exemptions. Nothing taken moves. testing:test-value now takes the paired-presence fix and the never-called shape as judgment-only checks. Decided 2026-11-12, landed in #15917: the exact-error-message line, an assertion that pins message wording when callers depend only on the error's type (pstack/skills/principle-test-behavior-not-implementation/SKILL.md:18), is a judgment-only change detector in testing:test-value, not a weak oracle: assert the error type, and keep a user-facing or published message with the existing cant-fail-ok reason (plugins/testing/skills/test-value/SKILL.md:187-191); no audit rule, and the eval case is plugins/testing/evals/tv-exact-error-message-detector/. |
none |
orchestrate |
session-flow:orchestrate, implementation:implement-dispatch, session-flow:keep-going |
Partial | Taken: the drain preset, as drain_cadence in implementation (on-arrival by default, or batched), which decides when implementation:implement-dispatch reads worker returns (plugins/implementation/reference/config.md:12). In implementation:implement-dispatch: an unfillable brief item holds the row back, leaves the unit unscoped and undispatched, and routes as a divergence (plugins/implementation/skills/implement-dispatch/SKILL.md:116); phase verdicts are recorded against the head SHA, and a rebase or new commit before [DONE] voids them (:323). In session-flow:orchestrate: auditing one sampled brief per wave beside the wave, never as a gate in front of it, in the tiered-delegation section (plugins/session-flow/skills/orchestrate/SKILL.md:262-265); landing each unit as a local commit as soon as it is verified, never held for the end of the run (imperative 4, :103-105); and reading a writing worker's liveness from its side effects, never by messaging it, with a stuck worker stopped and respawned (imperative 4, :96-100). Declined: the roughly 70% budget cutoff for stopping new spawns, because imperative 1 forbids estimating the remaining window and imperative 4 applies no budget-based cutoff. The retry ladder is in session-flow:keep-going: a retry is matched to why the work failed, at most two per item, then the item is abandoned and the rest replanned (plugins/session-flow/skills/keep-going/SKILL.md:92-104), and lost work is found again by its branch, PR, worktree and commits, not by an agent id (plugins/session-flow/skills/keep-going/SKILL.md:62-64). Re-checked at 23e4138: holds; unchanged at e43c7ee26e. Re-checked at 88d47514b8: holds. Graphite is gone: the frontier is computed from the PR base chain through the forge, one stacker per stack owns every topology change, and workers never retarget or merge; the coordinator line gained the subagent model-rule line. The 70% spawn cutoff we decline remains, and every take is unchanged. Re-checked at ae5650febf: changed. The playbook was rewritten: its state CLI, store layout, brief template, drain rules, stacker role and 70% spawn cutoff are gone. Still upstream: a verdict per head SHA, the sampled brief check, reading a worker from its side effects and never messaging it to check, and retries by cause capped at two. Now ours alone: drain_cadence, the full-brief refusal (upstream keeps only a rule for missing evidence), landing each unit as soon as it is verified, and finding lost work by branch and PR. The 70% decline is moot. New candidate, not taken: forking a worker when new work needs its built state; owner to decide in session-flow. One owner per PR, merging on a non-author verifier's pass, sits with the autopilot-full row's decline. Decided 2026-11-12: a verifier that is not the author and runs on another model family for every PR (pstack/skills/poteto-mode/playbooks/orchestrate.md:13) is not taken as mandatory. Non-author is built into verification:confirm, whose outcome verdict comes from a fresh-context agent that did not produce the change (plugins/verification/skills/confirm/SKILL.md:195); another vendor is a setting tied to risk class, cross_vendor_required in the verification topology's floors (plugins/autonomy/reference/guardrails/verification-topology.md:99-126), with /review:council adapters for a cross-vendor panel. No work item. |
none |
feature, bug-fix |
planning:plan, implementation:implement-dispatch, implementation:implement, debugging:debug |
Partial | Taken: before the fix commit, debugging:debug takes out every edit made for a hypothesis now marked refuted (plugins/debugging/skills/debug/SKILL.md:250) and puts the failing and passing loop runs in the PR description (:269); code_writing in implementation (inline by default, or dispatch) (plugins/implementation/reference/config.md:13). Brief item 15 in implementation:implement-dispatch, a Structure: line naming the data structure the new logic is built on (plugins/implementation/skills/implement-dispatch/SKILL.md:114). planning:plan Step 4.5 opens a multi-phase plan with a four-row table (gates, shared state, disjoint streams, worker count); a row that does not apply reads n/a: <reason>, and shared state is split before phases are serialized (plugins/planning/skills/plan/SKILL.md "Step 4.5: Execution-Shape Analysis", the four-row table; line hint :284-291). Re-checked at 23e4138: holds. Re-checked at 88d47514b8: holds; the only change to each playbook is the subagent model-rule line, machinery we drop. Re-checked at bf0c7cd80a: holds; both playbooks trace the subsystem inline instead of running how, and each delegate's default model changed. The rules we took are unchanged. |
none |
prototype |
prototype:explore-directions |
Partial | Taken: an optional reference pass, offered while the look is still open, that seeds two to four directions from comparable screens (plugins/prototype/skills/explore-directions/SKILL.md:194-200), and a handover that pairs each variant with a screenshot when /playwright:playwright is available and closes with the variant we would start from (:247-263). Re-checked at 23e4138: holds. |
none |
authoring-a-skill |
playbooks:skill-authoring |
Partial | Taken: the reply shape, as a three-part closing report for any run that creates or changes a skill (plugins/playbooks/skills/skill-authoring/SKILL.md:219-228). Rejected: skipping subjective tests and dropping the reason behind a rule; we keep both. Re-checked at 23e4138: holds. |
none |
eval |
evals:design, evals:plugin-eval |
Partial | Rejected: a runner and a mandatory cross-family judge. Re-checked at 23e4138: holds. Taken: evals:design case authoring states the user's goal, keeps evaluation words out of prompts, fixture paths and slugs, and never grades from the model's own account (plugins/evals/skills/design/SKILL.md:87-97). evals:plugin-eval spot-reads outputs against verdicts before trusting a delta and ends with a promote or do not promote line (plugins/evals/skills/plugin-eval/SKILL.md:360, :383). Re-checked at 88d47514b8: holds; the only change is the subagent model-rule line, machinery we drop. |
none |
runtime-forensics, trace-forensics |
debugging:analyze-profile, debugging:debug |
Adopted (new skill) | Taken: the capture-by-symptom table is guidance for a control skill's ## Capture (plugins/app-control/reference/control-skill.md), and /app-control:drive capture records the artifact; debugging:analyze-profile turns a trace, CPU profile or heap snapshot into a SQLite file it queries read-only, and hands a large read to a subagent that returns only the result (plugins/debugging/skills/analyze-profile/SKILL.md:40-91). Diagnose-only is a fixed rule, not a setting: the skill reports the mechanism at file:line and edits nothing (:88). Re-checked at 23e4138: holds. Re-checked at 118e73b0e1: holds; trace-forensics reworded its paired-capture step, and runtime-forensics is unchanged. Re-checked at 88d47514b8: holds; the only change to each playbook is the subagent model-rule line on its parsing step, machinery we drop. |
none |
session-pickup |
session-flow:find-handoff, session-flow:keep-going |
Partial | Taken: transcript_scope in session-flow (worktree by default, repo, all, the narrower of the user and repository values winning), which session-flow:find-handoff and session-flow:recall read before a transcript scan (plugins/session-flow/reference/config.md:18, plugins/session-flow/skills/find-handoff/SKILL.md:111). session-flow:keep-going names one of four verdicts for another agent's trail (continue, ship the finished recommendation, ratify, restart) and reports what was inherited versus redone (plugins/session-flow/skills/keep-going/SKILL.md:118-140). Re-checked at 23e4138: holds. Re-checked at 88d47514b8: holds; the only change is the subagent model-rule line, machinery we drop. |
none |
visual-parity |
testing:check-visual-parity, implementation:implement refactor mode |
Adopted (new skill) | Taken: testing:check-visual-parity (plugins/testing/skills/check-visual-parity/SKILL.md:19, baseline never edited; :95 compare) and refactor mode's parity step (plugins/implementation/skills/implement/context/refactor.md:17), with a pixel_tolerance map in testing.yaml. Re-checked at 23e4138: holds. Refactor mode never edits a test, recorded baseline or harness to turn a run green and asks the user instead (plugins/implementation/skills/implement/context/refactor.md:21). |
none |
correct |
corrections:codify, the review plugin's enforcement ladder, toolchain:ratchet |
Not taken (skill omitted) | New at e43c7ee26e. Decision: no skill, and nothing taken. Upstream reads a repository's history for mistakes agents keep repeating, groups them into classes once a class recurs, fixes each class at the strongest of five levels (structure, types, a lint whose message names the fix, tests, written rules last), and keeps a table that names the enforcing check beside every rule. Each part already has an owner here. The level order is our enforcement ladder, strongest first from make-impossible to human-review (plugins/review/context/enforcement-ladder.md:19-89). corrections:codify maps one described mistake, or each row of a findings file, to the strongest rung that can decide it and proposes make-impossible without building it (plugins/corrections/skills/codify/SKILL.md:72-92); the ladder's guardrail criterion asks for a message that names the fix (plugins/review/context/enforcement-ladder.md:120-121). Encoding a repeated correction is corrections:codify's selection rule on the same ladder (plugins/review/context/enforcement-ladder.md:146), and failing a change only for new instances of a widespread pattern is toolchain:ratchet's count ceiling (plugins/toolchain/skills/ratchet/SKILL.md:2). Not taken, with reasons: building the fix inside the audit, because corrections:codify proves a proposed check against the old code but never implements a rung, leaving the build to the rung's owner (plugins/corrections/skills/codify/SKILL.md:116-140,180-181); a rule-to-enforcer table in the agent instructions, because our AGENTS.md lists only rules CI does not enforce (AGENTS.md:93) and prose that repeats what a check already shows is removed (plugins/review/context/enforcement-ladder.md:163-166). Mining commits and review comments, first declined here, is now corrections:codify's count over recent commits and the session (plugins/corrections/skills/codify/SKILL.md:39-70) and its reviews action over merged pull requests' review comments (:31). Re-checked at 88d47514b8 (cursor/plugins#633, landed by 99743e9c15): holds; only the test level changed, which now fixes a test that passes when its behavior breaks or fails when only internals change (see the principle-test-behavior-not-implementation row). The five levels, the class grouping and the rule-to-enforcer table are unchanged, so the owners named above stand. |
none |
ketchup |
session-flow:orient catch-up |
Partial | New at 118e73b0e1. A typed-only skill that briefs the user on what happened in a chat after they last wrote in it: it re-checks each pull request, CI run and agent the chat touched, opens with the number of action items, and gives each a recommendation and what happens with no answer, shown as a before and after. No decision is recorded. Closest owners: session-flow:orient, which reads durable state rather than the chat since one message, and adhd:clarify, which reshapes one message. Re-checked at 88d47514b8: changed: it now explains each change in plain words with the screenshots, videos and small diagrams the work produced, keeps the story under a word budget, and lists every item waiting on the user with a pick, the evidence and the risk, so each call can be decided from the reply alone; it never drops an item to meet a word count. The autopilot-full and autopilot-stack ticks and replies now end with it. Still no decision; the closest owners are unchanged. Corrected 2026-11-12: "not taken" was wrong. /session-flow:orient catch-up already implements this skill: anchored on the user's last own message, it re-checks each pull request, CI run and agent the chat touched, opens Needs you with the number of action items, and gives each its options, recommendation and what happens with no answer (plugins/session-flow/skills/orient/SKILL.md:78-117). Decided 2026-11-12: its word cap is already a setting (:86-88) but covers the whole reply, so the never-cut-an-item rule is taken: no action item is dropped or merged into a count to meet the cap (pending adoption: session-flow:orient, #15834). Declined: embedding produced screenshots and videos in the catch-up; a visual stays one built from what is in hand (:116-117), and a video is an offer (see the Map's ketchup-video line). |
none |
poteto-help |
session-flow:recommend-skills, session-flow:workflow |
Partial | New at 118e73b0e1. A typed-only help router: it maps a question about pstack to the skill, playbook or guide page that answers it, hands back a prompt to send rather than starting the work, and asks about /setup-pstack when no model rule exists. No decision is recorded. Closest owners: session-flow:recommend-skills and harness-ops:inventory. Re-checked at 88d47514b8: changed: it routes to the four skills new at 88d47514b8, mentions the formal-verification pair when a question is about a race or an invariant, and says pstack also installs in Claude Code under a pstack: prefix (see the docs/harness/claude-code.md row). Still no decision; the closest owners are unchanged. Re-checked at ae5650febf: changed: it routes to ketchup-video and statechart and rewords the no-merge autopilot prompt. Still no decision; the closest owners are unchanged. Decided 2026-11-12, landed in #16359: already have; /session-flow:recommend-skills answers a free-text question with one skill and an example prompt. The one part taken is the map from a kind of task to its recipe (her poteto-mode playbooks): which stages a kind of task enters is the workflow skill's job, so it is one "Task kinds" table in the workflow skill's steps reference, naming no skills (plugins/session-flow/skills/workflow/context/steps.md:226-242), which the generated cheat sheet joins with each skill's stage and arguments in its "By kind of task" table (docs/skill-cheat-sheet.md:43-54), in place of the hand-kept "Start from your problem" table; the eval case is plugins/session-flow/evals/wf-bug-takes-task-kinds-stages/. Re-checked at bf0c7cd80a: holds; one sentence no longer names /why among the skills that spawn subagents with per-role models. The task-kind map we took is unchanged. |
none |
in-your-own-words |
playbooks:fable-5 problem framing, work-items:work |
Partial | New at 88d47514b8. A typed-only skill whose whole body asks the agent to restate, in its own words, the user's goals and the problem they are solving. No decision is recorded. Closest owner: planning:interview, whose rounds bring a task to shared understanding before work starts. Decided 2026-11-12, with the poteto-mode goal restatement: already owned. playbooks:fable-5's problem-framing chapter restates the outcome, not the request (plugins/playbooks/skills/fable-5/context/problem-framing.md:7-13), and work-items:work restates an item as "I read this item as" before claiming it. The one gap, naming the smallest change expected, is taken into that restatement, which counts only a change that removes the cause (plugins/playbooks/skills/fable-5/context/problem-framing.md:11). |
none |
sharpen |
none | Not adopted (deferred; see below) | New at 88d47514b8. Deferred under #7566. Re-checked at bf0c7cd80a: changed; the last sentence now has the rewritten prompt ask for a prototype of any choice an experiment can settle and a recommendation from what it showed, leaving only product and preference calls to the user. Reject: the deferral stands, because its reopen trigger, a repeated ask in the owner's sessions, has not fired. | none |
create-formal-verification-skill, maintain-formal-verification-skill |
none yet | Pending adoption (new formal-verification plugin) |
New at 88d47514b8. A formal-methods skill is parked by tlaplus-agentskills.md. Re-checked at ae5650febf: holds; a generated skill is now told to read TLA+ results with statechart, which is parked with this pair (see below). Decided 2026-11-12 (map #15651): the park is lifted. A new formal-verification plugin absorbs her three formal-verification skills, this pair and statechart, in three slices, each with its interview, and their pull requests update this row in detail: the chart skill, setup and consumer contract, proven on the loop-lane claim protocol, the specify skill's propose, create and check forms, and the specify skill's refresh action. |
none |
docs/harness/claude-code.md |
none | Not taken (nothing to reauthor) | New at 88d47514b8. An adapter page that maps pstack's Cursor tool names, subagent fields, model slugs, paths and Cursor-only features to Claude Code, so pstack itself runs there as /pstack:<name>. Every entry on this page is a reauthored skill of ours, not a run of an upstream skill, so the mapping applies to none of them. Re-checked at bf0c7cd80a: holds; only the default-model mapping lines changed. |
none |
Principle parts check (the principle-* rows above). Decision 58's 36 parts: 28 placed in the
skill that makes each decision, 2 recorded (1b, 2h), 6 deferred, and unplaced parts: 0. The read
at origin/main b000bb5ce7 (#12155) found 25 placed and 9 deferred; #12716 then placed 4c, 4e and
4f in /corrections:codify:
- 4c:
plugins/corrections/skills/codify/actions/note.md, Consent section. - 4e:
plugins/corrections/skills/codify/SKILL.md, section 3 (Place the lesson by scope). - 4f:
plugins/corrections/skills/codify/actions/note.md, opening paragraph.
The deferred: list:
- 1c, 3h, 6h, 8j, 19d: gated by decision 58, no stumble or eval evidence.
- 24i: gated per cursor/plugins#634 (Q557).
This port admits no inference from the code's own shape, at any tier. This is where we depart
from why's
inference tier as it stood before bf0c7cd80a. At bf0c7cd80a upstream's one-paragraph why says to
answer from the record, not the code, so upstream no longer holds the tier we departed from; the row
above records the re-check.
/discovery:trace-intent excludes code shape from the scale and records it as a gap. The reason is
operational. Code is the one source that is always present and free to read, so a weak but
admissible rung for it gets used exactly when the real record is thin, which is when the reader most
needs to be told the record is thin.
Version-control behavior is not code shape and stays admissible at Inferred. Change coupling,
churn and hotspot data locate a relationship the code alone cannot show, but they do not explain the
decision behind it, so they never reach Direct.
These sections give the reasoning behind the rows above, so a later reader can see what was considered and declined.
Four ideas from
arena
were absorbed (see the row above), and competing attempts for a plan phase now run inside
implementation:implement-dispatch as a setting. We ship no standalone runner skill for competing
candidate solutions, because of the Rule of Three. Re-checked at ae5650febf: the upstream changes
are runner and judge fallbacks, so this reasoning stands. Re-checked at bf0c7cd80a: the changes are
default runners and a vendor rule for the judge, so it still stands.
There is no second consumer. Every candidate-competition site here was read first:
naming:name-it-better: blind generators from distinct lenses, plus independent judges intournamentmode.architecture:improveDesign-It-Twice: three or four subagents under orthogonal design constraints; the parent compares, with no judge.prototype:explore-directions: no fan-out. It builds variants and a switcher, and a person clicks through them.implementation:implement-dispatchCompeting attempts: one worker per stated constraint, each in its own worktree with a phase-verifier, and the plan's selection rule; no judge.
The two real consumers disagree on what a shared runner would have to fix: blind versus constrained
generation, independent judges versus none, elimination rounds versus one comparison. A runner
spanning that gap would make all of it configurable, which moves complexity rather than removing it,
and migration-playbook.md calls extraction without a second consumer speculative generality.
Every trigger it would claim is taken. Competing names reach naming, competing interfaces
architecture:improve, competing layouts prototype:explore-directions, and competing implementations of a plan phase
implementation:implement-dispatch. A skill with no trigger
vocabulary of its own would cost a listing line and never run.
The precedent for this shape is absorb. The Pocock record shows an interface-design skill folded into Design-It-Twice rather than shipped alone.
- No general runner. A task that is not a name, an interface or a layout has
session-flow:orchestrate's posture andmulti-agent:assess's routing, but nothing that runs the loop. - No comparative judgment in general. Every judging site here checks one artifact against a standard; only naming's tournament picks among several, and only for names.
Recheck trigger: a second real consumer asks for fan-out over an arbitrary task. Route it through
multi-agent:assess then, not a new skill.
The record first omitted recall, citing four session-flow incumbents (orient,
find-handoff, reanchor, reconcile) and declining three absorbs. That decision is reversed:
session-flow:recall now rebuilds a topic's
history from past sessions in the current project, which none of the four does, since orient
reads durable state only.
Two of the earlier declines still hold and carry into the new skill:
- The closed status-tag set stays out.
source-control/skills/worktree/context/status.mdalready defines a closed two-axis vocabulary, and a second, differently worded set would be a silent second way. - Cutting detail before cutting threads stays with
session-flow:recommend-skills, which already states it.
The scope statement before a scan, declined earlier for want of observed stumble evidence, now has
a consumer: the recall
skill's transcript scan. The cross-project privacy question is settled by the transcript_scope
setting, which recall shares with find-handoff, rather than by a fixed rule, because find-handoff scans other projects deliberately when
asked.
education:teach and education:explain hold the trigger vocabulary, so no third skill. Two
mechanics were absent and now sit in education:teach's lesson contract:
- Incremental diagrams. The lesson contract asks for a sequence of diagrams for a system with several interacting parts, growing by one component per step.
- Inventories are not lessons. The contract rejects a lesson whose body is mostly an inventory of identifiers, because the learner gains no model of how the parts behave.
Both were re-homed on audit. The diagram rule was planned for visualization:visualize, which
routes form and medium and does no comprehension work; the names rule was planned for
education:explain, which never lists functions. The lesson unit is where a reference dump happens.
Corrected: this section said the marketplace has no how. discovery:explore and
discovery:trace-intent are that pair, and education:explain now calls both for a codebase subject.
Reversed: the anti-pacing rejection. education:explain now bars checks on the reader, labels
naming a sentence's role and commentary on its own tempo; education:teach is unchanged. An
explain_starting_rung setting lets a user or a repository start explain at peer level.
Still declined: the image-generation branch (our visualization skills have no image-generation
route) and the voice paragraph (docs-hygiene:write-for-humans owns voice guidance here).
Opposition noted: education:teach keeps coaching with one question at a time and with
quizzes; the row's Conflict cell records that decision. The absorbed rules are the ones that hold
under it.
We ship no
tdd
skill (decision in the row above): debugging:debug phase 5 covers the bug-fix moment, and a skill
that fires only when named would not earn a listing line. Most of the file is owned
twice: tdd:principles sources the bad-test definition from Khorikov and testing:audit enforces
it; implementation:implement already forbids gaming tests.
Two things survived:
- The cost branch, into
testing:write's decline list. That list covered code needing no test; this one covers a test not worth writing. Declining requires naming the trigger and the substitute check, a rule upstream no longer states and we keep. - Red for the right reason, into
debugphase 5. A test that fails on a typo is red too, and making it green does not touch the bug.
Re-homed on audit: both were planned for debug, but testing:write already owned a decline
list. Reopened and taken: the before-and-after evidence report, earlier rejected as marginal, is
now part of debugging:debug phase 6.
session-flow:retro and session-flow:running-retro own this axis, and reflect claims no trigger
they lack. Three of its mechanics were declined: the structural-enforcement check
(harness-config:audit-automation-gaps owns it with a default-reject posture), the
capture-don't-apply gate (stated twice already, and its backlog items are filed automatically where
running-retro never does), and the missed-trigger finding class (discipline:use-your-skills
owns it).
The gap closed: both retro skills produce skill candidates and named no destination. Both now
hand one to /playbooks:skill-authoring, gated on /skill-quality:check, presence-gated with a
fallback.
The recorded dissent is now resolved. An audit argued for absorbing the three orthogonal lenses;
the record declined them for lack of observed stumble evidence. That decision is reversed:
session-flow:retro session mode reads retro_lenses, 1 by default; at 3 it gives the
analysis to three subagents run side by side and merges their rows before scoring.
We take nothing from
bro:
restating the last reply more plainly already ships here twice:
education:explain, whose empty argument targets the previous reply, and discipline:wait-what,
the interjection-fired re-pitch. discipline:tighten-your-output covers the brevity half.
wait-what holds an argued entry on the naming grammar's closed exception list
(docs/plugin-philosophy.md); a bro skill would need a second entry with the same argument.
Nothing was absorbed and nothing is lost. Re-checked at 23e4138: the upstream file is unchanged.
We ship neither a skill nor a convention for
show-me-your-work;
the row above lists what we took and where it went.
The lane first proposed absorbing it into session-flow:running-retro. That failed because the
retro ledger is defect-shaped and written after the fact by a transcript-reading subagent, and it
cannot produce a result cell. The trail we wanted is decision-shaped and written by the acting agent at
decision time.
An audit then argued for a shared capture convention across eight audit-trail surfaces. Checked one by one, only one is a decision trail:
| Surface | Is it a decision trail? |
|---|---|
implementation:implement-dispatch DEVIATIONS.md |
Yes: planned versus done, why, blast radius, written by the acting agent at the time |
session-flow:handoff settled decisions and abandoned approaches |
Related, different shape: written once at pause time for the next session |
code-tidying:tidy's PR follow-up comment |
No: a change inventory with no decision or reason |
work-items:work-loop loop-state@2 |
No: counters and loop state |
work-items:work-loop escalation record |
No: a notification on the governed escalation schema |
work-items:attend-queue lane telemetry |
No: one status comment per lane, edited in place |
autonomy transition telemetry |
No: OTel spans, with no parallel schema |
autonomy return-accounting |
No: it captures return, not activity |
One real adopter, two if handoff counts, and those two disagree on the point a shared format would
fix: append at decision time versus synthesize at pause time. A convention with one adopter would be
a shallow module.
So the substance went to its one consumer. DEVIATIONS.md gained append-and-supersede,
evidence as a pointer (preferably one a committed script produced), an explicit outcome or
unverified, and one entry per decision. The formula-injection guard went separately into
harness-ops:audit-install-state, whose CSV writer passed names read from disk straight into a file
the reader opens in a spreadsheet. It is fixed, with a test that fails when the guard is disabled.
Still rejected: the per-reply summary section (a session-wide output posture with one example
here, adhd:shape) and a second log-versus-transcript self-audit (the loop lanes already
ground every cycle claim in a tool result). Reversed: the mandatory cross-model review; a
fresh-context review of the trail now runs in implementation:implement-dispatch for autonomous runs
(plugins/implementation/skills/implement-dispatch/SKILL.md:228-235).
Recheck trigger: a second surface starts appending decisions at decision time. At two consumers
that agree on shape, evaluate an owner doc under docs/conventions/, diffing against the table
above.
plugins/code-tidying/skills/tidy/lanes/docs-prose.md edits docs/**.md and README.md prose
under six structural patterns. Adding Diátaxis, Google-style or Global-English patterns to it was
declined: the lane edits files and cannot resolve the consuming project's own style guide, so it
would enforce a guide the project may have rejected. The style question stays with
docs-hygiene:write-for-humans, which resolves the guide first.
sharpenasks, before rewriting a prompt, what a top professional would add, the rawest material they would start from, the methods and tools they would use, and what outside check would show the result is right; it returns only the rewritten prompt. Sincebf0c7cd80athat prompt asks for a prototype of any choice an experiment can settle and a recommendation from the result, and leaves only product and preference choices to the user. It is deferred under #7566, which records that, if reopened, it ships as a one-paragraph, typed-onlydisciplineskill. Reopen when the owner's sessions show a repeated ask for it.create-formal-verification-skillgenerates a project-local skill that verifies one property where a bug costs most and tests reach least, chooses the tool by the question (TLA+ by default for whether events can happen in a bad order), shows each property failing on a broken variant, and records the code it covers with a sync commit.maintain-formal-verification-skillbrings such a skill back in line after the covered code changes and labels each check in sync, updated, bug or blocked. A formal-methods skill is parked:tlaplus-agentskills.mdrecords that no plugin here ships model checking or TLA+ authoring. Reopen on that page's recheck trigger (a loop-lane concurrency bug a model of the lane protocol would have caught, or a consumer asking for model checking), and read both upstreams at their then-current HEAD before deciding. Re-checked at ae5650febf: the first now tells a generated skill to read TLA+ results withstatechart; the parked decision stands. Decided 2026-11-12: the park is lifted for this pair andstatechart; see the pair's row for the threeformal-verificationplugin work items.statechart(linked in the Map) checks a TLA+ model and draws it as a chart with the shortest path to each state that breaks a rule. It is parked with the pair above, under the same page and trigger, and needs a TLC, Java and renderer install we have no consumer for. Read at ae5650febf: itsSKILL.md,references/render.md,package.jsonand launcher. Its compiler sources, tests and corpus were listed, not read, because the decision rests on what the skill is for. Decided 2026-11-12: it becomes theformal-verificationplugin's chart skill (#15648).
We credit why for applying evidence discipline to a software repository's own record. The rule
against inferring intent from code is ours in its hardened form (see the departure above).
discovery:trace-intent recovers rationale that nobody recorded. Recording rationale when a
decision is made is architecture:record-decision's job, so we treat design-rationale methods
(ADRs, IBIS, QOC, DRL) as outside this skill's scope rather than as its prior art.
We keep our own intent-evidence tier names. We considered ICD 203, GRADE and the Admiralty code as substitutes and adopted none, because our scale measures distance from a stated intent and we chose not to map it onto another axis.
One line per unit at the pin: 58 skills, 23 poteto-mode playbooks, 2 agents, the benny
automation pack and the Claude Code harness page. A unit with a row above names its owner there; this map names the owner or
nothing taken.
architect:planning:design,implementation:implement,architecture:improvearena:architecture:improve,prototype,naming:name-it-betterautomate-me:session-flow:audit-sessions,session-flow:retrobenchmark-checklist:performance:climb,performance:verifyblast-radius:review:quality-gatebro: nothing takencorrect: nothing taken (owners named in its row)create-formal-verification-skill: pending adoption, theformal-verificationplugin (see its row)create-verification-skill:app-control:map createfigure-it-out:planning:plan,implementation:implement,implementation:implement-dispatchhow:discovery:explorein-your-own-words:playbooks:fable-5,work-items:work(see its row)interrogate:review:fanout,review:code-reviewer,review:quality-gateketchup:session-flow:orientcatch-up (see its row)ketchup-video: decided 2026-11-12, taken as a one-line offer insession-flow:orient's catch-up, presence-gated like its existing visualize hand-off (pending adoption: #15834)maintain-formal-verification-skill: pending adoption, theformal-verificationplugin (see its row)maintain-verification-skill:app-control:map refreshmake-bot-ui: nothing taken (Cursor runtime only)no-comments:code-tidying:remove-comments,code-metrics:audit-suppressionspoteto-help:session-flow:recommend-skills,session-flow:workflow(see its row)poteto-mode:session-flow:orchestrate,session-flow:workflowprinciple-attack-the-premise:debugging:debugprinciple-boundary-discipline:implementation:implement,planning:design,review:code-reviewer,architecture:improveprinciple-build-the-lever:discipline:script-the-deterministic-work,implementation:implement-dispatchprinciple-encode-lessons-in-structure: enforcement-ladder convention,session-flow:retro,corrections:codifyprinciple-exhaust-the-design-space:planning:designprinciple-experience-first:planning:prd,planning:design,testing:run-e2eprinciple-explain-the-number:performance:verifyprinciple-fix-root-causes:debugging:debugprinciple-foundational-thinking:planning:design,planning:planprinciple-guard-the-context-window: nothing taken (covered byplaybooks:fable-5andsession-flow:orchestrate)principle-laziness-protocol:implementation:implement,architecture:improveprinciple-make-operations-idempotent:planning:devils-advocate,review:code-reviewer, loop-lane conventionprinciple-migrate-callers-then-delete-legacy-apis:implementation:implement,playbooks:fable-5principle-minimize-reader-load:architecture:improve,review:code-reviewerprinciple-model-the-domain:implementation:implement,review:code-reviewer,architecture:improveprinciple-never-block-on-the-human:playbooks:fable-5principle-outcome-oriented-execution:planning:plan,implementation:implementprinciple-prove-it-works:verification:confirmprinciple-redesign-from-first-principles:planning:design,implementation:implement,planning:planprinciple-separate-before-serializing-shared-state:coupling:reduceprinciple-sequence-verifiable-units:implementation:implementprinciple-subtract-before-you-add:implementation:implementprinciple-test-behavior-not-implementation:testing:write,testing:audit,testing:test-valueprinciple-type-system-discipline:planning:design,review:code-reviewerrecall:session-flow:recallreflect:session-flow:running-retro,session-flow:retrosetup-pstack:multi-agent:routesharpen: nothing taken (deferred; see its row)show-me-your-work:implementation:implement-dispatch,harness-ops:audit-install-statestatechart: pending adoption, theformal-verificationplugin's chart skill (see the formal-verification pair's row)swarm:session-flow:orchestrate,multi-agent:assess,multi-agent:routetdd:testing:write,debugging:debugteach:education:teach,education:explaintechnical-writing:docs-hygiene:write-for-humanstypescript-best-practices:planning:design,review:code-reviewerunslop:ai-slop:auditwhy:discovery:trace-intentauthoring-a-skill(playbook):playbooks:skill-authoringautonomous-run(playbook):implementation:implement,planning:draft-goal-conditionautopilot-full(playbook):review:fanout,implementation:implement-dispatch,work-items:work-loopautopilot-stack(playbook):source-control:pull-requestbabysit(playbook):source-control:babysit-prs,source-control:pull-requestbug-fix(playbook):debugging:debug,implementation:implementeval(playbook):evals:design,evals:plugin-evalfeature(playbook):planning:plan,implementation:implement-dispatch,implementation:implementhillclimb(playbook):performance:climb,performance:goal,performance:snapshotinvestigation(playbook):discovery:exploremulti-phase-plan(playbook):planning:plan,source-control:pull-request,verification:confirmopening-a-pr(playbook):source-control:pull-requestorchestrate(playbook):session-flow:orchestrate,implementation:implement-dispatch,session-flow:keep-goingpause-safely(playbook):session-flow:handoffperf-issue(playbook):performance:target,performance:climbprototype(playbook):prototype:explore-directionsrefactoring(playbook):implementation:implement,verification:confirmruntime-forensics(playbook):debugging:analyze-profile,debugging:debugsession-pickup(playbook):session-flow:find-handoff,session-flow:keep-goingshipping(playbook):source-control:pull-requestmergetrace-forensics(playbook):debugging:analyze-profilevisual-parity(playbook):testing:check-visual-parity,implementation:implementrefactor modeworktree-cleanup(playbook):source-control:worktreecomment-sicko(agent):code-tidying:remove-commentspoteto-agent(agent):session-flow:orchestratebenny(automation pack): nothing takendocs/harness/claude-code.md(harness page): nothing taken (see its row)