axstack 0.21.0 → 0.23.0

package/README.md

@@ -14,7 +14,8 @@ coordination. You can start at the phase you need.
 
 | Layer | Skill | What it does |
 | --- | --- | --- |
-| Plan | [axstack-align](skills/axstack-align/SKILL.md) | Settle scope through questions, a design lens sketch, and a four-family arena for hard choices. |
+| Plan | [axstack-align](skills/axstack-align/SKILL.md) | Settle scope through questions, a design lens sketch, and inline brainstorm validation. |
+| Plan | [axstack-brainstorm](skills/axstack-brainstorm/SKILL.md) | Validate an approach with independent candidates; judges for hard choices. |
 | Plan | [axstack-spec](skills/axstack-spec/SKILL.md) | Write and approve an observable specification. |
 | Plan | [axstack-tickets](skills/axstack-tickets/SKILL.md) | Break approved scope into executable tasks. |
 | Build | [axstack-implement](skills/axstack-implement/SKILL.md) | Build with strict TDD and an author → review → repair loop. |
@@ -22,6 +23,7 @@ coordination. You can start at the phase you need.
 | Verify | [axstack-review](skills/axstack-review/SKILL.md) | Review a PR or bounded codebase at an exact revision. |
 | Verify | [axstack-improve](skills/axstack-improve/SKILL.md) | Find evidenced codebase improvements without editing code. |
 | Verify | [axstack-audit](skills/axstack-audit/SKILL.md) | Measure a run's outcomes and evidence gaps. |
+| Verify | [axstack-correct](skills/axstack-correct/SKILL.md) | Report repeated mistakes and propose stronger checks when invoked by the user. |
 | Operate | [axstack-watch](skills/axstack-watch/SKILL.md) | Observe or maintain an existing PR within its authority. |
 | Operate | [axstack-cleanup](skills/axstack-cleanup/SKILL.md) | Retire eligible completed agent resources. |
 | Operate | [axstack-relay](skills/axstack-relay/SKILL.md) | Send an explicit message or authorized notification. |
@@ -36,7 +38,7 @@ implementation. Research, explanation, and peer review can start directly.
 
 | Failure mode | How Axstack responds |
 | --- | --- |
-| Wrong thing built | Align rounds clarify the request; a four-family arena compares approaches for hard choices. |
+| Wrong thing built | Align rounds clarify the request; every brainstorm runs a light arena across configured families. Use judges only for Rung 2 hard-to-reverse choices. |
 | Nobody really reviewed it | Strict TDD checks behavior first; with the mixed preset, cross-provider review checks the exact revision. |
 | Design rot | The design lens sketches boundaries before a build; Improve surfaces evidenced changes later. |
 | Agents left a mess | T3 makes delegation visible, one writer owns each PR, cleanup stays bounded, and a human merges. |

package/docs/installation.md

@@ -220,16 +220,20 @@ reconciles their findings.
 
 The mixed checker and `axstack-research-web-google` have provider
 `antigravity`; mixed `axstack-research-x` has provider `grok`. All three use
-`model: null` with notes authorizing their agent-ID routes; T3 resolves the exact
-model from the first entry for that provider in saved capabilities. Empty
-Antigravity model catalogs hold. The single-provider presets configure the checker and keep
+`model: null` with notes authorizing their agent-ID routes; T3 resolves Grok's
+exact model from the first entry for that provider in saved capabilities.
+For Antigravity `model:null` roles, select the first listed model whose ID ends
+with `-<effort>` from saved capabilities and record its exact ID.
+Missing effort-suffix matches hold resolution for Antigravity, including empty
+model catalogs. The single-provider presets configure the checker and keep
 both cross-provider research routes as intentional absences. Their
 unavailable adviser and round-2 seat remain explicit same-provider
 `model: null` roles, which do not make installation unready;
 Align and Spec still hold until both Astra and Opus can return independent
-receipts. For an arena-grade Align question, round 1 needs Opus; round 2, if
-invoked, needs escalation Fable and Astra; a required seat that is unavailable holds that
-round. The current chat drives on whatever
+receipts. Every brainstorm runs a light arena across configured families.
+Use judges only for Rung 2 hard-to-reverse choices: round 1 needs Opus; round 2,
+if invoked, needs escalation Fable and Astra; a required seat that is unavailable
+holds that round. The current chat drives on whatever
 model runs it; no preset carries a driver role. Every other missing, invalid, unsupported, or unavailable role value holds only
 the affected work. Codex and Claude class resolution reads the saved T3 capabilities catalog via
 `skills/axstack/scripts/resolve-models.js --provider`; missing or malformed

package/docs/workflows.md

@@ -18,6 +18,8 @@ and PR shape.
 Direct routes need no spec ceremony:
 
 - `axstack-research` answers one bounded source-backed question.
+- `axstack-correct` reports repeated mistakes and proposes stronger checks.
+  Only the user invokes it.
 - `axstack-explain` separates implemented, intended, tested, live, and unknown
   behavior; complex visuals receive exact-artifact QA where applicable.
 - `axstack-improve` returns a small ranked set of evidenced improvement
@@ -162,15 +164,17 @@ for explicit ownership transfer.
 - `axstack-align` maps facts and dependencies, asks prioritized questions, and
   consults Astra and Opus independently with the same bounded evidence and
   question. It synthesizes disagreements and reuses unchanged receipts. For a
-  hard-to-reverse design choice it runs an arena instead: Astra, Opus, Grok,
-  and Antigravity each author a candidate. `axstack-arena-judge-opus` scores
-  them in round 1; the driver compares its own pick with that verdict. If they
-  disagree on the base or the user rejects the round-1 synthesis,
-  `axstack-escalation-fable` and `axstack-arena-judge-astra` independently
-  score the same anonymized candidates and rubric in round 2. The driver
-  picks a base, grafts strong ideas, and records judge verdicts per round in
-  the `Decisions` rows without averaging. Fable escalation uses a fresh
-  session for round 2, high-stakes agreement, or the bounded trigger in
+  Rung 1 or 2 design question it loads `axstack-brainstorm` inline and reuses
+  its receipt instead of consulting twice; Align owns the interview.
+- `axstack-brainstorm` validates an approach standalone or inline in the driver,
+  report-only. Every invocation compares independent Astra, Opus, Grok and
+  Antigravity candidates, including premise and smallest-change/do-nothing
+  checks. The driver scores, picks and grafts; Rung 1 uses no judges. At Rung 2,
+  `axstack-arena-judge-opus` judges round 1; disagreement on the base or caller
+  re-invocation with the user's rejection triggers fresh Fable/Astra round 2.
+  It returns a verdict, sketch and proposed questions, with no interview,
+  prototype or execution approval. Required seats hold; optional dropouts are
+  fenced. Fable also serves the bounded triggers in
   [Standing contracts](../skills/axstack/references/contracts.md).
 - `axstack-spec` writes observable acceptance, exclusions, decisions, and one
   user-approved revision baseline. Linear through the executor MCP is the

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "axstack",
-  "version": "0.21.0",
+  "version": "0.23.0",
   "description": "Axstack installer and setup CLI: installs owned chat skills and role data, configures supported harness settings, and checks T3 Code capabilities.",
   "keywords": [
     "claude-code",

package/skills/axstack/references/automations.md

@@ -271,6 +271,15 @@ The orphan sweep covers the run record's repositories plus registered repositori
 `t3_thread_organize` settle/archive changes metadata only; exact guarded Git
 worktree removal remains separate. Unknown, active or user-taken-over threads,
 ambiguous publication and failed salvage stay preserved.
+
+A review-manager pass may retire a predecessor pass's local `t3code/*` branch
+with exact `git branch -D <branch>` only when ALL hold: the predecessor pass is
+settled and proven run/lane-owned by recorded identity, its descendants and
+evidence are settled, no worktree has the branch checked out, no remote
+counterpart exists, its tip equals the recorded tip, and its tip is an ancestor
+of the verified remote default branch. Otherwise hold; all other `git branch -d`
+rules remain unchanged.
+
 Past the authorized storage limit (default 20 retained lane worktrees), disable the schedule
 with `update_scheduled_task` using `enabled:false` and hold.
 Read back the disabled schedule with `list_scheduled_tasks`; uncertainty holds.

package/skills/axstack/references/candidate-publication.md

@@ -17,7 +17,16 @@ protection; a mismatch holds publication. If the push outcome is ambiguous,
 inspect remote state before retrying.
 
 Before reviewer dispatch, read the remote ref back and confirm that it resolves
-to the candidate SHA; also pin the current base. Record:
+to the candidate SHA; also pin the current base.
+After every push, before post-push diligence or merge-ready, compare the PR body's
+stated head SHA with `Confirmed remote SHA` and record `PR body head SHA: <sha or none>`.
+Accept `none` for a PR body without a stated head SHA.
+Accept a stated full head SHA only when it equals `Confirmed remote SHA`.
+Accept a stated abbreviated head SHA only when it is a matching prefix of
+`Confirmed remote SHA` with at least 7 hexadecimal characters.
+Hold on a mismatched head SHA or an abbreviation shorter than 7 characters.
+Leave other SHAs in the PR body unchanged.
+Record:
 
 ```text
 Candidate: <sha>
@@ -25,6 +34,7 @@ Base: <sha>
 Remote ref: <branch>
 Expected-old remote SHA: <sha | absent>
 Confirmed remote SHA: <sha>
+PR body head SHA: <sha or none>
 PR: <url>
 CI: <run ID or URL and triggered/pending/completed status>
 ```

package/skills/axstack/references/contracts.md

@@ -2,10 +2,12 @@
 
 Apply these authority, scope, and model rules before consequential action.
 
+For user-facing output, follow [STE-inspired writing](ste-writing.md).
+
 ## Required lifecycle load
 
-Except for `axstack-audit` and `axstack-relay`, every independently called phase
-must load and follow [Shared lifecycle](lifecycle.md) before acting. When a
+Except for `axstack-audit`, `axstack-relay`, and `axstack-correct`, every
+independently called phase must load and follow [Shared lifecycle](lifecycle.md) before acting. When a
 substantive run ends or reaches a meaningful checkpoint, apply the lifecycle
 audit hook. The audit phase loads these contracts, writes its assigned record,
 and stops; it never audits itself.
@@ -55,10 +57,9 @@ profile. Record the driver's provider and model in the run record.
 
 For Align and Spec, the driver forms an independent assessment first, then
 consults `axstack-advisor-astra` and `axstack-advisor-opus` independently with
-the same bounded evidence and question. The one exception is an arena-grade
-Align question: there the driver frames the brief and rubric, the advisers
-author candidates, and the driver assesses only after the candidates and judge
-verdicts return. The driver synthesizes disagreements,
+the same bounded evidence and question. The exception covers every brainstorm: the driver frames
+its brief and rubric and assesses only after the candidates and required judge
+rounds return. The driver synthesizes disagreements,
 owns the decision, and the user still approves the spec. Reuse each valid
 unchanged receipt; changed evidence, scope, or question requires a fresh
 receipt. If either adviser is unavailable, Align and Spec hold without model or

package/skills/axstack/references/design-lens.md

@@ -15,15 +15,15 @@ sketch through the scope identity.
   not reclassify work: Rung 1 can stay small. An unsettled material design
   question still makes routing reassess size.
 - **Rung 2 — arena.** A Rung 1 design that also meets the existing ADR test:
-  a meaningful, hard-to-reverse, non-obvious trade-off. Use Align's all-family
-  arena for that question.
+  a meaningful, hard-to-reverse, non-obvious trade-off. Use [Brainstorm](../../axstack-brainstorm/SKILL.md)
+  with its Rung 2 judge rounds for that question.
 
 There is no numeric threshold, file-count gate, or class-count gate.
 
 ## Questions in order
 
 Ask only unresolved areas in numbered `Qn` rounds of one to three. Give each a
-recommendation, reason, and trade-off; use Align's adviser critique. Design and
+recommendation, reason, and trade-off; use Align's inline brainstorm synthesis. Design and
 arena questions share its unchanged budget: 20 normally, a justified extension
 to 35, then opted-in refinement of at most five.
 

package/skills/axstack/references/diligence.md

@@ -3,7 +3,7 @@
 Read the [T3 runtime boundary](t3-runtime.md) before dispatch.
 Dispatch `axstack-diligence` with async `delegate_task` in the driver worktree,
 using a pinned brief and private evidence paths.
-It is read-only, never authors or edits, and returns `PASS` or `FINDINGS`
+It is read-only, never authors or edits, and returns `PASS`, `FINDINGS`, or `UNKNOWN`
 with locations, observed evidence, and limits. A stale or missing receipt is
 not a pass. Keep its first pass independent of other reviewers and workers.
 
@@ -21,5 +21,14 @@ publication, compare the author receipt with its evidence folder: red/green
 logs exist, and counts, SHAs, and paths match. For release preparation, compare
 the release PR body with the merged PRs.
 
+Retain the full output of every full-suite run as a named log in the dispatch's
+private evidence folder.
+For an unattributed full-suite failure, report `UNKNOWN` and return the failure to the driver.
+Never report `PASS` for an unattributed full-suite failure.
+The driver records the disposition of each full-suite failure in the run record.
+Report an attributed failure with its retained output normally.
+An observed full-suite failure still fails the suite, including when attributed or reported `UNKNOWN`.
+No reruns are required.
+
 `FINDINGS` identifies a mismatch for the driver to resolve at the owning phase;
 it does not edit the artifact or create another review round by itself.

package/skills/axstack/references/evidence-archive.md

@@ -118,6 +118,6 @@ After preservation and salvage checks, archive the exact eligible T3 thread
 with `t3_thread_organize`, then remove its exact recorded checkout path using
 `git worktree remove <path>` without force. Verify absence with
 `git worktree list --porcelain`; retire only eligible local-only branches with
-`git branch -d` under workspace hygiene. Never use shell recursive deletion or
+`git branch -d` under workspace hygiene. Never use shell recursive deletion to remove a whole worktree or
 treat archive success as ownership, settlement, liveness or cleanup proof.
 Failure or uncertainty preserves the resource.

package/skills/axstack/references/lifecycle.md

@@ -143,6 +143,8 @@ nothing without tested independent review.
 
 ## Close-out
 
+Close-out requires the [close-out acceptance table](run-record.md#close-out-acceptance).
+
 PRs merge by forge state; close out: (1) settle every T3 worker run and archive eligible threads; (2) compact record with counts and denominators—user
 interventions/deviations from plan/repairs; (3) `axstack-auditor`: settle
 non-zero/requested, else `counts zero`. A base auditor preflight rejection
@@ -153,7 +155,7 @@ use `axstack-cleanup`, remove the run's own scheduled tasks under
 [Workspace hygiene](workspace-hygiene.md), and close selected external-tracker tickets;
 (5) mark the
 [Run record](run-record.md) `Archived`. `Archived`—one each:
-settlement receipt; compact record path; auditor decision plus settlement
+settlement receipt; compact record path; close-out acceptance table; auditor decision plus settlement
 receipt, `counts zero`, or the unlaunchable UNKNOWN archive receipt; scheduled-task,
 release, and ticket receipts; archive timestamp.
 `active`/receipt-incomplete record: close-out pending, never done. One-step

package/skills/axstack/references/performance-checklist.md

@@ -0,0 +1,22 @@
+# Performance checklist
+
+Use this checklist only for performance claims.
+Keep workflow count accounting with axstack-audit.
+A limiter is the resource or code path that bounds performance.
+
+1. **Limiter: What bounds the result?**
+   Identify the resource or code path that limits the result from profiles or counters captured during a run.
+2. **Tuning: Did each option use suitable settings?**
+   Check each option with comparable production settings, versions, and data.
+3. **Limits: Is the result physically plausible?**
+   Compare the result with hardware limits and the changed code's share of total elapsed time.
+4. **Errors: Did the run succeed?**
+   Verify successful work and correct outputs using error counts from the measurement run.
+5. **Reproducibility: Does the difference repeat?**
+   Report the median and range from repeated, alternating runs of each option.
+6. **Relevance: Does the result matter to users?**
+   Measure the end-to-end user path with realistic data and concurrency.
+7. **Work happened: Did the timed work finish?**
+   Verify that the intended work completed inside the timed region.
+
+Ideas paraphrased from [pstack benchmark-checklist](https://github.com/cursor/plugins/blob/e43c7ee26e0038c6c1fa8380dd34ce86ff94cb2a/pstack/skills/benchmark-checklist/SKILL.md) (MIT), using Brendan Gregg's seven benchmark questions.

package/skills/axstack/references/routing.md

@@ -23,9 +23,11 @@ Resolve Codex and Claude classes with
 `skills/axstack/scripts/resolve-models.js --provider <provider> --capabilities <path>`
 using saved T3 capabilities JSON; missing or malformed capabilities holds.
 Claude exact IDs come from capabilities, replacing transcript read-back.
-Use an explicit model as given; for `model:null` without a class, only grok and
-antigravity use the provider's first listed model from saved capabilities and
-record its exact ID.
+Use an explicit model as given; for `model:null` without a class, grok uses the
+provider's first listed model from saved capabilities and records its exact ID.
+For `model:null` without a class, Antigravity must select the first listed model
+whose ID ends with `-<effort>` from saved capabilities and record its exact ID.
+Missing effort-suffix matches hold resolution for Antigravity.
 A Codex or Claude role with neither model nor class is an intentional absence
 and holds.
 Resume must reuse the saved capabilities and role snapshot with no re-resolution; changes require the user’s explicit decision.
@@ -50,6 +52,8 @@ step (3) for user routing: no substitution or same-provider review.
 
 ## Direct routes (no spec ceremony)
 
+- Validate an approach -> `axstack-brainstorm`: inline, report-only independent
+  candidates; light arena always, judges only at Rung 2; return to the caller.
 - Bounded research -> `axstack-research`: verify primary sources and code,
   cite limits, and fan out distinct questions.
 - Understand a system or gap -> `axstack-explain`:
@@ -61,6 +65,8 @@ step (3) for user routing: no substitution or same-provider review.
   off a classified repair (explain: how; debug: what's wrong).
 - Code quality/refactor discovery -> `axstack-improve`: rank bounded
   candidates with evidence; report only, no source edits.
+- Repeated mistakes need evidence and stronger checks -> `axstack-correct`:
+  user-invoked, report only.
 - Accepted worker/task/run completion or bounded backlog request -> driver invokes
   `axstack-cleanup` inline; never dispatch it.
 - Preparation completion, watch expiry, resume, or reconciliation -> the

package/skills/axstack/references/run-record.md

@@ -119,6 +119,21 @@ by the next owner:
   evidence lives. Read it on resume before reconciling; append, never rewrite,
   and keep entries as short as the evidence pointer allows.
 
+## Close-out acceptance
+
+Use one row per acceptance check, including each clause of a compound check.
+Record in each row a passing evidence pointer or a user-accepted hold with its `Decisions` row.
+If neither passing evidence nor a `Decisions` row with a user-accepted hold exists for an acceptance clause, hold close-out.
+A recorded user-accepted hold in `Decisions` satisfies that clause for close-out; keep the unmet result explicit.
+Record the reason for each driver-elected repair in `Decisions`.
+
+```markdown
+| Acceptance check | Result | Evidence or user-accepted hold | Repair election (Decisions row + reason, or none) |
+| --- | --- | --- | --- |
+| <check + clause> | passed | <SHA + check/log pointer> | <Decisions row + reason, or none> |
+| <check + unmet clause> | held (user accepted) | <Decisions row + user acceptance receipt> | <Decisions row + reason, or none> |
+```
+
 ## Privacy
 
 Record concise IDs, SHAs, URLs, status, timestamps, next actions, and evidence

package/skills/axstack/references/ste-writing.md

@@ -0,0 +1,22 @@
+# STE-inspired writing
+
+STE means Simplified Technical English.
+This reference borrows clarity rules from STE.
+
+Follow these rules only for new or materially revised user-facing output.
+User-facing output includes reports, PR bodies, briefs, read-backs, and packaged guidance.
+Do not rewrite existing prose for style.
+Keep exact identifiers unchanged.
+Keep quotes unchanged.
+Keep safety contracts unchanged.
+Keep prompt-byte contracts unchanged.
+
+Give each sentence one instruction.
+Keep sentences short.
+Write in active voice.
+If a step has a condition, state the condition before the step.
+Define technical terms before their first use.
+Use one name consistently for each term.
+Use words instead of slashes for "and" or "or".
+
+Ideas drawn from [pstack technical-writing](https://github.com/cursor/plugins/blob/e43c7ee26e0038c6c1fa8380dd34ce86ff94cb2a/pstack/skills/technical-writing/SKILL.md#L65-L88) (MIT).

package/skills/axstack/references/t3-runtime.md

@@ -32,6 +32,9 @@ A `model:null` role lacking a class must use the first model listed for its
 provider in saved capabilities only for grok and antigravity (launch-by-agent-id
 providers); record the exact ID, rather than an unresolved provider default.
 Antigravity must hold when saved capabilities advertise zero models.
+For Antigravity, first-listed selection must use the first model ID ending in `-<effort>` because its model ID encodes effort.
+No matching effort suffix holds resolution for Antigravity.
+For Antigravity, pass no effort option; effort read-back uses the model ID suffix.
 For codex or claude, a role lacking both model and class is an intentional
 absence and must hold; never use a provider default for that role.
 

package/skills/axstack/references/workspace-hygiene.md

@@ -14,7 +14,7 @@ Validate its real path, absence of symlinks and ownership before use and cleanup
 Evidence files still go to the private `<run>/evidence/<key>/` folder.
 
 Every shell deletion targets a literal absolute path or a `${VAR:?}`-guarded expansion, only inside the worker's own evidence folder, `TMPDIR`, or worktree.
-For example, `rm -rf -- "${EV:?}/mut"` requires a validated owned evidence path.
+For validated owned scratch, use `rm -r /tmp/<dispatch-key>/scratch` on a literal absolute path inside the evidence folder, `TMPDIR`, or worktree.
 Never use a bare `$VAR`, a glob on a variable, `/`, `HOME`, or a shared root as a deletion target.
 Prefer `git clean -- <exact prefix>` or tool-native cleanup. A safety prompt that
 still appears is a hold; agents do not answer it.

package/skills/axstack/scripts/resolve-models.js

@@ -45,7 +45,7 @@ function providerFrom(catalog, provider) {
   return instance;
 }
 
-function modelFrom(models, { provider, model: pin, class: modelClass, excluded }) {
+function modelFrom(models, { provider, model: pin, class: modelClass, excluded, effort }) {
   if (pin && pin !== 'null') {
     const model = models.find((model) => model.id === pin && !excluded.includes(model.id));
     if (!model) throw new Error(`missing requested model ${pin}`);
@@ -56,8 +56,9 @@ function modelFrom(models, { provider, model: pin, class: modelClass, excluded }
     if (['codex', 'claude'].includes(provider)) {
       throw new Error(`intentional absence for ${provider}: missing model and class`);
     }
-    // Launch-by-agent-ID providers bind the first listed ID; exclusions never select a substitute.
-    const model = models[0];
+    // Antigravity encodes effort in its ID; exclusions never select a substitute.
+    const model = provider === 'antigravity'
+      ? models.find((model) => model.id.endsWith(`-${effort}`)) : models[0];
     if (!model || excluded.includes(model.id)) throw new Error('missing first listed model');
     return model;
   }
@@ -88,14 +89,21 @@ try {
   const catalog = JSON.parse(readFileSync(path, 'utf8'));
   const instance = providerFrom(catalog, provider);
   const model = modelFrom(instance.models, options);
-  const effortOptions = model.options.filter((option) => effortIds[provider]
-    ? option.id === effortIds[provider] : ['reasoningEffort', 'effort'].includes(option.id));
-  if (effortOptions.length !== 1 || !effortOptions[0].options?.some((value) => value.id === effort)
-    || (provider === 'grok' && effort === 'max')) {
-    throw new Error(`unsupported effort ${effort} for ${provider}/${model.id}`);
+  let effortOption = null;
+  if (provider === 'antigravity') {
+    if (!model.id.endsWith(`-${effort}`)) {
+      throw new Error(`unsupported effort ${effort} for ${provider}/${model.id}`);
+    }
+  } else {
+    const effortOptions = model.options.filter((option) => option.id === effortIds[provider]);
+    if (effortOptions.length !== 1 || !effortOptions[0].options?.some((value) => value.id === effort)
+      || (provider === 'grok' && effort === 'max')) {
+      throw new Error(`unsupported effort ${effort} for ${provider}/${model.id}`);
+    }
+    effortOption = { id: effortOptions[0].id, value: effort };
   }
   console.log(JSON.stringify({ provider, providerInstanceId: instance.providerInstanceId,
-    model: model.id, effortOption: { id: effortOptions[0].id, value: effort }, source: 'capabilities', path }));
+    model: model.id, effortOption, source: 'capabilities', path }));
 } catch (error) {
   console.error(`model capabilities resolution hold: ${error.message}`);
   process.exitCode = 1;

package/skills/axstack-align/SKILL.md

@@ -26,7 +26,9 @@ Set the rung from researched facts; never ask the user to choose it. A change
 inside one module's existing interface, ownership, data flow, and failure
 guarantees is Rung 0: no design questions or sketch. Otherwise load the
 [design lens ladder](../axstack/references/design-lens.md) for Rung 1 or 2
-and settle only unresolved areas in its order within the existing budget. Carry a
+and settle only unresolved areas in its order within the existing budget.
+For unresolved Rung 1 or 2 design questions, load
+[Brainstorm](../axstack-brainstorm/SKILL.md) inline. Carry a
 Rung 1 or 2 sketch in the substantial spec's `Design` section or the returned
 small-change intent. A design question alone does not make small work
 substantial; apply routing's existing size reassessment rule.
@@ -65,9 +67,10 @@ substantial; apply routing's existing size reassessment rule.
 The current chat remains the driver under
 [Standing contracts](../axstack/references/contracts.md). For each new
 user round, the driver independently drafts the prioritized frontier and
-recommendations, except for an arena-grade question (below), where the driver
-writes the brief and rubric but drafts no recommendation until the candidates
-and judge verdicts return, so nothing anchors them. Then consult `axstack-advisor-astra` and
+recommendations, except for brainstorm questions, where the driver frames the
+brief and rubric and assesses after the candidates and required judge rounds
+return. Reuse valid brainstorm receipts to replace the adviser consult for
+that question. For other questions, consult `axstack-advisor-astra` and
 `axstack-advisor-opus` independently, without cross-reading, using the same
 bounded evidence and question. Each adviser challenges assumptions, edges,
 omissions, and alternatives; the driver synthesizes disagreements and accepts
@@ -87,60 +90,10 @@ the draft unchanged; it needs no new adviser pair. Changed draft text, a
 blocking finding, or a high-stakes decision requires fresh receipts on the new
 revision.
 
-## Arena for hard-to-reverse design choices
-
-Critique of one draft anchors every reader to that draft's shape. Rung 2 designs
-alone enter the arena: they meet the same test as for an ADR (a meaningful,
-hard-to-reverse, non-obvious trade-off: architecture, module boundaries, data
-model, migration strategy). Replace the critique round for that question with
-an arena. Small or routine questions never enter the arena.
-
-1. **Frame.** The driver writes the brief (the artifact, its constraints, the
-   settled decisions it must respect) and three to six gradeable rubric
-   criteria. Candidates receive only the brief; the rubric is for judging.
-2. **Fan out.** Produce one candidate per configured family independently from the same brief,
-   without cross-reading: `axstack-advisor-astra`, `axstack-advisor-opus`,
-   `axstack-arena-candidate-grok`, and `axstack-arena-candidate-antigravity`.
-   Each gives a design, rationale, and rejected alternatives. The driver authors no candidate.
-3. **Cross-judge.** After every candidate completes, give round 1 judge
-   `axstack-arena-judge-opus` the anonymized, relabeled candidates and rubric to
-   score every candidate per criterion and recommend a base with a reason.
-   The driver compares its own pick with the Opus verdict. Only if the driver
-   and the Opus judge disagree on the base, or the user rejects the round-1
-   synthesis,
-   run round 2 with fresh sessions: `axstack-escalation-fable` and `axstack-arena-judge-astra`
-   independently score the same anonymized candidates and rubric. Judges never
-   author, never cross-read each other, and never average verdicts. After round-2
-   verdicts return, the driver re-picks in step 4 and re-presents in step 6.
-4. **Pick.** The driver reads every candidate end to end and scores per
-   criterion, not on holistic feel, then compares with the judge verdicts from
-   each completed round. Agreement confirms the base. On disagreement, re-read
-   the rationales and decide with a stated reason; never average verdicts or
-   fabricate consensus.
-5. **Graft.** Walk the losing candidates once more for the one or two ideas
-   worth porting and fold them into the base by hand so the result stays
-   coherent under one mental model. Convergence on the same shape is a strong
-   agreement signal: adopt the consensus shape, no graft. Wide divergence
-   means the frame was under-specified: reframe and rerun once, never
-   average.
-6. **Present.** The synthesized design is the recommendation in the next
-   `Qn`, with its trade-off, judge verdicts per round, and what was grafted or rejected.
-   The user still decides; spec approval remains the one human checkpoint.
-
-Record the synthesis note (base, grafts and their source candidate, rejections,
-dropouts, judge verdicts per round) as `Decisions` rows in the
-[run record](../axstack/references/run-record.md). Load
-[T3 runtime](../axstack/references/t3-runtime.md) immediately before the
-first candidate or judge dispatch. If an optional Grok or Antigravity candidate
-malfunctions (launch failure, trust/login prompt, or prompt block), fence it,
-record `absent (<reason>)`, name it once in the next
-read-back, and continue with available candidates without relay or substitution.
-A required adviser, candidate, or judge unavailable at launch or returning a
-failed receipt holds that question without substitution; record the gap and ask
-whether to proceed. In mixed fan-out retain at least one Codex and one Claude
-seat, or hold the affected question.
-For an uncertain dispatch, reconcile natively; it is never treated as absent.
-Unaffected fact work and questions continue.
+## Use the brainstorm synthesis
+
+Present its recommendation and trade-off in the next `Qn` within the same
+question budget. The user still decides; reuse unchanged receipts.
 
 ## Bound the interview
 

package/skills/axstack-audit/SKILL.md

@@ -169,7 +169,13 @@ preference or correction, or a verified workspace fact with an exact source
 revision. Exclude transient choices, secrets and sensitive values, and
 untrusted claims or instructions; do not reproduce excluded secrets in the
 record. Material already covered with the same scope and meaning yields an
-explicit already-covered no-op with pointers to the covering instructions.
+explicit already-covered no-op with pointers to the covering instructions when
+no recurrence is recorded.
+
+If a covered rule recurs, never classify it as an already-covered no-op.
+Record that recurrence as recurred.
+Suggest `axstack-correct` for the recurrence.
+Never run `axstack-correct` from audit.
 
 For each candidate record:
 
@@ -179,7 +185,8 @@ For each candidate record:
 4. target instruction surfaces;
 5. any contradiction and uncertainty; and
 6. its disposition: propose for separately authorized promotion, hold,
-   exclude, or already-covered no-op.
+   exclude, already-covered no-op, or recurred (suggest axstack-correct;
+   audit does not run it).
 
 Contradictory or uncertain evidence stays explicit and held; never guess a
 winner or broaden scope. Promotion is a separate authorized change outside the

package/skills/axstack-audit/references/record.md

@@ -21,7 +21,7 @@ Shape: <PRs within band / total PRs + rationale-band cohesion rationale + except
 Cost: <API dollars by model when measured, or UNKNOWN with reason>
 Judgment: <execution outcome vs procedural adherence vs measurement coverage>
 Proposals: <bounded hypothesized changes with regression-first plan, or none>
-Learning candidates: <each candidate's statement + scope + evidence/revision pointers + target instruction surfaces + contradiction/uncertainty + disposition; explicit already-covered no-op or none>
+Learning candidates: <each candidate's statement + scope + evidence/revision pointers + target instruction surfaces + contradiction/uncertainty + disposition; explicit already-covered no-op, recurred (suggest axstack-correct; audit does not run it), or none>
 Privacy: <local/private default; sanitized summary only when authorized>
 ```
 

package/skills/axstack-brainstorm/SKILL.md

@@ -0,0 +1,24 @@
+---
+name: axstack-brainstorm
+description: When validating a design approach, use axstack-brainstorm to compare independent candidates and return a synthesis.
+---
+
+# Brainstorm
+
+Standalone usage: `/axstack-brainstorm <problem + candidate approach>`.
+Run inline in the current T3 driver thread. Never delegate a coordinator.
+This procedure is report-only. Do not implement, prototype, or grant execution
+or spec approval. Never interview the user or invoke Align.
+
+Load [Standing contracts](../axstack/references/contracts.md) before acting.
+Set the rung from researched facts under the
+[design lens](../axstack/references/design-lens.md); Rung 2 meets its ADR test.
+Follow the [arena procedure](references/arena.md), preserving settled decisions
+and naming missing evidence. Candidates challenge the supplied approach as
+well as proposing alternatives.
+
+Return `recommend | revise | reject | unresolved` with the
+[sketch](../axstack/references/design-lens.md#sketch), including `Rejected` and
+`Open`, plus base, criterion scores, grafts, dropouts and completed judge rounds.
+Return proposed questions to the caller. The caller owns preferences, next
+steps and any approval; an open blocker stays unresolved.

package/skills/axstack-brainstorm/references/arena.md

@@ -0,0 +1,56 @@
+## Arena
+
+Every brainstorm runs a light arena. At Rung 1, the driver scores, picks and
+grafts without a judge round. Only at Rung 2, run the judge rounds below for
+hard-to-reverse choices; they replace the critique round for that question.
+
+1. **Frame.** The driver writes the brief (the artifact, its constraints, the
+   settled decisions it must respect) and three to six gradeable rubric
+   criteria. Every brief requires a premise check and comparison with the
+   smallest change and doing nothing. The rubric uses agent-contributor red
+   flags as evidence prompts: seeing only opened files, copying the nearest
+   example, taking the shortest path that compiles. Treat them as prompts, not defects. Candidates receive only the brief; the rubric is for judging.
+2. **Fan out.** Produce one candidate per configured family independently from the same brief,
+   without cross-reading: `axstack-advisor-astra`, `axstack-advisor-opus`,
+   `axstack-arena-candidate-grok`, and `axstack-arena-candidate-antigravity`.
+   Each gives a design, rationale, and rejected alternatives. The driver authors no candidate.
+   The driver drafts no recommendation until every candidate returns.
+3. **Cross-judge (Rung 2 only).** After every candidate completes, give round 1 judge
+   `axstack-arena-judge-opus` the anonymized, relabeled candidates and rubric to
+   score every candidate per criterion and recommend a base with a reason.
+   The driver compares its own pick with the Opus verdict. Only if the driver
+   and the Opus judge disagree on the base, or the caller re-invokes with the user's rejection of the round-1
+   synthesis,
+   run round 2 with fresh sessions: `axstack-escalation-fable` and `axstack-arena-judge-astra`
+   independently score the same anonymized candidates and rubric. Judges never
+   author, never cross-read each other, and never average verdicts. After round-2
+   verdicts return, the driver re-picks in step 4 and re-presents in step 6.
+4. **Pick.** The driver reads every candidate end to end and scores per
+   criterion, not on holistic feel, then compares with the judge verdicts from
+   each completed round. Agreement confirms the base. On disagreement, re-read
+   the rationales and decide with a stated reason; never average verdicts or
+   fabricate consensus.
+5. **Graft.** Revisit losing candidates once; graft one or two ideas by hand
+   into the coherent base under one mental model. Convergence on the same shape is a strong
+   agreement signal: adopt the consensus shape, no graft. Wide divergence
+   means the frame was under-specified: reframe and rerun once, never
+   average.
+6. **Present.** Return the synthesized design to the caller with its
+   trade-off, judge verdicts per round, and what was grafted or rejected.
+   The user still decides; spec approval remains the one human checkpoint.
+
+Record the synthesis note (base, graft sources, rejections, dropouts, judge
+verdicts per round) as `Decisions` rows in the
+[caller's run record](../../axstack/references/run-record.md) when one exists,
+otherwise include it in the returned verdict. Load
+[T3 runtime](../../axstack/references/t3-runtime.md) immediately before the
+first candidate or judge dispatch. If an optional Grok or Antigravity candidate
+malfunctions (launch failure, trust/login prompt, or prompt block), fence it,
+record `absent (<reason>)`, name it once in the next
+read-back, and continue with available candidates without relay or substitution.
+A required adviser, candidate, or judge needed for this rung unavailable at
+launch or returning a failed receipt holds that question without substitution;
+record the gap and return a proposed question to the caller about whether to proceed. In mixed fan-out retain at least one Codex and one Claude
+seat, or hold the affected question.
+For an uncertain dispatch, reconcile natively; it is never treated as absent.
+Unaffected fact work and questions continue.

package/skills/axstack-correct/SKILL.md

@@ -0,0 +1,41 @@
+---
+name: axstack-correct
+description: When repeated mistakes need evidence and stronger checks, use axstack-correct to report correction proposals.
+disable-model-invocation: true
+---
+
+# Correct
+
+Usage: /axstack-correct ["<correction>"] [runs=<ids> | last=<N, default 10>]
+
+Load [Standing contracts](../axstack/references/contracts.md) before acting.
+
+Run only at the user's request.
+Write a report only.
+Never edit files.
+Never dispatch workers.
+Never merge.
+Never read transcripts.
+
+Read [run records](../axstack/references/run-record.md): Decisions, deviations, Learnings, FAILED receipts, audit proposals.
+Read git history bounded by selected runs.
+Read only PR reviews named in those records.
+Record selected runs and git bounds.
+Report inaccessible evidence.
+
+A class needs at least two distinct incidents.
+Each incident needs a distinct pointer: run id + attempt or file:line, SHA, or PR or review URL.
+Count an incident echoed in several records once.
+If a pointer is missing, report recurrence UNKNOWN and keep the class open.
+
+Try fixes in this order: remove copied pattern > script or helper error naming the fix > brief or receipt template > bun test > prose.
+Label each rule `shipped` or `runtime: advisory`.
+Use `shipped` only when a check fails in CI on the mistake itself.
+When only audit can observe a rule, label it `runtime: advisory`.
+A prose-contract test alone leaves a rule `runtime: advisory`.
+Propose fixes through a small-change intent and [axstack-implement](../axstack-implement/SKILL.md).
+
+The `## Enforced rules` table in the target repo's AGENTS.md changes only in the PR that adds its check.
+The first row's PR adds a test that fails when any row's enforcement path disappears.
+If a `shipped` check missed a later recorded recurrence, report `enforcement failed`.
+For that failure, propose relabelling the row `runtime: advisory` in a separately reviewed PR.

package/skills/axstack-debug/SKILL.md

@@ -22,6 +22,8 @@ read it before phase 7 and before any fan-out.
 Redact secrets before showing any command, output, or artifact; build loops
 against environment variables so credentials never appear in what is shown.
 
+For performance claims only, load [Performance checklist](../axstack/references/performance-checklist.md).
+
 ## Phases
 
 Each phase has an observable completion criterion. Skip one only with a

package/skills/axstack-implement/SKILL.md

@@ -230,6 +230,11 @@ For each PR:
    to step 1.
    Merge-ready also requires a current diligence `PASS` at that head; diligence
    `FINDINGS` return to the same author within the review round.
+   Diligence `UNKNOWN` records the PR as `held` with the reason in the run record
+   and blocks merge-ready pending the driver's recorded disposition.
+   Before merge-ready, treat a run-record `Learning` that contradicts a shipped
+   rule as a finding on the owning PR and hold merge-ready until the contradiction is resolved.
+   An unrelated run-record `Learning` leaves merge-ready eligibility unchanged.
    A round with reviewer `REQUEST_CHANGES` and/or diligence `FINDINGS` increments
    `repairs` once and counts once toward the third-round hold.
 

package/skills/axstack-improve/SKILL.md

@@ -27,6 +27,8 @@ Sol pair that fails to launch is fenced, recorded `absent (<reason>)`, and
 named once in the next read-back, then skipped without relay or substitution.
 In mixed fan-out retain a Codex and a Claude seat or hold the affected work.
 
+For performance claims only, load [Performance checklist](../axstack/references/performance-checklist.md).
+
 ## 1. Bound discovery
 
 1. Start with the user's named subsystem or pain. Otherwise inspect recent

package/skills/axstack-review/SKILL.md

@@ -33,6 +33,8 @@ When the caller is a bounded review-manager PR job, load
 carry the required escalation field and every eligible peer PR takes a binding
 `APPROVE` or `REQUEST_CHANGES` verdict under the automation exception below.
 
+For performance claims only, load [Performance checklist](../axstack/references/performance-checklist.md).
+
 ## Codebase findings mode
 
 Use this manual mode for existing code at a pinned exact source revision and a

package/skills/axstack-spec/SKILL.md

@@ -62,6 +62,12 @@ and the lifecycle's [audit skill](../axstack-audit/SKILL.md) hook.
    session to return plain AGREE. Present one
    reviewable, identified revision for this checkpoint. Its user approval
    creates the execution baseline.
+   Name the draft revision covered by each adviser receipt at the checkpoint.
+   If draft text changed and an adviser receipt covers an older revision, hold
+   approval until fresh receipts cover the presented revision.
+   Except for high-stakes decisions, a change confined to a `Decisions` row
+   reuses adviser receipts only while draft text, evidence, scope, and question
+   remain unchanged.
    Before user approval, dispatch `axstack-diligence` under
    [Diligence](../axstack/references/diligence.md) to check the draft against
    the Align decisions for anything dropped, added, or softened.