axstack 0.13.1 → 0.14.1

package/README.md

@@ -77,7 +77,7 @@ targets are reported without normal-path adoption. Install exits nonzero when
 an instruction conflict is preserved, while clean and idempotent installs exit
 successfully.
 
-The public bundle preserves three canonical 21-role inputs:
+The public bundle preserves three canonical 23-role inputs:
 [mixed](profiles/presets/mixed.json),
 [codex-only](profiles/presets/codex-only.json), and
 [claude-only](profiles/presets/claude-only.json). Each is exactly

package/docs/installation.md

@@ -51,7 +51,7 @@ profiles/presets/codex-only.json
 profiles/presets/claude-only.json
 ```
 
-Each has exactly `{ "version": 1, "roles": [...] }` with the same 21 stable
+Each has exactly `{ "version": 1, "roles": [...] }` with the same 23 stable
 role IDs. Installation writes `<skills-dir>/axstack/roles.json` as
 `{ "version": 1, "preset": "<selected preset>", "roles": [...] }` and records
 its ownership hash like every other installed skill asset. There is no second
@@ -118,9 +118,9 @@ The complete bundle is validated before writes:
 - each preset is a real JSON file with version 1, a non-empty `roles` array,
   the filename's selected identity supplied by the caller, and the same role-ID
   set as its peers;
-- every role has valid preserved fields, while the mixed checker and the
-  unavailable adviser in each single-provider preset explicitly permit
-  `model: null`;
+- every role has valid preserved fields, while the mixed checker and, in each
+  single-provider preset, the unavailable adviser and its matching arena judge
+  seat explicitly permit `model: null`;
 - obsolete runtime configuration flags fail before mutation with migration
   guidance.
 
@@ -148,15 +148,16 @@ to rewrite them.
 ## Role behavior after installation
 
 The runtime reads `roles.json` relative to the actually loaded `axstack` skill.
-A new run records the selected preset plus all 21 role rows. An active run keeps
+A new run records the selected preset plus all 23 role rows. An active run keeps
 that snapshot after a later preset install unless the user explicitly changes
 it and accepts the resulting evidence invalidation.
 
 The mixed checker has `model: null`; checker dispatch is held and never inherits
 a provider default. The single-provider presets configure the checker. Their
-unavailable adviser remains an explicit same-provider `model: null` role, which
-does not make installation unready; Align and Spec still hold until both Astra
-and Fable can return independent receipts. The current chat drives on whatever
+unavailable adviser and its matching arena judge seat remain explicit
+same-provider `model: null` roles, which do not make installation unready;
+Align and Spec still hold until both Astra and Fable can return independent
+receipts, and an arena-grade Align question holds until both judge seats can. 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. There is no model substitution, subscription inference, or
 quota routing.

package/docs/workflows.md

@@ -40,7 +40,7 @@ only affected work.
 
 Installation requires one explicit canonical preset. The three bundle files
 under `profiles/presets/` each contain exactly
-`{ "version": 1, "roles": [...] }` and the same 21 stable IDs.
+`{ "version": 1, "roles": [...] }` and the same 23 stable IDs.
 
 The current chat drives on whatever model runs it; no preset carries a driver
 role.
@@ -108,7 +108,13 @@ session and evidence remain valid.
 
 - `axstack-align` maps facts and dependencies, asks prioritized questions, and
   consults Astra and Fable independently with the same bounded evidence and
-  question. It synthesizes disagreements and reuses unchanged receipts.
+  question. It synthesizes disagreements and reuses unchanged receipts. For a
+  hard-to-reverse design choice it runs one arena round instead: Astra and
+  Fable each author a candidate, `axstack-arena-judge-astra` and
+  `axstack-arena-judge-fable` score both against the driver's rubric, and the
+  driver picks a base, grafts the losers' strong ideas, and presents the
+  synthesis as the recommendation; the note lands as `Decisions` rows in the
+  run record.
 - `axstack-spec` writes observable acceptance, exclusions, decisions, and one
   user-approved revision baseline.
 - `axstack-tickets` maps user-visible capabilities to dependency-aware internal

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "axstack",
-  "version": "0.13.1",
+  "version": "0.14.1",
   "description": "Axstack installer and setup CLI: installs owned chat skills and role data, configures supported harness settings, and checks Orca capabilities.",
   "keywords": [
     "claude-code",

package/profiles/presets/claude-only.json

@@ -189,6 +189,24 @@
       "modeId": "bypassPermissions",
       "thinkingOptionId": "xhigh",
       "notes": "Debug investigator seat 4. Dispatched only by axstack-debug at L1 with the shared evidence packet and one distinct brief; never reads another investigator's output. Works in its own disposable worktree at the pinned revision plus the recorded dirty patch; may instrument there for probes; never commits, pushes, publishes, or creates children. Returns one receipt per brief. Independence comes from brief isolation, not model diversity. This preset repeats claude-sonnet-5 at xhigh effort because it has fewer model families; independence comes from brief isolation."
+    },
+    {
+      "id": "axstack-arena-judge-astra",
+      "name": "Axstack arena judge Astra",
+      "provider": "claude",
+      "model": null,
+      "modeId": "bypassPermissions",
+      "thinkingOptionId": "xhigh",
+      "notes": "Required Arena judge (Astra seat). Read-only cross-judge for an arena-grade Align decision: receives the rubric and both candidates by label, scores each criterion, and recommends a base with rationale. Never authors a candidate, never cross-reads the other judge, never mutates. Disagreement between judges is surfaced by the driver, never averaged. Intentionally absent in the claude-only preset; the arena-grade decision holds without substitution."
+    },
+    {
+      "id": "axstack-arena-judge-fable",
+      "name": "Axstack arena judge Fable",
+      "provider": "claude",
+      "model": "claude-fable-5-1",
+      "modeId": "bypassPermissions",
+      "thinkingOptionId": "xhigh",
+      "notes": "Arena judge (Fable seat). Read-only cross-judge for an arena-grade Align decision: receives the rubric and both candidates by label, scores each criterion, and recommends a base with rationale. Never authors a candidate, never cross-reads the other judge, never mutates. Disagreement between judges is surfaced by the driver, never averaged."
     }
   ]
 }

package/profiles/presets/codex-only.json

@@ -189,6 +189,24 @@
       "modeId": "full-access",
       "thinkingOptionId": "xhigh",
       "notes": "Debug investigator seat 4. Dispatched only by axstack-debug at L1 with the shared evidence packet and one distinct brief; never reads another investigator's output. Works in its own disposable worktree at the pinned revision plus the recorded dirty patch; may instrument there for probes; never commits, pushes, publishes, or creates children. Returns one receipt per brief. Independence comes from brief isolation, not model diversity. This preset repeats gpt-5.6-terra at xhigh effort because it has fewer model families."
+    },
+    {
+      "id": "axstack-arena-judge-astra",
+      "name": "Axstack arena judge Astra",
+      "provider": "codex",
+      "model": "gpt-6-astra",
+      "modeId": "full-access",
+      "thinkingOptionId": "xhigh",
+      "notes": "Arena judge (Astra seat). Read-only cross-judge for an arena-grade Align decision: receives the rubric and both candidates by label, scores each criterion, and recommends a base with rationale. Never authors a candidate, never cross-reads the other judge, never mutates. Disagreement between judges is surfaced by the driver, never averaged."
+    },
+    {
+      "id": "axstack-arena-judge-fable",
+      "name": "Axstack arena judge Fable",
+      "provider": "codex",
+      "model": null,
+      "modeId": "full-access",
+      "thinkingOptionId": "xhigh",
+      "notes": "Required Arena judge (Fable seat). Read-only cross-judge for an arena-grade Align decision: receives the rubric and both candidates by label, scores each criterion, and recommends a base with rationale. Never authors a candidate, never cross-reads the other judge, never mutates. Disagreement between judges is surfaced by the driver, never averaged. Intentionally absent in the codex-only preset; the arena-grade decision holds without substitution."
     }
   ]
 }

package/profiles/presets/mixed.json

@@ -189,6 +189,24 @@
       "modeId": "full-access",
       "thinkingOptionId": "low",
       "notes": "Debug investigator seat 4. Dispatched only by axstack-debug at L1 with the shared evidence packet and one distinct brief; never reads another investigator's output. Works in its own disposable worktree at the pinned revision plus the recorded dirty patch; may instrument there for probes; never commits, pushes, publishes, or creates children. Returns one receipt per brief. Independence comes from brief isolation, not model diversity."
+    },
+    {
+      "id": "axstack-arena-judge-astra",
+      "name": "Axstack arena judge Astra",
+      "provider": "codex",
+      "model": "gpt-6-astra",
+      "modeId": "full-access",
+      "thinkingOptionId": "xhigh",
+      "notes": "Arena judge (Astra seat). Read-only cross-judge for an arena-grade Align decision: receives the rubric and both candidates by label, scores each criterion, and recommends a base with rationale. Never authors a candidate, never cross-reads the other judge, never mutates. Disagreement between judges is surfaced by the driver, never averaged."
+    },
+    {
+      "id": "axstack-arena-judge-fable",
+      "name": "Axstack arena judge Fable",
+      "provider": "claude",
+      "model": "claude-fable-5-1",
+      "modeId": "bypassPermissions",
+      "thinkingOptionId": "xhigh",
+      "notes": "Arena judge (Fable seat). Read-only cross-judge for an arena-grade Align decision: receives the rubric and both candidates by label, scores each criterion, and recommends a base with rationale. Never authors a candidate, never cross-reads the other judge, never mutates. Disagreement between judges is surfaced by the driver, never averaged."
     }
   ]
 }

package/skills/axstack/references/automations.md

@@ -2,7 +2,7 @@
 
 Read this when the current session is the native Orca PR **driver** or its
 **watchdog**. The approved contract is
-`docs/specs/pr-automations.md` revision 4; this reference restates the parts an
+`docs/specs/pr-automations.md` revision 5; this reference restates the parts an
 automation session must execute and does not widen them.
 
 Pair A/B is retired for this contract; its artefacts remain untouched.
@@ -127,32 +127,33 @@ The driver performs this order and exits:
 Claude Code trusts a folder per git toplevel and stops at its "Quick safety
 check" dialog otherwise, and the driver never answers that dialog for a
 worker. So workers run in a fixed pool: every allowlisted project has
-exactly two slot worktrees, `slot-1` and `slot-2`, its Orca child worktrees
-created once, parented to the project's primary worktree, and trusted once
+exactly five slot worktrees, `slot-1` through `slot-5`, its Orca child
+worktrees created once, parented to the project's primary worktree, and trusted once
 by the user through that dialog. The driver never creates or removes a
 worktree and never writes `~/.claude.json`. A slot is free when no live
 dispatch marker names it, it is not in `retained_slots[]`, no terminal is
 listed in it, and its tree is clean; no free slot in the project defers the
 PR to `deferred[]` like a budget, not a hold. Taking a slot fetches the head into the project clone,
 checks the slot out detached at the pinned head, and verifies HEAD equals
-it; the marker's worktree is the slot path. Trusting a folder activates the
-full project surface: its `.claude/settings.json` and the hooks it defines,
-its `.mcp.json` servers, marketplace plugin auto-install, and `CLAUDE.md`;
-a hostile branch's hooks or MCP servers would run the moment the folder
-opens. So the worker is launched in Claude Code's own isolation mode,
-`--safe-mode`, through `terminal create` and `worker-start --terminal`,
-since `worker-start` cannot pass argv. Safe mode is the binary's sanctioned
-"all customizations disabled" path: no `CLAUDE.md`, skills, plugins, hooks,
-MCP servers, custom commands or agents load from anywhere, project or user;
-built-in tools and authentication are untouched, and the brief loads the
-skill files it needs by path. The driver confirms readiness from the
-rendered frame — `wait.satisfied`, the prompt marker present, the dialog
-absent — before dispatching; a dialog on a slot means the slot is not
-trusted: the slot is named in a health line, the PR is deferred, nothing is
-dispatched. What remains live is the repository's files as data the worker
-reads and the commands the worker itself chooses to run, which it already
-runs today; nothing from the branch loads, executes, or is offered for
-invocation on its own.
+it; the marker's worktree is the slot path. The worker is launched by Orca
+itself — `worker-start --agent claude --model claude-opus-5 --effort medium`
+on that slot — so the runtime owns the process: `worker-release` ends it and
+`worker-show` proves it exited, which is what returns the slot to the pool.
+Never pre-create the worker's terminal or hand a terminal handle to
+`worker-start`: a reused handle is a resource Orca labels `external`, one it
+can neither stop nor prove exited, so every such slot ends retained. After
+`worker-start` run `worker-show` on the receipt's dispatch id and require
+`projection.resource.state == owned`; anything else is a launch Orca does not
+own: apply the runtime-refusal recovery rules under "Safety holds" (the
+`worker-list` row's `nextAction` argv verbatim; `none` means inspect and
+retain), append a `worker not owned` health line and a `retained_slots[]`
+entry for the slot, defer the PR, and record no marker. Orca's per-agent default arguments supply
+`--dangerously-skip-permissions`; the brief loads the skill files it needs by
+path. If `worker-start` reports a failed stage or a visible hold (the "Quick
+safety check" trust dialog) the slot is not trusted: name the slot in a health
+line, defer the PR, dispatch nothing, never answer the dialog. Project
+customizations load as they would for the user; the allowlist is defi-com
+only and the user accepted that surface on 2026-09-18.
 
 Every selected PR receives one dispatch marker with task id, dispatch id,
 worktree, head, `started_at`, reservation (`verdict` or `repair`), and trigger:

package/skills/axstack/references/blast-radius.md

@@ -0,0 +1,64 @@
+# Blast radius
+
+Find what a change breaks somewhere else before it ships, beyond the diff, and
+prove the one fact it is safe because of by running real code instead of
+writing it up. Loaded by `axstack-review` under angle 3 before any verdict and
+by `axstack-explain` for a direct "what could this break" question.
+
+Listing the callers is not the job; a symbol search does that in a second. The
+job is the breakage the search does not show.
+
+## Evidence ladder
+
+A writeup that sounds right is worthless: it reads as convincing whether or not
+it is true. For each fact the change's safety depends on, get it as far down
+this ladder as is cheap and say where it stopped:
+
+1. **Said so.** Worthless on its own.
+2. **Pointed at the line.** A real `file:line`, or the library's own source at
+   the pinned version.
+3. **Walked the failure.** The bad case was traced step by step and does not
+   reach.
+4. **Ran it.** A script or test that calls the real code and fails loud if the
+   claim is wrong.
+5. **Reproduced it in the running app.**
+
+A safety fact below step 4 is written as **unproven**, never as settled. Step 4
+is usually one small script that imports what the app ships and calls the exact
+function in question. "Not found" is an answer for the inspected scope only.
+
+## Steps
+
+1. **Read the change.** The diff, the symbols it adds, changes, and deletes,
+   and what it now does differently, including the part the diff does not
+   spell out. Pull the PR body and commit messages for the stated intent.
+2. **Find the one fact it is safe because of.** Most changes that look risky
+   are safe because of a single fact, such as "this call only drops
+   already-dead cache entries". Spend the time here, not on a long list of
+   maybes. If that fact holds, most risky cases clear at once.
+3. **Look where the search stops.** Library source at the pinned version and
+   any local patch; when things run (microtasks, unmount and teardown,
+   reactive frameworks); what a symbol search misses: the JSON an API returns,
+   a database column, a wire or on-chain format, another language reading the
+   same bytes, a feature flag, code three hops downstream.
+4. **Be honest about each risk.** Give it a real chance of happening and a
+   real cost if it does. Keep the confirmed risks; list the checked and cleared
+   ones separately. Cite a real `file:line`; never invent a caller or an API.
+5. **Prove the one fact.** Write the script or test, run it against the real
+   code, and paste what happened. If it cannot be proven cheaply, mark it
+   unproven; do not overstate.
+
+## What to hand back
+
+- **What it does.** What changed, including the part that is not obvious.
+- **The one fact it is safe because of.** State it, the ladder step reached,
+  and the proof; or `unproven`.
+- **Risks.** Only the real ones, each with how it breaks, the `file:line`, how
+  likely and how bad, and how to check.
+- **Cleared.** What was checked and why it is fine.
+- **Before merge.** The cheapest test or repro that catches the real bug,
+  including the script written for step 5.
+
+Strip anything private before the writeup leaves the run. In review, an
+unproven safety fact for a consequential change is a finding with its
+consequence stated, not a footnote.

package/skills/axstack/references/contracts.md

@@ -49,7 +49,10 @@ 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-fable` independently with
-the same bounded evidence and question. The driver synthesizes disagreements,
+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,
 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/orca-runtime.md

@@ -26,7 +26,7 @@ Read `roles.json` relative to the actually loaded `axstack` skill. The installed
 shape is `{ "version": 1, "preset": "<name>", "roles": [...] }`. Bundled
 profiles are setup inputs shaped as
 `{ "version": 1, "roles": [...] }`. A new run records the selected preset and
-all 21 role rows once. An active run keeps the exact snapshot until the user
+all 23 role rows once. An active run keeps the exact snapshot until the user
 explicitly changes it.
 
 Select the requested role by stable ID. A missing or null model holds only that role;

package/skills/axstack/references/routing.md

@@ -11,7 +11,7 @@ only with exactly one unambiguous preset; missing or contradictory sources are
 a setup gap: hold. Never infer from live profiles or `list_profiles`, harness,
 tools, credentials, quota, subscription, or default to `mixed`.
 
-At run start, capture one **routing snapshot**: the complete map of all 21 role
+At run start, capture one **routing snapshot**: the complete map of all 23 role
 IDs with provider/model/mode/effort, absent or unconfigured roles recorded
 explicitly, and no invented provider default. An absent or unconfigured role
 holds only that role's work, not the run. A role installed or changed later
@@ -38,8 +38,10 @@ Role IDs:
   | `mixed` | Claude / Opus (`claude/claude-opus-5`) | `axstack-reviewer-primary` (`codex/gpt-5.6-sol` medium) |
   | `codex-only` | Codex / Sol (`codex/gpt-5.6-sol`) | `axstack-reviewer-secondary` (`codex/gpt-5.6-terra` xhigh) |
   | `claude-only` | Claude / Opus (`claude/claude-opus-5`) | `axstack-reviewer-secondary` (`claude/claude-sonnet-5` xhigh) |
-- `axstack-advisor-astra` and `axstack-advisor-fable` advise independently;
-  `axstack-auditor` audits; `axstack-checker` reports discrepancies.
+- `axstack-advisor-astra` and `axstack-advisor-fable` advise independently
+  and author align arena candidates; `axstack-arena-judge-astra` and
+  `axstack-arena-judge-fable` judge them. `axstack-auditor` audits;
+  `axstack-checker` reports discrepancies.
 - `axstack-explainer` authors explanations; `axstack-explainer-review`
   reviews them. `axstack-monitor` observes only; `axstack-watchdog` sends
   only gate-authorized health escalations.
@@ -47,10 +49,9 @@ Role IDs:
 
 Provenance is matched on provider/model ID; record effort but never use it to
 create a mapping. Provenance absent from the preset's table row is
-unsupported and `INCOMPLETE` (including its secondary reviewer model, Astra,
-Luna, or Fable); report the exact gap and ask the user. Never derive a reverse
-pairing from slot position, driver, owner, or provider. Author and owner never
-review their own work.
+unsupported and `INCOMPLETE`; report the exact gap and ask the user. Never
+derive a reverse pairing from slot position, driver, owner, or provider.
+Author and owner never review their own work.
 
 ## Direct routes (no spec ceremony)
 
@@ -58,37 +59,37 @@ review their own work.
   and code, return a cited note with limitations. Fan out only distinct
   questions.
 - Understanding a system, change, or implementation gap -> `axstack-explain`:
-  show current/intended behavior, evidence dimensions, and bounded gaps; use
-  project docs and verify rendered behavior when applicable. Stale axstack-docs
-  is superseded. Publication needs separate authority.
+  current/intended behavior, evidence dimensions, and bounded gaps from project
+  docs and rendered behavior. "What could this break" follows
+  [Blast radius](blast-radius.md). Publication needs separate authority.
 - A bug, failing test, regression, or wrong behavior, red loop wanted ->
   `axstack-debug`: diagnose, escalate via adviser-directed investigators, hand
   off a classified repair (explain: how; debug: what's wrong).
 - Codebase-quality or refactor discovery -> `axstack-improve`: inspect bounded
   scope, rank evidenced candidates, report only; no spec, tickets, or source
-  edits. Selected changes return via preparation or execution.
+  edits.
 - Preparation completion, watch expiry, resume, or reconciliation -> the
   [lifecycle](lifecycle.md#native-handoff-and-resume): reconcile the run
   record, keep its owner, launch no native handoff.
 - Explicit user-requested ownership transfer -> the same lifecycle section.
   Load the [Orca runtime boundary](orca-runtime.md), follow the runtime-owned
   handoff guide, and require explicit recipient acceptance before ownership
-  changes. Missing capability is a setup gap, never license to invent one.
+  changes. Missing capability is a setup gap; never invent one.
 - Colleague PR review -> `axstack-review`, peer mode.
 - Own PR maintenance or monitoring -> `axstack-review` in authored mode,
   `axstack-watch` for adoption.
 
 Research, explanation, improvement discovery, debugging, handoff, peer review,
-and adopted maintenance need no alignment, approved spec, or ticket map;
-authority and intent boundaries still apply.
+and adopted maintenance need no alignment, spec, or ticket map; authority and
+intent boundaries still apply.
 
 ## Proportional scope identity
 
 Classify new work as substantial, small, or unclear; record it with brief
 reason in the run record, or in the brief for tiny direct work.
 
-- **Substantial:** substantial features, multi-PR work, or stacked work. A
-  bounded small feature is not substantial merely because it is labelled one.
+- **Substantial:** substantial features, multi-PR work, or stacked work; a
+  bounded small feature is not substantial because it is labelled one.
   Require an approved spec plus a ticket map tied to that exact spec
   revision, with acceptance checks and dependencies in the explicitly selected
   Markdown or Linear store. Prepare via `axstack-align` -> `axstack-spec`
@@ -107,8 +108,8 @@ reason in the run record, or in the brief for tiny direct work.
 Reassess size when growth adds an additional PR, a new execution dependency
 that materially expands scope, an unsettled material design question, or a
 security/infrastructure boundary crossing. An ordinary
-test-then-code sequence is not multi-task growth; a minor file dependency does
-not alone need a formal spec. Hold affected unsafe work while reassessing.
+test-then-code sequence is not multi-task growth; a minor file dependency is
+not alone a formal spec trigger. Hold affected unsafe work while reassessing.
 
 ## Lifecycle routes (mode-specific scope identity required)
 
@@ -125,5 +126,5 @@ not alone need a formal spec. Hold affected unsafe work while reassessing.
   author provenance — then use `axstack-review` and `axstack-watch` without
   repeated approval or new spec ceremony. Never infer the author from the
   orchestrator or assume an imported own PR's author.
-- Direct later phase: start there and pass that phase's identity check. Entry
-  never admits work a deeper phase would reject.
+- Direct later phase: start there and pass that phase's identity check; entry
+  never admits work a deeper phase rejects.

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

@@ -74,6 +74,23 @@ Reconcile named sessions, revisions, PR state, watches, and deliveries before
 creating or redelivering anything. Touch only this run; no global sweep, new
 runtime database, or scheduler follows from the record.
 
+## Decision trail and learnings
+
+Two fields make the record reviewable by a human who stepped away and usable
+by the next owner:
+
+- `Decisions` holds one row per consequential choice: when, what was chosen,
+  why in plain words, the evidence pointer that proves it (SHA, PR, receipt,
+  `file:line`, or artifact path, never a paragraph), and the result
+  (`tests green`, `reverted`, `held`, `open`). A choice that a reviewer could
+  not reconstruct from Git or PR state belongs here; routine mechanics do not.
+  An arena synthesis note (base, grafts and their source candidate,
+  rejections, judge verdicts) is recorded as `Decisions` rows.
+- `Learnings` holds what the next owner needs that the code and history do not
+  show: root causes, gotchas, patterns that held or failed, and where the
+  evidence lives. Read it on resume before reconciling; append, never rewrite,
+  and keep entries as short as the evidence pointer allows.
+
 ## Privacy
 
 Record concise IDs, SHAs, URLs, status, timestamps, next actions, and evidence
@@ -98,8 +115,13 @@ IDs: <repo/project + workspace/agent receipt pointers>
 Evidence: <check/review/submission/audit receipt pointers>
 Pending: <launch/acceptance/external receipts + timer execution heartbeat actual ID + handshake + deadline>
 Unresolved: <decision -> next owner + next action>
+Learnings: <root cause, gotcha, or pattern -> evidence ref>
 Resume: <commands or evidence refs bound to exact revisions>
 
+| When | Decision | Why | Evidence | Result |
+| --- | --- | --- | --- | --- |
+| <UTC timestamp> | <what was chosen> | <plain reason> | <SHA/PR/receipt/file:line> | <tests green/reverted/held/open> |
+
 | Task | Dependencies | Owner | State | Revision evidence | Next action |
 | --- | --- | --- | --- | --- | --- |
 | <task> | <task IDs or none> | <role + session ID + worktree, or receipt ref> | <pending/in progress/complete/blocked> | <SHA + check/receipt refs> | <action + owner> |

package/skills/axstack-align/SKILL.md

@@ -41,7 +41,9 @@ This preserves the required contracts -> lifecycle -> audit load edge.
 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. Then consult `axstack-advisor-astra` 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
 `axstack-advisor-fable` 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
@@ -56,6 +58,48 @@ unchanged receipts. Record compact adviser evidence, the driver's assessment,
 and user-resolved choices for `axstack-spec`. If either adviser is unavailable,
 hold Align; safe fact work may continue without substitution.
 
+## Arena for hard-to-reverse design choices
+
+Critique of one draft anchors every reader to that draft's shape. When a
+question is arena-grade, 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
+one arena round. 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.** `axstack-advisor-astra` and `axstack-advisor-fable` each
+   independently produce one candidate design plus a short rationale naming
+   the alternatives considered and rejected, from the same bounded evidence and
+   question, without cross-reading. The driver authors no candidate.
+3. **Cross-judge.** After both candidates are complete, `axstack-arena-judge-astra`
+   and `axstack-arena-judge-fable` each independently score every candidate
+   per criterion from the rubric and candidates by label, and recommend a base
+   with a reason. Judges never author, never cross-read each other.
+4. **Pick.** The driver reads every candidate end to end and scores per
+   criterion, not on holistic feel, then compares with both judges. Agreement
+   confirms the base. Disagreement between judges or with the driver means one
+   reading is biased or the rubric was ambiguous: re-read both rationales and
+   decide with a stated reason; never average verdicts or fabricate consensus.
+5. **Graft.** Walk the losing candidate 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, 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, both judge verdicts) as `Decisions` rows in the
+[run record](../axstack/references/run-record.md). Load
+[Orca runtime](../axstack/references/orca-runtime.md) immediately before the
+first candidate or judge dispatch. If either adviser or judge seat is
+unavailable, hold that question without substitution; unaffected fact work
+and questions continue.
+
 ## Bound the interview
 
 Twenty cumulative presented questions is the normal ceiling, not a target.

package/skills/axstack-explain/SKILL.md

@@ -33,6 +33,18 @@ immediately before an actual profile dispatch.
 4. For every gap, cite its inspected scope and applicable revision, stable
    source identity, or content hash. “Not found” never means app-wide missing
    without app-wide evidence; anything outside the inspected scope is unknown.
+5. For "what could this break" or "blast radius of X", follow
+   [Blast radius](../axstack/references/blast-radius.md): find the breakage
+   beyond the diff and prove the one safety fact by running real code, or
+   mark it unproven.
+6. For "show me your work" on existing work (a run, PR, branch, or change),
+   reconstruct the decision trail rather than re-describing the diff: read the
+   run record's `Decisions` and `Learnings`
+   ([run record](../axstack/references/run-record.md)), then the commits, PR
+   body, review receipts, and ADRs. Present each consequential choice as what
+   was chosen, why, the evidence pointer, and its result; separate what the
+   record proves from what is inferred, and list choices with no recorded
+   reason as open rather than inventing one.
 
 ## 2. Choose proportional output
 

package/skills/axstack-review/SKILL.md

@@ -134,7 +134,9 @@ session is the owner for every PR it handles.
    in authored mode the one reviewer covers all six angles:
    1. Security and trust boundaries.
    2. Correctness, failures, and edge cases.
-   3. Integration and regressions.
+   3. Integration and regressions: load [Blast radius](../axstack/references/blast-radius.md),
+      find what the change breaks beyond the diff, and grade the one fact it
+      is safe because of on the evidence ladder; below "ran it" is unproven.
    4. Requirements, acceptance, and user behavior.
    5. Architecture and solution design, including SOLID and credible simpler
       alternatives.
@@ -233,6 +235,7 @@ Verdict: <APPROVE | REQUEST_CHANGES | INCOMPLETE>
 Coverage: <angles + acceptance + executable evidence checked>
 Limitations: <unverified boundaries + why>
 Findings: <evidence + consequence each>
+Safety fact: <the one fact the change is safe because of> — <ladder step + proof | unproven>
 Escalate to user: <yes | no> — <criterion> — <reason>
 ```
 

package/src/roles.js

@@ -64,8 +64,8 @@ export function assessRoleReadiness(roles, preset) {
   const gaps = [];
   const isIntentionalAbsence = (role) => role.model === null && (
     (preset === 'mixed' && role.id === 'axstack-checker') ||
-    (preset === 'codex-only' && role.id === 'axstack-advisor-fable') ||
-    (preset === 'claude-only' && role.id === 'axstack-advisor-astra')
+    (preset === 'codex-only' && ['axstack-advisor-fable', 'axstack-arena-judge-fable'].includes(role.id)) ||
+    (preset === 'claude-only' && ['axstack-advisor-astra', 'axstack-arena-judge-astra'].includes(role.id))
   );
   for (const role of roles) {
     if (!bounds.has(role.provider)) {