feat: add FableLoom autopilot self-improvement

client/src/components/fableloom/LoomEditorialAutomation.jsx

@@ -81,6 +81,7 @@ function FindingLink({ loom, finding }) {
 export default function LoomEditorialAutomation({ loom, dirty, onLoomUpdate }) {
   const [route, setRoute] = useState({ providerId: '', model: '', effort: '' });
   const [maxRounds, setMaxRounds] = useState(3);
+  const [selfImprove, setSelfImprove] = useState(false);
   const [result, setResult] = useState(null);
   const [autopilotRun, setAutopilotRun] = useState(null);
   const handledTerminalRunRef = useRef(null);
@@ -165,6 +166,7 @@ export default function LoomEditorialAutomation({ loom, dirty, onLoomUpdate }) {
     const run = await startLoomEditorialAutopilot(loom.id, {
       ...routeBody,
       maxRounds,
+      ...(selfImprove ? { selfImprove: true } : {}),
     }, { silent: true });
     handledTerminalRunRef.current = null;
     setAutopilotRun(run);
@@ -194,7 +196,7 @@ export default function LoomEditorialAutomation({ loom, dirty, onLoomUpdate }) {
   const findings = result?.type === 'autopilot' || shownRun?.status
     ? shownRun?.residualFindings || []
     : review?.findings?.length ? review.findings : evaluation?.findings || [];
-  const summary = review?.summary || evaluation?.summary || shownRun?.message;
+  const summary = shownRun?.message || review?.summary || evaluation?.summary;
   const passed = result?.type === 'playtest'
     ? response?.passed
     : shownRun?.status === 'completed' || diagnostics?.passed;
@@ -262,6 +264,25 @@ export default function LoomEditorialAutomation({ loom, dirty, onLoomUpdate }) {
         </p>
       ) : null}
 
+      <div className="rounded border border-port-border bg-port-bg/30 px-3 py-2">
+        <label htmlFor="fableloom-editorial-self-improve" className="flex items-start gap-2 text-xs text-port-text">
+          <input
+            id="fableloom-editorial-self-improve"
+            type="checkbox"
+            checked={selfImprove}
+            onChange={(event) => setSelfImprove(event.target.checked)}
+            disabled={busy}
+            className="mt-0.5"
+          />
+          <span>
+            <span className="font-medium">Improve FableLoom itself when autopilot stalls or fails</span>
+            <span className="mt-1 block text-[11px] text-port-text-muted">
+              Runs one content-free, budget-gated post-mortem to distinguish story work from a broken or inefficient editor/reviewer workflow. A confident PortOS defect queues a deduplicated worktree + PR CoS task in the approval queue; it never starts by itself. Healthy and canceled runs spend nothing.
+            </span>
+          </span>
+        </label>
+      </div>
+
       <div className="grid gap-2 sm:grid-cols-3">
         <button
           type="button"
@@ -377,6 +398,25 @@ export default function LoomEditorialAutomation({ loom, dirty, onLoomUpdate }) {
               {shownRun.rounds.length} bounded editorial round{shownRun.rounds.length === 1 ? '' : 's'} recorded. The run stops on success, plateau, cancellation, provider failure, or the selected round limit.
             </p>
           ) : null}
+          {shownRun?.selfImprove?.verdict === 'pipeline' ? (
+            <div className="rounded border border-port-accent/30 bg-port-accent/5 px-3 py-2 text-xs text-port-accent">
+              <p>
+                {shownRun.selfImprove.duplicate
+                  ? `FableLoom improvement already tracked (${shownRun.selfImprove.area}): ${shownRun.selfImprove.title}`
+                  : shownRun.selfImprove.filed
+                    ? `Queued a FableLoom improvement (${shownRun.selfImprove.area}): ${shownRun.selfImprove.title}`
+                    : `Diagnosed a FableLoom workflow defect (${shownRun.selfImprove.area}), but task filing failed: ${shownRun.selfImprove.title}`}
+              </p>
+              {shownRun.selfImprove.taskId ? (
+                <Link
+                  to={`/cos/tasks?task=${encodeURIComponent(shownRun.selfImprove.taskId)}&source=internal`}
+                  className="mt-1 inline-block font-medium hover:underline"
+                >
+                  Review CoS task
+                </Link>
+              ) : null}
+            </div>
+          ) : null}
         </div>
       ) : null}
     </section>

client/src/components/fableloom/LoomEditorialAutomation.test.jsx

@@ -127,6 +127,63 @@ describe('LoomEditorialAutomation', () => {
     expect(screen.getAllByText(/Round 1: evaluating and remediating/).length).toBeGreaterThan(0);
   });
 
+  it('opts a run into approval-gated FableLoom workflow diagnosis', async () => {
+    const user = userEvent.setup();
+    api.startLoomEditorialAutopilot.mockResolvedValue({
+      id: 'editorial-run-1', loomId: 'loom-1', status: 'running', round: 0, maxRounds: 3,
+      message: 'Starting FableLoom editorial autopilot…', rounds: [], residualFindings: [],
+    });
+    renderPanel();
+
+    const toggle = await screen.findByLabelText(/improve fableloom itself/i);
+    expect(toggle).not.toBeChecked();
+    await user.click(toggle);
+    expect(screen.getByText(/queues a deduplicated worktree \+ PR CoS task in the approval queue/i)).toBeInTheDocument();
+    await user.click(screen.getByRole('button', { name: 'Start editor autopilot' }));
+
+    await waitFor(() => expect(api.startLoomEditorialAutopilot).toHaveBeenCalledWith(
+      'loom-1', { maxRounds: 3, selfImprove: true }, { silent: true },
+    ));
+  });
+
+  it('links a filed workflow diagnosis to its CoS approval task', async () => {
+    api.getLoom.mockResolvedValue(loom);
+    api.getLoomEditorialAutopilotStatus.mockResolvedValue({
+      run: {
+        id: 'editorial-run-1', loomId: 'loom-1', status: 'paused', round: 2, maxRounds: 3,
+        pauseReason: 'plateau', message: 'Editorial autopilot paused.', rounds: [],
+        residualFindings: [],
+        selfImprove: {
+          verdict: 'pipeline', area: 'prompt', title: 'Tighten the remediation contract',
+          taskId: 'sys-example', filed: true, duplicate: false,
+        },
+      },
+    });
+    renderPanel();
+
+    expect(await screen.findByText(/Queued a FableLoom improvement \(prompt\)/)).toBeInTheDocument();
+    expect(screen.getByRole('link', { name: 'Review CoS task' })).toHaveAttribute(
+      'href',
+      '/cos/tasks?task=sys-example&source=internal',
+    );
+  });
+
+  it('keeps the decided pause explanation visible when cancellation lands during diagnosis', async () => {
+    api.getLoomEditorialAutopilotStatus.mockResolvedValue({
+      run: {
+        id: 'editorial-run-1', loomId: 'loom-1', status: 'canceled', round: 1, maxRounds: 1,
+        pauseReason: 'round-limit',
+        message: 'Editorial autopilot reached its 1-round limit with review findings still open.',
+        rounds: [], residualFindings: [],
+        lastReview: { summary: 'The story review still has open findings.', findings: [] },
+      },
+    });
+    renderPanel();
+
+    expect(await screen.findByText(/reached its 1-round limit with review findings still open/i))
+      .toBeInTheDocument();
+  });
+
   it('blocks every mutating AI action while the series plan has unsaved edits', async () => {
     renderPanel({ dirty: true });
 

client/src/services/README.md

@@ -103,7 +103,7 @@ toasts on throw). **Custom catch ⇒ `silent: true`** — otherwise toasts fire
 | `apiMediaJobs.js` | Media generation job tracking + `refineMediaPrompt` / `promptFromMedia` (vision reverse-prompt). |
 | `apiCreativeDirector.js` | Creative Director (video production). |
 | `apiCreativeCommission.js` | Creative Commissions (Autonomous Creation Engine — standing recurring briefs). |
-| `apiFableLoom.js` | FableLoom branching narratives — loom/episode/scene-node/transition CRUD, deterministic graph validation, and the AI lanes (weave, branch, review, play turns, per-episode scene reformat). |
+| `apiFableLoom.js` | FableLoom branching narratives — loom/episode/scene-node/transition CRUD, deterministic graph validation, AI authoring lanes, and the bounded editorial/playthrough autopilot lifecycle. |
 | `apiGames.js` | Game studio records, managed-app binding, reusable sprite/music bindings, deterministic asset-bundle compilation/integrity preflight, and AI feedback history. |
 | `apiMusicVideo.js` | Music Video projects + scene board + audio analysis. |
 | `apiSprites.js` | Sprite Manager records, asset library, production-set import (#2895), reference workflow: create/generate/lock (#2896), directional walk and per-track generation/approval, animation-type definition CRUD (#3153), trim/postprocess, and per-run source-frame listing for the Loop Trimmer's re-derive (#2980), and animation render-provider readiness (#4876). |

client/src/services/apiFableLoom.test.js

@@ -70,9 +70,9 @@ describe('apiFableLoom', () => {
       method: 'POST', body: JSON.stringify({ aiReview: true }),
     });
 
-    await api.startLoomEditorialAutopilot('loom-1', { maxRounds: 3 });
+    await api.startLoomEditorialAutopilot('loom-1', { maxRounds: 3, selfImprove: true });
     expect(request).toHaveBeenCalledWith('/fableloom/loom-1/editorial/autopilot/start', {
-      method: 'POST', body: JSON.stringify({ maxRounds: 3 }),
+      method: 'POST', body: JSON.stringify({ maxRounds: 3, selfImprove: true }),
     });
 
     await api.getLoomEditorialAutopilotStatus('loom-1', { silent: true });

data.reference/prompts/stage-config.json

@@ -1080,6 +1080,13 @@
       "returnsJson": true,
       "variables": []
     },
+    "fableloom-editorial-self-improve": {
+      "name": "FableLoom — Editorial Autopilot Self-Improvement",
+      "description": "Opt-in, content-free post-mortem for a paused or failed FableLoom editor/reviewer loop. Distinguishes unfinished story work from an inefficient, broken, or incomplete PortOS workflow and returns a bounded software diagnosis; a confident PortOS verdict queues a deduplicated, approval-gated CoS task.",
+      "model": "heavy",
+      "returnsJson": true,
+      "variables": []
+    },
     "fableloom-review-playthroughs": {
       "name": "FableLoom — Review Playthrough Variations",
       "description": "Reviews deterministically enumerated playthrough variations for mechanics, causal branch coherence, audience agency, continuity, pacing, ending payoff, canon, and ship-quality storytelling.",

data.reference/prompts/stages/fableloom-editorial-self-improve.md

@@ -0,0 +1,83 @@
+# FableLoom Editorial Autopilot — Self-Improvement Diagnosis
+
+You are a **systems engineer** doing a post-mortem on PortOS's automated FableLoom editor/reviewer loop. Decide whether the run's trouble came from the **story** (the interactive series still needs editorial work, while the automation behaved correctly) or from **PortOS** (its workflow, prompt, validation, runner, configuration, or UI is broken, wasteful, or unable to apply the repair it claims to perform).
+
+You are NOT reviewing the story. You receive no story prose, titles, record ids, or finding text. Do not invent or request them. Your output may become a CoS coding-task brief and a public pull request, so it must describe **software behavior only**.
+
+## The workflow
+
+Each bounded round performs these steps in order:
+
+1. `evaluate-remediate` — inspect the complete existing series and return a safe sparse patch. It may repair series-plan fields, episode outlines, existing scene metadata, and existing transition labels/targets, but it must preserve episode, scene, and transition membership and ids. PortOS rejects invalid output and graph/outline/playthrough/continuity regressions before saving.
+2. `playthrough-review` — deterministically enumerate bounded path variations, run structural and continuity diagnostics, and ask an independent story editor for a quality verdict.
+3. If findings remain, translate them into guidance for the next remediation pass. Stop when the story passes, the same findings survive a no-change pass (`plateau`), the selected round limit is reached, the user cancels, or a step throws.
+
+The post-mortem is opt-in and runs only after a pause or failure. A confident `pipeline` verdict (PortOS is at fault) queues a deduplicated, worktree-isolated, PR-opening CoS task. The task waits for human approval before an agent starts. Reporting `content` or `none` is useful and files nothing.
+
+## The run
+
+- **Outcome:** `{{outcome}}`
+- **Reason:** `{{outcomeReason}}`
+- **Step active when the run stopped:** `{{currentStep}}`
+- **Error code:** `{{errorCode}}`
+- **Round:** {{round}} / {{maxRounds}}
+- **Playthrough path cap:** {{maxPaths}}
+
+Content-free round telemetry:
+
+```json
+{{telemetryJson}}
+```
+
+Each round reports only counters and booleans:
+
+- `remediation.changed` / `changeCount` — whether a sparse patch actually changed the loom and how many change notes the stage returned.
+- `before` / `after` — deterministic blocker counts around the accepted patch.
+- `evaluationFindingCount` — how many issues the remediation stage itself identified.
+- `diagnosticsPassed` / `diagnosticFindingCount` — whether non-narrative editorial diagnostics cleared.
+- `deterministicPassed` / `deterministicComplete` / `stats` — path enumeration health and coverage.
+- `reviewPassed` / `qualityScore` / finding counts by category and severity — the independent quality verdict, without any story text.
+
+## What counts as a PORTOS problem
+
+Return `pipeline` only when the counters, terminal reason, or error code support a concrete software change. Strong cases include:
+
+1. **An ineffective loop step.** Remediation repeatedly reports no applicable change while its own findings remain, accepted edits do not reduce the relevant deterministic blockers, or next-round guidance cannot target what the review produced.
+2. **A broken output contract or validator.** A stage returns unusable data, claims edits that do not apply, produces a shape its consumer cannot read, or a safe intended repair is rejected for a contract PortOS should support.
+3. **A missing or late gate.** Expensive full-series editing discovers a structural condition that a cheaper outline/graph preflight could have surfaced before expansion.
+4. **A runner or lifecycle defect.** The wrong provider route is used, cancellation/status handling is inconsistent, a failure is swallowed, or a terminal run is left active.
+5. **A bad bound or missing control.** A round/path limit makes the workflow unable to do its job, or the editor needs a specific user-facing option that does not exist.
+6. **A wasteful ordering.** The workflow performs costly review before required deterministic evidence is available, or repeats work that an earlier retained result should have made unnecessary.
+
+## What is NOT a PortOS problem
+
+- A story genuinely needs another authored choice, consequence, scene, or ending, and the loop correctly pauses instead of changing graph membership → `content`.
+- The round limit is reached while distinct story findings continue to improve from round to round → `content`.
+- A plateau alone, without counters showing an automation mismatch → `content` or `none`; do not assume every hard editorial problem is a software bug.
+- A one-off provider or network failure with no software evidence → `none`.
+- Any theory that requires story text or source inspection not present here → `none` with low confidence.
+
+## Output contract
+
+Return ONLY valid JSON matching this shape — no prose, no markdown fence, no commentary.
+
+```json
+{
+  "verdict": "pipeline",
+  "confidence": 0.0,
+  "area": "pipeline-step",
+  "title": "string — imperative, one line, names the software defect",
+  "problem": "string — what is wrong in PortOS and which counters/status values show it",
+  "evidence": ["string — a specific counter, status, reason, or error code from the telemetry above"],
+  "proposedChange": "string — the smallest concrete source/prompt/check/config/UI change that fixes it",
+  "risks": "string — behavior that must keep working"
+}
+```
+
+- `verdict` must be `pipeline` (PortOS is at fault), `content` (the story needs work), or `none`.
+- `confidence` is 0–1. Diagnoses below 0.6 are discarded; a guess should be discarded.
+- `area` must be `editorial-check`, `pipeline-step`, `prompt`, `runner`, `config`, or `ui`.
+- `evidence` has at most 8 entries and may cite only the supplied content-free telemetry.
+- `proposedChange` must be actionable without access to this run or its story.
+
+When the story needs work or the evidence is inconclusive, return the same shape with `verdict` set to `content` or `none` and empty change fields.

docs/features/fableloom.md

@@ -98,6 +98,13 @@ scene-local clothes cannot silently reappear between episodes.
   are client-side state (restart is free; nothing persists server-side).
 - **Story settings drawer** — audience role and communication medium, scene
   format (plus the rewrite pass), and the narrator's provider/model/effort pin.
+- **AI editor, reviewer & playtest** — a whole-series remediation pass and a
+  bounded autopilot that alternates safe edits with deterministic + narrative
+  path review. Its optional **Improve FableLoom itself** post-mortem runs only
+  after a pause or failure, sends content-free counters (never story records)
+  through one budget-gated diagnosis, and queues a deduplicated PortOS CoS task
+  when the workflow rather than the story is at fault. The task remains in the
+  approval queue; the per-run checkbox does not grant unattended source edits.
 - **Series detail page** (`/pipeline/series/:seriesId`) — a "Branching
   narratives" card lists the looms linked to that series (counts + a link into
   the editor) and spawns a new one pre-linked to the series and its universe.

scripts/migrations/322-fableloom-editorial-self-improve-stage.js

@@ -0,0 +1,5 @@
+/** Seed the FableLoom editorial-autopilot self-improvement stage. */
+
+import { makeSeedMigration } from './_seedStageHelpers.js';
+
+export default makeSeedMigration('fableloom-editorial-self-improve');

scripts/migrations/322-fableloom-editorial-self-improve-stage.test.js

@@ -0,0 +1,12 @@
+import { describe } from 'vitest';
+
+import migration from './322-fableloom-editorial-self-improve-stage.js';
+import { runSeedStageMigrationTests } from './_seedStageTestHelpers.js';
+
+describe('migration 322 — seed the FableLoom editorial self-improvement stage', () => {
+  runSeedStageMigrationTests({
+    migration,
+    stages: ['fableloom-editorial-self-improve'],
+    prefix: 'migration-322-',
+  });
+});

server/lib/fableLoomValidation.js

@@ -358,6 +358,10 @@ export const playthroughReviewSchema = z.object({
 export const editorialAutopilotStartSchema = z.object({
   maxRounds: z.number().int().min(1).max(LOOM_LIMITS.EDITORIAL_AUTOPILOT_ROUNDS_MAX).optional(),
   maxPaths: z.number().int().min(1).max(FABLELOOM_PLAYTEST_LIMITS.MAX_PATHS).optional(),
+  // Opt-in post-mortem over content-free run counters. A confident PortOS
+  // verdict queues an approval-gated CoS task; healthy/canceled runs spend
+  // nothing, and the story itself never crosses into the task brief.
+  selfImprove: z.boolean().optional(),
   ...llmPickFields,
 });
 

server/lib/promptStageCallSites.generated.json

@@ -37,6 +37,9 @@
   "fableloom-editorial-remediate": [
     "server/services/fableLoom/editorial.js"
   ],
+  "fableloom-editorial-self-improve": [
+    "server/services/fableLoom/editorialSelfImprove.js"
+  ],
   "fableloom-feedback-episode": [
     "server/services/fableLoom/weave.js"
   ],

server/lib/socketEventCatalog.generated.json

@@ -2261,7 +2261,7 @@
         {
           "direction": "server-to-client",
           "source": "client/src/components/fableloom/LoomPlayPanel.jsx",
-          "line": 114
+          "line": 125
         }
       ]
     },
@@ -2300,7 +2300,7 @@
         {
           "direction": "server-to-client",
           "source": "client/src/components/fableloom/LoomPlayPanel.jsx",
-          "line": 145
+          "line": 156
         }
       ]
     },
@@ -2339,7 +2339,7 @@
         {
           "direction": "server-to-client",
           "source": "client/src/components/fableloom/LoomPlayPanel.jsx",
-          "line": 137
+          "line": 148
         }
       ]
     },
@@ -2365,7 +2365,7 @@
         {
           "direction": "server-to-client",
           "source": "client/src/components/fableloom/LoomPlayPanel.jsx",
-          "line": 118
+          "line": 129
         }
       ]
     },
@@ -2396,7 +2396,7 @@
         {
           "direction": "server-to-client",
           "source": "client/src/components/fableloom/LoomPlayPanel.jsx",
-          "line": 122
+          "line": 133
         }
       ]
     },
@@ -2409,7 +2409,7 @@
         {
           "direction": "server-to-client",
           "source": "client/src/components/fableloom/LoomPlayPanel.jsx",
-          "line": 126
+          "line": 137
         }
       ]
     },