fix: bound oversized task-check rewrites to one correction

## Model

GPT-6.1-Sol.

## Summary

Refs #521/#522/#489; FEATURE_SPEC5.3/5.4/5.5 respecting3/7. Actual local-c21ee2959a02b0ef consumed600s during repeated oversized TASK proposals. Give this specific length failure its existing single validation correction immediately, rather than spending the general turn budget rewriting the same invalid contract. Still-invalid correction explicitly stops before coder/publication. Preserve all existing changed-Ask, scope, locked-model and other rewrite behavior. No parsing waiver, requirement truncation or increased budget.

## Verification

New red20-turn regression confirmed prior repeated rewrite exhaustion. Final Node20 planner/acceptance/task/plan-mode/layout63/63pass,4.8s. Exactly two calls for unchanged invalid oversized draft, successful written and JSON corrections retained; existing Ask/protected-path/selected-model refusals unchanged. Last full trusted CI1720/1720green; new head requires CI. No private config, runtime deps/store, workflow/policy/vendor edit, dummy commit or invented evaluation.

node --test --test-concurrency=4 tests\planner.acceptance.test.mjs tests\planner.test.mjs tests\planner-task.test.mjs tests\plan-mode.test.mjs tests\test-layout.test.mjs

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

AI-Agent: r4yharris-agent-coder
AI-Model: GPT-6.1-Sol

Author: r4yharris-agent-coder[bot]

docs/MULTIAGENT.md

@@ -48,7 +48,10 @@ reject oversized checks explicitly rather than cutting off requirements.
 Reading existing task documents preserves complete checks, including legacy
 long checks, so source-validated acceptance continuation remains compatible.
 Model validation feedback requests concise complete wording within the existing
-turn budget. Title shortening is unchanged. This implements FEATURE_SPEC
+turn budget. Oversized written checks receive the single validation correction
+immediately, rather than repeating rewrites until the general budget expires.
+A still-invalid correction stops before coding and publication.
+Title shortening is unchanged. This implements FEATURE_SPEC
 sections 5.3-5.5; neither the original Ask nor its safety obligations are waived.
 Directory-document checks also recognize content phrasing such as "covering",
 "gives instructions" and "indexes"; operation-success assertions stay denied.

src/planner/stub.mjs

@@ -429,6 +429,16 @@ export async function planAsk(ask, {
           });
         } catch (error) {
           if (!(error instanceof Error)) throw error;
+          if (!/^[{\[]/.test(plannerPlanText(message.content)) &&
+              /^Acceptance check must be at most 240 characters/.test(error.message)) {
+            if (validationRepairUsed) return fallback(error.message, turn);
+            validationRepairUsed = true;
+            messages.push({ role: 'user', content:
+              `The written TASK.md is invalid (${redactSecrets(error.message, { env, apiKeyEnv: config.llm.api_key_env })}). ` +
+              'Use this single validation correction to return the complete required JSON plan. ' +
+              'Keep every requirement and the original Ask; the harness will write the validated artifacts.' });
+            continue;
+          }
           if (!/^[{\[]/.test(plannerPlanText(message.content)) && turn >= budget + Number(repairUsed)) {
             if (validationRepairUsed) return fallback(error.message, turn);
             if (!/^Acceptance checks must contain 1-8 entries/.test(error.message)) {

tests/planner.acceptance.test.mjs

@@ -64,6 +64,28 @@ test('an oversized written TASK stays invalid until a complete concise JSON corr
   assert.deepEqual(parseTaskDocument(result.task, { expectedAsk: ask }).acceptance_checks, [concise]);
 });
 
+test('a repeatedly invalid written TASK stops after one correction even with a large general turn budget', async () => {
+  const config = parseConfig(readFileSync(new URL('../roster.config.example.yml', import.meta.url), 'utf8')
+    .replace('base_url: ""', 'base_url: "http://localhost:1234/v1"').replace('model: ""', 'model: test-model'));
+  const valid = buildPlan(ask, { ...options, acceptanceChecks: ['The guide contains setup instructions'] }).task;
+  const invalid = valid.replace('The guide contains setup instructions', check(241));
+  let calls = 0;
+  const result = await planAsk(ask, { config: { ...config, planner: { ...config.planner, turn_budget: 20 } }, env: {},
+    tools: { write_file: async () => ({ ok: true }) },
+    fetchImpl: async () => {
+      calls += 1;
+      return Response.json({ choices: [{ finish_reason: 'tool_calls', message: {
+        role: 'assistant', tool_calls: [{ id: `invalid-${calls}`, type: 'function', function: {
+          name: 'write_file', arguments: JSON.stringify({ path: 'TASK.md', content: invalid }),
+        } }],
+      } }] });
+    } });
+  assert.equal(calls, 2);
+  assert.equal(result.mode, 'stub');
+  assert.match(result.error, /240.*shorten/);
+  assert.equal(parseTaskDocument(result.task, { expectedAsk: ask }).ask, ask);
+});
+
 test('slice JSON correction retains the original Ask and does not accept a truncated requirement', async () => {
   const config = parseConfig(readFileSync(new URL('../roster.config.example.yml', import.meta.url), 'utf8')
     .replace('base_url: ""', 'base_url: "http://localhost:1234/v1"').replace('model: ""', 'model: test-model'));