fix: preserve worker authority through start observation

config/reliability-gates.jsonc

@@ -13025,7 +13025,7 @@
         "https://github.com/stablyai/orca/issues/13821",
         "https://github.com/stablyai/orca/issues/14347"
       ],
-      "invariant": "Injected orchestration task prompts for recognized agent CLIs must send the prompt body inside one bracketed-paste frame, sanitize embedded ESC bytes, preserve chunk boundaries without losing the frame, and submit exactly once only after the agent can accept Enter. A successful orchestration.workerStart must durably record exactly one accepted and started turn; a swallowed Enter must fail with agent_prompt_stalled and never trigger a blind rescue Enter. Claude and Codex must emit a post-paste composer marker and then settle, or reach the bounded fallback first; every other agent retains the platform delay.",
+      "invariant": "Injected orchestration task prompts for recognized agent CLIs must send the prompt body inside one bracketed-paste frame, sanitize embedded ESC bytes, preserve chunk boundaries without losing the frame, and submit exactly once only after the agent can accept Enter. Local worker-start with supported observation must preserve an unobserved turn as start_unknown without revoking authority, closing questions, or triggering a rescue Enter; a worker report during observation must settle normally. Claude and Codex must emit a post-paste composer marker and then settle, or reach the bounded fallback first; every other agent retains the platform delay.",
       "oracle": "Runtime tests assert the exact PTY write sequence, failure cleanup, Claude/Codex marker-gated multi-frame renders, and the legacy platform delay for every other configured agent. The candidate resets settlement on later frames, gives a late marker a fresh bounded window, and still submits once at the hard deadline if output never settles. The worker-start contract drives the production RPC through a delayed fake Codex composer and independently checks exact turn/Enter counts plus reopened SQLite Task, Dispatch, worker receipt, and mutation receipt state for accepted and swallowed outcomes. Other orchestration tests assert dispatch/coordinator use the agent prompt path; the live CLI harness covers long Codex-like framing.",
       "commands": [
         "pnpm exec vitest run --config config/vitest.config.ts src/shared/agent-prompt-injection.test.ts src/main/runtime/orca-runtime.test.ts src/main/runtime/rpc/methods/orchestration/runs/tasks-dispatch.test.ts src/main/runtime/orchestration/coordinator.test.ts",
@@ -13075,7 +13075,8 @@
           "file": "src/main/runtime/rpc/methods/orchestration/worker/worker-start-prompt-contract.test.ts",
           "assertions": [
             "delayed composer readiness produces exactly one submitted and started turn with no premature Enter and durable ready receipts",
-            "a swallowed Enter records agent_prompt_stalled across Task, Dispatch, worker, and mutation receipts without a rescue Enter"
+            "a swallowed Enter durably records start_unknown without a rescue Enter or capability revocation",
+            "early worker reports settle during observation, and outstanding questions survive observation uncertainty"
           ]
         },
         {

src/cli/handlers/orchestration-worker-cli.test.ts

@@ -104,6 +104,34 @@ describe('orchestration worker-start CLI contract', () => {
     expect(process.exitCode).toBeUndefined()
   })
 
+  it.each(['succeeded', 'failed'])(
+    'accepts a successful start whose task already %s',
+    async (workerOutcome) => {
+      const receipt = {
+        taskId: 'task_1',
+        dispatchId: 'ctx_1',
+        state: 'ready',
+        stage: 'settled',
+        workerOutcome,
+        effects: [],
+        residualResources: []
+      }
+      callMock.mockResolvedValue({ result: receipt })
+      await invokeWorkerStart(
+        new Map([
+          ['task', 'task_1'],
+          ['from', 'term_coord']
+        ])
+      )
+      expect(process.exitCode).toBeUndefined()
+      expect(printResult).toHaveBeenCalledWith(
+        expect.objectContaining({ result: receipt }),
+        true,
+        expect.any(Function)
+      )
+    }
+  )
+
   it('capability-gates and forwards per-invocation launch preferences', async () => {
     callMock
       .mockResolvedValueOnce({

src/main/runtime/orchestration/db/dispatch-context/worker-report-settlement.ts

@@ -109,9 +109,13 @@ export function settleWorkerReportInTransaction(
     (dispatch.status === 'pending' || dispatch.status === 'dispatched') &&
     task.status === 'blocked' &&
     reportingWorker?.state === 'start_unknown'
+  const reportingStart =
+    dispatch.status === 'pending' &&
+    task.status === 'dispatched' &&
+    reportingWorker?.state === 'starting'
   const previousDispatchStatus = settledByUnobservedPrompt
     ? 'failed'
-    : reconnectingStart
+    : reconnectingStart || reportingStart
       ? dispatch.status
       : 'dispatched'
   const previousTaskStatus = settledByUnobservedPrompt
@@ -198,7 +202,7 @@ export function settleWorkerReportInTransaction(
     const dispatchTransition = transitionLifecycleWithDb(this.db, {
       entity: 'dispatch',
       id: params.dispatchId,
-      from: reconnectingStart ? ['pending', 'dispatched'] : 'dispatched',
+      from: reconnectingStart || reportingStart ? ['pending', 'dispatched'] : 'dispatched',
       to: expectedDispatchStatus,
       projection: {
         completed_at: new Date().toISOString(),
@@ -234,11 +238,11 @@ export function settleWorkerReportInTransaction(
       projection: { stage: 'settled', updated_at: new Date().toISOString() },
       correction: 'unobserved_prompt_report'
     })
-  } else if (reconnectingStart && params.outcome === 'succeeded') {
+  } else if ((reconnectingStart || reportingStart) && params.outcome === 'succeeded') {
     transitionLifecycleWithDb(this.db, {
       entity: 'worker',
       id: params.dispatchId,
-      from: 'start_unknown',
+      from: reportingStart ? 'starting' : 'start_unknown',
       to: 'ready'
     })
     transitionLifecycleWithDb(this.db, {
@@ -253,7 +257,7 @@ export function settleWorkerReportInTransaction(
       entity: 'worker',
       id: params.dispatchId,
       // A start_unknown success report reconnects through 'ready' above; only failure settles here.
-      from: params.outcome === 'succeeded' ? 'ready' : ['ready', 'start_unknown'],
+      from: params.outcome === 'succeeded' ? 'ready' : ['ready', 'start_unknown', 'starting'],
       to: params.outcome === 'succeeded' ? 'succeeded' : 'failed',
       projection: { stage: 'settled', updated_at: new Date().toISOString() }
     })

src/main/runtime/orchestration/db/worker-dispatch/worker-dispatch-outcome.ts

@@ -155,7 +155,7 @@ export function markWorkerStartUnknown(
       from: 'dispatched',
       to: 'blocked'
     })
-    this.closeQuestionsForDispatch(dispatchId)
+    // Authority survives uncertainty, so its outstanding questions must remain answerable.
     this.db.exec('COMMIT')
     return this.getWorkerDispatch(dispatchId) as WorkerDispatchRow
   } catch (error) {

src/main/runtime/rpc/methods/orchestration/worker/worker-start-prompt-contract.test.ts

@@ -41,6 +41,8 @@ const openDatabases: OrchestrationDb[] = []
 const temporaryRoots: string[] = []
 
 type PromptContractHarness = {
+  runtime: Awaited<ReturnType<typeof createAgentPromptSubmissionRuntime>>['runtime']
+  handle: string
   db: OrchestrationDb
   dbPath: string
   dispatcher: RpcDispatcher
@@ -131,6 +133,8 @@ async function createPromptContractHarness(
   vi.spyOn(runtime, 'getTerminalOrchestrationCliCommand').mockReturnValue('orca')
 
   return {
+    runtime,
+    handle,
     db,
     dbPath,
     dispatcher: new RpcDispatcher({ runtime, methods: ORCHESTRATION_METHODS }),
@@ -227,6 +231,80 @@ describe('orchestration worker-start prompt contract', () => {
     })
   })
 
+  it.each([
+    ['succeeded', false],
+    ['succeeded', true],
+    ['failed', false],
+    ['failed', true]
+  ] as const)('preserves an early %s report with turn evidence=%s', async (outcome, observed) => {
+    vi.useFakeTimers()
+    const harness = await createPromptContractHarness('swallowed')
+    vi.spyOn(harness.runtime, 'observeTerminalAgentPrompt').mockImplementation(
+      async (_handle, prompt) => {
+        const dispatch = harness.db.findActiveDispatchForAssignee(harness.handle)
+        expect(dispatch).toBeDefined()
+        expect(
+          harness.db.settleWorkerReport({
+            taskId: harness.taskId,
+            dispatchId: dispatch!.id,
+            outcome,
+            result: 'Finished before the hook arrived'
+          })
+        ).toMatchObject({ action: 'settled', outcome })
+        return observed ? { ...prompt, stages: ['input_accepted', 'turn_started'] } : prompt
+      }
+    )
+    const pending = harness.dispatcher.dispatch(harness.request)
+    await vi.runAllTimersAsync()
+    expect(await pending).toMatchObject({
+      ok: true,
+      result: { state: 'ready', stage: 'settled', workerOutcome: outcome }
+    })
+    expect(harness.db.getTask(harness.taskId)?.status).toBe(
+      outcome === 'succeeded' ? 'completed' : 'failed'
+    )
+  })
+
+  it('retains accepted authority when the observation binding becomes stale', async () => {
+    vi.useFakeTimers()
+    const harness = await createPromptContractHarness('swallowed')
+    vi.spyOn(harness.runtime, 'observeTerminalAgentPrompt').mockRejectedValue(
+      new Error('terminal_handle_stale')
+    )
+    const pending = harness.dispatcher.dispatch(harness.request)
+    await vi.runAllTimersAsync()
+    expect(await pending).toMatchObject({ ok: true, result: { state: 'outcome_unknown' } })
+    expect(harness.db.findActiveDispatchForAssignee(harness.handle)).toMatchObject({
+      status: 'pending',
+      capability_hash: expect.any(String),
+      capability_revoked_at: null
+    })
+    expect(harness.submittedTurns()).toBe(1)
+    expect(vi.getTimerCount()).toBe(0)
+  })
+
+  it('keeps a worker question answerable after turn observation times out', async () => {
+    vi.useFakeTimers()
+    const harness = await createPromptContractHarness('swallowed')
+    let questionId = ''
+    vi.spyOn(harness.runtime, 'observeTerminalAgentPrompt').mockImplementation(
+      async (_handle, prompt) => {
+        const dispatch = harness.db.findActiveDispatchForAssignee(harness.handle)!
+        questionId = harness.db.createQuestion({
+          runId: dispatch.run_id!,
+          dispatchId: dispatch.id,
+          askerHandle: harness.handle,
+          question: 'Which target should I use?'
+        }).question.message_id
+        return prompt
+      }
+    )
+    const pending = harness.dispatcher.dispatch(harness.request)
+    await vi.runAllTimersAsync()
+    expect(await pending).toMatchObject({ ok: true, result: { state: 'outcome_unknown' } })
+    expect(harness.db.getQuestion(questionId)?.status).toBe('pending')
+  })
+
   it('reports a swallowed Enter as start_unknown while keeping the worker and its capability', async () => {
     vi.useFakeTimers()
     const harness = await createPromptContractHarness('swallowed')
@@ -273,7 +351,7 @@ describe('orchestration worker-start prompt contract', () => {
     expect(persisted.getWorkerDispatch(dispatchId)).toMatchObject({
       state: 'start_unknown',
       stage: 'turn_start_unobserved',
-      last_error: expect.stringContaining('never started a turn')
+      last_error: expect.stringContaining('turn start could not be verified')
     })
     const persistedEffects = JSON.parse(
       persisted.getWorkerDispatch(dispatchId)?.effects ?? '[]'

src/main/runtime/rpc/methods/orchestration/worker/worker-start-readiness-settlement.ts

@@ -85,7 +85,10 @@ export async function deliverAndSettleWorkerStartReadiness(args: {
     setupReceipt: args.setupReceipt,
     effects
   })
-  if (turnStart.verdict === 'unobserved') {
+  // A worker report can settle the dispatch while turn observation is outstanding.
+  const currentWorker = db.getWorkerDispatch(args.dispatchId)
+  const alreadySettled = currentWorker && currentWorker.state !== 'starting'
+  if (turnStart.verdict === 'unobserved' && !alreadySettled) {
     // Honest `unverifiable`: keep the dispatch capability and the terminal — the worker may
     // still recover and report (worker-report settlement reconnects a start_unknown worker) —
     // but never claim ready for a turn nobody observed.
@@ -125,13 +128,21 @@ export async function deliverAndSettleWorkerStartReadiness(args: {
       ...(args.terminalRevealWarning ? { warning: args.terminalRevealWarning } : {})
     }
   }
-  const worker = db.markWorkerDispatchReady(args.dispatchId, effects)
+  const worker = alreadySettled
+    ? currentWorker
+    : db.markWorkerDispatchReady(args.dispatchId, effects)
+  // A completed task proves start succeeded; older callers use only 'ready' as start success.
+  const reportedOutcome =
+    worker.stage === 'settled' && (worker.state === 'succeeded' || worker.state === 'failed')
+      ? worker.state
+      : undefined
   return {
     runId: run.id,
     taskId: task.id,
     dispatchId: args.dispatchId,
-    state: worker.state,
+    state: reportedOutcome ? 'ready' : worker.state,
     stage: worker.stage,
+    ...(reportedOutcome ? { workerOutcome: reportedOutcome } : {}),
     turnStart: turnStart.verdict,
     setup: args.setupReceipt,
     launch: args.launchReceipt,

src/main/runtime/rpc/methods/orchestration/worker/worker-start-turn-observation.test.ts

@@ -107,4 +107,14 @@ describe('observeWorkerTurnStart', () => {
     ).resolves.toEqual({ verdict: 'unsupported', prompt })
     expect(observe).not.toHaveBeenCalled()
   })
+
+  it('preserves uncertainty when observation loses the terminal binding', async () => {
+    const prompt = delivery()
+    const { runtime, observe } = runtimeObserving(prompt)
+    observe.mockRejectedValue(new Error('terminal_handle_stale'))
+    await expect(
+      observeWorkerTurnStart({ runtime, terminalHandle: 'term_w', prompt })
+    ).resolves.toEqual({ verdict: 'unobserved', prompt })
+    expect(observe).toHaveBeenCalledTimes(1)
+  })
 })

src/main/runtime/rpc/methods/orchestration/worker/worker-start-turn-observation.ts

@@ -56,11 +56,17 @@ export async function observeWorkerTurnStart(args: {
   if (verdict !== 'unobserved') {
     return { verdict, prompt: args.prompt }
   }
-  const observed = await args.runtime.observeTerminalAgentPrompt(
-    args.terminalHandle,
-    args.prompt,
-    args.timeoutMs ?? AGENT_PROMPT_EFFECT_TIMEOUT_MS
-  )
+  let observed: RuntimeTerminalPromptDelivery
+  try {
+    observed = await args.runtime.observeTerminalAgentPrompt(
+      args.terminalHandle,
+      args.prompt,
+      args.timeoutMs ?? AGENT_PROMPT_EFFECT_TIMEOUT_MS
+    )
+  } catch {
+    // Observation failure cannot revoke authority for input that was already accepted.
+    return { verdict: 'unobserved', prompt: args.prompt }
+  }
   if (observed.observation === 'incarnation_replaced') {
     // The PTY under this handle changed mid-observation; the accepted write is unproven.
     return { verdict: 'unobserved', prompt: observed }
@@ -71,8 +77,8 @@ export async function observeWorkerTurnStart(args: {
 export function describeUnobservedWorkerTurnStart(agent: string | null): string {
   const name = agent ?? 'the agent'
   return (
-    `Dispatch input was written and submitted, but ${name} never started a turn within ` +
-    `${Math.round(AGENT_PROMPT_EFFECT_TIMEOUT_MS / 1000)}s. This is unverifiable, not proof the ` +
+    `Dispatch input was written and submitted, but ${name}'s turn start could not be verified ` +
+    `during observation (up to ${Math.round(AGENT_PROMPT_EFFECT_TIMEOUT_MS / 1000)}s). This is unverifiable, not proof the ` +
     'worker is dead: the agent may still be starting, may be wedged (for example waiting on ' +
     'network), or may be holding the task unsent in its composer. If the worker recovers and ' +
     'reports, this Dispatch settles normally.'