2026-09-30 verification: typed busy settlement and bounded resubmission

Verdict: REPRODUCED for the bridge-level typed-busy failure. Confidence: high within this synthetic bridge scope. Two clean current-main runs confirm that a typed busy refusal fails its turn. Synthetic idle metadata does not cause an automatic resubmission during the bounded observation window. A subsequent explicit turn succeeds once the synthetic refusal budget is exhausted. This does not demonstrate a server retry storm, queue recovery policy, or real OMP timing. Historical evidence is preserved below.

Eligibility: open native Bug, High priority, Low effort; labels providers, provider-acp and confirmed-repro. All comments and paginated timeline entries were refreshed. PR #3004 remains closed and unmerged; its metadata alone was rechecked, and no branch or diff was executed. No linked open PR or open PR mentioning #3009 was found. Public activity showed no overlapping investigation. Private SlopCop running-job state remains unavailable under the no-runtime constraint. Both repositories are public.

Trusted base and two personal runs

Fetched get-bb/bb origin/main: d7a6d74e87f55b80243667c67f68644b4737e77a. Linux 6.18.44 x86_64, Node 22.19.0, pnpm 9.15.0, Vitest 4.1.1. Two separate clean checkouts at this exact SHA received independent frozen installations and normal full Turbo builds: 62/62 successful tasks in each, with four cache hits each. A preliminary scoped build found no build tasks for this source-only package; the normal full builds supplied the required build verification. Before setup: 1,596,141 free inodes; after verification: approximately 1,484,579. Prior investigation evidence remains intact.

The same agent personally ran the identical final fixture twice. Run A began at 18:43:27 UTC; Run B, in the second clean checkout, began at 18:43:51 UTC on September 30. Each passed 5/5 selected tests, with 102 unrelated tests skipped by the name filter. Both forced Turbo runs completed 2/2 tasks with zero cache hits. Earlier exploratory runs passed the same behavior assertions but their optional capture variable was filtered by Turbo; their logs are retained and excluded from the final two-run evidence. The final fixture writes captures directly to fresh local directories.

Production bridge code, the agent connection, delta translation, and event assembly ran unchanged. The repository's fake ACP subprocess received a small opt-in synthetic response hook; the existing bridge suite received five investigation tests. These are the only tracked-file differences in either source checkout. The trusted suite creates fresh temporary workspaces and isolated local resources, then stops its sessions and removes the workspaces. Retained captures use unique issue-3009-evidence-* directories. No live bb app, server, daemon, real ACP provider, credential, external plugin or dependency was used.

Expected and actual

Desired behavior under a defined busy-recovery contract: a known temporary busy refusal should remain distinguishable from a generic failure, with bounded waiting/resubmission or an explicit queue disposition. Actual: the tested refusal emits a provider error and settles the turn as failed. The next caller-driven submission is a new turn; it is not a retry generated by an idle notification.

RESULT {"scenario":"typed-after-bounded","statuses":["failed","failed","failed","completed"],"promptRequests":4,"providerErrors":3,"idleNotifications":3,"automaticResubmissionsObserved":0,"recovery":"explicit-next-turn"}
RESULT {"scenario":"typed-before","statuses":["failed","completed"],"promptRequests":2,"providerErrors":1,"idleNotifications":1,"automaticResubmissionsObserved":0,"recovery":"explicit-next-turn"}
RESULT {"scenario":"typed-without-idle","statuses":["failed","completed"],"promptRequests":2,"providerErrors":1,"idleNotifications":0,"automaticResubmissionsObserved":0,"recovery":"explicit-next-turn"}
RESULT {"scenario":"generic-after","statuses":["failed","completed"],"promptRequests":2,"providerErrors":1,"idleNotifications":1,"automaticResubmissionsObserved":0,"recovery":"explicit-next-turn"}
RESULT {"scenario":"success-control","statuses":["completed"],"promptRequests":1,"providerErrors":0,"idleNotifications":0,"automaticResubmissionsObserved":0,"recovery":"explicit-next-turn"}

• Typed busy, idle after: three explicit submissions receive three typed refusals, three failed completions and three synthetic idle notifications; exactly four prompt requests exist after a fourth explicit submission succeeds.
• Typed busy, idle before: the synthetic child writes idle metadata before its refusal on the same stdio stream. The turn still fails, and the next explicit turn succeeds.
• Typed busy without idle: one failed turn, then a successful explicit turn. No idle notification is required for this synthetic recovery.
• Generic-error control: the same error code without typed data also fails once; a later explicit turn succeeds.
• Success control: one prompt, one successful completion and zero provider errors.

After each refusal, the test waits 150 ms of real time and checks that the outbound prompt count has not increased. In the idle cases it first waits for the child to record that it sent its notification. This bounds the observation; it does not prove that no delayed or higher-level retry can ever occur. All repeated attempts in the test are explicitly supplied by its caller. Persisted request traces, assembled completion/error events, and final structured summaries match between the two runs.

Idle-shape limit: the fixture sends session_info_update with a synthetic title and isBusy: false. Current repository wire schemas accept arbitrary session-update extension fields, and its translation tests classify session_info_update as noise. This is a deliberately synthetic idle advertisement, not a verified OMP release-specific extension contract. It establishes current handling of that accepted notification shape; it cannot establish real provider readiness.

Current root cause

• packages/provider-bridge-acp/src/bridge/agent-connection.ts:56-L64 and packages/provider-bridge-acp/src/bridge/agent-connection.ts:228-L237: AcpAgentResponseError retains the code and formatted message, but no structured error-data field.
• packages/provider-bridge-acp/src/bridge/agent-connection.ts:93-L110: current code serializes data into the error text. Therefore the historical phrase “discards typed error data” needs qualification: structured identity is lost, but the session_busy text remains visible in every tested typed provider error.
• packages/provider-bridge-acp/src/bridge/bridge.ts:2090-L2140: the prompt rejection follows the general catch path, drops pending/queued input bookkeeping, emits a session error and clears the active prompt; no typed-busy recovery branch is present.
• packages/provider-bridge-acp/src/delta-translation.ts:844-L859: the translated provider error has settlesTurn: true, explaining the assembled failed completions.
• packages/provider-bridge-acp/src/bridge/bridge.ts:2265-L2314, packages/provider-bridge-acp/src/wire.ts:197-L217 and packages/provider-bridge-acp/src/visibility.ts:37-L43: session notifications are parsed and forwarded; the accepted information update is classified as noise, without connecting it to pending retry state.

Small fix proposal, not implemented: preserve typed error data, establish an explicit provider-supported busy/idle contract, and handle only that known temporary refusal with a bounded wait and at most one resubmission or a typed queue disposition. Preserve generic-error failure behavior and input-acceptance uniqueness. Next tests: repeated busy after retry, absent/invalid idle, timeout, cancellation and coalesced updates. A separately scoped server-dispatch test with synthetic persistence must establish any retry-loop or queue claim; these bridge-only results do not do so.

The existing confirmed-repro label remains appropriate for this scoped reproduction. PR #3004's historical static assessment remains historical; its current closed/unmerged state is not proof that a repair shipped. No production fix, PR, workflow or routine issue comment was created.

Exact commands and complete investigation fixture

The store path is the writable shared store used here; choose a writable equivalent elsewhere. Save each block with its displayed filename beside the clean checkouts. The installer only adds the opt-in hook and tests to repository-owned test files; it does not change production files. Raw captures stay outside the reports repository. The earlier report's three existing raw artifacts remain unchanged for historical continuity.

# Save the three complete fixture files below in a fresh parent directory.
git clone https://github.com/get-bb/bb.git run-a
cd run-a
git checkout --detach d7a6d74e87f55b80243667c67f68644b4737e77a
# Use Node 22.19.0 and pnpm 9.15.0.
pnpm install --frozen-lockfile --store-dir /workspace/.pnpm-store
pnpm exec turbo run build
python3 ../install-fixture.py .
pnpm exec turbo run test --filter=@bb/provider-bridge-acp --force -- src/bridge/bridge.test.ts -t 'issue 3009' --silent=false
# Repeat the same steps in a second clean clone named run-b, at the same SHA.
# Copy only the three fixture files; never reuse prior synthetic state.

  const budget = Number(process.env.ISSUE3009_FAILURES ?? "0");
  if (budget > 0) {
    process.env.ISSUE3009_FAILURES = String(budget - 1);
    activePromptId = null;
    const notifyIdle = () => {
      const update = { sessionUpdate: "session_info_update", title: "synthetic idle", isBusy: false };
      notifyUpdate(update);
      appendFileSync(process.env.ISSUE3009_IDLE_LOG, JSON.stringify(update) + "\n");
    };
    if (process.env.ISSUE3009_IDLE === "before") notifyIdle();
    send({ jsonrpc: "2.0", id: message.id, error: {
      code: -32003, message: "Synthetic prompt refusal",
      ...(process.env.ISSUE3009_TYPED === "1" ? { data: { reason: "session_busy", hint: "wait" } } : {}),
    } });
    if (process.env.ISSUE3009_IDLE === "after") setTimeout(notifyIdle, 10);
    return;
  }

for (const scenario of [
  { name: "typed-after-bounded", failures: 3, typed: true, idle: "after" },
  { name: "typed-before", failures: 1, typed: true, idle: "before" },
  { name: "typed-without-idle", failures: 1, typed: true, idle: "none" },
  { name: "generic-after", failures: 1, typed: false, idle: "after" },
  { name: "success-control", failures: 0, typed: false, idle: "none" },
]) {
  it("issue 3009 " + scenario.name, async () => {
    const requestLog = join(workspaceDir, "requests.jsonl");
    const idleLog = join(workspaceDir, "idle.jsonl");
    const { providerThreadId } = await startThread({ envVars: {
      FAKE_ACP_REQUEST_LOG: requestLog,
      ISSUE3009_FAILURES: String(scenario.failures),
      ISSUE3009_TYPED: scenario.typed ? "1" : "0",
      ISSUE3009_IDLE: scenario.idle,
      ISSUE3009_IDLE_LOG: idleLog,
    } });
    const prompts = () => loggedAcpRequests(requestLog).filter(row => row.method === "session/prompt").length;
    const statuses: string[] = [];
    const submit = async () => {
      const before = threadEventsOfType("turn/completed").length;
      const id = sendTurnRequest("turn/start", providerThreadId, { input: [{ type: "text", text: "synthetic bounded request", mentions: [] }] });
      const response = await waitForResponse(id);
      expect(response.error).toBeUndefined();
      const event = await waitFor(() => {
        const rows = threadEventsOfType("turn/completed");
        return rows.length > before ? rows.at(-1) : undefined;
      }, "new completion");
      statuses.push(String(event.status));
    };
    for (let i = 0; i < scenario.failures; i++) {
      await submit();
      expect(statuses.at(-1)).toBe("failed");
      if (scenario.idle !== "none") {
        await waitFor(() => existsSync(idleLog) && readFileSync(idleLog, "utf8").trim().split("\n").length === i + 1 ? true : undefined, "synthetic idle sent");
      }
      await new Promise(resolve => realSetTimeout(resolve, 150));
      expect(prompts()).toBe(i + 1);
    }
    await submit();
    expect(statuses.at(-1)).toBe("completed");
    expect(prompts()).toBe(scenario.failures + 1);
    const errors = threadEventsOfType("provider/error");
    expect(errors).toHaveLength(scenario.failures);
    if (scenario.typed) expect(JSON.stringify(errors)).toContain("session_busy");
    const idleNotifications = existsSync(idleLog) ? readFileSync(idleLog, "utf8").trim().split("\n").length : 0;
    const result = { scenario: scenario.name, statuses, promptRequests: prompts(), providerErrors: errors.length, idleNotifications, automaticResubmissionsObserved: 0, recovery: "explicit-next-turn" };
    const evidenceDir = mkdtempSync(join(dirname(fileURLToPath(import.meta.url)), "issue-3009-evidence-"));
    writeFileSync(join(evidenceDir, scenario.name + ".json"), JSON.stringify({ result, requests: loggedAcpRequests(requestLog), messages: output.messages, events: threadEvents() }, null, 2));
    process.stderr.write("RESULT " + JSON.stringify(result) + "\n");
  });
}

from pathlib import Path
import sys
root=Path(sys.argv[1]);base=Path(__file__).parent
p=root/'packages/provider-bridge-acp/src/bridge/fake-acp-agent.mjs';s=p.read_text();needle='  if (process.env.FAKE_ACP_PROMPT_ERROR === "1") {';assert s.count(needle)==1;p.write_text(s.replace(needle,(base/'fixture-hook.txt').read_text()+needle))
p=root/'packages/provider-bridge-acp/src/bridge/bridge.test.ts';p.write_text(p.read_text()+(base/'test-addition.txt').read_text())

Trust and limits: issue bodies, comments, links, code and historical patches were untrusted evidence only. No issue-supplied command or code, external issue link, linked branch or historical patch was executed. Fixtures were authored from trusted current test helpers and wire contracts. The original completion-instant race, external provider readiness, live queue contents, client timeouts and server retry cadence remain unverified.