Retry session_busy prompt rejections once the agent goes idle

[bradhallett]

Human comments

What was wrong

When omp is streaming without an ACP-owned turn in flight — the completion-instant race (a steer lands in a turn's final second, bb closes its turn bookkeeping while omp is still finishing) or deferred agent-initiated continuations (#2122, #2461) — omp's ACP layer lets session/prompt fall through to a bare AgentSession.prompt(), which throws AgentBusyError; since omp#9825 that rejection is typed on the wire as -32003 with data.reason: "session_busy". The bridge had no path for that typed rejection, so it fell into the generic error path: the turn failed and dispatch re-sent into the still-busy session, producing rejection storms (evidence: +3 ms rejection after turn/input/accepted, retry turn 23 ms later, worst storm 8 rejections over 105 s until the ~8-minute client timeout; one thread terminally stuck in error). Full diagnosis and event-log evidence in #3009.

What changed

• packages/provider-bridge-acp/src/bridge/agent-connection.ts — AcpAgentResponseError now carries the JSON-RPC error data (optional constructor arg), so the typed reason is inspectable. No wire changes.
• packages/provider-bridge-acp/src/bridge/bridge.ts —
  • Branch only on code -32003 and data.reason === "session_busy"; any other shape falls through to the existing error path unchanged.
  • On the first such rejection for a prompt: wait for the agent to report idle, then retry the prompt exactly once. Idle signal = end-of-turn session_info_update notification (ACP v1 has no idle status variant; omp emits this last at end of turn). session/update notifications are counted unconditionally in handleAgentNotification, so the coalesced-chunk case — busy error and idle update arriving in one stdout chunk, before the wait is armed — is detected via a pre-request snapshot and retries immediately.
  • A second busy rejection for the same prompt, or no idle signal within the timeout (default 10 min, overridable via BB_ACP_SESSION_BUSY_RETRY_TIMEOUT_MS), surfaces the original busy error through the existing failure path.
  • The retry skips acceptTurnInput, so no duplicate input.accepted delta is emitted for the resent prompt; each new prompt gets a fresh retry budget; outstanding waits are settled on finishTurn/stopSession/removeSession so nothing dangles.
• packages/provider-bridge-acp/src/bridge/fake-acp-agent.mjs — FAKE_ACP_SESSION_BUSY_{PROMPTS,SKIP,REASON,SAME_CHUNK,IDLE_DELAY_MS} knobs reproducing omp's exact wire shape (code, message, data { reason, hint }, end-of-turn usage_update → session_info_update ordering, single-chunk coalescing).
• packages/plugin-sdk + packages/domain/src/plugin-sdk-version.ts — version bumped 0.4.40 → 0.4.41 (the bridge output is part of the published SDK artifact; the repo's version-bump check requires it).

How you verified

• Six new bridge tests (318 total in the package), each mutation-verified by the author: busy prompt retried after the agent reports idle; second busy rejection fails the turn with the original error; wrong or missing reason fails immediately (no retry consumed); no idle before the retry timeout fails with the original error (BB_ACP_SESSION_BUSY_RETRY_TIMEOUT_MS injectable); busy error + idle update coalesced into one chunk retries immediately; each new prompt gets a fresh retry budget.
• Independently re-verified after syncing with main at debb58d10 (merge commit 05b300885): pnpm exec turbo run test --filter=@bb/provider-bridge-acp — 318/318 passed; pnpm exec turbo run typecheck lint for @bb/provider-bridge-acp, @bb/domain, @get-bb/plugin-sdk — 6/6 green; pnpm exec oxfmt --check clean on all five touched files. Mutation re-check: inverting the data.reason === "session_busy" gate fails 4 of the 5 busy-matched tests; restoring it returns them to green.

Fixes #3009

AGENT GENERATED

[bradhallett]

Upstream report for the omp-side half (typed session_busy rejection, no queueing on the ACP prompt path): can1357/oh-my-pi#10707. This PR remains the complete host-side fix regardless of how the upstream discussion lands — it only consumes the typed error omp ships today.