[Bug] daemon catch-up drops tool_execution_end after large tool results

[ponythewhite]

Version

• Prime Agent: 0.8.1
• Reproduced from tag v0.8.1 (514633727bf26d74f39f3119c2b0e31a5ceb2a9d)
• The same filtering path is present on main at 5b6c0e94e11a97fcfdd7a9fc9dc4f7acbda9c853
• Node.js: 22.22.1

Problem

After a large tool result, an attached daemon/RPC client can receive tool_execution_start and the complete result update but never receive the matching tool_execution_end. The successful result remains durable in the session JSONL, so the tool itself finished; only the terminal stream boundary is missing.

Consumers that wait for tool_execution_end cannot safely advance and eventually hit their deadline. The failures that led to this report used result frames around 219 KiB and had no provider, network, Docker, authentication, or verifier error.

Root cause

There are two independent filtering points:

1. AgentDaemon.broadcastToSession()
2. DaemonSupervisor.handleWorkerFrame()

A large socket.write() can return false after Node has accepted and queued the frame. Prime Agent marks that client as backpressured. A subsequent terminal event is then skipped in favor of snapshot catch-up. The same skip occurs while the active session has a snapshot in progress.

Snapshot catch-up restores session state, but it does not replay the omitted tool_execution_end event boundary.

Deterministic reproduction

The proposed branch adds focused cases to the existing daemon tests. Each case:

1. makes a large write return false;
2. marks the same session snapshot-active;
3. sends tool_execution_end;
4. sends an ordinary event afterward.

Without the source changes, the terminal frame is replaced by catch-up and the cases fail. With the change, the terminal event is delivered while ordinary traffic still uses catch-up.

git clone https://github.com/ponythewhite/prime-agent.git
cd prime-agent
git switch fix/lossless-daemon-terminal-events
npm ci
cd packages/coding-agent
npx tsx ../../node_modules/vitest/dist/cli.js --run \
  test/daemon-mode.test.ts \
  test/daemon-supervisor-monitor.test.ts \
  -t 'preserves (recovery checkpoints|terminal tool events)'

Proposed fix

• Let sequenced recovery checkpoints pass the worker-daemon snapshot/backpressure filters.
• Let tool_execution_end pass the supervisor snapshot/backpressure filters.
• Keep ordinary high-volume traffic on the existing snapshot catch-up path.
• Do not change the daemon schema, capabilities, or protocol version.

The focused tests pass, both affected test files pass (268/268), and npm run check passes.

Code and prior PR

• Fork branch: https://github.com/ponythewhite/prime-agent/tree/fix/lossless-daemon-terminal-events
• Closed PR: fix(coding-agent): preserve terminal events under daemon backpressure #1901

PR #1901 was closed automatically by the contribution gate because I am not yet a vouched contributor. I am opening this Discussion first as required by CONTRIBUTING.md. I would appreciate maintainer feedback on the diagnosis and patch. If the change fits the project, please invite the implementation or advise how you would prefer it to be submitted.