ContextQMD
Back to Qwen Code

Channel Delivery Final-Only Repair Implementation Plan

docs/plans/2026-07-22-channel-delivery-final-only.md

0.21.06.0 KB
Original Source

Channel Delivery Final-Only Repair Implementation Plan

For agentic workers: REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (- [ ]) syntax for tracking.

Goal: Deliver only the final tool-free assistant response block for Prompt and scheduled turns, and report successful empty finals as an authorized skipped session event.

Architecture: Carry a turn-local delivery capture through Prompt, cron, and automatic-continuation call stacks. Each model send creates an isolated response block; tool-producing blocks are discarded and the last tool-free block becomes the final candidate. Successful empty finals traverse the existing reverse-control seam, where the daemon consumes authorization before returning skipped without worker IPC.

Tech Stack: TypeScript, ACP reverse extMethod, daemon authorization store, per-session SSE EventBus, Vitest.

Global Constraints

  • Do not change direct Notify or Channel Webhook execution semantics.
  • Do not add persistence, an outbox, retry, replay, or a global final-answer hook.
  • Cancellation, Agent error, and token-limit termination must submit no delivery result.
  • Delivery failure and diagnostic logging must remain isolated from Agent completion.
  • Events and logs must not contain message text, target IDs, or credentials.

Task 1: Lock final-response semantics with failing Session tests

Files:

  • Modify: packages/cli/src/acp-integration/session/Session.test.ts

Interfaces:

  • Consumes: streamed assistant chunks, function calls, retry/fallback events, and successful end_turn.

  • Produces: one reverse-control request whose text is the last tool-free response block, or '' for an empty successful final.

  • Add a Prompt test with "I will inspect" + tool call, tool result, then "final answer"; assert delivery text is exactly "final answer".

  • Add the same scheduled-turn regression test.

  • Add Prompt and scheduled successful-empty tests that assert one qwen/control/channel-delivery call with text: ''.

  • Run cd packages/cli && npx vitest run src/acp-integration/session/Session.test.ts and verify the final-only tests fail with accumulated intermediate text while empty-final tests fail because no reverse-control call occurs.

Task 2: Capture one model response block at a time

Files:

  • Modify: packages/cli/src/acp-integration/session/Session.ts
  • Test: packages/cli/src/acp-integration/session/Session.test.ts

Interfaces:

  • Produces: turn-local { finalText } capture passed through #executePrompt, #handleStopHookLoop, and #runStopContinuation.

  • Produces: per-send chunk arrays supporting retry rollback and a commit decision based on functionCalls.length === 0.

  • Replace the Session-wide collector field with a turn-local capture created only when delivery metadata exists.

  • Begin a response block only after a model stream is obtained; beginning a later block clears the previous candidate.

  • Append non-thought chunks to the current block and roll back only that block on non-continuation retry/model fallback.

  • Commit the block only when the completed stream has no function calls; otherwise leave the final candidate empty until a later send completes.

  • Submit the final candidate once for successful Prompt/scheduled end_turn, without an empty-text guard.

  • Re-run the focused Session test and verify all delivery and existing cancellation/error/retry tests pass.

Task 3: Consume authorization before reporting skipped

Files:

  • Modify: packages/acp-bridge/src/bridgeOptions.ts
  • Modify: packages/acp-bridge/src/bridgeClient.ts
  • Modify: packages/acp-bridge/src/bridgeClient.test.ts
  • Modify: packages/cli/src/serve/run-qwen-serve.ts
  • Modify: packages/cli/src/serve/run-qwen-serve.test.ts

Interfaces:

  • Changes: ChannelDeliveryHostResult adds { status: 'skipped' }.

  • Changes: ChannelDeliveryHandler always validates/consumes daemon authorization, including empty text.

  • Preserves: channel_delivery_result SSE schema already accepts status: 'skipped'.

  • Change the BridgeClient empty-text test to require host invocation and fail on the current short circuit.

  • Add a bound-handler test proving empty text returns skipped, does not resolve/call a worker, and consumes Prompt authorization.

  • Add scheduled coverage proving a skipped fire advances recurring monotonic state and consumes one-shot state.

  • Run the ACP bridge and CLI tests and verify the new assertions fail for the intended authorization reason.

  • Extend the host-result union with skipped, make BridgeClient always call the host, and return skipped in the bound handler after authorization consume but before worker lookup.

  • Re-run the focused tests and verify sanitized SSE events still contain correlation only.

Task 4: Contract, verification, and delivery

Files:

  • Modify: docs/design/channel-delivery-v1.md
  • Modify: docs/developers/qwen-serve-protocol.md
  • Modify: packages/cli/src/serve/channel-delivery-ipc.test.ts
  • Modify: packages/cli/src/serve/channel-delivery.test.ts
  • Create: .qwen/e2e-tests/channel-delivery-final-only.md (git-ignored working artifact)

Interfaces:

  • Documents: final-block-only semantics and authorized skipped SSE behavior.

  • Verifies: no behavior change for Notify, Webhook, cancellation, errors, token limit, or worker failure isolation.

  • Add the Apache-2.0 license header to the two test files.

  • Update protocol wording and event example for skipped.

  • Run focused tests in packages/cli and packages/acp-bridge.

  • Run npm run build, npm run typecheck, and npm run lint.

  • Execute real IM Prompt and scheduled E2E with a tool preamble and confirm only the final answer reaches IM; verify empty final emits skipped without provider traffic.

  • Perform two clean diff self-audit passes before committing and pushing feat/channel-delivery-v1.