fix: acp-loop auto-retry on no-content upstream truncation (#221)

[ranxianglei]

Context

Follow-up to #221 (agent stops at the acp tool). PR #239 made the upstream-truncation failure visible ([acp-proxy: upstream stream truncated ...]), and the user's logs confirm that is the failure mode:

[error] [acp-loop] upstream stream truncated (no completion event; round 1, 0 text chars + 0 reasoning chars received)

But the agent still stopped, because the proxy surfaced the error and terminated the turn.

Root cause

The relay cuts a long-running request (MiniMax-M3 thinking server-side for minutes) before any token is emitted. The upstream SSE ends with no completion event and zero content — nothing reached the client.

Fix

A truncation that produced no text/reasoning is a clean re-send: the loop now re-requests the same body and continues instead of erroring out. The retry is treated as a later round, so message_start is not re-emitted and no prefix is duplicated.

• Bounded by BILI_TRUNCATION_RETRY_MAX (default 2 retries; 0 restores the fail-fast error).
• A partial-content truncation is deliberately NOT retried (the model would regenerate from scratch and duplicate the already-streamed prefix) and still surfaces the visible error.

Tests

tests/upstream-truncation.test.ts — 3 new cases:

• no-content truncation auto-retries and delivers the full response (single message_start, no error marker)
• no-content truncation that keeps truncating surfaces the error after retries are exhausted
• partial-content truncation is NOT auto-retried

Full suite: 644 pass. typecheck + build clean.

[ranxianglei]

[bot] Reviewed the PR, verified the core mechanism, and found (and fixed) two edge-case holes in the retry guard. Pushed a follow-up commit to the branch: 03b7f8f.

Verified as designed

• No-content truncation → clean re-send of the same body: coreMessages is untouched when no proxy tool ran, so the re-request is byte-identical; the retry flows through the existing re-request machinery, is treated as a later round, and message_start (a firstRoundOnly meta event) is not re-emitted. ✓
• Partial-content truncation still surfaces the visible error. ✓
• BILI_TRUNCATION_RETRY_MAX default 2 / 0 = fail-fast. ✓

Bug 1 — retry after a flushed real tool call (the important one). A truncation can carry a completed tool call: the Anthropic adapter flushes tool_use at content_block_stop and the Responses adapter at output_item.done — both before the completion event. (OpenAI chat is safe: it settles tool calls only at finish_reason/[DONE].) With the original guard (!sawDone && zeroContent && budget), a truncated stream that had zero text/reasoning but one closed real tool call would (1) emit that tool call to the client, then (2) auto-retry — streaming the regenerated response behind the tool call in the same client response. The client would see one assistant message containing the tool call plus a duplicate regeneration, and its tool-result follow-up would be out of sync with the stream. Fix: canRetryTruncation now also requires realCalls === 0; that case surfaces the visible error as before. (Proxy-tool calls are unaffected — they take the normal reRequest path, and the retry counter/log no longer misattribute that re-request.)

Bug 2 — final-round truncation swallowed the error. On round MAX_LOOP_ROUNDS (10), a zero-content truncation passed the error check (retry allowed) and fell into the loop-limit branch, ending the turn with a degenerate empty stop_reason: "length" completion instead of the visible error. Fix: canRetryTruncation now also requires round < MAX_LOOP_ROUNDS.

Tests — added 2 regression cases to tests/upstream-truncation.test.ts (both confirmed to fail on the pre-fix code):

• truncation with a fully closed real tool_use block → error surfaces, no retry content behind the tool call
• 9 proxy-tool rounds + zero-content truncation on round 10 → error surfaces, no length completion

Full suite: 646 pass (was 644). typecheck + build clean.

PR is ready for merge whenever you are — I can't merge it myself (AGENTS.md: merges are human-only).

_{🤖 ework agent · vllm-qwen/qwen3.8-27b}

[github-actions]

📦 Built Package Artifact

Branch: 2026-08-26_acp-truncation-autoretry (4d87ab2)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-261

Each push to this PR publishes a new version under the pr-261 npm tag.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf billion-context-pr261.tgz
npm install -g package

---

This comment is automatically updated on each push.

[ranxianglei]

📦 Built Package Artifact

Branch: 2026-08-26_acp-truncation-autoretry (4d87ab2)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-261

Each push to this PR publishes a new version under the pr-261 npm tag.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf billion-context-pr261.tgz
npm install -g package

---

This comment is automatically updated on each push.