[ranxianglei]

Summary

Fixes #377 — with 2+ active ACP compressed blocks, bili rendered each block's summary as its own role:"system" message, so the OpenAI wire carried 2+ system messages. SGLang (and vLLM with the Qwen3-family "system-first" chat template) reject that with a hard 400:

{"object":"error","message":"System message must be at the beginning.","type":"BadRequest","code":400}

The earlier #355 fix (hoistMidSystemMessages) hoisted mid-conversation systems to a contiguous leading prefix but kept them separate — which SGLang still rejects, because its Qwen3 template guard is if message.role == 'system' and not loop.first: raise, i.e. it requires exactly one system message at index 0, not merely a leading prefix.

Change

Replace hoistMidSystemMessages with mergeSystemMessages (src/util.ts): it collapses all system/developer messages into a single leading system message, preserving the relative order of every other message. sysParts (the client's original system + the static compress prompt) is placed first so the stable prefix-cache anchor stays at the head across compress turns; the folded block summaries follow in their original anchor order.

• src/server.ts (prepareOpenai) and src/loop/adapter-openai.ts (buildRequest) now call coreToOpenai(...) then mergeSystemMessages(msgs, sysParts); the now-unused injectOpenaiSystem import is dropped from both.
• The Responses API path is unaffected (it already injects a single joined developer message).

Tests

• tests/hoist-mid-system.test.ts → tests/merge-system.test.ts: unit tests rewritten for the merge (no-op, single/mid/multiple systems, sysParts-first ordering, the [Bug] ACP: ≥2 compressed blocks render as multiple system messages → SGLang (vLLM-compatible) upstreams return 400 #377 wire shape collapsing to exactly one system, non-string content JSON-stringified).
• tests/e2e-mid-system.test.ts: the real-proxy e2e now asserts exactly one system/developer message at index 0 in every upstream request (the [Bug] ACP: ≥2 compressed blocks render as multiple system messages → SGLang (vLLM-compatible) upstreams return 400 #377 requirement), in addition to the existing summary-marker checks.

Pre-flight

• npm run typecheck — clean
• npm test — 863 pass, 0 fail (incl. the e2e 多段 compress 夹 user 消息时 T1 摘要 (role:system) 落在对话中部, sglang 报 400 "System message must be at the beginning" #355 real-proxy run)
• npm run build — success

[github-actions]

📦 Built Package Artifact

Branch: 2026-08-31_merge-system-messages (2d44dd4)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-379

Each push to this PR publishes a new version under the pr-379 npm tag.

Option B — Download artifact

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

tar xzf billion-context-pr379.tgz
npm install -g package

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 ework agent · vllm-qwen/qwen3.8-27b

[bot] Reviewing this PR now. Diff reviewed: mergeSystemMessages in src/util.ts replaces hoistMidSystemMessages, both call sites (prepareOpenai in src/server.ts, buildRequest in src/loop/adapter-openai.ts) updated, unused injectOpenaiSystem imports dropped, tests rewritten + e2e assertion tightened to "exactly one system at index 0". Running the full pre-flight (typecheck / test / build) on the branch now — will report results here.

[ranxianglei]

🤖 ework agent · vllm-qwen/qwen3.8-27b

[bot] Review complete — verified independently on branch 2026-08-31_merge-system-messages (57cd966). All pre-flight claims check out:

Pre-flight (re-run locally)

• npm run typecheck — clean
• npm test — 863 pass, 0 fail (matches the PR claim)
• npm run build — success
• tests/e2e-mid-system.test.ts (real-proxy e2e) — passes standalone with the new "exactly one system/developer at index 0" assertion on every upstream request

Code review

• mergeSystemMessages (src/util.ts:194): correct — collects all system/developer messages, sysParts first (stable prefix-cache anchor), block summaries in original anchor order, non-system relative order preserved, no-op (same array ref) when there's nothing to merge, non-string content JSON-stringified. The "\n\n---\n\n" separator and the sysParts-first ordering match the stated cache-stability rationale.
• Both call sites updated (prepareOpenai in src/server.ts:1442, buildRequest in src/loop/adapter-openai.ts:176); unused injectOpenaiSystem imports dropped from both. No lingering references to hoistMidSystemMessages anywhere. Remaining injectOpenaiSystem hits are only kernel-API unit tests in tests/basic.test.ts / tests/proxy-basic.test.ts, which are legitimate.
• Tests: tests/merge-system.test.ts covers no-op, single/mid/multiple systems, sysParts-first, the [Bug] ACP: ≥2 compressed blocks render as multiple system messages → SGLang (vLLM-compatible) upstreams return 400 #377 wire shape collapsing to exactly one system, and non-string content. E2e assertion correctly tightened from "contiguous leading prefix" to "exactly one at index 0" — that is sglang's actual guard.
• Responses API path untouched, as claimed.

One caveat

• The full codex e2e regression (tests/e2e/e2e-codex.test.ts) could not be run in this environment: no codex binary and no reachable upstream (127.0.0.1:8199 refused, E2E_UPSTREAM_URL unset). Per AGENTS.md it should be exercised before merging request-pipeline changes — the PR's own real-proxy e2e covers the OpenAI wire path through the actual server code, but if you have a codex-capable box, a run of the 4-phase core there would be the last mile.

LGTM from my side — merge is yours to make (I can't merge PRs).