[ranxianglei]

Problem

Strict OpenAI-compatible backends (SGLang, Qwen3-family "system-first" chat template) require exactly one system message at index 0 and reject a mid-conversation or second system message with 400 {"message":"System message must be at the beginning"} (#377).

acp-kernel renders each compressed block's summary as role:"system" anchored mid-conversation. In proxy mode (plain client) the compress tool call is ephemeral (server-side loop, never in the client's re-sent history) and preflight blocks have no tool call at all — so the acp_summary system message is the only carrier, and ≥2 active blocks produce ≥2 system messages → 400.

Fix

Replace hoistMidSystemMessages (which hoisted mid-stream system messages to the head as separate system messages) with systemToUser — it converts the mid-stream acp_summary system/developer messages to role:"user", leaving them at their anchor position.

This satisfies both constraints at once:

• SGLang compatibility — only one system message remains (the head, injected by injectOpenaiSystem); summaries ride on user messages, which are allowed anywhere.
• Prefix-cache stability — the head system message (the cache anchor) stays byte-stable across compress turns, so a new block no longer invalidates the whole-conversation prefix. (Hoisting/merging to the head would have broken this.)

In plugin/launcher mode the summary carrier is the compress tool call (in the agent's own re-sent history), so the kernel's acp_summary is stripped by stripKernelSummaries and systemToUser is a no-op there.

Why user, not a forged tool call: a forged compress tool call would be the "pure" carrier, but in proxy mode it requires fabricating an assistant tool_calls + user tool_result pair by id, declaring the tool in the request, and handling preflight blocks that have no authentic call — far more invasive than re-voicing a standalone note. A summary is a stand-in for the folded history; re-voicing it as a user turn (clearly marked [Compressed conversation section]) is the accepted trade-off.

Changes

• src/util.ts: hoistMidSystemMessages → systemToUser
• src/server.ts, src/loop/adapter-openai.ts: call sites + two-mode comments
• README.md: new "Two compression modes" section (carriers, why they differ, why user not system/tool-call, coexistence determination)
• AGENTS.md: Key Design Decisions item 6 (two modes, different carriers) + Before Making Changes item 4 (every wire-touching change must be reasoned in BOTH modes)
• tests: new system-to-user.test.ts (7 tests), updated e2e-mid-system.test.ts (summaries must ride on user messages, not leak into system), deleted hoist-mid-system.test.ts

Verification

• system-to-user.test.ts: 7/7 pass
• e2e-mid-system.test.ts: pass (real startServer vs mock upstream; asserts zero system/developer after the leading prefix + summaries on user messages)
• npm run build: success
• npm test: 891/892 (1 pre-existing failure compress-reject-reasons.test.ts [bug] 0.1.63 起 Windows persist EPERM 风暴：会话持久化静默丢数据（4.9k 次/117 会话）+ compress no-valid-ranges 拒绝 #362 — installed acp-kernel lacks invalidReasons; fails on origin/master too, unrelated)
• npm run typecheck: 6 pre-existing errors in src/compress-tool.ts / src/stream.ts (invalidReasons), none in changed files; identical on origin/master

Supersedes the abandoned 2026-09-01_merge-system-messages-377 branch (merge-to-system approach, rejected because it breaks prefix-cache stability).

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-01_user-carrier-summary-377 (499f314)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-428

Each push to this PR publishes a new version under the pr-428 npm tag.

Option B — Download artifact

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

tar xzf billion-context-pr428.tgz
npm install -g package

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 ework agent · qwen3.8-27b

Reviewed PR #428 (branch 2026-09-01_user-carrier-summary-377, commit 4f7e603). The fix is correct, and the verification is actually in better shape than the issue body claims.

Verification (clean npm ci from the committed lockfile):

• tests/system-to-user.test.ts — 7/7 pass
• tests/e2e-mid-system.test.ts — pass (real startServer vs mock upstream; asserts the summaries ride on user messages and leak zero into system/developer)
• npm run build — success
• npm test — 892/892 pass
• npm run typecheck — clean, exit 0

The two "pre-existing" problems the issue body calls out no longer reproduce: the lockfile pins acp-kernel@0.0.47 (integrity-locked), and that artifact now ships invalidReasons, so a fresh npm ci gives a fully green tree — the compress-reject-reasons #362 test failure and the 6 src/compress-tool.ts/src/stream.ts type errors are both gone. CI (npm ci) will see the same.

Code review — logic is sound:

• Both call sites run systemToUser before injectOpenaiSystem (src/server.ts:1536→1547, src/loop/adapter-openai.ts:175→176), so the head system is re-injected after the re-voicing → exactly one system at index 0, byte-stable for the prefix cache.
• In proxy mode openaiToCore hoists the client's leading system out first (src/server.ts:1511), so by the time systemToUser runs it only ever sees the mid-conversation acp_summary messages and converts them to user in place — never touching the head.
• Plugin mode: stripKernelSummaries removes the acp_summary (the tool call is the carrier), so systemToUser is a no-op — matches the documented two-mode model.
• No leftover hoistMidSystemMessages references; tests/hoist-mid-system.test.ts deleted; README "Two compression modes" + AGENTS.md items 6 & 4 document the carrier distinction and the "reason in BOTH modes" rule.

One follow-up (human decision): PR #379 (2026-08-31_merge-system-messages, "merge all system/developer into one leading message") is the superseded merge-to-head approach this PR replaces. It should be closed in favor of #428 — I've left it open since closing/merging is human-only.

LGTM — ready to merge (merge is human-only per AGENTS.md).

[ranxianglei]

确认是否有回归问题

[ranxianglei]

🤖 ework agent · qwen3.8-27b

已确认，无回归问题。我做了 master 基线对比 + 全量测试（clean npm ci，acp-kernel 0.0.47）：

全量测试

• master 基线：893/893 通过，0 失败
• 本 PR 分支：892/892 通过，0 失败
• 没有任何一个在 master 上通过的测试在 PR 上失败

−1 个测试是有意为之，不是漏掉回归

• PR 删除 tests/hoist-mid-system.test.ts（8 个用例），新增 tests/system-to-user.test.ts（7 个用例）
• 7 个新用例与旧场景一一对应（no-op、单个/多个 mid-system、wire shape），断言按新的"改写为 user"行为更新
• 净减的 1 个是旧用例 feat: bili CLI + debug in config + README rewrite #7「全 system 列表保持顺序」——退化场景（真实会话不可能全是 system），其行为在新实现下根本改变（全部转成 user），已被新通用转换用例（feat: OpenAI Responses API support (/v1/responses) #4/feat: persistent + namespaced sessions, multi-tenant session isolation, stream robustness #5/feat: bili CLI + debug in config + README rewrite #7）覆盖
• 新文件还补充了旧文件没有的覆盖：developer 角色处理 + 对象身份保持（non-system 消息不复制）

本 bug 的专属回归测试通过

• tests/e2e-mid-system.test.ts 驱动真实 startServer 代理对 mock upstream，断言新 wire 形态：summary 落在 user 消息上、system/developer 零泄漏、index 0 恰好一个 system

无残留引用

• 代码/文档里剩余的 "hoist" 全部指 kernel openaiToCore 的前缀 hoist（kernel 0.0.37，仍在用），与被删除的 hoistMidSystemMessages 函数无关

一个说明（非回归，是本沙箱的覆盖限制）

• AGENTS.md 要求请求管线改动合并前跑真实客户端 E2E tests/e2e/e2e-codex.test.ts，但它需要 codex 二进制 + 本地 Responses upstream（默认 127.0.0.1:8199）。本沙箱两者都不可用（codex 未安装、upstream 不可达），且被 ACP_TEST_E2E=1 门控。合并前建议由维护者或 manual-dispatch CI 跑一遍——这是环境限制，不是回归信号。

结论：无回归。 行为变更（hoist-to-head → 原地改写为 user）是有意设计且覆盖完整；唯一在本沙箱没跑到的只有需要本地 upstream 的真实客户端 e2e。

[ranxianglei]

新方案发到原issue

[ranxianglei]

🤖 ework agent · qwen3.8-27b

[bot] 🏷 汇总本轮在 #428 上完成的工作：

1. 代码评审 — systemToUser（src/util.ts）逻辑正确：两个调用点（src/server.ts:1536→1547、src/loop/adapter-openai.ts:175→176）都在 injectOpenaiSystem 之前执行，head system 在改写后重新注入 → index 0 恰好一条 system、字节稳定利于 prefix-cache；proxy 模式下 openaiToCore 先把 client 的 leading system 提走，所以 systemToUser 只碰 mid-stream 的 acp_summary。plugin 模式 stripKernelSummaries 剥掉 acp_summary，systemToUser 是 no-op。

2. 验证（clean npm ci，acp-kernel 0.0.47）— system-to-user.test.ts 7/7、e2e-mid-system.test.ts 通过、全量 npm test 892/892、typecheck clean、build 成功。

3. 回归确认 — master 基线 893/893，PR 分支 892/892，0 失败；净 −1 是删除的旧 hoist-mid-system.test.ts 里一个退化用例（全 system 列表），已被新通用用例覆盖，非漏覆盖。唯一本沙箱跑不到的是需要本地 upstream 的真实客户端 e2e（e2e-codex：codex 未装 + 127.0.0.1:8199 不可达），建议合并前由维护者/CI 跑一遍。

4. 已把新方案同步到原 issue [Bug] ACP: ≥2 compressed blocks render as multiple system messages → SGLang (vLLM-compatible) upstreams return 400 #377，说明它取代 PR fix: merge all system/developer messages into one leading message (SGLang #377) #379（merge-to-system，会破坏 prefix-cache 稳定性），并建议关闭 PR fix: merge all system/developer messages into one leading message (SGLang #377) #379。

结论：PR #428 LGTM，无回归，可合并（合并为人工操作）。