[ranxianglei]

What changed

withConversationIdNote() (src/compress-tool.ts) embedded the live per-conversation id in the static system prompt:

[Your bili conversation id: pfa-…. When calling the bili compression tools, pass this value as the conversation_id argument so a shared MCP process can route the call to THIS session.]

A fork / resume / session-switch continues under a new canonical id while the message history stays byte-identical (#1486). That one system line was then the only cross-session divergence, so the upstream prompt cache missed entirely (cached_tokens=0) even though the whole history was identical — nullifying #1486's byte-exact resume inheritance.

Fix

• src/compress-tool.ts — withConversationIdNote(text) now emits a constant, byte-stable pointer (CONVERSATION_ID_POINTER_NOTE, no id value); new conversationIdTrailingNote(id) returns [Your bili conversation id: <id>].
• src/server.ts — each of the four client lanes appends that trailing note as an ephemeral user message, under the same gate the old id-in-system used: Anthropic opts.compress.injectTool, OpenAI/Google shouldInject, Responses shouldInject && !isCompactionTrigger && !ACP_NO_COMPRESS_PROMPT. The compress-loop re-request keeps the constant system pointer (no trailing note — it's proxy-driven; tools execute server-side).

This mirrors the existing nudge / imgNote / retrieveNote pattern ("not persisted, never enters the agent's re-sent history, safe for the prefix-cache anchor"). MCP routing is unchanged: the model still reads the value from context and echoes it as the conversation_id argument (src/mcp.ts routes on the argument).

Behavior change (flagged per AGENTS.md §4)

• System prompt text changed. Old: the full [Your bili conversation id: <id>. When calling…THIS session.] line sat in the static system. New: the static system carries only the constant pointer ("Your bili conversation id is provided in a short ephemeral message near the end of this conversation…") and the live value appears as a trailing user message [Your bili conversation id: <id>].
  • Why: the id value is per-conversation, not stable across forks/resumes; embedding it in system made every cross-session continuation diverge at exactly one line → 0% cache. Moving it to the ephemeral tail keeps the system — and therefore the whole-history prefix — byte-identical across sessions.
• Wire shape: +1 trailing user message on every tool-injecting request (same class of change as the already-shipped nudge/imgNote/retrieveNote tails). Within-session cache hits are unchanged (the note sits past the divergence point); cross-session / fork / resume now hit through the entire history instead of 0%.

Both compression modes were reasoned: plugin mode (agent owns compression) and proxy mode (proxy executes compress) both keep the note present under the identical gate, so no mode-specific regression.

Pre-flight

• npm run typecheck — clean
• npm test — 3123 tests, 0 fail (two consecutive runs)
• npm run build — success
• E2E preflight (E2E_CHECK=1 tests/e2e/e2e-codex.test.ts) — green (codex-cli 0.147.0 + dist OK). The full token-costing e2e needs a live upstream not reachable from this sandbox; the authoritative e2e runs in CI (ci-e2e.yml) since this touches the request pipeline (src/server.ts).

Test updates reflect the intended +1 trailing message: count/shape assertions bumped by one, and the #1548 append-stability check now excludes the convnote tail slot alongside the chain tag (the byte-level cacheable history prefix is unchanged by this fix).

Closes #1611

Refs: #760 (per-call conversation_id), #1486 (byte-exact resume inheritance), #1541/#1542 (chain stamps), #1592 (0% signature quantified).

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-29_conv-id-trailing-prefix-cache (3586673)

Option A — Install this exact build (recommended)

npm install -g billion-context@0.0.1562-pr.1620.c6af033

This exact version is immutable and reproducible. The rolling pr tag tracks the newest test build across all PRs.

Option B — Download artifact

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

tar xzf billion-context-pr1620.tgz
npm install -g package

---

Install instructions are refreshed on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Review of PR #1620 (fix for #1611)

① Issue triage — valid, root cause correctly targeted. #1611's premise checks out: the live per-conversation id sat in the static system line, so any fork/resume/session-switch (new canonical id, byte-identical history, #1486) diverged at exactly one system line → provider prefix cache missed entirely (cached_tokens=0, quantified in #1592's S2 fork evidence). The fix removes the producer of the divergence rather than masking it: system now carries a byte-stable pointer const (CONVERSATION_ID_POINTER_NOTE, src/compress-tool.ts), and the live value rides an ephemeral trailing user message under the same gate the id-in-system used. MCP routing is untouched — the model still reads the value from context and echoes it as conversation_id (src/mcp.ts). I also verified the one subtle path: the compress-loop re-request deliberately drops the trailing note (keeps only the constant pointer) — safe, because conversation_id is consumed exclusively by the MCP shim, never by the server-side loop.

② Regression ledger (existing behaviors on touched paths):

• System prompt text — old: full [Your bili conversation id: <id>. When calling…THIS session.] line; new: constant pointer, live value as [Your bili conversation id: <id>] tail. Flagged by author ✓ — this is the intended fix, not drift.
• Wire shape — +1 trailing user message on every tool-injecting request across all four lanes (Anthropic/OpenAI/Google/Responses), same class as the already-shipped nudge/imgNote/retrieveNote tails. Within-session cache hits unchanged (tail sits past the divergence point); cross-session/fork/resume now hit through the entire history — the point of the PR.
• Responses lane invariant (fix: keep Codex compaction_trigger final when injecting the compress nudge #283/bug(OpenAI SSE): proxy reorders tool_call events + changes response ID → client loses tool results #209) — note is gated shouldInject && !isCompactionTrigger && !ACP_NO_COMPRESS_PROMPT, so compaction_trigger stays the last input item. ✓
• Both compression modes — plugin mode: gate verified to have no !pluginMode term (shouldInject = injectTool && !isTitleGen), matching the old "id in system in BOTH modes" behavior; proxy mode: acp_summary carrier path untouched. No mode-specific regression.
• Test quality — the new conversation-id line in the static system prompt breaks the upstream prefix cache on fork/resume/session-switch (cached≈0 despite byte-identical history) #1611 test in tests/issue760-percall-conversation.test.ts is a genuine repro-with-teeth: two sessions, identical filler history, different ids, asserting system + shared history are byte-identical and only the tail differs. The other 8 updated files are honest count/slot bumps for the +1 message (incl. the codex 版本压缩后缓存大幅度降低问题 #1548 append-stability exclusion and the cache-friendly slot taxonomy gaining convnote) — nothing weakened.

③ CI red on the original head — root-caused and fixed (2 commits pushed).

• real pi package-native lane failed with compress must have been called and its result re-sent; real opencode native lane with filler message must carry its ACP ref tag, got null. All ~9 prior runs of both workflows (other branches off current master) were green → caused by this PR, not flake. Root cause: tests/e2e/fake-upstream-chat.mjs scripts the model off the last user message (directive queue source) and reads lastUserRef from it — the new convnote tail lands after the client's real prompt and masked it: the scripted compress directive never fired (pi) and the filler-ref oracle read null (opencode). Fix (commit efdca14d): skip convnote messages in both filters, same transport-metadata treatment as the [design] chain checkpoint step 3: carrier stamp shape on role-alternation-strict wires (anthropic/google) #1421 chain carrier. Reproduced locally first: pi suite 3/4 → 4/4, opencode (V1 lane, host 1.14.46) 4/4.
• The branch was 13 commits behind master; I rebased onto current master. That surfaced a semantic collision the pre-rebase run couldn't see: master added tests/fold-round2-shape.test.ts (dsh 缓存失效 #1592, febbc162) after this branch point, whose divergence model hardcodes a 3-slot volatile tail (nudge / chain stamp / new turn). conversation-id line in the static system prompt breaks the upstream prefix cache on fork/resume/session-switch (cached≈0 despite byte-identical history) #1611 adds a 4th, so steady bodies end […, prompt, convnote, chain] and the first growth pair was misclassified as a mid-history rewrite (pair 0->1 diverges at element #3/7 — verified against raw bodies on both trees: master passes at exactly the idxB >= count-3 boundary; the branch needs -4). Fix (commit 35866736): constant widened to 4 with the slot list documented; all pairs pass incl. both fold round-2s.

Pre-flight on the final head 35866736: typecheck clean · 3139 unit tests / 0 fail · build OK · pi e2e 4/4 · opencode e2e 4/4 (V1 locally; V2 2.0.3 lane runs in CI). Local dsh lane not executed here (sandbox dsh 0.1.1-rc.2 vs CI pin 0.1.5-rc.3 — profile-boot mismatch unrelated to this diff); the dsh run on the previous head was green and the fake change only restores pre-#1611 last-user semantics for it. CI re-runs were triggered by the push.

Verdict: direction correct, root cause (not symptom), wire-shape change properly flagged, tests have teeth. Per §7.4 this stays a human-merge (wire/message-shape change affecting both modes). Ready to merge once the triggered CI suite is green.

中文摘要：修复了 #1611「会话 id 写进静态 system 导致 fork/resume 后前缀缓存全 miss」的问题（id 值改为随临时尾部消息下发，system 只留常量指针）；我定位并修掉了它引起的两处 CI 红（e2e fake 的"最后一条用户消息"语义被新尾注遮蔽、以及 rebase 后与 master 新增 #1592 折叠形状测试的 3→4 尾槽冲突），已补两个测试修复 commit 并 rebase 到最新 master，本地全套预检通过，等 CI 绿后可合并。

[ranxianglei]

Superseded by #1685 / #1692 (zero-injection: the tail note this PR adds is exactly what the approved design rules out). Suggest closing.