feat: session id is the client-provided conversation value verbatim; 400 without identity (#286)

[ranxianglei]

#286：session ID = client-provided conversation 值原文（无 hash、无其他维度）

设计（owner 确认）

• session ID = client-provided conversation 值原文：codex session-id/thread-id、claude x-claude-code-session-id、opencode x-opencode-session、Responses body session_id / metadata.session_id、prompt_cache_key 等。
• 无强信号 → 400（错误信息列出可用身份字段）；content-fingerprint 会话禁用（指纹有真实碰撞面，静默建会话只会孤儿化状态）。
• credential（OAuth bearer 轮换）、upstream（中转切换）、protocol（中转翻译 / 跨协议切模型）都是可变的，均不参与身份。会话状态本身协议中立（kernel CompressionState + 文本 blocks + CoreMessage），客户端每轮重发全量历史，协议切换不断状态。
• meta.protocol / upstreamOrigin 保持 first-wins（持久化文件名命名空间 sessions/<proto>/<host>_<sha>.json 由此派生，更新会孤儿化旧文件）；persist.ts 无需改动（文件名已是 id 的 sha256，id 从文件体读取）。

变更

• src/session-id.ts：删除 deriveSessionId / extractKey；保留 clientConversationHeader / affinityToken / preferPromptCacheKeyIdentity。
• src/server.ts：
  • 无强信号 → 400（anthropic 与 openai/responses 各自的错误格式）。
  • sessionId = conversation（原文）。
  • 修复本设计暴露的潜在 bug：route 的 compressProtocol 配置原先按 session.meta.upstreamOrigin（first-wins）解析——切换中转后新中转的配置会被静默忽略；现按请求实际 upstream 解析。
  • debug dump 文件名 sanitize（safeSessionId）——id 现在是客户端原始值。
• 测试：e2e-session-identity.test.ts 重写（10 例：bearer 轮换 / 中转切换 / 跨协议续接 / 匿名 400 ×3 / 隔离）；7 处匿名测试调用点补强信号；2 个 headless-binding 测试首请求带与 register 不同的强信号（保持走 pending 路径）。

代价

• 一次性：存量持久化会话重新定键（旧记录孤儿化）——可接受（owner 确认）。
• 无原生 session id 的客户端（hermes/dsh 无 plugin、通用客户端）现在会收到 400——显式失败，而非静默丢状态。

验证

• typecheck clean；657/657 tests pass。

[github-actions]

📦 Built Package Artifact

Branch: 2026-08-27_strong-signal-session-id (950c89b)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-288

Each push to this PR publishes a new version under the pr-288 npm tag.

Option B — Download artifact

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

tar xzf billion-context-pr288.tgz
npm install -g package

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 ework agent · vllm-qwen/qwen3.8-27b

[bot] Review of PR #288 (branch 2026-08-27_strong-signal-session-id, commit 6215446) — LGTM, ready for human merge.

Verified locally (reproduced CI checks on the branch)

• npm run typecheck — clean
• npm test — 670/670 pass (incl. the 6 new strong-tier unit tests in tests/proxy-session-id.test.ts and all 7 tests in new tests/e2e-session-identity.test.ts, which I also ran in isolation: 7/7)
• npm run build — success

Code-level verification

• Kernel semantics match the PR description. acp-kernel 0.0.46 conversationIdentityResponses: header / body session_id / metadata.session_id → clientProvided: true; previous_response_id / content-fingerprint → false. Proxy-side preferPromptCacheKeyIdentity promotes fingerprint→pck to true exactly when source === "content-fingerprint", so it can never override a stronger signal.
• Flag computation per protocol in src/server.ts:760-769 is correct:
  • responses: responsesIdentity?.clientProvided ?? false — identity always exists, tier matches the conversation value source.
  • openai: openaiIdentity?.clientProvided ?? false — header → true, fingerprint → false, pck promotion → true.
  • anthropic: !!convHeader — convHeader = clientConv ?? sessionHeader, both are client-sent request headers, so true ⟺ the client actually provided a conversation signal.
• Single session-derivation site (src/server.ts:770 deriveProxySessionId); all callers of the changed deriveSessionId signature are updated; no other getSession/session-creation paths bypass the flag.
• Weak tier unchanged: fingerprint sessions still hash protocol|upstream|credential|conversation; the e2e control test confirms same-content cross-credential isolation is preserved (3 sessions for [no-auth ×2, keyX ×1, different-content ×1]).
• AGENTS.md compliance: no as any / @ts-ignore / console.* in the diff; version untouched; branch naming + feat: commit convention OK.
• One-time re-key cost: persist layer is lazy per-key, so orphaned old-keyed records are inert (never loaded); matches the accepted trade-off in 会话身份不稳定：credential/upstream 维度导致长会话状态孤儿化（sessions 缺失） #286.

Non-blocking notes

1. Tier flip: a client that intermittently sends/omits its conversation header will create two sessions (strong + weak) for one conversation. The "tiers never collide" test documents this as intended — no silent state migration. Fine.
2. Strong tier drops the credential dimension, so two different accounts on one proxy instance sharing an identical client-provided conversation id would share compression state. In practice the ids in scope are UUIDs (codex thread-id, claude session-id, opencode session) — negligible collision surface, and the proxy is single-user local. Accepted in 会话身份不稳定：credential/upstream 维度导致长会话状态孤儿化（sessions 缺失） #286 discussion; noting for the record.
3. previous_response_id-only clients remain weak tier, so their session id still churns per turn — pre-existing behavior, not a regression, out of scope here.

Coordination

Confirmed: #249's chatgpt-account-id credential-dimension part is superseded by this fix (account switches must not break session continuity); thread-id subagent identity + compaction fast path remain complementary. #281/#283 unaffected.

Per repo rules I will not merge — please merge manually: #288