[ranxianglei]

Problem (#553)

Anonymous prefix-affinity sessions minted after an ACP compression (forks) start with lastInputTokens = 0. Preflight then decided solely on the optimistic defaultCountTokens estimate (non-CJK 4 chars/token), which undercounts dense code/JSON replays by up to ~4x — so over-window payloads were forwarded raw upstream: no compression, no interception, doom-looping into #551's 300s transfer cutoffs and exceed_context_size_error 400s.

Fix

When a session has no measured baseline AND arrived anonymously (prefix-affinity resolution), judge payload size by a char-count upper bound — 1 token per char, which never undershoots BPE token counts since every token covers at least one character. Applied at three points:

• Trigger: tokenCount = char upper instead of max(baseline, optimistic), so over-window dense replays enter the compress loop.
• Nothing-compressible gate: such sessions fail fast (structured 502) instead of gambling a raw forward past the window.
• Loop internals: char-based chunk packing (min-compress-range integrity preserved via a minUnits guard so sub-minimum chunks are never emitted), char-based netting, and a regime-aware PreflightResult.fitsWindow verdict — optimistic for measured-baseline sessions (preserves the fix: stitched-stream usage double-count after compress re-request (adapter-anthropic >0 guard) #300 stale-baseline contract byte-for-byte), char upper for unknown-baseline ones (closes the residual gap where the loop exits with failure while the optimistic figure still says "fits").

Explicit-identity zero-baseline sessions keep today's optimistic path: they are genuinely-new or post-native-compaction populations whose payloads are small or shrink, and they self-heal via the learned-window overflow path (persist.ts persists lastInputTokens across restarts, so zero-baseline anonymous sessions are exactly the fork/reload population).

Why not an estimator multiplier? Any fixed k<4 false-fires on repetitive prose that BPE-compresses at ~4+ chars/token (the codex-compact-e2e setup sits at 3.5x window in chars while truly fitting at 87% of window); k=4 is exactly the bug. Content is indistinguishable by char-ratio, so the conservative measure is scoped structurally (anonymous + zero-baseline) instead.

Tests

• New tests/preflight-fork-estimate.test.ts: anonymous fork replay whose optimistic estimate fits but char upper does not → preflight fires, summarizes, forwards within window; small anonymous fresh session → no preflight (no over-fire). Both verified RED on old code / GREEN on fix.
• Full suite: 948/949 (single failure is pre-existing env-only in launcher.test.ts, fails identically on clean tree — sandbox lacks client binaries).
• npm run typecheck clean; npm run build clean.
• E2E: E2E_CHECK=1 zero-token preflight passes against built dist. Full ACP_TEST_E2E=1 run is not completable in this sandbox (codex warmup fails identically on the clean tree before any change lands) — flagged for CI/manual run.

[ranxianglei]

🤖 ework agent · qwen3.8-27b

[bot] 🏷 On it — checking out 2026-09-05_fork-preflight-upper-bound, will verify the diff, re-run typecheck/test/build, and confirm the new test is RED on old code / GREEN on fix before reporting back.

[ranxianglei]

🤖 ework agent · qwen3.8-27b

[bot] 🏷 Review complete — LGTM with minor nits. I checked out the PR branch and verified every claim independently.

Verification (all executed, not taken on trust)

Check	Result
npm run typecheck	✅ clean
Full suite	✅ 948/949 — the single failure is launcher.test.ts "resolveClientCommand: codex/claude resolve to themselves"; I re-ran it on a clean origin/master worktree and it fails identically there → pre-existing env-only (sandbox has a codex binary on PATH), as claimed
npm run build	✅ clean (dist/index.js 2.60 MB)
RED/GREEN	✅ new fork-replay test fails on master code, passes on the fix; the over-fire guard passes on both (correct regression-guard behavior)
E2E_CHECK=1 zero-token preflight	✅ passes against built dist
Full ACP_TEST_E2E=1	⚠️ not completable in this sandbox (real upstream + codex warmup required; warmup fails identically before any change lands). Flagging for CI/manual-dispatch run before merge

Triage conclusions

1. Bug is real and correctly layered. The root cause is in the measurement regime, not in chunking or folding: anonymous + lastInputTokens <= 0 sessions are exactly the fork/reload population (prefix-affinity resolution mints a fresh session id after ACP compression; persist.ts only persists a baseline that was ever measured), and for them the optimistic 4-chars/token estimate has no floor to lean on. Fixing the estimator multiplier was rightly rejected — any k<4 false-fires on prose, k=4 is the bug itself; content is indistinguishable by char-ratio, so structural scoping is the right layer.
2. fix: stitched-stream usage double-count after compress re-request (adapter-anthropic >0 guard) #300 contract preserved. For measured-baseline sessions every measure reduces byte-for-byte to the old code path (countText = defaultCountTokens, credit-based netting, fitsWindow = payloadEstimate < limit) — I traced this line by line in src/preflight.ts. The stale-baseline forward contract is untouched.
3. Both compression modes considered. Plugin-mode Anthropic clients early-return empty processedMessages in pluginMode, so the trigger can't fire there; opencode sends x-session-affinity (not anonymous). The new regime effectively covers proxy-mode anonymous clients — matching the design scope.
4. Mechanics check out: applyRanges doesn't mutate originals in place (so post-fold char netting is correct); the minUnits guard can't infinite-loop (chunks always advance ≥1 message; sub-minimum trailing chunks hit the existing rangeChars < minChars skip → clean exhausted failure); explicit-identity zero-baseline sessions keep today's optimistic path as documented.

Nits / observations (non-blocking)

• Cosmetic: src/server.ts:2264 — if (result.fitsWindow) { is indented 12 spaces instead of 4.
• Accepted tradeoff, stated for the record: low-density-prose anonymous forks (BPE ≥ ~4 chars/token — e.g. the codex-compact-e2e setup at 3.5× window in chars while fitting at 87%) will now over-fire preflight, including the nothing-compressible → structured 502 case for a giant single-message conversation that upstream might have accepted. That's the deliberate conservative direction (fail-fast with a diagnostic over guaranteed doom-loop into 上游长时间 prefill(首 token 前无输出)被 undici 默认 bodyTimeout=300s 掐断,UPSTREAM_TIMEOUT_MS=10min 配置形同虚设,大上下文请求必超时 #551 cutoffs) — fine by me, just noting the edge.
• Transient baseline inflation: after a char-regime preflight, lastInputTokens briefly holds the char-based upper bound until the forwarded response's usage report lands. Self-heals (the test asserts correction to the reported value); interim requests can only trigger extra conservative rounds, never a raw past-window forward.

Housekeeping

• Branch is based on an older master (predates fix: starved tool-carrying main requests no longer demote to side requests (#546) #548 and fix: drop previous .old before logger rotation rename (Windows-safe) #555 merges); those touch src/logger.ts / tests/logger-rotate.test.ts — no file overlap, clean merge expected.
• Per repo rules I won't merge — please merge yourself once you're happy with the CI E2E run: fix: gate fork sessions on char-count upper bound in preflight (#553) #558

[ranxianglei]

已在 9630396 将本 PR 语义合并到当前 master 并重写为线性单提交(修 #575 同样的 diff-爆炸问题)。全量 1146/1146 通过、typecheck 0 错、build 成功。语义合并要点:#553 fork 上界:char regime 移植到 #575 多 range walk 上,上界补 #470/#488 floors;unknownBaseline 用 fitsWindow、known 用 master 重建实测;unknown+无range 才提前 failFast,known 落穿保 #330 relax;测试按新 kernel 阈值(minPressureBenefit 5k)与 #388 side-request 形态调整。CI 绿后请人工合并。

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-05_fork-preflight-upper-bound (9630396)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-558

Each push to this PR publishes a new version under the pr-558 npm tag.

Option B — Download artifact

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

tar xzf billion-context-pr558.tgz
npm install -g package

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 ework agent · qwen3.8-27b

[bot] 🏷 重写后的 9630396 已独立复核完毕，全部通过。

验证结果（均实际执行）

检查	结果
线性单提交	✅ 父提交 = #591 merge（3f57ce4）；master 此后仅前进 v0.1.88 release commit（#592，只动 package.json/lock）→ 与 PR 文件零重叠，干净合并
npm run typecheck	✅ 0 错
全量测试	✅ 本环境 1145/1146 —— 唯一失败是 launcher.test.ts 的 resolveClientCommand（本沙箱 PATH 上有 codex 二进制导致解析出绝对路径），在干净 master 2b510af 上同样失败 → env-only，与 PR 无关；无 codex 环境下即你的 1146/1146
npm run build	✅ 成功
E2E_CHECK=1 零 token preflight	✅ 对构建产物通过
完整 ACP_TEST_E2E	⚠️ 沙箱内不可完成（需真实 upstream + codex warmup），维持 CI/手动跑标记

Diff 审计（对照合并说明逐条核实）

• preflight.ts：char regime 已正确移植到 fix: preflight tries all viable ranges before declaring exhaustion #575 多 range walk 上；上界现携带 image+wire floors（track: plugin-mode proxy-side emergency truncation (last-resort backstop, deferred from #467) #470/多模态会话上下文控制失效 + preflight 救援 502（codex responses store:false） #488）——这点比原 PR 更完整（原版的 finalUpper 未含 floors）。minUnits 在 baselineKnown 时恒为 0，乐观路径的 chunk packing 逐字节保持原样；known 的 netting 仍走 credit-based。✅
• server.ts：effectiveTokenCount 喂入 anthropic/openai/responses 三个 prepare 分支；trigger 的 tokenCount 仅对 anonymous zero-baseline 用 upper+floors；post-loop 判定 unknown 走 result.fitsWindow、known 走 preflight fail-fast × soft-protection livelock: over-window payload with nothing foldable outside the protected zone 502s forever #330 重建 payload 再估计；unknown+无 compressible range → failFast(502, retryable=false)，known 落穿保留 preflight fail-fast × soft-protection livelock: over-window payload with nothing foldable outside the protected zone 502s forever #330 relax 路径。✅
• 新测试：tests/fork-nudge-trigger.test.ts 两个用例（窗口下匿名 fork → nudge 注入且 preflight 静默；同对话 explicit-identity → 无 nudge，验证 regime scoping）均通过。✅
• anonymousPrefixAffinity 元数据在 prefix-affinity 解析处一次性写入 session（sticky），与 trigger 的 anonymous 参数同源（anonAffinity !== null），口径一致。✅

一点范围观察（非阻塞）

effectiveTokenCount 不只改了 preflight trigger，也改变了喂给 kernel processTurn 的 tokenCount——匿名 zero-baseline fork 现在会在窗口下就收到 nudge，而不是"usage 0%、直到 overflow 才触发"。这超出了原 PR 范围，但正对 #553 "什么都不触发"的症状，作用域与原 regime 完全一致（anonymousPrefixAffinity && lastInputTokens == 0），方向保守（只可能提前压缩，不会放行超窗 payload），且有新测试覆盖。记录在案即可。

CI 绿后请人工合并（按仓库规则我不执行 merge）：#558