[ranxianglei]

Fixes #987

What

The context window is a deployment property — owned by declarations (config / routes / registry / runtime-info), never adjusted from session traffic. This removes the runtime window-learning mechanism entirely, along with the accumulated patch-on-patch machinery that grew around it (#570 retraction guards, #572 confirmed-channel, #857 provenance tiers, #969 operator-owns-window — each a fix for the previous fix).

Removed (~1,700 lines gross)

• src/weak-overflow.ts — the whole file (armFailureShrink / armedAtLeast ladder). An upstream overflow now arms the one-shot emergency shrink directly:
  • stated window in the 400 body → arm at exactly it (source=usage, [Question] 非白名单上游（如 DeepSeek）图片按 b64/4 估算，实测高估约 100×；且该估算被写回 lastInputTokens 并自愈成 confirmed 窗口，使 #496 逃生门永久失效 #857 semantics kept)
  • no parseable number → arm at min(declared, payload-estimate) — the rejection is evidence at the size actually sent, so the PR #489 后续: 默认图片计费 b64/4 在像素 tile 计费上游(官方 Anthropic/OpenAI)会硬 502 误杀多模态会话 #496 forward-once gate still breaks the 多模态会话上下文控制失效 + preflight 救援 502（codex responses store:false） #488 400 loop after one rejected forward, without ever re-centering the declared window
• server.ts self-heal application ladder — learnedLimit / operatorOwnsWindow / nativeFromFallback learn paths. The declared window resolves once per request and keeps governing; a wrong declaration self-heals only through the one-shot emergency (next turn compresses below the failing size).
• confirmedContextLimit / confirmedContextLimits session metadata fields — gone from session.ts and all call sites.
• sideRequestGuard learnedLimit parameter — the guard now intersects the declared window with the usage-grounded arm (lastInputTokensSource === "usage") only.
• Tests: tests/weak-overflow.test.ts + tests/plugin-truncation-weak-overflow.test.ts deleted; tests/e2e-overflow-selfheal.test.ts rewritten to three context window problem continues #987 cells (T1 stated → arms, learns nothing; T2 no number → arms at payload size, no fold; T3 correct declaration still recovers); exit-matrix references updated.

Kept (explicitly)

• One-shot emergency shrink (per model window problem(I'm using opencode2 and 9router) #969 / context window problem continues #987 discussion): arms from the rejection itself, source=usage, overwritten by the next real usage report.
• PR #489 后续: 默认图片计费 b64/4 在像素 tile 计费上游(官方 Anthropic/OpenAI)会硬 502 误杀多模态会话 #496 image-relay forward-once gate — now keyed on the rejection arm.
• [Question] 非白名单上游（如 DeepSeek）图片按 b64/4 估算，实测高估约 100×；且该估算被写回 lastInputTokens 并自愈成 confirmed 窗口，使 #496 逃生门永久失效 #857: a window number stated by the upstream remains usage-grade.
• Learned context window (self-heal) treats any truncated upstream stream as a window overflow: false positives permanently shrink the session window, persisting across restarts #570: trivially guaranteed — there is nothing left to retract.

Why not keep the learner

Field evidence (#987): a session window silently re-centered to 16898 tokens and the session was stuck compressing forever. Every learned value was a guess from one error body; wrong guesses compound (the #969 incident chain). Declarations are explicit, inspectable, and owned by the operator — the upstream's own overflow 400 is still surfaced verbatim so misconfigurations are visible immediately.

Pre-flight

• npm run typecheck ✅
• npm test — 1950 pass / 0 fail / 2 skipped (E2E gates) ✅
• npm run build ✅

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-19_remove-learned-window (8955887)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-992

Each push to this PR publishes a new version under the pr-992 npm tag.

Option B — Download artifact

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

tar xzf billion-context-pr992.tgz
npm install -g package

---

Install instructions are refreshed on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Reviewing now: fetching 2026-09-19_remove-learned-window, re-running typecheck/tests/build on the branch head and checking CI status before giving a verdict. Will follow up with findings.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Review complete — code verified, one small leftover found and fixed directly on the branch, all checks green. Ready for you to merge (human-only per repo rules).

What I verified

Base freshness: single commit 721c4a5 sitting directly on current origin/master (ea2b33f) — no stale-base risk, no rebase needed.

Removals match the description (read the full diff, grepped for leftovers):

• src/weak-overflow.ts fully deleted; no remaining imports or references anywhere in src/ or tests/.
• Self-heal application ladder (learnedLimit / operatorOwnsWindow / nativeFromFallback) gone from server.ts; declared window resolves once per request and keeps governing.
• confirmedContextLimit(s) gone from session.ts and all call sites. Remaining test mentions are intentional negative assertions (assert.equal(…, undefined)) or legacy persisted-file compat checks.
• sideRequestGuard now takes armedLimit instead of learnedLimit and intersects declared ∩ usage-grounded arm (lastInputTokensSource === "usage") only — all call sites (incl. tests) updated consistently.
• Overflow-400 path: stated number → arms at exactly it with source="usage" ([Question] 非白名单上游（如 DeepSeek）图片按 b64/4 估算，实测高估约 100×；且该估算被写回 lastInputTokens 并自愈成 confirmed 窗口，使 #496 逃生门永久失效 #857 semantics kept); no number → arms at min(declared, estimateTokensFast(body)), so the PR #489 后续: 默认图片计费 b64/4 在像素 tile 计费上游(官方 Anthropic/OpenAI)会硬 502 误杀多模态会话 #496 forward-once gate still breaks the 多模态会话上下文控制失效 + preflight 救援 502（codex responses store:false） #488 loop after one rejected forward without ever re-centering the declaration.
• The pre-existing arm emergency shrink on upstream failure to break relay-5xx deadlock #604 armFailureShrink in server.ts is untouched and tags source="estimate", so it correctly stays out of the usage-only gates — no interaction bug.
• e2e rewrite covers the three context window problem continues #987 cells exactly as described (T1 stated→arms/learns nothing, T2 no-number→arms at payload size/no fold, T3 correct declaration recovers via preflight fold under the declared window).
• Docs: no stale references to the removed learner in README(zh)/CONFIGURATION(zh) — the surviving "self-heal" mentions describe the kept one-shot emergency mechanism (or unrelated features).

One gap found — fixed on the branch (commit 8955887)

The web UI still consumed the removed field: src/web/client.ts rendered fmtW(item.contextWindow)+"/"+fmtW(item.provenMaxInput) while the stats endpoint no longer sends provenMaxInput, so the column would permanently show window/—; the page.ts header tooltip ("相信窗口 / 近期成功最大输入（差距大 = provider 虚标）") was stale too. Fixed both (2 lines total, tooltips now say 声明窗口（不再从会话流量学习）). Re-ran the full pre-flight after the fix — all green. Per review discipline I committed it straight onto the PR branch rather than filing a separate round-trip.

Evidence

• Local on final head 8955887: npm run typecheck ✅ · npm test 1950 pass / 0 fail / 2 skipped (E2E gates) ✅ · npm run build ✅
• CI on 8955887: CI, e2e-fake-codex, PR Build Artifact, CI Windows codex overlay — all 4 workflows success.

Mergeability note

Per §7.4 this stays human-merge: it changes wire behavior in both compression modes and the session persistence format (metadata fields removed), both of which are outside the auto-merge gate. Nothing blocks merge; please merge when ready: #992

中文摘要：完整核对了移除窗口学习机制的 PR（基线最新、无残留引用、#857/#496 语义保留、e2e 三格覆盖到位），发现并直接在分支上修掉了 Web UI 仍在渲染已删除字段 provenMaxInput 的小遗漏（2 行改动，已补 commit 8955887），本地与 CI 全绿——可以合并。