[ranxianglei]

Problem (#570)

The weak-overflow heuristic (3 high-usage stream truncations in 15 min) and the strong path (a real non-2xx overflow rejection) both wrote into the same per-model learnedContextLimits. Any mid-stream death — llama.cpp shared-KV exhaustion (failed to find free space in the KV cache), OOM, network reset — looked like an overflow, so one misjudged session permanently shrank the model's window for every later session, across restarts, with no correction path: the only upward self-heal is gated on fallback windows, never on the authoritative configured value.

Fix (three parts)

1. Provenance split. Strong evidence (an actual non-2xx overflow rejection, incl. the payload-size upper bound) now lands in metadata.confirmedContextLimits / confirmedContextLimit; the weak heuristic keeps writing only to learnedContextLimits / learnedContextLimit. Resolution order is confirmed > speculative, so a weak guess can no longer clobber a confirmed window.

2. Retraction. A successful turn whose upstream-reported input exceeds a learned window beyond a margin (max(256 tokens, 3%)) proves it stale — KV-pressure false positive, resized server, … — and removes it before this request's config is resolved. True overflows cannot be contradicted (a turn cannot succeed above the real window); the armed emergency value (== the learned value exactly) sits below the margin and never retracts itself.

3. Ground truth for llama.cpp. inspectContextOverflow now recognizes exceed_context_size_error and parses the real limit from the (N / M > W) body shape, so a genuine overflow learns W instead of letting the weak path guess.

Both compression modes considered: plugin mode (agent-owned compression; proxy suppresses injection) and proxy mode — the changed paths are upstream-facing (error inspection, per-request limit resolution, session metadata) and identical on both; no wire-format change.

Pre-flight

• npm run typecheck — clean
• npm test — 1056/1057 pass; the single failure (launcher.test.ts: resolveClientCommand codex/claude resolve to themselves) is pre-existing on clean master in this environment (sandbox stub at /usr/bin/codex makes the resolver return an absolute path) and unrelated to this change
• npm run build — success
• New unit coverage: provenance precedence (confirmed governs; weak can't clobber), retraction (stale removed, margin respected, armed no-op, confirmed retracts too, other models untouched), llama.cpp body parsing → window 150528

Note: the real-client E2E suite (tests/e2e/, needs ACP_TEST_E2E=1 + live Responses upstream) was not runnable in this sandbox; the local e2e-style self-heal tests (tests/e2e-overflow-selfheal.test.ts, issue-393, issue496) all pass.

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-06_weak-overflow-self-correction (35d2809)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-572

Each push to this PR publishes a new version under the pr-572 npm tag.

Option B — Download artifact

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

tar xzf billion-context-pr572.tgz
npm install -g package

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 ework agent · qwen3.8-27b

[bot] 🏷 Review complete — preflight independently verified, one defect found in review and fixed on the branch.

Preflight (all three re-run by me):

• npm run typecheck — clean ✓
• npm test — 1056/1057; the single failure (launcher.test.ts: resolveClientCommand codex/claude resolve to themselves) reproduces identically on clean master in this environment, so it's environmental, not from this change ✓
• npm run build — success ✓

Design review: provenance split, the retraction margin (max(256 tokens, 3%)), and the llama.cpp ground-truth parsing all check out against the code. One deliberate divergence worth noting (not a bug): handle() resolves limits in order cPM→cS→sPM→sS while resolveLearnedLimit uses cPM→sPM→cS→sS — the former drives the provenance log label, the latter drives noteWeakOverflow's usage-% denominator where the tighter bound is preferred. Harmless; left as-is.

Defect found in review (fixed): retraction reads session.stats.lastInputTokens, which is a high-water mark that failed turns also write — noteWeakOverflow arms it at the failing input size, and forward()'s overflow floor takes max(prev, floor). Consequence: three mid-stream deaths at 185k input (KV-pressure pattern) → mark at 185k; then a genuine parseable overflow learns W=150528 into confirmedContextLimits; on the next request, retraction sees 185000 − 150528 ≥ margin and deletes the just-learned ground truth, logging "a later turn succeeded at 185000" when nothing succeeded. That contradicts the PR's own stated invariant ("True overflows cannot be contradicted") — and a second variant (kills above an already-governing confirmed window) deletes a correct value. Impact is narrow and transient (one wasted overflow cycle + misleading log pair), so non-blocking — but it undermines the guarantee this PR exists to provide, so I fixed it on the branch:

1. forward(): when the overflow body parses a window, arm at exactly it — no turn can succeed above the real window, so any higher armed value provably came from a failed turn; unparsed cases keep the max-floor.
2. noteWeakOverflow(): while a confirmed window governs, cap the armed value at it — a mid-stream death above the real window is a non-window kill mode (KV pressure / OOM / reset), not evidence about the window.
3. Regression coverage: unit tests for both caps; a new e2e rig (three 370k KV-pressure deaths + a parsed 128k overflow, asserting the confirmed window survives the next request's retraction check — fails without the fix); tightened the existing armed-value assertion to exact equality.

Re-ran everything after the fix: typecheck clean, npm test 1058/1059 (same sole pre-existing failure), build success. Commit pushed to the PR branch. Merge remains yours: #572

[ranxianglei]

已在 35d2809 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:#570 weak-overflow:util.ts 取 master #554 的超集正则;server 的保守估计保留 raw-body 兜底但走 confirmed 通道命名;side-request guard 改读 resolveLearnedLimit(confirmed→speculative),老 #554 测试断言对齐 confirmed 通道。CI 绿后请人工合并。

[ranxianglei]

🤖 ework agent · qwen3.8-27b

[bot] 🏷 Verified the rewritten branch independently:

• Linearity: 35d2809 is a single linear commit on current master (parent 3f57ce4) ✓
• Guard survival: the review-fix (exact arm at parsed window in forward(), armInput cap in noteWeakOverflow(), plus both regression tests incl. the e2e rig) is present in the semantic merge ✓
• Preflight (fresh npm ci — note the base now carries acp-kernel 0.0.54, stale deps initially surfaced two unrelated acp_status 会计缺陷:tool 结果全计为 text + 同一报告两套互斥的 tool 定义 + 双估计器混印不标尺 + pct 地板 #386 CJK failures that cleared after install): typecheck clean ✓, full suite 1152/1153 — sole failure is the environmental launcher.test.ts: resolveClientCommand codex/claude sandbox-stub issue that reproduces identically on clean master here ✓, build success ✓

Everything checks out. On the merge itself: I can't merge PRs — system rules forbid it. Please merge yourself: #572