[Bug]: chunked compaction deadlocks on thinking-model summaries and unnumbered provider overflows / 分段压缩在思考型模型摘要与无数字超窗错误上死锁

[Linearl]

[Bug]: chunked compaction deadlocks on thinking-model summaries and unnumbered provider overflows / 分段压缩在思考型模型摘要与无数字超窗错误上死锁

TL;DR(中文摘要)

问题：超长会话（实测 2M tokens）的 pressure 触发压缩会进入无法收敛的失败循环，最终 turn 以 HTTP 400 超窗死亡。两条独立根因：① summarize 采集器丢弃 ChunkReasoning，思考型模型（DeepSeek vision SKUs）把整段摘要放进 reasoning_content、content 为空 → summarizer returned empty output，chunked 分块回退在 fragment N/M 上重复同一失败；② ParseContextLimitError 的正则族不识别无数字的超窗错误（智谱 GLM 1261 "Prompt exceeds max length"），AsContextLimitError 返回 nil → chunked 回退不触发 → uncapped 请求每次透传失败。

影响面：任何使用 thinking 模型（DeepSeek vision 系）或 GLM 网关的会话，一旦上下文超过单次摘要容量即进入死循环；#9572 un-deadlock follow-up 的"失败走 overflow recovery"前提被 ② 破坏。

期望行为：① 纯 reasoning-only 摘要应被 surface（对齐 #9679 对 boundedllm 的处理）；② 无数字的 provider 确认超窗应被 trust 为 ContextLimitError（窗口未知），让 chunked 回退接管。

修复：PR 附上（两个定点修复 + 回归测试，已在 fork 实证）。

Background

Observed on a real 2,032,885-token session (deepseek/deepseek-v4-flash-vision-exp, mode yolo, 86 turns over 8 days). The pressure-triggered maintenance sequence was: prune OK (2,032,885 → 1,960,242) → compaction_started → compaction_done → context_maintenance failed, action=summary → turn_done failed with HTTP 400: Prompt exceeds max length. Two distinct failures, captured at different times on the same session:

Failure A — thinking-model summary is misread as empty (manual summary receipt failed-summary-1, reason context summary failed: fragment 2/14: summarizer returned empty output):

The single-shot summary fails, the chunked fallback (#9082 family) correctly kicks in and streams "compacting fragment N/M", but summarize() accumulates only ChunkText. DeepSeek vision thinking models put the whole answer in reasoning_content with an empty content block, so every fragment hits the same empty-output check and the loop never converges. This is the exact shape boundedllm learned to surface in #9679 (goal evaluator) — the compaction summarizer has the same blind spot, but the fix was never extended here. Note the intent contrast: the goal evaluator's verdict must be surfaced; a summary turn that also attempted tool calls legitimately keeps the rejection (that reasoning is private chain-of-thought, not digest material).

Failure B — unnumbered provider overflow is invisible to the chunked fallback (automatic receipt failed-summary-14, reason context summary failed: glm-cn: status 400: {"error":{"code":"1261","message":"Prompt exceeds max length"}}):

ParseContextLimitError (internal/provider/context_limit.go) matches three numeric shapes (OpenAI-style "maximum context length is N tokens...", Anthropic-style "prompt is too long: N > M", "input length and max_tokens exceed...") and a JSON field family (context_length, prompt_tokens, ...). GLM's overflow carries no numbers at all — contextLimitInvariant requires positiveToken(window) and rejects it, so AsContextLimitError returns nil. Consequences:

1. The chunked fallback condition (errors.Is(err, errSummaryOutputTruncated) || errors.Is(err, ErrCompactionRequired) || provider.AsContextLimitError(err) != nil) does not trigger — the oversized request fails transparently on every pressure retry.
2. The [Bug]: Compaction cannot rebuild a folded projection once invalidated — uncapped summary requests overflow shared-window models / 投影失效后压缩无法重建折叠投影：无安全前缀上限的摘要请求必然溢出共享窗口模型 #9572 un-deadlock follow-up deliberately keeps an uncapped fold when no balanced prefix remains, relying on "if the provider rejects the uncapped request, the existing overflow recovery takes over". That handoff requires the rejection to be recognized as a context-limit error; with GLM it never is, so the designed recovery path is dead on arrival for this provider.

Fix (attached PR)

Two minimal fixes + regression tests (also validated against a live 2M session):

1. internal/agent/compact.go — accumulate ChunkReasoning alongside ChunkText; when text is empty and no tool call was attempted, surface the reasoning (clamped to 32 KiB, matching the summaryOutputMaxTokens envelope). A reasoning+tool-call turn still gets summarizer returned empty output, preserving the existing shape contract in TestSummaryCollectorRejectsEmptyAndLengthLimitedOutput.
2. internal/provider/context_limit.go — isUnnumberedPromptTooLong(message, body) matches the GLM shape case-insensitively; ParseContextLimitError returns a trusted ContextLimitError with zero token fields. Consumers already treat window=0 as "learn nothing, fall back to the configured window" (learnContextBudget only stores window > 0), and the chunked fallback now triggers.

总结

• 两条根因均在当前 main-v2 复现成立（代码级核对）；fix(goaleval): parse reasoning-only evaluator responses / 修复Goal评估器在思考型模型下空响应致目标暂停 #9679 修复了 boundedllm 的同形态问题，本次把同样的处理带到压缩摘要，并补上 GLM 无数字超窗的识别。
• 风险低：两处均为定点修复 + 保守语义（带 tool call 仍拒绝；窗口未知时一切消费方维持既有 fallback 行为）。
• 跟进：PR 附回归测试；已在 2M tokens 真实会话验证不再死循环。

Refs

• [Feature]: Extract key points from an over-length/unusable legacy session into a fresh conversation / 对超长/无法继续的旧会话：一键结构化提取要点并生成新会话 #9082 (chunked summary fallback), [Bug]: Compaction cannot rebuild a folded projection once invalidated — uncapped summary requests overflow shared-window models / 投影失效后压缩无法重建折叠投影：无安全前缀上限的摘要请求必然溢出共享窗口模型 #9572 (summary cap + un-deadlock follow-up), [Bug]: Goal completion evaluator fails with empty response on reasoning-only models / Goal完成评估器在思考型模型下空响应致目标暂停 #9678/fix(goaleval): parse reasoning-only evaluator responses / 修复Goal评估器在思考型模型下空响应致目标暂停 #9679 (reasoning-only evaluator responses — same shape, boundedllm side)

[Linearl]

Clarification: this is not a regression from a version update / 澄清：这不是版本更新引入的回归

We verified the history before filing: summarize() in internal/agent/compact.go has never accumulated ChunkReasoning — the check goes back to at least v1.24.1 (where summarizer returned empty output was introduced) and is still absent on current main-v2. Likewise, context_limit.go has never recognized the GLM unnumbered shape.

The relationship to #9679: that fix surfaced reasoning-only responses for the boundedllm channel (goal evaluator / recovery reviewer). The compaction summarizer is a separate collector with the same blind spot that was simply not covered by that fix — same failure shape, different code path, so this issue is a sibling rather than a duplicate or a regression of #9679.

Why it only surfaced now: each ingredient alone degrades gracefully. It takes the combination of (a) a 2M-token session that only the chunked fallback can fold, (b) a thinking model that puts the entire answer in reasoning_content with an empty content block, and (c) a gateway whose overflow error carries no token numbers (GLM 1261) to turn the failure into a non-converging loop. The attached PR (#9882) keeps the #9679 semantics: pure reasoning-only summaries are surfaced; a summary turn that also attempted tool calls still gets the empty-output rejection.

[Linearl]

Follow-up: parallel fragment summaries + context-limit half-splitting / 后续：分块摘要并行化 + 超窗半切

End-to-end validation on the 2M-token session passed after the fix PR, and surfaced two wall-time/robustness follow-ups now prepared as a second PR:

1. summarizeExtractChunks ran fragments strictly serially — the 2M fold spent minutes paying N × per-fragment latency on a thinking model. Fragments are independent LLM calls; the follow-up runs them with a bounded 4-worker pool (fail-fast cancel, index-ordered merge). The shared chunkedSummaryRun gains a mutex; fragment reservations drop the serial future-plan reserve (meaningless out of order) and reserve only themselves, with half-split recursion and merge steps still reserving against the same maxChunkedSummaryCalls hard budget.
2. extractFragmentResilient / mergeGroup retried only output truncation and hard compaction requirements; a fragment (or merge group) that overflowed the summarizer window failed the whole fold instead of splitting into pieces that fit. A trusted ContextLimitError now triggers the same bounded half-split — depth caps unchanged.

PR incoming (separate, perf-titled, adapted to the current main-v2 chunkedSummaryRun architecture).

[SivanCola]

Fixed by #9882, merged into main-v2 as 6361ba2 (original fix by @Linearl plus maintainer review fixes in 41594e9).

What changed

• Thinking-model summaries that arrive as reasoning_content with an empty content block are now surfaced as the compaction briefing instead of failing with "summarizer returned empty output" on every fragment; a reply that also opened a tool call keeps the rejection, and the clamp stays valid UTF-8.
• A provider overflow without token numbers (Zhipu GLM 1261 "Prompt exceeds max length") is recognized as a context-limit error. The sampling path now recovers it through overflow compaction instead of resending the same prompt, and the error text no longer quotes zero token counts. The summary-path handling for such overflows (skip re-plan, transcript form, then fragments) lands with Recover from summary-request overflow and rescue over-ceiling context / 摘要请求超窗自恢复与硬上限截断救援 #9879.

Verification

• go test ./internal/agent/ ./internal/provider/... ./internal/control/ — includes TestSummarizerReasoningOnlyIsSurfacedNotEmptied, TestSummarizerReasoningWithToolCallStaysEmpty, TestSummarizerReasoningClampKeepsValidUTF8, TestParseContextLimitErrorGLMUnnumbered1261, TestUnnumberedContextLimitSkipsIdenticalRetry; make lint clean; PR CI 24/24 green.

Release availability: not in a released build yet (latest is v1.38.1). It ships with the next release from main-v2.