fix(agent,provider): un-stick compaction on thinking-model summaries and unnumbered overflows / 修复思考型模型摘要与无数字超窗导致的分段压缩死锁

[Linearl]

Summary

Fixes #9878 — two compaction failure shapes observed on a real 2M-token session, where pressure-triggered summary retries never converged and the turn died with an HTTP 400 overflow:

1. Thinking-model summaries were misread as empty (internal/agent/compact.go): summarize() accumulated only ChunkText; DeepSeek vision thinking models put the whole briefing in reasoning_content with an empty content block, so every chunked fragment died on summarizer returned empty output (observed: fragment 2/14). Now a pure reasoning-only summary is surfaced (clamped to 32 KiB, matching the summaryOutputMaxTokens envelope), mirroring the boundedllm treatment in 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. A reasoning + tool-call turn keeps the empty-output rejection — that reasoning is private chain-of-thought, not digest material, preserving the shape contract in TestSummaryCollectorRejectsEmptyAndLengthLimitedOutput.
2. Unnumbered provider overflows were invisible to the chunked fallback (internal/provider/context_limit.go): Zhipu GLM reports {"error":{"code":"1261","message":"Prompt exceeds max length"}} with no token numbers — none of the numeric regexes match and contextLimitInvariant requires window > 0, so AsContextLimitError returned nil and the chunked fallback ([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 family) never triggered; the uncapped fold then failed transparently on every retry, defeating 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 handoff. ParseContextLimitError now trusts the provider-confirmed overflow with zero token fields (window unknown); consumers already treat window=0 as "learn nothing, fall back to the configured window" (learnContextBudget only stores window > 0).

Testing

• go test ./internal/provider/ -run TestParseContextLimit -count=1 — includes new TestParseContextLimitErrorGLMUnnumbered1261 (trusted overflow with zero token fields, case-insensitive message variant, 401 still rejected).
• go test ./internal/agent/ -run 'TestSummaryCollector|TestSummarizer|TestOverflow' -count=1 — includes new TestSummarizerReasoningOnlyIsSurfacedNotEmptied; existing empty-output shape intent preserved.
• Live validation: 2,032,885-token session no longer deadlocks on pressure compaction with both failure shapes active.

Cache Impact / 缓存影响

Cache-impact: none - summarize surface and context-limit parsing do not touch promptCacheKey, projection keys, or cache-key derivation; reasoning-only surfacing feeds the compaction digest path only. / 仅压缩摘要与超窗识别路径，不触及 promptCacheKey/会话投影/缓存键逻辑。
Cache-guard: go test ./internal/agent/ -run 'TestSummaryCollector|TestSummarizer' -count=1 && go test ./internal/provider/ -run TestParseContextLimit -count=1 / 两条路径均有定点回归守护。
Documentation-impact: none - internal-only fix, no docs or config surface changed.
System-prompt-review: none - no system prompt text changes.

Refs: #9878, #9678/#9679 (same reasoning-only shape, boundedllm side), #9082, #9572

Review fixes (41594e9) / 评审修复

Pushed by the maintainer on top of the original commit:

• Recovery: an overflow without token numbers no longer resends the same prompt under a clipped max_tokens; recoverContextLimit treats it as physically over the window and goes straight to overflow compaction (TestUnnumberedContextLimitSkipsIdenticalRetry).
• Error text: a zero-field limit falls through to the generic 400 message with the provider's own reason instead of "prompt 0 + completion 0 = 0 tokens, window 0".
• Reasoning clamp cuts on rune boundaries (truncateUTF8Bytes) and ChunkToolCallStart counts as a tool attempt (Copilot findings), each with a regression test.
• Three comment blocks trimmed to repolint's limit so make lint passes.

Correction to the mechanism note above: on current main-v2 the chunked-fallback gates key only on errSummaryOutputTruncated / ErrCompactionRequired, so recognizing the GLM overflow does not by itself route the summary path into chunked recovery; that wiring lands with #9879, whose ladder skips the re-plan for zero-field limits and goes to the transcript form, then the fragment path. This PR is independently valuable regardless: reasoning-only summaries no longer dead-loop, and the sampling path now recovers a GLM overflow through compaction instead of an identical retry.

[Copilot]

🟡 Changes recommended

summarize() currently (a) may mis-detect tool-call attempts because it only counts ChunkToolCall (not start/args-delta), and (b) clamps reasoning by raw byte slicing which can produce invalid UTF-8.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR fixes two compaction deadlock / overflow-recognition failure modes seen in very long sessions: (1) thinking-model summaries where the “answer” is emitted only via reasoning_content (empty text), and (2) provider context-limit overflows that don’t include any token counts (GLM 1261 “Prompt exceeds max length”), preventing chunked fallback from triggering.

Changes:

• Treat “unnumbered” provider-confirmed overflows as ContextLimitError (window unknown / zero token fields) so overflow recovery and chunked compaction fallback can engage.
• Surface reasoning-only summaries for compaction when text is empty and no tool call was attempted, with a 32KiB clamp.
• Add focused regression tests for both scenarios.

File summaries

File	Description
internal/provider/context_limit.go	Recognizes GLM-style unnumbered overflows as trusted context-limit errors.
internal/provider/context_limit_test.go	Adds coverage for GLM unnumbered 1261 overflow parsing and status gating.
internal/agent/compact.go	Accumulates reasoning chunks and conditionally surfaces reasoning-only summary output (clamped).
internal/agent/compact_test.go	Extends fake provider to simulate reasoning-only streaming shape for tests.
internal/agent/compact_summary_failure_test.go	Adds regression ensuring reasoning-only summaries can install a projection instead of retry-looping on “empty output”.

Review details

• Files reviewed: 5/5 changed files
• Comments generated: 2
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.