fix(runtime): retry mid-turn compaction after a no_safe_completed_span miss

[Colafornia]

Summary

Fixes #5790.

A mid-turn compaction attempt that finds nothing safe to fold (no_safe_completed_span) never reaches the summarizer, yet it latched state.summarizerFailure for the rest of the turn. Every later attempt short-circuited before re-reading the ledger — including the reactive recovery meant to rescue a real provider context_overflow — so a turn could die on input capacity even though a grown pool would have folded.

The latch now covers only outcomes that actually invoked the summarizer (provider_error, malformed_*, output_length, input_too_large, empty summary), preserving the bounded-retry behavior added for #4634. no_safe_completed_span is a property of the event pool at that step, so the next attempt re-reads the ledger and can fold once a completed tool pair appears.

Verification

• node --test on the rebuilt @maka/runtime dist: mid-turn-capacity-backend.test.js, overflow-reactive-recovery.test.js, history-compaction.test.js — 157 tests pass.
• Two new tests reproduce the issue deterministically through the backend fixtures (the proactive trigger and the reactive recovery path); reverting the fix makes exactly those fail.
• npm run lint and npm run format:check pass repo-wide; tsc passes via npm --workspace @maka/runtime run build.
• Not run: the full npm test matrix across all workspaces, knip, and the desktop e2e budget check.

AI use

• Generative tooling made a substantive contribution

Tool(s) and scope: Devin authored the fix and the tests.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No

中文版本

摘要

修复 #5790。

一次轮中压缩尝试如果找不到可安全折叠的区间（no_safe_completed_span），并没有发起任何 summarizer 调用，却把 state.summarizerFailure 置位到本轮结束。之后的每次尝试都在重读账本前被短路——包括本应用于抢救真实 provider context_overflow 的响应式恢复——因此即使增长后的事件池本来可以折叠，turn 仍可能死于输入超限。

现在只有真正调用过 summarizer 的结果才置位（provider_error、malformed_*、output_length、input_too_large、空摘要），#4634 引入的有界重试行为不变。no_safe_completed_span 只是该步事件池的属性，下一次尝试会重读账本，在出现完整的工具调用对后即可折叠。

验证

• 在重新构建的 @maka/runtime dist 上运行 node --test：mid-turn-capacity-backend.test.js、overflow-reactive-recovery.test.js、history-compaction.test.js —— 157 个测试通过。
• 新增两个测试，通过后端 fixture 确定性复现该问题（主动触发路径与响应式恢复路径）；还原修复代码后恰好这两个测试失败。
• npm run lint 与 npm run format:check 仓库级通过；npm --workspace @maka/runtime run build 的 tsc 通过。
• 未运行：全 workspace 的 npm test 矩阵、knip、desktop e2e 预算检查。

[jackwener]

Reviewed head 3ca4ea16cab22445a51b4108ab3c3cd9a54de7a2. The diagnosis is right: a plain no_safe_completed_span never reaches the summarizer, so it should not spend the Turn-wide latch. plan.reason has only two fail-open values (summarizer_failed, no_safe_completed_span), so the new condition is exhaustive. 1×P2 (inline).

P2: after the one-step retreat, no_safe_completed_span can follow a real summarizer call, and the latch no longer catches it.

• In planHistoryCompaction (history-compaction.ts, around 380-395), a summarizer rejection with a proven accepted-input boundary sets maxCoveredCount = proven and continues.
• If the smaller window then has no safe coverage, the same call returns { reason: 'no_safe_completed_span' } (:316, or :464 when the loop runs out) even though the summarizer was already dispatched and failed.
• With this change that outcome does not latch. So every later step re-plans the same prefix, repeats the same doomed summarizer call, and retreats into the same miss. That is the repeated-summarizer-call loop #4634 bounded, and a slow provider that fails makes each iteration expensive.
• Suggested fix: have the planner report that it invoked the summarizer, and latch on that rather than on reason. Alternatively, return summarizer_failed (with the original diagnosticReason) when the retreat itself finds no safe span. Please add a test: an input_too_large rejection whose proven boundary leaves no safe coverage, then a second step, asserting that no second summarizer call is made.

The new tests cover the proactive trigger and the reactive-recovery path for the plain miss, and CI test is green. Not run locally: the suites. The P2 comes from reading the code paths above; I did not reproduce it end to end.

[hqhq1025]

Reviewed commit 826b419. The change lets a structural no_safe_completed_span retry when the event pool grows, while retaining the Turn-wide latch after a real summarizer failure. A later commit removes all mid-turn input_too_large retreats. I found one P2 regression in that removal, inline.

Node 24 build:test and the three focused compaction suites passed (160 tests). I also exercised the planner with a prior-run, same-turn model reply after the head anchor: it made one summarizer call and returned summarizer_failed without testing the proven smaller prefix. Fresh main 71bc045 merges cleanly and git diff --check passes. Current-head hosted test was queued at review time. I did not run the full suite or a real provider/handoff process; the concrete handoff path and synthetic rejection establish the reachable boundary, not its production frequency. Please address the inline regression before merge. This is not merge approval.

Automated review notice: This comment was posted by an automated review agent operated by hqhq1025. It is not an independent human review and does not replace one.

[jackwener]

Approving at 826b41976134c2a2a80840d411a78993b88df1f0. The P2 from review 5339998409 is closed, and no P0–P2 remain.

• a6b6c2227: once a retreat has happened (retreated = true), a smaller window with no safe coverage, or a loop that runs out, now returns summarizer_failed / input_too_large instead of no_safe_completed_span. The Turn latch therefore catches a failure the summarizer actually produced. A plain miss that never reached the summarizer still returns no_safe_completed_span, which is #5790's fix.
• 826b41976: mid_turn no longer retreats. This checks out: the mid-turn planner receives state.priorInvocations, so acceptedInputBoundary can only land on a prior-run reply, which precedes the head anchor that mid_turn coverage must pass. The retreated cut could never fold. input_too_large now fails open as summarizer_failed directly and latches. standalone and pre_turn keep the retreat.
• New tests cover the retreat-then-miss case in the planner and the mid-turn backend path.

CI test is green on this head.

Not run locally: the suites.

[hqhq1025]

Reviewed commit fa32a3f. The new delta removes the unconditional mid-turn failure on input_too_large (packages/runtime/src/history-compaction.ts:389-405) and restores the one-step proven-boundary retreat. The added planner regression covers a handoff predecessor reply after the current Turn's head anchor, asserting a second, smaller coverage attempt; the retreat-then-no-safe-span case still reports summarizer_failed, so the Turn latch bounds repeated calls. The earlier P2 is addressed. No new substantiated P0-P3 in the inspected change. Node 24 build:test and three focused compaction files passed (161 tests); current-head hosted test is green. Fresh main 2f32205 merges cleanly and git diff --check passes. I did not run the entire local suite or a real provider/handoff process. This is not merge approval.

Automated review notice: This comment was posted by an automated review agent operated by hqhq1025. It is not an independent human review and does not replace one.