fix(agent-runtime): retry compaction summary before retained-tail fallback

[ZxlDragonDoctor]

Problem

Automatic context compaction often fails to produce a real summary. On failure the runtime immediately writes a retained_tail fallback checkpoint (~112 tokens of recovery prose) and drops older messages from the next model request. There is no retry-until-a-real-summary-exists loop (#543).

Two concrete gaps in packages/agent-runtime/src/runtime.ts:

1. buildCheckpoint calls generateCompaction once; transient provider failures go straight to recoverCompactionFailure.
2. compactionSummaryWouldExceedBudget returns recoverable: true without calling the model, so oversized history never attempts a reduced payload.

Solution

• Retry loop in performCompaction: up to COMPACTION_SUMMARY_MAX_ATTEMPTS = 3 with backoff (2s → 4s → 8s). Only after retries are exhausted does retained-tail recovery run.
• Reduced payload via shrinkPreparationForSummary: when the budget guard trips, drop oldest summarize-eligible messages until the input fits, then still call the model. Keep at least one message.
• Honest failure label: fallback checkpoint summary now says "Summary generation failed after retries" instead of implying a successful tiny summary.

Classification, provider-retry backoff, abort/cancel paths, and the fresh_window rollover strategy are unchanged.

Testing

• New unit file packages/agent-runtime/src/compaction-retry.test.ts:
  • retry cap = 3
  • backoff schedule 2s/4s/8s
  • shrink until budget clears
  • undefined when even minimal payload cannot fit
  • never shrink below one message
• Full vitest / desktop suite not run in this environment (fork behind upstream main; rebasing after review).

Issue alignment

Addresses #543. Claimed on the issue before implementation. Distinct from #224 / PR #249 (fallback recoverability on a later compaction) — this PR retries the failing attempt.

Non-goals

• Map-reduce multi-pass summarization
• Changing Codex-shaped keepRecent survival rules ([Feature] 设置页增加上下文压缩阈值（keepRecent / reserve） / Expose compaction keepRecentTokens and reserveTokens in Settings #203)
• Settings UI beyond the failure wording

Core value: Long sessions keep a real model summary when the provider flaps, instead of silently dropping history behind a 112-token stub.

[vastsa]

谢谢，方向是对的：先重试、再缩减、最后回退。这个 PR 与当前 main 冲突，且有几处我没法直接合：

• 外层循环对所有 recoverable: true 的失败都重试，包括「无新内容可压缩」、401、配额耗尽这类确定性失败，每次白等 2s+4s。pi-ai 的 retryAssistantCall 已经自带瞬时错误分类和 abort 归一化，compact() 的 retry 参数就是为这个留的。
• shrinkPreparationForSummary 直接丢掉最旧的消息，但 checkpoint 的 throughMessageId 仍覆盖它们，等于摘要静默宣称了它没看过的范围。
• compaction-retry.test.ts 在测试文件里重新声明常量和一个「镜像」shrink 函数，没有 import runtime，对真实接线没有覆盖。
• 缺 ADR/spec：ADR 0049 第 1 条决策（预检守卫）被改了。

另外根因还多一层：预检守卫对原始消息求和，而 pi 序列化时每条 tool result 已截到 2 000 字符，所以它高估了好几倍——这是「不调模型就失败」的主要来源。

#563 保留了你的三段式方向，用 RetryPolicy + 按序列化 prompt 预检 + 只截工具输出不删消息 + UI 失败标签 + ADR 0282 落地，已在 issue 里 credit 这个 PR。这个 PR 我会在 #563 合入后关闭。

---

Thanks — the direction (retry, then reduce, then fall back) is right and is kept in #563. What blocked landing this as-is: the outer loop retries deterministic failures too; shrinking by dropping the oldest messages leaves the checkpoint claiming a range the summary never saw; the test file mirrors the constants instead of importing the runtime; the ADR 0049 preflight decision changes without an ADR; and the branch conflicts with main. The bigger root cause was also missing: the guard summed raw messages while pi caps tool results at 2 000 chars when serializing, so it overstated tool-heavy inputs several-fold. #563 credits this PR and will close it when merged.

[vastsa]

Superseded by #563, now merged into main (c36d246). Thanks for the direction — credited in the ADR 0282 record and on #543.