Repository navigation
fix: reconcile compaction restart and custom-storage fork recovery #407
Description
Problems
Two pre-existing state-recovery paths can restore stale or incomplete ACP state:
- During session initialization, ACP records the newest compaction timestamp from current history before loading persisted state. It then restores persisted nudge/message-ref/tool-cache state without comparing the current history boundary to the persisted boundary. Because
lastCompactionis already current, the next transform does not reset those stale transient fields. - Fork recovery loads the parent state from the default storage location instead of the child session's resolved
storagePath. It also consumes legacy four-digit parent message refs without normalizing them to the current five-digit format, so parent-to-fork mapping can fail.
Impact
A restart immediately after native compaction can retain stale refs and nudge baselines. A fork using custom storage or legacy state can lose inherited compression blocks when copied history no longer contains replayable compress inputs.
Proposed fix
- Compare the current completed-compaction boundary with the persisted boundary after state load. When history is newer, reset transient refs/nudges/tool cache while preserving compression blocks and stats, then persist the corrected state.
- Load parent fork state from the resolved active storage directory. Normalize legacy parent
byRef/byRawIdaliases before translating IDs. - Add restart-after-compaction, custom-storage fork, and legacy-ref fork regression tests.
Related OpenCode V2 migration: #395. Concurrency work: #404.
Activity
ranxianglei commentedon Sep 16, 2026
[bot] 🏷 Acknowledged — triaging #407 now.
Plan (will verify each claim against the code before implementing):
lib/state/(compaction-restart reconciliation + fork storage/ref normalization).Will report findings as soon as verification is done.
ranxianglei commentedon Sep 16, 2026
Both problems verified real, fixed, and covered by regression tests → PR #409
Verification (before implementing):
ensureSessionInitialized(lib/state/state.ts),state.lastCompaction = findLastCompactionTimestamp(messages)runs before the persisted load, then nudge anchors/baselines and message refs are restored unconditionally. After a restart between native compaction and the next transform,updatePerTurnState's reset trigger (lastCompactionTimestamp > state.lastCompaction) can never fire because both sides already equal the new boundary — stale transients survive indefinitely.state.ts:348omitted the child's resolvedstorageDir, so with a customstoragePaththe parent file was never found; (b)mapForkIds(lib/state/rebuild.ts) consumed parentbyRefkeys verbatim — legacy pre-1.1.0 four-digit refs (m0001) can never match fork five-digit refs (m00001), so translation yields an empty map and inherited blocks are lost whenever the copied history lacks replayable compress inputs.Fixes:
_persistedLastCompaction; when history is newer, reset transient refs/nudges/tool cache via the existingresetOnCompaction(preserves compression blocks and stats by design — parity with the live-compaction path) and persist the corrected state.storagePath" transition (found in code review — without it that scenario degraded from transfer to replay).normalizeParentMessageIds()normalizes legacy four-digit parentbyRef/byRawIdbefore translation (mirrors the own-session load-time migration).Tests (
tests/restart-compaction-fork-recovery.test.ts, 6 tests): restart-after-compaction reset + negative control (no over-resetting), custom-storage fork e2e, storagePath-transition fallback, legacy-ref fork (unit + full-init e2e). Per §5.7.3 discipline, every bug-repro test was verified to fail against unfixed code (stash fixes → fail; restore → pass); the transition test likewise fails without the fallback.Quality gates: dual-agent review completed (§5.3 code + §5.6 tests, both APPROVE; findings fixed on-branch — see PR comment); full suite 1269/1269; typecheck/build/Prettier clean; CI on the PR: pr-validation, build, test (Node 22+24) all passing, e2e job running. Devlog:
devlog/2026-09-16_compaction-fork-recovery/.📎 PR: #409 (merge is a human-only operation per project rules — please merge when satisfied.)
中文摘要：修复了重启落在原生压缩之后导致过期消息引用/nudge 基线被无条件恢复、以及自定义 storagePath 或旧版 4 位引用下 fork 恢复丢失继承压缩块的问题（含审查发现的存储路径过渡场景父状态回退缺失），已加 6 个回归测试并验证修复前必失败，PR #409 可以合并。
fix: reconcile compaction restart and custom-storage fork recovery
fix: reconcile compaction restart and custom-storage fork recovery
ranxianglei commentedon Sep 16, 2026
Status update — PR #409 is now fully green and up-to-date with master.
While finalizing I noticed master had moved (v1.18.1 release + #403 Windows path fix merged upstream), so I rebased the branch onto the current master tip (
06efd39). Along the way an external commit (style: prettier-format new devlog files, by ranxianglei) had been pushed directly to the PR branch — it was preserved via cherry-pick, nothing was clobbered.Final state of #409 :
1576ed4, rebased onto latest master, no conflictsmergeable: true,mergeable_state: cleanReady for human merge whenever you are (merge stays a human-only operation per project rules).
中文摘要：PR #409 已 rebase 到最新 master（含 v1.18.1 与 #403），保留了他人直接推到分支上的 devlog 格式化提交，全部 CI（含 e2e）通过、本地 1280/1280 测试通过，状态 clean，等待人工合并。