fix: reconcile compaction restart and custom-storage fork recovery

[ranxianglei]

Summary

Fixes #407 — two pre-existing state-recovery paths could restore stale or incomplete ACP state.

1. Restart immediately after native compaction (lib/state/state.ts)

ensureSessionInitialized set state.lastCompaction from the current history before loading persisted state, then restored persisted nudge anchors/baselines, message refs, and tool-cache state unconditionally. If native compaction completed after the last persist (restart in between), those fields were stale — and updatePerTurnState could never reset them because its trigger compares against an already-current lastCompaction.

Fix: after the load, compare the current completed-compaction boundary with the persisted _persistedLastCompaction. When history is newer, call the existing resetOnCompaction(state) — which resets message refs, all nudge anchors/baselines, and the tool cache while preserving compression blocks and stats (parity with the live-compaction path) — and let the existing end-of-init save persist the corrected state.

2. Fork recovery (lib/state/state.ts + lib/state/rebuild.ts)

• Parent state was loaded without the child's resolved storageDir, so with a custom storagePath the parent file was never found → inherited blocks lost whenever the copied history lacked replayable compress inputs. Now passes state.storageDir (undefined falls back to the default location — backward compatible).
• mapForkIds consumed parent byRef keys verbatim; legacy pre-1.1.0 four-digit refs (m0001) never match fork five-digit refs (m00001). New normalizeParentMessageIds() mirrors the own-session load-time migration (byRef rebuilt from byRawId).

Tests

New tests/restart-compaction-fork-recovery.test.ts (5 tests):

1. restart after newer compaction → transients reset, blocks/stats preserved, corrected boundary persisted
2. negative control: restart without newer compaction → transients still restored (no over-resetting)
3. custom-storage fork e2e → parent found under storagePath, block transferred, fork persisted there
4. legacy 4-digit parent refs (unit via restoreForkCompressionState)
5. legacy 4-digit parent refs (full-init e2e through ensureSessionInitialized)

Regression validity verified per AGENTS.md §5.7.3: with the lib fixes stashed, tests 1/3/4/5 FAIL and the control passes; with fixes, 5/5 pass.

Full suite: 1268 tests, 0 failures. Typecheck + build clean.

Devlog: devlog/2026-09-16_compaction-fork-recovery/

[github-actions]

📦 Built Plugin Artifact

Branch: 2026-09-16_compaction-fork-recovery (1576ed4)

Option A — Install from npm PR tag (recommended)

opencode plugin opencode-acp@pr-409 --global

Each push to this PR publishes a new version under the pr-409 npm tag.

Option B — Install from GitHub

opencode plugin "github:ranxianglei/opencode-acp#2026-09-16_compaction-fork-recovery" --global

Option C — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf opencode-acp-pr409.tgz
cp -r package/dist ~/.cache/opencode/packages/opencode-acp@latest/node_modules/opencode-acp/dist

3. Restart opencode to pick up changes.

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Received — this is a completed agent PR (ework-agent-pr marker), so I'm verifying rather than redoing work.

Status so far:

• Branch 2026-09-16_compaction-fork-recovery fetched; 2 commits (419045b fix + f30da03 devlog).
• Diff is clean and scoped: lib/state/state.ts (+27/-4), lib/state/rebuild.ts (+39), new tests/restart-compaction-fork-recovery.test.ts (370 lines), devlog REQ/WORKLOG. No version field change, no unrelated files.
• CI on head f30da03: pr-validation ✅, test (24) ✅, build-artifact ✅, test (22) still running.

I'm now checking the code changes against issue #407 (root-cause validation) and will report findings shortly.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Review complete — PR verified, ready for human merge.

Root-cause validation (against #407, per triage rules):

• Bug 1 confirmed: lib/state/state.ts:320 sets state.lastCompaction from current history before loading persisted state, so the live trigger in updatePerTurnState (state.ts:41, compares findLastCompactionTimestamp > state.lastCompaction) can never fire for an already-completed compaction — stale refs/nudge baselines/tool cache would be restored forever. The fix's placement (after load+restore at state.ts:459-469, before the end-of-init save at state.ts:484) is correct: after the Math.max merge, lastCompaction > persistedBoundary ⇔ history has a newer compaction than what was persisted.
• Bug 2 confirmed (both halves): parent was loaded without storageDir (loadSessionState already accepted the optional 3rd param — backward compatible, undefined → default location); and mapForkIds (rebuild.ts:202-203) looks up fork refs via state.messageIds.byRef.get(ref) where fork refs are always 5-digit — legacy 4-digit parent keys could never match, silently dropping inherited blocks. normalizeParentMessageIds mirrors the own-session migration (state.ts:432-443), rebuilds byRef from byRawId (authoritative direction), carries byRef-only entries defensively.

Fix semantics checked:

• resetOnCompaction (lib/state/utils.ts:402) preserves prune.messages + stats by design (Bug 2 comment) — parity with the live-compaction path holds.
• Legacy persisted state without _persistedLastCompaction → boundary treated as 0 → reset fires if history contains any completed compaction. This is the safe direction (stale refs reset rather than restored); sessions never compacted have lastCompaction = 0 → behavior unchanged. No regression risk identified.

Verification performed on branch 2026-09-16_compaction-fork-recovery:

• Full suite locally: 1268 passed / 0 failed; tsc --noEmit clean; build clean.
• Regression validity re-verified independently (§5.7.3): reverted both lib files to master versions → tests 1/3/4/5 fail, negative control (test 2) passes; restored → 5/5 pass. The new tests genuinely catch the bugs.
• Test file passes §5.6 checklist: imports real source modules (no local reimplementation), names match assertions, buildConfig() complete incl. gc, legacy-ref inputs actually rewritten to 4-digit form (regex-asserted /^m\d{4}$/), precondition sanity assertions make the reset checks meaningful, no tautologies. Fork-copy uses inputless compress parts — exactly the scenario where replay can't reconstruct and only parent transfer works.
• Diff cleanliness: 5 files only (lib/state/state.ts, lib/state/rebuild.ts, new test file, devlog REQ/WORKLOG). No version bump, no unrelated changes.
• CI on head f30da03: pr-validation ✅ · test (22) ✅ · test (24) ✅ · build ✅ · build-artifact ✅ · e2e still running (Docker suite).

No issues found that need fixing in this PR. Merge is a human-only operation — please merge when e2e goes green: #409

---

中文摘要：已独立验证 #407 两个根因均真实存在（compaction 后重启时 lastCompaction 先于持久化状态加载导致重置永不触发；fork 恢复未传 storageDir 且旧版 4 位 ref 无法匹配 5 位 ref），修复逻辑正确且向后兼容，本地 1268 测试全过、无修复时新测试确实失败（回归有效性成立），CI 除 e2e 外全绿——可以合并。

[ranxianglei]

Dual-agent review (AGENTS.md §5.3 + §5.6) — completed, findings fixed on-branch

Reviewer 1 (tests) — verdict: APPROVE.

• Fixed: buildConfig() in the new test file is now type-conformant to PluginConfig (added logLevel, top-level allowSubAgents, qualityGate, messageFilters; removed mis-nested experimental.allowSubAgents — an inherited gap from tests/rebuild.test.ts).
• Fixed: pinned with a comment why toolParameters.size === 0 is asserted (the cache is transient by design, re-derived each turn via syncToolCache; it was never persisted).
• Fixed: added a persisted-side-effect reload assertion to the legacy-ref e2e test.
• Verified: inputs force the exact code paths under test (input-less compress part blocks replay → parent-state transfer path; summary message triggers findLastCompactionTimestamp), and T1/T3/T4/T5 fail against unfixed code while the negative control passes.

Reviewer 2 (lib code) — verdict: APPROVE, one minor finding fixed on-branch:

• Parent-state lookup now falls back to the default storage location when the custom-dir load misses (state.ts, fork branch). Without it, the "just configured storagePath" transition (child state still at the default location — the scenario warned about ~15 lines above) would degrade fork recovery from transfer to replay. New regression test added for this path and verified to fail without the fallback.
• Verified clean: persisted format unchanged, own-session migration loop untouched, no as any/assertion hacks, init-only paths (no hot-path cost), Math.max boundary ordering correct, resetOnCompaction preserves blocks/stats as claimed.

Verification: 6/6 new regression tests pass (all bug repros verified to fail pre-fix); full suite 1269/1269; typecheck + build + Prettier clean.

中文摘要：修复了重启落在原生压缩之后导致过期瞬态状态被无条件恢复、以及自定义 storagePath / 旧版 4 位引用下 fork 恢复丢失继承压缩块的问题；双 agent 审查发现的 storagePath 过渡场景父状态回退缺失已直接在分支上补上并加了回归测试，可以合并。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Re-verified the updated branch (head 5d12bb4) after the dual-agent review round, plus one small fix I pushed on-branch.

Verification of 5d12bb4 (review-fix commit)

Diff reviewed line-by-line — all claimed changes are present and correct:

• Fallback fix (lib/state/state.ts, fork branch): parent load is now loadSessionState(parentSessionId, logger, state.storageDir) with a follow-up if (!parent && state.storageDir) parent = await loadSessionState(parentSessionId, logger). Correct and minimal — the fallback only fires when a custom storageDir is resolved, so default-location behavior is untouched, and it covers exactly the "just configured storagePath" transition the warning above describes.
• Test file: buildConfig() is now type-conformant (logLevel, top-level allowSubAgents, qualityGate, messageFilters added; mis-nested experimental.allowSubAgents removed), the toolParameters.size === 0 assertion carries the pinning comment, the legacy-ref e2e test asserts the persisted-side-effect reload, and the new transition-fallback test is present.

Local runs on 5d12bb4:

• Full suite: 1269/1269 pass; tsc --noEmit clean.
• Code/test files Prettier-clean (see one nit below).
• Regression validity re-checked both ways: against master's lib, tests 1/3/5/6 fail and the negative control passes; against f30da03's lib (has the storageDir pass but not the fallback), test 4 fails while the other five pass — so the new transition test is a genuine regression test for the fallback, confirming the "verified to fail without the fallback" claim.

One nit found → fixed on-branch as 6d94b7a

The two new devlog markdowns (devlog/2026-09-16_compaction-fork-recovery/REQ.md, WORKLOG.md) were not Prettier-formatted — npm run format:check flags them (list indentation, table padding, *em*→_em_). CI doesn't run format:check so nothing blocked, but AGENTS.md §5.1 requires formatting consistency, so I ran prettier --write on just those two files and pushed: 6d94b7a style: prettier-format new devlog files. Markdown-only reflow, no content change.

CI

• 5d12bb4: pr-validation / test(22) / test(24) / build / build-artifact / Docker e2e — all ✅ (the e2e that was still running in my first review has passed).
• 6d94b7a: all six checks ✅ after my docs-only commit.

No open findings. Branch is ready — merge is a human-only operation per AGENTS.md §5.1.1.2: #409

中文摘要：复核了双 agent 审查后的分支，fallback 修复与全部测试改动均属实并本地验证（含"无 fallback 时新测试必失败"的回归有效性）；发现两个新 devlog 文件未过 Prettier（CI 不查格式所以未拦截），已在分支上直接补了格式化提交 6d94b7a，CI 全绿，可以合并。