fix: migrate legacy 4-digit CompressionBlock boundary refs at load and fork transfer

[ranxianglei]

Addresses the applicable item of #415 (R4: "Legacy fork state normalizes parent message-ref maps but not four-digit CompressionBlock.startId/endId boundaries"). The other nine items of #415 were triaged as not applicable to this V1-only repo — see the issue comment with per-item evidence.

Problem

Pre-1.1.0 persisted states carry 4-digit message refs (m0001). At session load, lib/state/state.ts migrates the messageIds ref maps to canonical 5-digit form (m00001) and re-saves — but CompressionBlock.startId/endId boundaries are never migrated:

1. Own-session load: loadPruneMessagesState passes boundaries through verbatim, so a legacy session's saved file keeps mixed-width refs permanently.
2. Fork transfer: restoreForkCompressionState reads the parent's raw persisted record (loadSessionState does not normalize blocks) and spreads ...block, copying 4-digit boundaries into otherwise canonical fork state.

Impact is low severity (no functional breakage — boundary resolution uses raw IDs + blockId, never byRef), but concrete: hideConsumedCompressCalls range-key dedup (lib/compress/hide-consumed.ts:84) stops matching consumed compress calls for legacy blocks → minor context duplication; export/recap display carries inconsistent widths.

Fix

• lib/message-ids.ts: new exported migrateMessageRef(ref) — 4→5 digit via existing parseMessageRef/formatMessageRef; non-message values (empty strings, bN block refs, free text, out-of-range indices) pass through unchanged.
• lib/state/utils.ts: loadPruneMessagesState migrates block startId/endId.
• lib/state/rebuild.ts: restoreForkCompressionState normalizes translated block boundaries.
• Self-healing: after one load+save cycle under new code, files are canonical on disk.

Tests

tests/legacy-block-boundary-migration.test.ts — 5 tests:

• migrateMessageRef unit coverage incl. edge cases (m00000, 6-digit, b3, whitespace/case)
• pass-through guard for non-message values (green before fix by design)
• own-load migration, fork-transfer migration, full ensureSessionInitialized integration with self-healing re-save assertion

Red/green proof (§5.7.3): behavioral tests written first fail with actual: 'm0001' vs expected: 'm00001' at all three production paths; all green after the fix.

Verification

• typecheck ✓, build ✓, prettier clean on touched files ✓
• Full suite: 1272 tests, 1270 pass, 2 fail — both failures reproduce identically on clean master 06efd39 (sandbox /tmp is read-only: soft-block.test.ts EACCES mkdir, inactive-block-decompress.test.ts toFile path guard); pre-existing environment incompatibility, unrelated to this change.
• No version bump (non-release branch).

Related: #415 (item R4); related code path introduced alongside #409.

[github-actions]

📦 Built Plugin Artifact

Branch: 2026-09-17_legacy-block-ref-migration (2f54a14)

Option A — Install from npm PR tag (recommended)

opencode plugin opencode-acp@pr-416 --global

Each push to this PR publishes a new version under the pr-416 npm tag.

Option B — Install from GitHub

opencode plugin "github:ranxianglei/opencode-acp#2026-09-17_legacy-block-ref-migration" --global

Option C — Download artifact

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

tar xzf opencode-acp-pr416.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] 🏷 Picking this up. This issue tracks the already-open PR #416 (head 2026-09-17_legacy-block-ref-migration → master); the fix is implemented and pushed as commit 1a68306. Per our PR-handling rules I'm not re-doing the work — I'm independently verifying the branch against the description and checking CI, then I'll post a concrete conclusion. Initial CI snapshot: pr-validation ✅, build ✅, test (22) ✅, test (24) ✅, build-artifact ✅, e2e still running.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Verification complete — PR #416 (1a68306, 2026-09-17_legacy-block-ref-migration → master) checked out fresh and independently verified against the description. Conclusion: code is correct, CI fully green, ready for the second independent review + human merge.

Code review (per file)

• lib/message-ids.ts — new export migrateMessageRef(ref): 4→5 digit via existing parseMessageRef/formatMessageRef; non-message values pass through unchanged. Edge cases verified against the actual regex/range (/^m(\d{4,5})$/, index ∈ [1, 99999], width-5 padding): m0001→m00001, m00999→m00999, 5-digit idempotent, m00000 / m123456 / b3 / empty / free text pass through, whitespace + case handled by parseMessageRef.
• lib/state/utils.ts — loadPruneMessagesState migrates block startId/endId, preserving the existing typeof === "string" guards.
• lib/state/rebuild.ts — restoreForkCompressionState overrides the spread startId/endId with migrated values (later properties win over ...block). Non-blocking nit: no typeof guard here (unlike utils.ts) — acceptable since input is typed PersistedSessionState and sibling fields (anchorMessageId, …) rely on the same shape trust; flagging only for the record.
• Impact mechanism confirmed as described: hideConsumedCompressCalls builds live range keys from block boundaries (rangeKey(block.startId, block.endId)), so mixed-width refs break dedup matching → consumed calls stay visible. Low severity assessment holds: boundary resolution itself never goes through byRef.

Independent local runs (fresh checkout of 1a68306)

• npm run typecheck ✓
• Targeted: 5/5 in tests/legacy-block-boundary-migration.test.ts pass
• Full suite: 1272 tests, 1270 pass, 2 fail — exactly matching the PR body. Both failures reproduce identically on clean master 06efd39:
  • soft-block.test.ts: EACCES: permission denied, mkdir '/tmp/opencode-dcp-dangerous-*'
  • inactive-block-decompress.test.ts: toFile guard rejects the hardcoded /tmp/test-inactive-block-decompress.txt path
  • Both are pre-existing environment incompatibilities (tests hardcode /tmp paths not writable in this sandbox), unrelated to this change.
• Test file meets §5.6: imports real sources (no reimplementation), complete config factory incl. gc, behavioral red/green paths through all three production sites, self-healing test reads persisted JSON back after init.

CI & process

• All 6 checks on head 1a68306 SUCCESS: pr-validation, test (22), test (24), build, build-artifact, e2e. mergeable_state: clean.
• Devlog REQ.md/WORKLOG.md present and compliant (folder = branch name, refs fix: close final OpenCode V2 lifecycle, projection, and artifact verification gaps #415-R4/fix: reconcile compaction restart and custom-storage fork recovery #409, red/green proof recorded). No version bump on a non-release branch ✓.

Remaining before merge: dual-agent review per §5.3 (code) + §5.6 (tests) — this verification counts as one agent's pass; one more independent agent review is required per policy. Then human merge (merges are human-only).

中文摘要：已独立核验 PR #416 —— 代码逐文件正确（含边界用例）、本地全量 1272 测试中 2 个失败在干净 master 06efd39 上同样复现（沙箱 /tmp 只读的预存环境问题，与本改动无关）、CI 六项全绿、devlog 齐全、未动版本号；代码可合并，尚差第二个独立 agent review，合并由人工执行。

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 ## Dual-agent review record (§5.3 code + §5.6 tests)

Two independent reviews were run against the original commit (1a68306):

Code review: Request changes → fixed in c2651ca

• MAJOR: block-side-only migration would regress hideConsumedCompressCalls batch filtering for legacy pre-1.1.0 sessions — pre-fix both sides of the range key were consistently 4-digit; immutable historical tool inputs stay 4-digit forever, so a 5-digit-migrated block key would never match and consumed batch-mate summaries would leak back into context. Fixed by width-normalizing both sides of rangeKey via migrateMessageRef, plus non-string pass-through for legacy records that omit boundaries (first fix attempt crashed 3 existing test files with TypeError on undefined.trim() — caught by the suite, guarded).
• MINOR: missing typeof === "string" guard in restoreForkCompressionState (parent record is raw persisted JSON) — added.
• NITs: JSDoc now documents the inherited trim/lowercase matching semantics of migrateMessageRef; REQ.md impact analysis corrected (pre-fix consumer was width-matched — the risk was introduced by the migration itself, not present before).

Test review: Approve

• All 6 new tests verified against source; red/green independently reproduced in an isolated checkout.
• Minor, accepted as follow-up (not fixed here): buildConfig() factory omits optional PluginConfig fields — copied verbatim from existing tests/rebuild.test.ts; test files are excluded from tsconfig typecheck and gc is present.

Post-fix verification: typecheck + build pass; full suite 1273 tests / 1271 pass / 2 fail — the 2 failures reproduce identically on clean master (sandbox environment incompatibility, unrelated to this change). Prettier clean on all touched lines (two files carry pre-existing master drift on unrelated lines, left untouched to keep the diff minimal).