[nwparker]

	Files	Added	Deleted	Net
Test	1	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​24	0	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​24
Prod	3	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​12	$\color{#cf222e}{\Huge{\mathbf{−}}}$​24	$\color{#cf222e}{\Huge{\mathbf{−}}}$​12

ELI5

OpenCode 2 history survives a brief database write. When the database stays busy, Vault explains that the database could not be read and retries it on refresh.

What Changed

OpenCode 2 listing and preview reads now use the existing read-only database opener: a 1500 ms busy timeout for native paths and zero for WSL network paths. Errors carry the OpenCode 2 agent and database scope so the existing desktop and mobile notice can describe the unavailable history accurately.

Why

This reuses the policy already used by OpenCode 1 rather than giving the two readers different lock behavior. It tolerates short writes while keeping network paths from waiting on locks they cannot resolve locally.

Linked Issue

Continues the database-opening intent of #15671 and the focused OpenCode Vault work from #10708. This change depends on the native OpenCode core in #24647 and does not settle either original PR's full scope.

Visual Proof

Actual production worker reads use installed OpenCode 2.0.16 history copied into an isolated database. A controlled SQLite writer establishes contention; it is not a currently running CLI writer.

The hidden production encrypted mobile History route at390×812 uses the same genuine retained CLI session: before, V2 appears as a skipped transcript with a raw lock error; after, both OpenCode readers return typed source-writing notices, with zero skipped transcripts. Release and Refresh recover the same session (1→0→1). No synthetic notice amplification was used. The retained before already contains #24772's viewport fix; the after rebuild includes its equivalent copy extraction and this scanner correction. This is macOS rendered web-route proof, not physical phone coverage.

Actual writer lock	Before	After
Releases after 300 ms	No rows; immediate lock error	One genuine CLI-created row; no issue
Remains exclusive	Immediate generic lock error	Typed OpenCode 2 database scope after about 1.6 seconds
Writer released	One row	One row

Testing

• I manually tested these changes locally
• Automated tests added/updated

ORCA_BACKGROUND_LAUNCH=1 pnpm test src/main/ai-vault/session-scanner-opencode-sqlite-open.test.ts src/main/ai-vault/session-scanner-opencode2-sqlite.test.ts src/main/ai-vault/session-scanner-opencode-sqlite-paths.test.ts passed 39 distinct tests (24 opener/v2 and 15 related SQLite tests). Native macOS production worker reads covered both real lock cases and recovery; the zero-timeout WSL path is covered by tests, with physical Windows/WSL validation still required.

Review

Only four files change: the canonical opener, its tests, and the two OpenCode 2 readers. The opener preserves read-only access, existing error handling, and the native/WSL boundary.

Agent skill upstream boundary

• Not applicable; no upstream skill implementation is copied.

Notes

Credit @AmethystLiang for the original OpenCode AI Vault work in #10708 and Brennan Benson for the original #15671 database-opening intent. Cache bounds and scan ownership are separate focused changes.

Checklist

• This PR is small and focused
• I explained the user-facing change, mechanism, and reason for the approach
• Backend before/after behavior is recorded; rendered notice captures accompany the native validation
• Self-reviewed for correctness, security, and performance
• Cross-platform, SSH/remote, and compatibility impact considered
• Full lint and packaging build are left to CI; focused tests, typecheck, and changed-file gate pass

[pullfrog]

✅ No new issues found.

Reviewed changes

• Canonical opener for OpenCode 2 reads — both the v2 list leg and the v2 preview parse now open opencode-next.db through openOpenCodeDatabaseReadonly, dropping their private openReadonlyDatabase (which passed no timeout). They inherit the 1500 ms native / 0 ms WSL busy-timeout split already used by OpenCode 1, and the opener's close-on-pragma-failure guard.
• Typed issue for the v2 list leg — the catch now calls openCodeDatabaseScanIssue(dbPath, err, 'opencode2'), emitting kind: 'scope' and the OpenCode 2 is writing to … copy. Previously it pushed an unkinded issue, which the panel counts as a skipped transcript.
• openCodeDatabaseScanIssue accepts 'opencode2' — adds the agent to its union and the OpenCode 2 label; 'opencode'/'zcode' paths untouched.
• New test — covers a brief v2 lock (recovers, no issues) and a persistent lock (waits inside the busy timeout, returns [], emits the scope issue).

I verified the focused tests (session-scanner-opencode-sqlite-open.test.ts, session-scanner-opencode2-sqlite.test.ts — 24 passed), oxlint on the four changed files, and pnpm tc:node; all clean. The change is a faithful reuse of the existing v1 policy, and the new test genuinely fails against the pre-change default-zero timeout, so it pins the behavior rather than the current state.

  ^{｜ View workflow run ｜ Using deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏}