fix(runtime-host): surface root cause for session revision commit failure instead of generic message

[me2seeks]

What happened

Rewinding a TUI session via Esc Esc -> Rewind shows a transient Note: 正在回退到该轮之前… and then fails with the generic:

Error: Session conversation copy could not be committed

The message is produced by the single empty catch in packages/runtime-host/src/server/session-revision-coordinator.ts:540 / dist/.../session-revision-coordinator.js:382:

try {
  await artifacts.copyConversationArtifacts(...)
  await cloneConversationRuntimeLedger(...)
  await taskLedger.copyConversationTaskLedger(...)
  await sessionStore.appendMessages(...)
  await sessionStore.updateHeader({state:'committed'})
  await continuity.refreshCanonical(...)
} catch {
  return rollbackIncompleteCopy('Session conversation copy could not be committed')
}

The original exception (e.stack / e.code) is discarded. The TUI (packages/cli/src/pi-tui-runner.ts:1768 + runtime-host-session-driver.ts:684 session.revision.create) only sees persistence_failed and cannot distinguish:

• SQLITE_BUSY / wal contention during appendMessages
• archived tool result hash mismatch after copyConversationArtifacts
• owned references inside archived tool results / continuation authority rewrite failures
• SQLITE_FULL / permission errors

from the expected session_busy: Source Session has an active Turn that is correctly returned in the validation phase (coordinator.ts:355).

As a result, en already reproduces at the default panel measure min(680px, calc(100% - 32px)) without any narrowing — two graph_record_* ids are sufficient at 680px. zh (one character per line) is the louder symptom, but the root shrink is the same. Captured at both 430px and 980px in #3970 for the related UI issue, but the Host-side swallowing is independent and affects all locales (UI_LOCALES = ['zh','en']).

How to reproduce

1. Start TUI maka on main (07e40fe, Linux, Node 26.3.0) with workspace ~/.config/Maka/workspaces/default.
2. Use a session that has at least two user turns and at least one large archived tool result (e.g., e2dbda6e-336b-... with turns c64163d9 and 1252e2dd / graph_record_* artifacts, or any session with Agent Graph 8/8 operators).
3. Ensure the session is idle (session.catalog.query kind:get shows status: idle and empty liveRunState.runningTurnIds). If running, the expected session_busy is returned — this is not the bug.
4. In TUI, press Esc Esc, select any earlier turn (e.g., 1252e2dd before 创建issue).
5. Observe the transient 正在回退... notice, then Error: Session conversation copy could not be committed with no further details. With a patched Host that logs catch(e), /tmp/rewind-error.log shows the real e.stack (e.g., SQLITE_BUSY or hash mismatch).

Stable reproduction on main when the source session is idle but its copyTurnIds contain large archives or taskLedger contention. A fresh single-turn session with no archives rewinds successfully, proving the failure is data-dependent, not environment.

Environment

• Maka version or commit: main @ 07e40fe (2026-08-27)
• OS and version: Linux
• Surface: TUI (maka CLI) + Runtime Host
• Node.js version: 26.3.0
• Workspace: ~/.config/Maka/workspaces/default (runtime.sqlite 1.1G, runtime.sqlite-wal 17M)

Logs, screenshots, or additional context

• TUI screen: Note: 正在回退到该轮之前… followed by Error: Session conversation copy could not be committed (from pi-tui-runner.ts: rewindToTurn).
• Direct Host call also reproduces:

await conn.request('session.revision.create', {
  sourceSessionId: 'e2dbda6e-336b-4d65-8e7f-13453b106339',
  targetSessionId: 'test-rewind-...',
  sourceTurnId: '1252e2dd-d065-4784-9573-e07579c8efba',
  expectedSourceRevision: 98
})
// -> RuntimeHostOperationError: Session conversation copy could not be committed
// with swallowed cause

• Suggested fix: change catch { to catch (e) { logger.error({err:e, sourceSessionId, targetSessionId}, 'revision copy commit failed'); return rollbackIncompleteCopy('...: ' + e.message) } and split the 7 steps into distinct error codes (artifact_copy_failed, runtime_ledger_failed, task_ledger_failed) so the TUI can surface the specific reason. Keep session_busy as the early validation path.
• Related issues: fix(desktop): prevent Agent Graph Selected results label from shrinking to min-content #3966 (UI shrink, same Graph data) and fix(desktop): keep Agent Graph results label from shrinking #3970 (UI fix). This Host issue is independent — same data triggers both, but the swallowing is in session-revision-coordinator.
• Label suggestion: bug

[fxl112233]

I'd like to work on this issue.

Plan:

• preserve the existing persistence_failed / rollback semantics;
• retain a bounded, redacted diagnostic from the caught commit error so the client can distinguish failures such as SQLite contention and archive validation errors;
• keep commit-outcome-unknown handling unchanged;
• add focused Runtime Host regression coverage and run the package checks.

Automated disclosure: this claim was drafted and posted by OpenAI Codex at @fxl112233's direction. @fxl112233 remains the human contributor of record.