fix(cloud-agent-next): reconcile completed kilo turns on wrapper death

[eshurakov]

Problem

When the wrapper OOM-dies after emitting wrapper_finalizing but before sending complete, the DO liveness watchdog (handleUnhealthyWrapper via wrapper_ping_timeout) and the disconnect-grace expiry terminalize all accepted messages as wrapper_failure — even though the turn actually completed. Incident: 8 sessions posted their work but were reported as failed.

Why the DO can settle this itself

Ingest is a single FIFO channel, and the wrapper only emits wrapper_finalizing after message.updated(time.completed) and session.idle have already been sent. Kilo events with an entity id are upserted last-wins into the DO's events table, so receiving wrapper_finalizing implies the DO already holds the terminal assistant state. (Verified via deadline arithmetic on the incident: 7 of 8 sessions provably reached wrapper_finalizing.)

Change

Both wrapper-death paths (ping/no-output deadline, disconnect-grace expiry) now reconcile each accepted message against the DO's stored kilocode events before failing:

• assistant info.error → failed/assistant_error with the existing classifyAssistantFailure classification (extracted into a shared helper also used by settleSealedBatch)
• assistant info.time.completed present → completed with the assistant message id, completionSource: 'idle_reconciliation'
• no positive terminal evidence → previous wrapper_failure / wrapper_disconnected behavior, unchanged

Unlike the sealed-batch path, bare assistant-message presence is not sufficient here — no wrapper complete vouches for turn finality, so the same completion marker the wrapper itself uses (time.completed or terminal error) is required. Emits a warn log with reconciled/fallback counts for observability.

Residual gap: if the final message.updated never reached ingest at all, the store has no evidence and the outcome is wrapper_failure as today; closing that needs a live Kilo terminal query or an HTTP terminal assertion (separate work).

Verification

• 4 new unit tests: ping-timeout reconcile to completed (incl. callback fired), no-completion-marker fallback, assistant-error reconcile (not masked as wrapper failure), disconnect-grace reconcile
• Full cloud-agent-next unit suite: 2451 passed; death-path integration tests in Workers runtime: 30 passed
• tsgo --noEmit, oxlint, oxfmt clean
• services/cloud-agent-next/AGENTS.md invariant bullet updated

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (3 files)

• services/cloud-agent-next/AGENTS.md
• services/cloud-agent-next/src/session/wrapper-supervisor.ts
• services/cloud-agent-next/src/session/wrapper-supervisor.test.ts

Previous Review Summary (commit 114b3f5)

Current summary above is authoritative. Previous snapshots are kept for context only.

Previous review (commit 114b3f5)

Status: No Issues Found | Recommendation: Merge

Files Reviewed (3 files)

• services/cloud-agent-next/AGENTS.md
• services/cloud-agent-next/src/session/wrapper-supervisor.ts
• services/cloud-agent-next/src/session/wrapper-supervisor.test.ts

---

_{Reviewed by kimi-k3 · Input: 67K · Output: 6.4K · Cached: 528.4K}

_{Review guidance: REVIEW.md from base branch main}

[eshurakov]

Ran manual generic cloud agent verification