Agent mailbox seam migration

NanoClaw now routes per-session mailbox access through matching semantic registries in the Node host and Bun runner. The open-source composition still selects SQLite, and existing inbound.db and outbound.db files remain valid.

Detect affected customizations

Search custom source and installed skills for removed raw session helpers and mailbox calls or public types whose shape changed:

rg -n "openInboundDb|withInboundDb|openOutboundDb|openOutboundDbRw" src .claude/skills container/agent-runner/src
rg -n "writeSessionMessage|writeSessionRouting|writeOutboundDirect|writeMessageOut|createScheduledTask|restartAgentGroupContainers|createAgentFromTemplate" src .claude/skills container/agent-runner/src
rg -n "setContainerToolInFlight|clearContainerToolInFlight|clearStaleProcessingAcks|touchHeartbeat|DeliveryActionHandler|PostDeliveryHook" src .claude/skills container/agent-runner/src
rg -n "(trigger|onWake)\\s*:\\s*[01]\\b|kind\\s*:\\s*string" src .claude/skills container/agent-runner/src

Imports of the raw helpers from src/session-manager.ts require migration. Review custom calls from the second search and ensure their enclosing function awaits the result. Built-in matches are already migrated.

The moved and narrowed symbols map old to new as follows:

Why the interface changed

Session callers previously opened SQLite handles directly, making SQLite's schema, paths, and lifecycle part of every caller. The mailbox seam gives the selected implementation ownership of opening, refreshing, flushing, and closing session storage. That ownership requires an asynchronous operation boundary even though the default remains SQLite.

Two rules follow from that ownership:

• Reads never provision. AgentMailbox.exists() is a side-effect-free probe. session() operates only on storage already authorized by prepare(); it never creates storage by itself. Core read paths use withExistingMailboxSession() and treat a missing mailbox as empty.
• Never nest same-key sessions. An implementation may serialize session() per key while loading and committing its state. Opening another session() for the same session from inside an action (including via helpers like writeSessionMessage) can deadlock the implementation. withMailboxSession always throws on same-key nesting. Finish the open session, then call the helper.
• Sync methods are snapshot-scoped. The synchronous MailboxSession methods operate on implementation-managed session state; their effects must be durable only once session() resolves. The asynchronous writes (insertMessage, insertTask, writeDirect) must be durable when their own promise resolves — the host wakes containers on the strength of them.

Update custom host modules

Replace raw database access with the semantic operation that expresses the module's intent:

// Before: SQLite-specific.
const db = openInboundDb(agentGroupId, sessionId);
try {
  // query or update the session mailbox
} finally {
  db.close();
}

// After: works with the registered mailbox.
await withMailboxSession(agentGroupId, sessionId, (mailbox) => {
  return mailbox.countDueMessages();
});

Import withMailboxSession from src/session-manager.ts. Available semantic operations are declared in src/mailbox/types.ts. If custom SQL performs an operation absent from that contract, keep the customization explicitly SQLite-only by importing the low-level opener from src/mailbox/sqlite/session-db.ts, or add the smallest semantic operation required by every selected mailbox.

Add await to mailbox writes and propagate async through custom callers:

await writeSessionMessage(agentGroupId, sessionId, message);
await writeSessionRouting(agentGroupId, sessionId);
await writeOutboundDirect(agentGroupId, sessionId, message);

In the runner, continue using the compatibility modules under container/agent-runner/src/db/, but await writeMessageOut(). New code may use getAgentMailbox().operations directly. getAgentMailbox().run() scopes a logical unit of mailbox work (the MCP server wraps each tool call in it); note that the built-in poll loop and compatibility modules call operations outside run(), so a replacement mailbox must keep operations functional without an enclosing run() — treat run() as an optional scoping hint, not a gate.

AgentMailbox.start() may receive null only during the brief upgrade window where the shared runner source updates before the host restarts and writes its session context. SQLite accepts that legacy sentinel; an implementation that requires context should reject null explicitly.

Compose one mailbox implementation

There is one composition story in both runtimes: the real module barrel imports mailbox/compose.ts, and a capability skill replaces the registration inside that file. Keep the barrel and entrypoint imports unchanged. Never append a second mailbox registration import—the registry intentionally rejects it.

Keep the canonical model storage-neutral

src/mailbox/model.ts defines the exact records exchanged with mailbox implementations. Its values are JSON-native: camel-case fields, booleans, UTC ISO-8601 timestamps, and nonnegative safe integers. A storage implementation must translate its native representation at its boundary—for example SQLite 0/1 flags and snake-case columns—and validate the translated record before returning it.

Canonical records are intentionally flat: fields contain JSON primitives or null, while kind-specific structured payloads remain encoded in content. The shared parser rejects nested objects and arrays so a future field cannot silently bypass exact-field validation.

Serialization metadata belongs to the implementation that needs it and is not part of the canonical mailbox record. Each implementation should test its contract lifecycle alongside representation-specific round trips so its translations cannot silently drift.

Verify the migration

Build both runtimes and run their tests:

pnpm run build
pnpm test
pnpm exec tsc -p container/agent-runner/tsconfig.json --noEmit
cd container/agent-runner && bun test

With the default composition, send a message through a real channel and verify that the existing session's inbound.db receives it and outbound.db receives the reply. Re-run the searches above and confirm every affected custom write is awaited or deliberately SQLite-only.

Roll back

No stored-data migration occurs. Return NanoClaw and custom modules to the previous revision, rebuild the host and agent image, and restart the service. The same SQLite session files remain usable after rollback.