hermes - 💡(How to fix) Fix Wake-injected user turns: drain unconsumed wake-inbox records on session activation (durability layer) [1 participants]

Summary

Hermes Agent's Smith Crafts active-session wake transport uses a durable wake-inbox before attempting socket delivery. The live socket path can enqueue a wake into the active CLI when a session is present, but accepted wakes that were not delivered because no active session was available need a native drain path on the next session start/activity.

Related local wake-socket work:

• #28570

Desired behavior

• On runtime/socket startup and interactive session start, scan the Stitch wake-inbox for records with consumed_at:null.
• Match records to the active wake session id.
• Sort by emitted_at, falling back to file mtime.
• Deduplicate by wake_id; suppress unconsumed duplicates when a consumed record with the same wake_id already exists.
• Enqueue through the same external-user-message helper as the socket fast path.
• Mark consumed_at only after the model-facing enqueue succeeds.
• If enqueue fails partway through a batch, leave the failed and remaining records unconsumed for the next attempt.

Why this matters

The socket path should remain the fast path, but the wake-inbox should be the durability layer. A transient no-active-session state should not require manual replay if the record was accepted and written durably.

Local implementation note

Smith Crafts is carrying a protected smith-crafts/live commit that adds drain_wake_inbox(cli) in tools/agent_wake_socket.py, calls it after socket bind and from cli.py session startup, and adds focused pytest coverage for success, dedupe-vs-live-delivery, success-only consumption, and partial-failure retry behavior.