Summary

The Newgate-style memory path around the IDE companion currently has enough pieces to look plausible in code, but the live execution path is not proven end to end.

This issue tracks an execution-first investigation of the following path:

GH Copilot CLI gpt-5-mini worker -> automatic fleet log -> KI queue -> promoted knowledge -> indexed memory -> recall injection

Verified current state

• The bridge can be running while the GH Copilot CLI gpt-5-mini worker is not actually active.
• fleet_log is only written when it is called explicitly.
• KI promotion currently stops at file creation under knowledge/.
• Recall exists as a memory/tool surface, but it is not injected into the worker prompt path.
• The current Pipeline① packetizer is JavaScript-specific, so it cannot directly audit the Python/shell hook path.

Investigation goals

1. Prove worker boot with a real gpt-5-mini run.
2. Prove worker completion automatically emits structured fleet logs.
3. Prove success auto-generates a KI queue candidate.
4. Prove promotion auto-indexes promoted knowledge.
5. Prove recall is injected back into the next worker prompt.
6. Prove operator-visible status surfaces show the active stage and failures.

Proposed artifact

A design note for this investigation has been prepared and will be linked from the PR that addresses this issue.