Advisory @WRN on stderr is easy for LLM tool loops to ignore

[chouswei]

Summary

Operational signals (stale_in_store, near_cap*, ttl_expiring, mission_settled) are emitted on stderr as @WRN lines. MCP folds them into the JSON envelope, but agents typically reason over stdout wire only.

Evidence

• Emission: src/memnet/output.py; documented in README.md / LLM-GUIDE.md ("Warnings live on stderr — you must read them").
• MCP: MemNetResponse carries stderr plus errors[] (errors from @ERR); warnings are not first-class structured fields.
• Domain pipelines often call run_memnet and parse stdout (e.g. warm supplements) without branching on warnings.

Impact

Sessions expire, caps fill, and stale rows accumulate without the agent adjusting behaviour — exactly the failure mode the goldfish loop tries to prevent.

Suggested direction

Promote top warnings into structured MCP fields (warnings: [{code, msg}]) and/or echo critical warnings into warm stdout as @WRN rows when serving agents.

[chouswei]

Still valid: advisory @WRN remains on stderr; MCP JSON has stderr / errors but no first-class warnings[]. Unchanged by recent refactor.

[chouswei]

Refresh (v0.3.6): Serve protocol now returns framed errors for oversize/bad frames (3c537c7); advisory @WRN on stderr for session caps/TTL remains as described — still no structured warnings[] in MCP.

[chouswei]

Status (Hatch 0.19.1). @WRN on stderr / MCP envelope stderr is unchanged. Agents still typically reason over stdout wire. Leftover @TAG pipe is not agent teach (GQL only).

Structured warnings: [{code, msg}] on MCP remains Later. Engine path is parts/common/memnet/, not src/memnet/.