fix(broker): agent-reply auto-route as true fallback with dedup (#280)

[olegbrok]

What

Make the #279 agent-to-agent reply auto-route a true fallback instead of an unconditional second delivery path, eliminating duplicate inbox entries when an agent replies explicitly.

Why

#279 delivers a completed agent-injected turn's final text to the requester's inbox. But agents (Murzik especially) also reply explicitly via send_to_agent → comms.send, which writes straight to the inbox. A turn that did both delivered the reply twice.

The dup hadn't surfaced in the wild yet only because no agent-injected turn had completed since #279 deployed (verified against the live conversations_agent_comms.db: the only auto_routed row in 3 days is a pushok self-test; every Murzik verdict came via explicit send_to_agent and predates the deploy). The next review verdict would have doubled.

How

• inject_agent_message stamps the turn boundary: (responder, requester) -> start_ts.
• route_agent_reply pops the marker and asks comms.has_message_since(responder, requester, since, exclude_auto_routed=True) — i.e. did the responder already send an explicit reply to the requester during this turn? If yes → suppress (the explicit copy is already in the inbox). If no → deliver (genuine fallback, e.g. an interrupted turn that produced text but never sent it).
• Backend-agnostic — no reliance on tool_uses (codex/tmux turns omit them), which is why detection is done against the comms DB, not the turn payload.
• Never content-matches — so it can't drop a genuinely new reply.
• Fails safe — a missing marker (lost across a restart) or a probe error falls through to delivery. A stray dup is acceptable; a dropped review verdict is not.
• Orphan markers swept by TTL (1h) on insert → bounded memory.
• Kill-switch PINKY_AGENT_REPLY_DEDUP (default on); off restores the Server-side presence tracking: stamp last_seen agent-agnostic #279 unconditional delivery. Counted via the existing broker deduped stat.

Design note for review

This uses a timestamp window, not content equality. Rationale: the auto-route is a safety net for agents that produce a reply but don't send it. If an agent demonstrably knows how to send (it sent explicitly to the requester this turn), the net is off for that turn — it should send any additional content explicitly too. Trade-off: if an agent sends an unrelated explicit message mid-turn (e.g. a clarifying question) and then relies on the auto-route for its actual answer, the answer is suppressed (recoverable — it's in the transcript and the requester notices the gap). Considered acceptable vs. content-matching, which would silently miss real dups whenever a codex turn's final text differs from the sent content.

Tests

tests/test_broker.py — suppress-on-explicit-send, deliver-on-no-send, deliver-on-missing-marker (safety), killswitch-off-always-delivers, inject-marks-turn-start.
tests/test_agent_comms.py — has_message_since incl. exclude-auto-routed, window boundary, legacy/empty-metadata-counts.

pytest tests/test_broker.py tests/test_agent_comms.py → 126 passed. ruff clean.

Deploy

Held — not for immediate deploy. Touches the live agent-reply path; wants Murzik review + a supervised daemon bounce (recent fan-out instability on the Mini). Inert behind the existing #279 routing being on; the dedup defaults on but only changes behavior on the explicit-send-this-turn path.

🤖 Opened by Barsik