Evaluate CC 2.1.224 native cross-session messaging (SendMessage/ListAgents) as an inter-agent transport supplement

[olegbrok]

What shipped

Claude Code 2.1.224 (2026-08-07, already live on the Mini via last night's auto-update) added native cross-session messaging: running CC sessions can discover each other with ListAgents and message each other with SendMessage (macOS + Linux).

Why PinkyBot should evaluate it

Every inter-agent message today rides: daemon broker → tmux queue → pane paste → submit-verify. That chain is our top reliability sink — this morning alone produced the #991 pinned-head specimen (queue depth 32, schedule wakes starved 4h behind a phantom head turn) and the Pi's vanish-pre-delivery wakes; historically V17's silent worker death ("command too long"), #953's Enter loss, #1006's sticky metas. A first-party harness-level channel between sessions could bypass the pane-paste layer entirely for agent→agent traffic — and possibly for wake delivery.

Spike scope (feasibility, NOT a rewire commitment)

1. Discovery model — do daemon-spawned tmux CC sessions see each other in ListAgents? What identity/naming do they get? Cross-user/cross-box behavior?
2. Isolation audit (hard gate) — feat: Codex CLI as first-class agent provider #149 isolated=1 tenants must NOT be able to message or even enumerate other agents' sessions. If ListAgents has no scoping control, that alone confines this to non-isolated fleets or kills it.
3. Delivery semantics — does SendMessage inject mid-turn or queue-to-idle? Receipts/acks? What happens when the target session is busy/wedged/dead? (Exactly the failure modes Scheduler: proven-live outbox drain replay-storms — no attempt cap + FIFO head-of-line break re-pastes a never-confirming wake every tick and immortalizes zombie rows (live incident: 7+ dup deliveries/hr on Pi) #991 tracks — if SendMessage has better semantics, that's the win; if it has the same, document and stop.)
4. Fit points, cheapest first — (a) agent→agent chat (today's send_to_agent), (b) daemon→agent nudge ("check your queue") as a redundant wake channel, (c) full wake delivery (unlikely — needs receipts we control).
5. Version gating — fleet runs mixed CC versions; feature detection needed.

Sources

• CC 2.1.224 changelog (github.com/anthropics/claude-code/releases)

🤖 Opened by Barsik (#600 daily research)