[P1] Preserve line-start A2A handoffs when the target already has active work

[mindfn]

Problem

A valid line-start @target can be persisted but never invoke its target when that cat already has an active or queued agent entry. This drops an ordinary A2A handoff rather than retaining it until the target slot is free.

Observed in thread thread_mruayc4owlyzazbx:

• Fable posted a valid line-start @sol with the Q3/Q4 handoff (0001786435553718-000283-0311fbcd).
• Sol was completing an earlier turn in the same timing window; the first response did not consume Q3/Q4.
• A later manual @sol (0001786436054020-000300-40672525) did invoke Sol, which then answered Q3/Q4. This rules out handle parsing, availability, and provider failure.

Root cause

resolveInlineCat() returns skip: dedup_active whenever hasActiveAgent(cat) is true. routeSerial() then immediately continues, so the ordinary text-derived handoff never reaches claimOrDeferA2ATarget(), the lower path that is designed to durably enqueue an occupied target.

Relevant paths:

• packages/api/src/domains/cats/services/agents/routing/routing-decision.ts
• packages/api/src/domains/cats/services/agents/routing/route-serial.ts
• packages/api/src/domains/cats/services/agents/invocation/QueueProcessor.ts

The present unit test explicitly locks in the losing behavior: routing-decision.test.js expects skip: dedup_active for a processing target.

Expected behavior

A busy target must prevent a concurrent invocation, not erase the handoff. Preserve one durable A2A carrier using the original trigger message as a2aTriggerMessageId; when the active slot is released, invoke the target exactly once. Repeated identical handoffs may coalesce, but must remain observable and must not silently disappear.

Acceptance criteria

• A line-start mention to a target with an active/queued agent entry produces a durable queued/coalesced A2A carrier instead of skip: dedup_active.
• The carrier retains the originating message ID and target, survives target-slot release, and invokes the target exactly once.
• No concurrent duplicate target invocation is started.
• A regression test exercises the real queue-backed route path (not only routeSerial with the active-queue hook omitted).
• F167 custody/diagnostic projections observe a real queued handoff rather than a text-only false success.

Relation to #1312

Related to #1312 (F254 message delivery attempts), but not part of its scope. #1312 begins after a durable queue carrier exists and provides per-target failure/retry custody. This defect occurs before carrier admission, so there is nothing for #1312 to retry. Keep the fixes separate; both may share queue/custody primitives after the admission contract is corrected.

[zts212653]

Thanks — this is a valid P1 coordination bug, and the proposed boundary is the right one: a busy target should prevent concurrent execution, not erase an actionable line-start handoff.

I rechecked current main. The line-start A2A behavior is already part of the product contract, while the current routing path can still return dedup_active before the carrier reaches the durable claim/defer path. That makes this a bounded F167/F055 admission repair, separate from #1312's later per-target delivery-attempt custody.

We're marking this bug + triaged + accepted. The main implementation guard is to reuse the existing queue/custody primitives and prove exactly-once behavior on the real queue-backed path — no second queue or parallel lifecycle store.

Do you plan to send the PR for this one? If yes, we'll preserve your implementation custody and review it. If not, say so and we can route an internal fix.

One publication-hygiene request: the private-instance thread/message coordinates in the report are not verifiable from the public repository. Please redact them or replace them with a public/reproducible evidence description.

[小太阳·砚砚/GPT-5.6 Sol🐾]