Multi-agent direct edges send the successor a trailing-assistant prefill on Claude — empty or continuation-shaped successor turns #345

Symptom

In a multi-agent direct-edge flow (writer → critic) on Anthropic or Bedrock-Claude, the successor agent's turn can be empty, and even when non-empty it is a prefill continuation of the previous agent's output rather than a response in the successor's own voice.

Reproduced live 3/3 (empty critic) with the scenario runner shipped in #335:

npx tsx --env-file=.env src/scripts/preempt-scenarios.ts --scenario ma-continue --provider anthropic
# baseline (no preemption) for contrast — note the critic's output BEGINS mid-document:
npx tsx --env-file=.env src/scripts/preempt-scenarios.ts --scenario ma-baseline --provider anthropic

Mechanism

The successor's payload ends with the previous agent's assistant turn. Claude treats a trailing assistant message as a prefill, so the successor model continues it. When the trailing turn reads complete — e.g. a preemption steer's short resume ("Constantinople fell in 1453.") — the continuation has nothing to add and returns empty content. OpenAI is unaffected (no prefill semantics); the no-preemption baseline's critic output is itself continuation-shaped (\n\n**Critique:**…), showing the mechanism predates preemption — #335 only makes the trigger shape (short, complete trailing turns) more common.

Fix direction

When a direct-edge successor's payload would end assistant-role, frame the handoff as a user turn — the chain-prompt wrapper that promptInstructions edges already get is most of the machinery (MultiAgentGraph wrapper nodes, CHAIN_PROMPT_PREFIX). Options:

1. Apply chain-prompt framing by default for direct edges (behavior change for all providers — cleanest semantics).
2. Insert a minimal user-turn cue only for prefill-semantics providers (Anthropic, Bedrock-Claude) when the payload ends assistant-role (narrowest change).
3. Document that direct edges targeting Claude should carry promptInstructions (no code change).

Deliberately not fixed in #335: it is a pre-existing multi-agent framing behavior, and changing it affects existing flows that may rely on continuation semantics — it deserves its own decision and review.

Context: #335's PR description ("Known interaction" section) and the ma-continue/ma-baseline scenario replies there.