Model output & the leaked-reasoning guard

The model answers naturally. Its reasoning streams on the gateway's native reasoning_content channel (lifted by graph.llm._ReasoningChatOpenAI, surfaced as reasoning frames by server.chat), and its answer is plain text/markdown. There is no <scratch_pad>/<output> text protocol — the model is never told to wrap its response in tags.

Server-side, graph/output_format.py::extract_output is now just a thin pass-through that strips any leaked reasoning before the answer reaches consumers (A2A artifacts, the OpenAI-compatible /v1 endpoint, the console, subagent return values).

What survives: the leaked-reasoning guard

strip_reasoning (and the extract_output that wraps it) removes balanced or truncated <think> / <scratch_pad> blocks from the answer content channel. Two reasons it still earns its keep:

• Provider leaks. Some gateway/model combos leak raw <think>…</think> into the content channel instead of reasoning_content — notably MiniMax via LiteLLM #22392. Model selection is gateway config, not code, so the runtime can't guarantee a clean channel; a cheap, idempotent strip at the boundary is the insurance.
• Storage guardrail (ADR 0021). Leaked reasoning must never persist to the knowledge base or session memory. strip_reasoning is the guard the memory/knowledge write paths call before storing assistant text.

The strip is backtick-guarded: an answer that mentions a tag in inline code (`<think>`) is left intact — only a genuine leak (an unbacktick-ed tag) is removed.

History: the retired <scratch_pad>/<output> protocol

Earlier versions instructed the model to wrap deliberation in <scratch_pad> and the final answer in <output>, and extract_output parsed the <output> block out. That made sense before models exposed a native reasoning channel — it let strong models "think out loud" without the reasoning reaching the user.

It was retired once native reasoning landed:

• #1328 dropped the protocol for the lead agent (native reasoning_content).
• A follow-up de-protocoled the subagents (they now answer naturally too).
• The parser retirement then deleted the <output>-extraction, the dropped-turn "kicker" retry, the streaming <output> views, and the <confidence> self-report — none of which had a live producer anymore — leaving only the leaked-reasoning guard above.

Native reasoning is strictly better here: the channel separation is structural (no tag spanning chunk boundaries, no unclosed <output>, no stray </output> in scratch), and the console renders the model's real thinking above the answer with zero parsing.

Related

• Architecture — where in the runtime this lives
• graph/output_format.py — the implementation (now ~65 lines)