feat: --extra-body, a request field the server wants and quackd never sends

[rokbenko]

Closes #12.

Every request an OpenAI-compatible provider sends is built from a fixed set of keys, so a field quackd has no name for cannot be sent at all. @Vallhalen hit that with Qwen3 on vLLM: thinking is on unless the request body says otherwise, the switch is a chat template argument rather than a sampling parameter, and one step of find-and-kick spent 150 s and 1717 output tokens deliberating before a decision that was correct anyway.

What it does

QUACKD_EXTRA_BODY and --extra-body take one JSON object and hand it to the SDK's own extra_body, which merges it into the top of the request. A passthrough, not a new code path.

• Nine of the eleven cloud vendors and all five local presets. Anthropic and Gemini have other SDKs, and ignore it as they already ignore --base-url.
• Both APIs. step() moves a run from Chat Completions to Responses when the API asks it to, and a knob that quietly stops halfway through a run is worse than one the server ignores.
• The flag beats the variable. An empty object sends nothing, which silences a .env line for one run.
• Six keys refused: model, messages, input, instructions, tools, stream. instructions is the one that is easy to miss: on Chat Completions the system prompt is the first of messages, but on Responses it is a field of its own, so leaving it out would have let a passthrough replace the whole contract on one API and not the other. Everything else overrides what quackd sends, tool_choice included, because that is the point.
• A bad value stops the command before anything connects, and before the branch that would have ignored it, so a typo is refused the same way whichever provider was named. That also makes --provider fake the cheapest way to find out whether a shell mangled the quoting.
• run_start records it. A run whose model was told not to think reads nothing like one that was.

Two things found by running rather than reading

No single spelling of a JSON string survives bash, PowerShell 5.1 and cmd.exe, so the docs show all three and then show the .env line that sidesteps them. And in that file the quoting is not cosmetic: double quotes around JSON make python-dotenv drop the line and set nothing, with no warning, so the run reads as though it had never been written.

The issue said this could not be fixed server-side. It can: vllm serve ... --default-chat-template-kwargs '{"enable_thinking": false}'. That is the better answer for a server you run yourself and no answer at all for one you share, so the docs name both.

Gate

ruff check, ruff format --check against git archive bytes, mypy on 3.11 and 3.12, and the full suite. All green locally on Windows. This PR exists to get macOS, Linux and the physics job.

Reviewed adversarially before opening

A six-lens review over the diff found six things, all fixed here: the instructions gap above; the coordinator-flock path (method: auction) was threaded but only the pilot flock was tested; the scope decision and the run_start emit were both mutation-confirmed unguarded, so both now have tests; "eight cloud vendors" was nine; and docs/architecture.md's run_start row needed the new field.

🤖 Generated with Claude Code