docs(transcripts): Qwen3-32B on vLLM, with and without thinking

CHANGELOG.md

@@ -242,6 +242,17 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   from one run into the next. Thanks to [@Bayway](https://github.com/Bayway) (#7),
   whose runs and whose transcripts these are.
 
+- **Two transcripts from Qwen3-32B-AWQ on vLLM, and the first numbers anybody has for what
+  `--extra-body` actually stops.** Same build, same seed, same server, one variable: with the
+  flag off, all eight LLM calls deliberate in the open and the run spends 2,049 output tokens
+  over four steps; with it on, none of the five do and the run spends 263 over three. Neither
+  needed the JSON text fallback. The totals flatter it, because the two runs took different
+  paths — the thinking one spent two steps on `remember` and `quack` — so the per-call figures
+  are the ones to read: 165 to 402 output tokens a call becomes 19 to 130. `reasoning_tokens`
+  stays 0 in both, which is a property of a server running without `--reasoning-parser` rather
+  than evidence of a model that did not reason: the thinking sits in `content` and is billed as
+  output. The machine is an `aarch64` NVIDIA GB10, which nothing here had run on before.
+
 ### Changed
 
 - **The README and the social card wear the duck head quackd.org uses, and `logo.svg` is

README.md

@@ -217,7 +217,7 @@ Version 0.8, simulator and mocks. What has been built, and how far each piece ha
 | MCP server (`quackd serve-mcp`) | ✅ Claude Code and Claude Desktop, one robot or a flock with `--robots` or `--flock NAME` (nine `robot_*` tools, tested in process against the simulator and the mocks), no Claude Desktop session on record |
 | Memory between runs (`quackd memory`, `remember`) | ✅ one JSONL file per `adapter:backend`, or per registered robot name, notes and run outcomes into the next prompt, tested end to end offline, 🧪 the `remember` tool itself exercised by one local model on one machine and by no cloud model ([docs/memory.md](docs/memory.md)) |
 | Providers: eleven cloud vendors, fake | ✅ implemented, tested offline against stubbed SDK clients, with one hand curated catalogue of 115 model ids that `--model` is checked against before any call (`quackd list-models`), real model hero recording pending an API key |
-| Local models (Ollama, vLLM, llama.cpp, LM Studio, any OpenAI compatible server) | ✅ implemented and tested against the OpenAI wire format, 🧪 two live runs by a contributor (Qwen 2.5 Coder 14B on LM Studio, seeds 5 and 6), never on this machine, transcripts in [`docs/assets/transcripts/`](docs/assets/transcripts/), more welcome |
+| Local models (Ollama, vLLM, llama.cpp, LM Studio, any OpenAI compatible server) | ✅ implemented and tested against the OpenAI wire format, 🧪 four live runs by contributors (Qwen 2.5 Coder 14B on LM Studio, seeds 5 and 6, and Qwen3-32B-AWQ on vLLM on an `aarch64` NVIDIA GB10, seed 1, with and without thinking), never on this machine, transcripts in [`docs/assets/transcripts/`](docs/assets/transcripts/), more welcome |
 | Registered robots and flocks (`quackd robot`, `quackd flock`) | ✅ both command groups over `~/.quackd/robots.json` and `~/.quackd/flocks.json`, so `--robot NAME` means the same thing in every command that takes a robot and `--flock NAME` in `run` and `serve-mcp`, tested offline, `robot list --probe` answered by the mocks, 🧪 never pointed at hardware ([docs/registry.md](docs/registry.md)) |
 | Pilot flocks, several robots on one task (`--flock NAME`) | ✅ one whole pilot per body, any backend, same or different bodies, 2 to 8, each with its own executor, allowlist, budget, heartbeat, memory and verdict, dividing the work with `tell` over the bus, 🧪 experimental, exercised on mock and sim2d bodies with the scripted pilot, by no real model and on no hardware |
 | Coordinator flock, one referee instead (`--flock N`) | ✅ deterministic auction and bus, one planner LLM call at most, ground truth checked in tests, 🧪 experimental and sim2d Microducks only. Its capability aware role auction (spotter/kicker) is unit tested but has no bundled two-robot demo today |
@@ -632,7 +632,7 @@ The physics simulator costs what physics costs. Measured here on one Windows lap
 - `grab` is open loop upstream and unreliable here on purpose. `fetch` says so in its file.
 - A manifest can be smaller than the robot. The LeRobot arm's `real` backend claims no camera and no `pick` until it connects, and even then `pick` appears only when a policy object was injected in code. A rosbridge base over `ws` has no camera verbs unless the address names an image topic, and a ToddlerBot has no `move` unless a walk checkpoint is staged.
 - The model catalogue is hand curated. It was read off the eleven vendors' own documentation on 2026-09-12 and it is a snapshot of that day, not a live list. A vendor can retire, rename or add an id between quackd releases, and this build would then refuse an id that is real and offer one that is gone. `quackd list-models` prints exactly what this build knows, which is the only thing `--model` accepts for a cloud vendor.
-- Local model quality is unmeasured. The JSON text fallback and the one retry exist because small models often miss native tool calls. One contributor ran `find-and-kick` against Qwen 2.5 Coder 14B through LM Studio on two seeds, both successes, one of them reading an earlier run's memory. The two transcripts are in [`docs/assets/transcripts/`](docs/assets/transcripts/) and read in [docs/local-llms.md](docs/local-llms.md).
+- Local model quality is unmeasured. The JSON text fallback and the one retry exist because small models often miss native tool calls. One contributor ran `find-and-kick` against Qwen 2.5 Coder 14B through LM Studio on two seeds, both successes, one of them reading an earlier run's memory. Another ran it twice against Qwen3-32B-AWQ on vLLM on an `aarch64` NVIDIA GB10, same seed and same build, differing only in whether `--extra-body` turned the model's thinking off: 2,049 output tokens became 263, and neither run needed the text fallback. The four transcripts are in [`docs/assets/transcripts/`](docs/assets/transcripts/) and read in [docs/local-llms.md](docs/local-llms.md).
 - The pilot flock takes any body and has run on `mock` and `sim2d` ones only: N simulated members are N separate worlds with no shared arena, it costs one budget and one model call per member per turn, and `tell` has been exercised by the scripted pilot and by no real model. The coordinator flock is simulator only, ships one choreography and exactly two roles (spotter and kicker, unit tested but with no bundled multi-role starter), and knows only the Microduck. Separation uses sim ground truth, and two robots share no frame of reference on hardware.
 - LAN discovery and the MQTT bus have each been exercised once, on one machine. Nothing has crossed to a second machine, the MQTT bus is a library with no `--bus` flag, and a coordinator flock across machines also needs a clock across machines, which does not exist yet. A pilot flock needs no such clock and has simply never been tried across two.