[Feat] Served Laya as a decision model (--model laya-served)

.env.example

@@ -34,6 +34,13 @@ MODEL_NAME=google/gemini-2.5-flash
 # LAYA_MAX_LEN=1024
 # LAYA_HEAD_MAX_LEN=512              # raise for choice questions with many options
 
+# ---- Served Laya (behind --model laya-served; a system1-omni worker or laya-serve, no extra needed) ----
+# LAYA_SERVED_URL=http://127.0.0.1:8000   # the worker; :8080 for the omni-jev frontend
+# LAYA_SERVED_MODEL=english
+# LAYA_SERVED_API_KEY=                    # the worker's LAYA_API_KEY, when it sets one
+# LAYA_SERVED_TIMEOUT_S=5                 # per decision, the retry and any /health refresh included
+# LAYA_SERVED_MAX_LEN=512                 # the worker's LAYA_MAX_LEN
+
 # ---- Cua-S1 Nano (the in-process option scorer behind --model cua; needs `uv sync --extra cua`) ----
 # CUA_S1_CHECKPOINT=cua-ai/cua-s1-nano-0.1   # a Hugging Face id, or a local directory holding <subfolder>/
 # CUA_S1_SUBFOLDER=text                     # the text-only checkpoint; the window is 256 bytes of state

CHANGELOG.md

@@ -12,6 +12,12 @@ The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/); ver
 
 ### Added
 
+- `--model laya-served`: Laya served over HTTP by system1-omni's worker (or plain laya-serve), on every front that
+  takes `laya`, in `decide` and in MCP `decide`. Configured by `LAYA_SERVED_URL` and optional `LAYA_SERVED_*`
+  variables; no cloud key. Run records name the served checkpoint, revision and device. Design, API spec and the
+  run steps: `docs/served-laya.md`, `docs/api/`.
+- Tool-front ticks keep the answering model as `model`, and a served model's `served_by`, `url`, `request_id` and
+  `server_timing`.
 - The MCP `decide` tool accepts `model="jev"|"laya"|"cua"`, defaulting to `jev`. Local backends use their
   optional extras and need no Jev API key.
 - `docs/benchmarks.md`: the Google Flights driver comparison rerun on 2026-09-23 from Poland, every arm three times on
@@ -20,8 +26,8 @@ The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/); ver
 
 ### Changed
 
-- `--model laya` loads in about 3 s instead of about 35 s: the encoder is built with transformers' weight init
-  off, since the checkpoint replaces every weight. Weights and answers are unchanged.
+- `--model laya` loads in about 3 s instead of about 35 s: the `laya` extra now needs laya 0.3.9 or later, which
+  builds the encoder with transformers' weight init off, since the checkpoint replaces every weight.
 - `--model` picks the model on every agent, on `decide` and on `probe`: `jev`, `laya`, `cua`, `llm`, `random` or
   `rule`. The results table's column, the replay page's badge data and a browser run's `answer.json` name it
   `model` as well; the replay still reads the `slot` key of records written by 0.1.0.

CONTRIBUTING.md

@@ -71,7 +71,7 @@ Everything outside `openjiuwen` is an extra. An agent whose extra is missing say
 | `report` | pillow, playwright | `python -m evals.replay`, the showcase pages and GIFs; `--gif` also needs `uv run playwright install chromium` |
 | `laya` | laya (torch, transformers) | `--model laya` on every agent and on `decide` and `probe`: Laya in process, no Jev key; the checkpoint downloads into the Hugging Face cache (`HF_HOME`) on first use |
 | `cua` | cua-s1 (torch), huggingface-hub | `--model cua` on tool and browser agents and on `decide` and `probe`: Cua-S1 Nano in process; the 3 MB checkpoint downloads into the Hugging Face cache (`HF_HOME`) on first use |
-| `dev` | pytest, pytest-asyncio, ruff, ty | the test suite, `scripts/smoke.sh` and the lint and type checks |
+| `dev` | pytest, pytest-asyncio, ruff, ty, jsonschema, pyyaml, referencing | the test suite, `scripts/smoke.sh` and the lint and type checks; the last three check the served-Laya fixtures against `docs/api/` |
 
 `uv sync --all-extras` installs all seven. The CLI runs from a checkout; a wheel install (`uv tool install`,
 `pip install`) is unsupported, because the data folders (`evals/2048`, `evals/millionaire`, `evals/labelled`) sit

README.md

@@ -141,6 +141,7 @@ interface fits: [docs/architecture.md](docs/architecture.md), [docs/decision-mod
 - [docs/agents.md](docs/agents.md): every agent with its flags, run command and extra.
 - [docs/architecture.md](docs/architecture.md) and [docs/decision-models.md](docs/decision-models.md): the fronts, the model slot, the model interface, adding a backend.
 - [docs/browser-front.md](docs/browser-front.md): the browser policy, decision by decision.
+- [docs/served-laya.md](docs/served-laya.md): Laya served by system1-omni as a decision model over HTTP, with its [API spec](docs/api/laya-systemone.openapi.yaml).
 - [docs/configuration.md](docs/configuration.md): environment variables, defaults and reader subsystems in one table.
 - [docs/glossary.md](docs/glossary.md): terms the documentation glosses on first mention.
 - [docs/why.md](docs/why.md): the problem, the philosophy, the precedents.