Recipes

Short, tested paths from an intent to a working result. Each recipe lists its prerequisites, the commands, what they print, and what to do when they print something else. Every command block in these files is run by internal/srcgate.TestRecipesRun against a real jitllmd, so a recipe that stops working fails a test instead of a reader.

Recipe	What you end up with
Start a server and make the first request	jitllmd serving a model, and a completion and a chat reply from it
Connect an OpenAI-compatible client	the Python and JavaScript openai SDKs, and curl, talking to jitllmd
Tools and structured output	tool calls (chosen and forced) and replies held to a JSON schema
Choose a supported checkpoint and convert it	a .jlm container, and a way to tell an unsupported model before downloading it
Memory budgets and device placement	a model held to a memory budget, and blocks placed where you ask
Diagnose a slow or failed request	the server’s own account of where a request’s time and memory went

All of them need the two binaries, jitllm and jitllmd, on your PATH (install), and curl. They run on a machine with no GPU. Every server in them listens on 127.0.0.1:4242.

The capability manifest (docs/capabilities.json, jitllm capabilities) and jitllm doctor answer “does this build support my model and workload” and “what is wrong with this machine” in machine-readable form, where a build has them.

How the test runs a recipe

TestRecipesRun builds jitllm and jitllmd from the tree, then for each recipe makes an empty working directory, puts the two binaries first on PATH, and runs the fenced blocks in order:

• ```sh blocks run with bash -euo pipefail, each in a fresh shell (a block does not inherit another’s variables).
• ```python blocks run with JITLLM_RECIPE_PYTHON (default python3) when it can import openai, and are skipped by name when it cannot.
• ```js blocks run as ES modules with node when it can import 'openai' (from the working directory, which links JITLLM_RECIPE_NODE_MODULES as its node_modules when that is set), and are skipped by name when it cannot.
• ```text and ```json blocks are shown, not run.

Every 127.0.0.1:4242 and 127.0.0.1:80NN address is rewritten to a free port, the same one for the same address throughout a recipe. An HTML comment directly before or after a block steers it:

Comment	Meaning
<!-- test: background -->	before a block: start it and leave it running; it is stopped when the recipe ends
<!-- test: expect TEXT -->	after a block: its output must contain TEXT (several may follow one block)
<!-- test: exit N -->	before a block: it must exit with status N
<!-- test: instead COMMAND -->	before a block: run COMMAND in its place. Only for downloads: COMMAND copies the same file from the repository ($JITLLM_REPO) or the test model directory ($JITLLM_TEST_MODELS)
<!-- test: needs FILE -->	every later block needs FILE in the test model directory (JITLLM_MODELS); without it they are reported missing, which fails the test except in the model-free run, where it is a named skip
<!-- test: skip REASON -->	before a block: not run, and the reason is logged

Run them:

./scripts/cap 8G -- taskset -c 0,2,4,6,8,10 go test ./internal/srcgate -count=1 -run TestRecipesRun -v

JITLLM_RECIPE=diagnose runs one recipe. CI’s model-free run executes every block that needs only the repository’s own testdata/models/stories260K.gguf.