Immutable. This exact content is served forever at /api/v1/blob/85dff2ebe3e058fb.
# Agent Guidelines for Llama Stack
This file provides guidance for AI coding agents working on this codebase.
It complements [CONTRIBUTING.md](CONTRIBUTING.md) with agent-specific instructions.
Human contributors should follow the conventions in CONTRIBUTING.md.
## Project Overview
Llama Stack is an API server implementing the OpenAI Responses API, Chat Completions,
Embeddings, and supporting APIs (files, vector stores, batches, eval, safety). It supports
multiple inference backends (OpenAI, Azure, Bedrock, vLLM, Ollama, WatsonX, etc.) through
a provider architecture.
## Repository Layout
```text
src/llama_stack/ # Server implementation
core/ # Request routing, server, storage
providers/
inline/ # Built-in providers (responses, eval, vector_io, etc.)
remote/ # Remote provider adapters (OpenAI, Azure, vLLM, etc.)
registry/ # Provider registration specs
utils/ # Shared provider utilities (OpenAI mixin, MCP, etc.)
distributions/ # Distribution configs (starter, ci-tests, etc.)
src/llama_stack_api/ # API definitions, Pydantic models, FastAPI routes
tests/
unit/ # Unit tests
integration/ # Integration tests with recording/replay system
```
## Python & Tooling
- **Python 3.12** is required. Pre-commit hooks only work with 3.12.
- Use `uv` for all dependency management and script execution.
- Run scripts with `uv run`, never bare `python3` or `python`.
- Use standard library modules when possible.
- All function signatures must have type hints. Prefer Pydantic models for validation.
- Code must pass `mypy`. Check the exclude list in `pyproject.toml` for known exceptions.
- Use `def _function_name` for private functions.
- Prefer explicit top-level imports over inline imports.
- Do not use exceptions as control flow.
## Code Style
- Comments must add value. Do not write filler comments that describe the next line.
- Add comments only to clarify surprising behavior that is not obvious from the code.
- Good variable naming and clear code organization matters more than comments.
- Do NOT remove existing comments unless they are factually wrong.
- Error messages must be prefixed with "Failed to ...".
- Use structured logging via `from llama_stack.log import get_logger`. Always use
key-value style: `logger.info("Processing request", model=model_id, provider=provider)`
instead of f-strings or %-style formatting. The pre-commit hook
`Block f-string logging` enforces this.
- The pre-commit hook `Ensure 'llama_stack.log' usage for logging` enforces that all
logging uses the project's logger, not the standard library directly.
## Git Conventions
- Always use `--signoff` (`-s`) when creating commits.
- Do not amend commits and force push during PR review. Use new commits instead.
- Use `git merge main` to update branches, not `git rebase`.
- Commit messages must use [conventional commits](https://www.conventionalcommits.org/)
format and full sentences, not bullet points.
- Merge `upstream/main` before pushing a branch to avoid CI failures from stale code.
## Testing
- Unit tests: `./scripts/unit-tests.sh` or `uv run pytest tests/unit/ -x --tb=short`
- Run pre-commit checks: `uv run pre-commit run --all-files`
### Integration Tests
Integration tests use a recording/replay system. Recordings are JSON files in
`tests/integration/*/recordings/` keyed by SHA256 hashes of HTTP request bodies.
When modifying code that changes request bodies sent to providers, recordings may
need to be re-recorded.
Run integration tests via the CI script (same command used in GitHub Actions):
```bash
# Replay mode (default) — uses pre-recorded responses, no API keys needed
uv run --no-sync ./scripts/integration-tests.sh \
--stack-config server:ci-tests --setup gpt \
--file tests/integration/responses/test_compact_responses.py
# Re-record missing recordings (requires API keys, e.g. OPENAI_API_KEY)
uv run --no-sync ./scripts/integration-tests.sh \
--stack-config server:ci-tests --setup gpt \
--inference-mode record-if-missing \
--file tests/integration/responses/test_compact_responses.py
# Run a full suite
uv run --no-sync ./scripts/integration-tests.sh \
--stack-config server:ci-tests --setup gpt --suite responses
```
Key flags: `--stack-config` (required), `--setup` (`gpt`, `ollama`, `vllm`),
`--inference-mode` (`replay`, `record`, `record-if-missing`), `--file` (single file),
`--pattern` (pytest `-k` filter), `--suite` (`base`, `responses`, `vision`).
If a test fails in replay mode with "Recording not found", re-run with
`--inference-mode record-if-missing` and commit the new recording files.
Recording locally requires an API key for the provider under test (e.g.
`OPENAI_API_KEY` for `--setup gpt`). If you do not have a key, repository
maintainers can trigger the record workflow on GitHub via
`.github/workflows/record-integration-tests.yml`.
## Provider Architecture
Each provider implements a protocol (e.g., `Inference`, `Responses`, `VectorIO`) and is
registered in `src/llama_stack/providers/registry/`. Provider specs include:
- `provider_type`: e.g., `remote::openai`, `inline::builtin`
- `module`: Python module path
- `config_class`: Pydantic config class path
- `api_dependencies`: APIs this provider depends on
- `deprecation_warning`: For deprecated providers (triggers runtime warning)
- `toolgroup_id`: For tool_runtime providers that auto-register tool groups
Configuration classes must use Pydantic `Field` with `description` parameters —
these generate the provider documentation automatically.
## Distribution Configs
Distribution YAML files in `src/llama_stack/distributions/` are partially auto-generated.
After changing provider configs, run:
```bash
uv run ./scripts/distro_codegen.py
uv run ./scripts/provider_codegen.py
```
Do not edit generated files in `docs/docs/providers/` manually.
## API Changes
When modifying or extending APIs:
1. Update models in `src/llama_stack_api/`
2. Regenerate OpenAPI specs: `uv run ./scripts/run_openapi_generator.sh`
3. Check for breaking changes — the pre-commit hook `Check API spec for breaking changes`
enforces backward compatibility.
4. Include a test plan with a testing script and execution output in your PR description.
## Common Patterns
### Adding a new parameter to an existing API
1. Add the field to the Pydantic model in `src/llama_stack_api/`
2. Thread it through the provider protocol and implementation
3. Update affected distribution configs if needed
4. Regenerate specs and docs
5. Add unit test cases covering the new parameter
### Adding a deprecated alias for a renamed provider
Use the existing `deprecation_warning` field on `InlineProviderSpec` or `RemoteProviderSpec`.
Search for existing examples: `grep -r "deprecation_warning" src/llama_stack/providers/registry/`
### Before adding any new pattern
Search the codebase for existing examples of the same pattern first.
Use `grep` to find how deprecation, validation, configuration, or aliasing is already
handled elsewhere.
## Keeping Documentation Up to Date
When making code changes, check whether the following documentation needs updating:
- `README.md` — the root README contains an ASCII architecture diagram showing API
endpoints, inference providers, vector stores, tools/connectors, and file storage.
Update it when adding or removing providers, APIs, or backend integrations.
- `ARCHITECTURE.md` — system overview, request flow, provider architecture, API layer,
storage, configuration, and test recording system
- Module-level `README.md` files in key directories:
- `src/llama_stack/README.md`, `src/llama_stack/core/README.md`
- `src/llama_stack/core/server/README.md`, `src/llama_stack/core/storage/README.md`
- `src/llama_stack/core/routing_tables/README.md`
- `src/llama_stack/providers/README.md`, `src/llama_stack/providers/inline/README.md`
- `src/llama_stack/providers/remote/README.md`, `src/llama_stack/providers/registry/README.md`
- `src/llama_stack/providers/utils/README.md`, `src/llama_stack/providers/utils/inference/README.md`
- `src/llama_stack/providers/inline/agents/README.md`
- `src/llama_stack/providers/inline/tool_runtime/README.md`
- `src/llama_stack/providers/remote/inference/README.md`
- `src/llama_stack/distributions/README.md`
- `scripts/README.md`
- `tests/README.md`, `tests/unit/README.md`, `tests/integration/README.md`
These files help LLMs and new contributors navigate the codebase. If your change
adds, removes, or renames modules, providers, APIs, or storage backends, update
the relevant documentation to stay in sync.