litellm/litellm-rust/crates/python-bridge/AGENTS.md

• Target invariants, not completion claims; these supersede the crate guidance below where they conflict
• Keep this crate the product-specific PyO3 consumer of litellm-host-python
  • Own registration, input projection, the route host and the caller callables it answers operations with (file readers, token providers), public response/error construction and the per-call composition of machine, route host and callback contract
  • Legacy callback sharing (the caller's args, kwargs and request object, body/header roots, re-aliasing unchanged body keys) lives in litellm-callbacks-legacy-python behind PublicCall and run_legacy_call; the bridge hands the public call over and keeps no copy
  • Value-oriented execution, sync waiting, nested-runtime checks, signal polling and panic containment live in litellm-host-python; native async work uses pyo3-async-runtimes, Serde output uses Pythonized<T>
  • Core owns typed native state, the route machine, provider preparation/I/O and normalization; the host driver owns terminal events; the legacy adapter in litellm-callbacks-legacy-python owns Logging dispatch policy
  • Python, Rust SDK and gateway use one lifecycle-bearing core route entrypoint; provider helpers stay private, never bridge-accessible transport drivers
  • Built-in provider/config/secret/auth/document preparation stays in Rust; caller-authored callbacks and focused Python-file reads run only at core-selected points
• Target GIL-enabled CPython explicitly with #[pymodule(gil_used = true)]; detach Rust-only work
  • GIL and tokio invariants, each pinned by a test in host-python (execution.rs, gil.rs) so a regression fails there before it deadlocks a proxy:
    • Never hold the GIL while waiting on the runtime. A sync entrypoint releases it with release_gil around block_on, because every task that attaches would otherwise wait on the thread that is waiting on them (pyo3 parallelism)
    • Never block_on from a tokio worker; the sync entrypoints refuse with "cannot run from a Tokio context" instead of panicking inside the runtime (tokio Runtime::block_on)
    • Inside a future, Python::attach only for GIL-cheap work: cloning a Py<T>, building a small value, reading a settings snapshot. Anything that can block (a secret manager read, a callback that does I/O, an import, a network call) goes through litellm_host_python::attach_blocking, which runs it on the blocking pool so the async workers keep polling other calls (tokio spawn_blocking). block_in_place is not an alternative: it needs a multi-thread worker and still steals it
    • attach_blocking work runs on a thread the interpreter did not create (pinned by the threading.get_ident() test). Like any foreign-thread attach it therefore has no running asyncio loop and a fresh contextvars context: do not hand it a coroutine or anything bound to the caller's loop
    • Dropping the await (an asyncio cancel) does not interrupt the Python call; it runs to completion and its result is discarded. A panic in it reaches the awaiting task as a panic
    • Add a case to gil.rs when a new seam changes any of these; the tests are the spec
  • Free-threading requires separate runtime/concurrency validation; omitting the attribute does not opt out on PyO3 0.28+
• Preserve public argument binding and Python object provenance
  • Project only consumed fields at reference read points; no eager whole-graph serialization or equality-based alias reconstruction
  • Preserve provider-specific upload/submission/poll observation and encoding boundaries; signed/build-captured bytes must not be silently reserialized
• Conversion errors and every failure after the call starts are terminal
  • Disabled/unavailable native execution may select legacy once; callback exceptions never authorize fallback or replay
• Use one ordinary inline async def driver in litellm/rust_bridge/lifecycle.py, with the native handle and call driver in litellm-host-python
  • Contract: start, resume_value, resume_error, idempotent close; explicitly tagged Await/Complete preserve awaitable final values
  • Validate Created/Running/Suspended/Closed protocol states; the machine yields ops, the driver emits one terminal event, the adapter chooses dispatch policy
  • Defer effectful setup/context reads/timestamps until start; unstarted-handle destruction releases inputs independently of Python finally
  • Catch only the selected await's errors; start/resume errors propagate, GeneratorExit closes without further awaits
  • Inline hooks preserve caller task/thread/loop and context writes; into_future creates a separate task and cannot satisfy this contract
• Finalize fallible public response/error construction, replacements and metadata before terminal dispatch
• Make ownership safe across suspension, re-entry, cancellation and GC
  • Keep native provider state typed in core; do not shuttle it through opaque Python transport/response classes
  • Prefer one retained Py<PyBaseException> via PyErr::into_value(py); reconstruct transient PyErrs, preserving identity, traceback, cause and context
  • Traverse every owned Python edge, including duplicate references; traversal cannot call Python
  • Take state out and mark Running under a short borrow, release borrows/locks before Python invocation, publish terminal state before finalizer-capable drops
  • Close/GC/deferred release are idempotent and re-entry-safe, including during Rust unwinding; release only owned references, never clear caller containers or mask the selected error
  • The machine owns its in-flight provider future; interrupt drops it synchronously, so provider captures are released before the driver returns and no task outlives the call
• Verify behavior through a fresh, provenance-checked installed extension and positive native execution evidence before replacing the custom coroutine
  • Cover admitted provider workflows, binding/read-point/identity behavior, failure continuation, finalization, no replay, deferred gates, re-entry, GC and cancellation termination
  • Measure real conversion/copy costs before optimizing; preserve input contracts and capture lifetimes with PyBackedBytes, and lookup timing when interning names
  • Ship accurate _native.pyi declarations and typing markers; distinguish Future-returning bindings from coroutine-returning bindings
• References: ownership, GC, exception transfer, re-entry
  • GIL policy, experimental async limits, task conversion, native cancellation/delivery
  • performance, PyBackedBytes, typing

Rules for litellm-rust/crates/python-bridge.

Responsibility

python-bridge is the PyO3 boundary between Python LiteLLM and Rust transforms. Keep this crate thin. It exposes LiteLLM Rust APIs, assembles domain requests, maps domain errors to Python exceptions, and delegates generic conversion and GIL handling to litellm-host-python.

Bridge Shape

• Prefer one stable method per top-level LiteLLM route, for example messages(...), calling the matching litellm-core entrypoint.
• Do not add one exported PyO3 function per provider helper unless there is a measured reason.
• Provider dispatch belongs in the litellm-core route module (e.g. litellm_core::messages), not in this PyO3 crate.
• Python owns rollout state and fallback. Rust should return errors; Python decides whether to raise or fall back. For a rust-only provider/route (no Python reference), the Python side is a thin dispatch that calls Rust and raises when the bridge is unavailable, with no fallback.
  • Declare it by passing python=NO_PYTHON (litellm.rust_bridge.runtime) to PublicDispatch.run/arun or runtime.run/arun, never a stand-in callable that raises, and give every context of it a RUST_REQUIRED catalog rule
  • Any other decision, an unprojectable call, or a bypass raises NoPythonImplementationError before native runs, so a misdeclared route fails in tests instead of reaching deleted code. When deleting a route's Python implementation, switch its dispatch to NO_PYTHON in the same change
• Keep the Python interface minimal (well under 100 lines per route): it only marshals inputs and calls Rust. Do not add per-route feature flags, and do not put provider dispatch in litellm/main.py; it lives in a thin dispatch class under litellm/llms/<provider>/<route>/.

Data Handling

• OCR payloads can contain personal data and large base64 images. Do not log payloads or provider responses.
• Avoid copying large payloads more than needed. The current JSON round-trip is acceptable for the first scaffold, but future performance work should evaluate direct PyO3 conversion before expanding Rust coverage to image-heavy paths.
• Do not expose raw Rust errors that include document contents or upstream bodies.

Tests

• cargo test --workspace must compile this crate.
• Python tests must cover bridge disabled, bridge enabled, and module-missing fallback behavior for every exposed route.