feat: add livekit-plugins-thegrid

[CastilloLuis]

What

Adds livekit-plugins-thegrid, an LLM plugin for The Grid — an OpenAI-compatible inference marketplace.

Modeled on livekit-plugins-perplexity. The plugin subclasses the OpenAI plugin's LLM and sets the base URL and THEGRID_API_KEY handling; there's no bespoke client.

Registration points, matching how perplexity/cerebras are wired:

• livekit-plugins/livekit-plugins-thegrid/ (new package)
• livekit-plugins-thegrid = { workspace = true } in the root pyproject.toml
• thegrid = ["livekit-plugins-thegrid>=1.8.0"] extra in livekit-agents/pyproject.toml
• tests/test_plugin_thegrid.py

Left CHANGELOG.md and the version manifests alone per CONTRIBUTING.

One deliberate difference from the perplexity template

Perplexity passes _strict_tool_schema=False. I tested The Grid with strict: true tool schemas (and additionalProperties: false) and it returns well-formed tool calls, so I did not carry that flag over — the default applies.

Verification

Run against the live API, not mocked:

$ pytest tests/test_plugin_thegrid.py --plugin thegrid -q
...                                                                      [100%]
3 passed in 6.34s

That covers chat, a tool call round-trip on agent-standard, and streaming. Also checked ruff format --check and ruff check clean on the new files, that the missing-key path raises, and that GET /v1/models returns 200 directly with no redirect (relevant because _prewarm_impl() calls it with a 5s timeout).

Note on models

The Grid addresses capability tiers — text-standard, code-prime, agent-max each route to a current model for that tier, so an agent keeps working when the underlying model is swapped. models.py lists the tiers plus the lab-pinned *-latest ids; model still accepts any str, and GET /v1/models is authoritative.

Disclosure: I work on The Grid. Happy to adjust anything to house style.

[CLAassistant]

All committers have signed the CLA.

[CastilloLuis]

Good catch, and your suggestion was right. Applied, with one extra step it needed.

Guarding the raise on client is None alone was not sufficient: this class still forwarded api_key="" to the OpenAI plugin, and llm.py:152 rejects an explicitly-given empty string with "OpenAI API key is required". So an injected client failed one line further down instead.

It now resolves the key into a local and passes resolved_key or NOT_GIVEN, which leaves the OpenAI plugin free to use the supplied client's own authentication.

Verified all three paths:

supplied client without key: OK, model = text-standard
no client, no key: still raises ValueError OK
env key path still works: text-standard

pytest tests/test_plugin_thegrid.py --plugin thegrid still passes 3/3 against the live API, and ruff check / ruff format --check are clean on 0.15.22.

Separately, I have added the missing uv.lock entry for the new workspace member in this branch, which is what was failing the ruff job (it died at uv sync before lint ran).

[CastilloLuis]

Merged main in; the only conflict was uv.lock.

Resolved by taking upstream's lockfile wholesale and re-running uv lock, so the delta against main is just the new workspace member rather than a hand-merged lockfile:

Resolved 328 packages
Added livekit-plugins-thegrid (dynamic)

Verified after the merge:

• uv run ruff check on the plugin and its test: all checks passed
• uv run ruff format --check: 6 files already formatted
• uv run pytest tests/test_plugin_thegrid.py: 3 passed, run against the live API with a real THEGRID_API_KEY

Worth flagging for whoever reviews: those three tests are live integration tests, so they need a THEGRID_API_KEY secret in CI to run there. Without one they fail at construction with THEGRID_API_KEY is required. Happy to mark them skipif on a missing key if you would rather they no-op in CI, or to supply a key if you have a way to take one.