trustedrouter-provider-check

trustedrouter-provider-check runs provider-side compatibility checks for the OpenAI-compatible HTTP contract consumed by TrustedRouter. It checks the catalog, every advertised native model, non-streaming chat semantics, and the streaming behavior that most often fails behind gateways and reverse proxies. Opt-in Tiers 5–6 add tools, structured output, and advisory production-style performance sampling.

The implementation vendors production contract functions and a content-hashed snapshot from TrustedRouter. The local test suite joins those functions to a configurable mock HTTP server. Tiers 1–4 have been exercised against OpenAI, z.ai GLM, DeepSeek, and Ollama during development; that is point-in-time test evidence, not certification. Tiers 5–6 are mock-verified against the inspected enclave contract but have not yet been verified in this repository against a live third-party provider.

Running the checker

Install the project with uv, then pass the API root that owns /models and /chat/completions. Include /v1 when the endpoint uses that prefix. Authentication is optional: when neither --api-key nor TR_PROVIDER_API_KEY is set, the checker sends no Authorization header and records that fact in both report formats.

export TR_PROVIDER_API_KEY='your-provider-key'
uv run trustedrouter-provider-check \
  --base-url https://inference.example.com/v1 \
  --model deepseek-chat

--api-key is also accepted, but TR_PROVIDER_API_KEY keeps the key out of shell history. When supplied, the key is recursively redacted from human and JSON reports.

Available options:

--base-url URL       OpenAI-compatible API root; required for a check run
--api-key KEY        Optional provider key; falls back to TR_PROVIDER_API_KEY
--model MODEL        Native model id for Tiers 3-6; defaults to the first /models id
--catalog-url URL    Public Provider Contract v2 declaration; omitted means skip
--tier {1,2,3,4,5,6} Highest tier to run, including lower tiers; default 4
--perf-samples N     Billed Tier 6 completions; default 3
--json OUT           Also write redacted JSON to OUT; use - for JSON on stdout

Human-readable output is the default. The conformance gate and process exit status use Tiers 1–4 only: a Tier 1–4 hard failure exits 1; otherwise the run exits 0, even when an informational Tier 5 finding or advisory Tier 6 sample is red. Invalid or missing arguments exit 2. Running with no arguments prints help and exits 2; it never reports an empty success.

Examples:

# Catalog discovery and an optional public declaration only
uv run trustedrouter-provider-check \
  --base-url https://inference.example.com/v1 \
  --tier 1

# Keyless local Ollama/llama.cpp/vLLM setup: no Authorization header is sent
uv run trustedrouter-provider-check \
  --base-url http://127.0.0.1:11434/v1 \
  --model llama3.2 \
  --tier 4

# All tiers, plus a JSON artifact
uv run trustedrouter-provider-check \
  --base-url https://inference.example.com/v1 \
  --model qwen3-4b:latest \
  --catalog-url https://inference.example.com/catalog.v2.json \
  --tier 6 \
  --perf-samples 3 \
  --json provider-report.json

# Package and exact contract identities
uv run trustedrouter-provider-check --version
uv run trustedrouter-provider-check --print-contract-version

What the tiers check

Tier 4 passes the response through the vendored _observe_provider_stream. Role-only chunks and : ping comments do not count as first output; the clock stops only for a non-empty content, reasoning, or tool delta. Usage may appear in any chunk. The accepted content/reasoning delta shapes are the same multi-shape set exported by production.

Tier 5 mirrors two enclave-sensitive details. stream_translate.go keys tool fragments by index and starts a tool block from the first observed name, so a late function.name cannot be repaired after translation. byok.go replays a tool-only assistant turn with content: "" (not null) followed by one role: "tool" message per tool_call_id.

Tier 6 runs only when requested. Every --perf-samples unit is a billed long completion with stream: true, stream_options.include_usage: true, and a 512-token cap. A throughput sample needs at least 128 provider-reported output tokens. Effective throughput is output_tokens * 1000 / elapsed_ms, where elapsed time starts before the request; this prevents batched SSE delivery from inflating the result. The same observation is projected into vendored ProviderBenchmarkSample rows and passed to the vendored aggregate_leaderboard; declared time budgets are clamped with model_deadlines_declared. When a validated catalog price matches the model, the report includes both a cap estimate and an observed-token estimate.

JSON report

--json writes a versioned object with report_version, suite_version, contract_version, generated_at, target, summary, checks, performance, and a reserved unsigned submission block. The target records whether bearer authentication was used without recording the key. The submission block reserves JCS/RFC8785 canonicalization metadata; signing and portal upload are intentionally not implemented.

The checked-in schema is src/tr_provider_check/data/provider-report.schema.json. Summary totals cover all requested tiers, while summary.conformance_gate is computed only from Tiers 1–4. summary.provider_owned_failures counts hard failures attributed to the provider across the report; it is informational and does not redefine the gate.

The mock corpus includes isolated modes for non-SSE HTTP 200 responses, invalid data: framing, missing [DONE], ignored include_usage, early role plus late content, buffered output, finish-only streams, mid-stream error objects, bad usage, native/catalog ID disagreement, dead/flaky routes, request-field rejection, tool index/name/argument corruption, empty-string tool-turn rejection, structured-output mismatch, insufficient throughput usage, metadata drift, and every vendored stream delta shape. Tests assert the complete result map for each mode so dependent checks cannot pass vacuously.

Relationship to TrustedRouter

The canonical packaged snapshot is src/tr_provider_check/data/contract_snapshot.json. Selected functions in src/tr_provider_check/contract.py are byte-for-byte copies from the production Lore-Hex/quill-router checkout and are protected by normalized source hashes and behavioral replays. contract_version is a hash of the exported contract, not a branch name.

The HTTP and stream checks are justified against the enclave implementation in enclave-go/internal/llm/byok.go, stream_translate.go, and http_client.go, plus enclave-go/cmd/enclave/provider_stream.go for the pre-output retry and first-byte gate. In the supplied enclave reference checkout, provider_stream.go is under cmd/enclave, not internal/llm.

Run the cross-repository parity gate from the production checkout:

cd /path/to/quill-router
PROVIDER_CHECK_REPO_PATH=/path/to/trustedrouter-provider-check \
  uv run --no-sync pytest tests/test_provider_check_contract_parity.py -q

Synchronizing from production

1. Export the production contract with uv run python scripts/export_provider_check_contract.py in quill-router.
2. Copy src/trusted_router/data/provider_check_contract.json to src/tr_provider_check/data/contract_snapshot.json here.
3. Replace each source-hash mismatch with the exact production function body. Do not reformat vendored bodies.
4. Run the local and cross-repository gates below.

Development

uv run ruff check .
uv run ruff format --check .
uv run mypy src tests
uv run pytest -q
uv build

src/tr_provider_check/contract.py is excluded from Ruff formatting because its production function bodies are hash-guarded. See CONTRIBUTING.md for the snapshot sync checklist.

License

Apache License 2.0; see LICENSE and NOTICE.