LiteLLM compatibility gap matrix
This matrix tracks standalone Go Feather Route parity with the LiteLLM behavior currently consumed by Thingd Cloud. It is an engineering checklist, and migration checklist. Go Feather Route is the Cloud primary; LiteLLM remains the explicit rollback path.
| Area | Status | Evidence or remaining work | Priority |
|---|---|---|---|
| Chat endpoint and OpenAI request shape | Supported | Router forwards the bounded request body to /v1/chat/completions | P0 |
| JSON-object responses | Partial | response_format is forwarded; add Cloud-shaped fixture assertions for Ollama | P0 |
| Model aliases and provider routing | Supported | Configuration routes aliases and translates Ollama aliases to upstream model names | P0 |
| Gateway authentication | Supported | Protected API endpoints require the gateway bearer key | P0 |
| Provider authentication | Supported | Provider keys are injected at runtime and sent upstream | P0 |
| Non-streaming usage passthrough | Supported | Upstream response and headers are copied within response limits | P0 |
| Streaming SSE forwarding | Partial | Forwarding and termination exist; verify Cloud stream parsing and final usage with Ollama | P0 |
| Client cancellation | Partial | Request context is propagated; add provider-observed cancellation fixture | P0 |
| Bounded retries and ordered fallbacks | Supported | Retryable HTTP responses advance through configured providers; ambiguous transport failures are not replayed | P0 |
| Timeout behavior | Supported | Request context uses the configured upstream timeout | P1 |
| Embedding endpoint | Supported | Single and batch forwarding plus response validation tests exist; Ollama alias mapping is covered | P0 |
| Embedding worker payloads | Not verified | Replay the Cloud worker-shaped batch and failure cases | P0 |
| Project memory workloads | Not verified | Validate indexing, retrieval, dimensions, and rebuild behavior through the gateway | P0 |
| OpenAI-compatible errors | Partial | Error forwarding exists; compare exact Cloud error parsing and status expectations | P1 |
| Request ID propagation | Supported | Gateway request ID is forwarded as X-Request-ID | P1 |
| Usage and reservation accounting | Supported | Go Feather emits idempotent normalized usage events; Cloud deduplicates by request ID and owns quotas/settlement | P0 |
| Cloud caching and quotas | Cloud-owned | Must remain outside the gateway and be tested during a future integration phase | P0 |
| Health and readiness | Supported | Liveness, readiness, status, model status, and metrics endpoints exist | P1 |
| Low-memory operation | Measured baseline | Repeat representative workloads with immutable images and recorded cgroup data | P1 |
Priority meaning
- P0 — required before any Cloud integration review.
- P1 — required before a compatibility-focused release.
The matrix should be updated only when source, tests, or reproducible benchmark evidence changes the status. It must not be used to enable Go Feather Route in Thingd Cloud.