* feat(providers): add Google Vertex AI provider
Adds Vertex AI provider profiles, routing, credential refresh plumbing, CLI support, docs, and regression coverage. Keeps the related NETLINK_ROUTE seccomp allowance needed by Vertex client tooling that calls getifaddrs.
* docs: add Vertex AI sandbox usage for Claude Code and OpenCode
Cover the full end-to-end setup for running Claude Code and OpenCode
inside an OpenShell sandbox via inference.local with a Vertex AI backend:
- google-vertex-ai.mdx: add 'Use from a Sandbox' section with tabbed
examples for Claude Code (--bare flag, no /v1 suffix) and OpenCode
(/v1 suffix required). Add providers_v2_enabled prerequisite and
--no-verify note for global region. Document policy proposals table
covering metadata.google.internal (always blocked), downloads.claude.ai,
and storage.googleapis.com.
- inference-routing.mdx: expand 'Use the Local Endpoint' section with
tabbed examples for Claude Code, OpenCode, Python OpenAI SDK, and
Python Anthropic SDK. Add notes explaining the /v1 path suffix
difference between clients.
- supported-agents.mdx: update Claude Code and OpenCode rows to mention
inference.local support and correct base URL requirements.
* fix: address vertex review findings
* test(sandbox): retry on spurious Ok in fork-exec ambiguity test
On arm64 under heavy CI load, the /proc fd scan in
find_socket_inode_owners can transiently miss the parent process's
socket fd entry, returning only the child as an owner. This causes
resolve_process_identity to return Ok (single owner, no ambiguity
check fires) instead of the expected ambiguous-ownership Err.
Extend the retry loop to also handle unexpected Ok results, mirroring
the existing retry for transient Err results. 10 retries at 50ms gives
a 500ms settling window, which is sufficient for procfs to stabilize
on loaded arm64 runners.
* fix: address vertex review regressions
* docs(router): clarify stream_response semantics for Vertex rawPredict routing
Document the three call sites of prepare_backend_request and their
stream_response values in a caller table:
- send_backend_request: false → :rawPredict (unary endpoint)
- send_backend_request_streaming: true → :streamRawPredict
- verify_backend_endpoint: explicitly false to probe the unary endpoint
Cross-reference the table from build_provider_url and
is_vertex_anthropic_rawpredict_route so the stream_response=true guard
in the suffix upgrade branch is understood in full context.
Also note that is_vertex_anthropic_rawpredict_route is a structural
predicate (model_in_path + anthropic_messages + :rawPredict suffix),
not a named-provider check, so any future provider with the same route
shape inherits the transforms automatically.
* fix(inference): prevent silent truncation of large streaming responses
The L7 inference proxy silently dropped tool_calls from large streaming
responses due to an aggressive 30s per-chunk idle timeout and a reqwest
total-request timeout that capped the entire body stream. Reasoning models
that pause during "thinking" phases triggered these timeouts, producing
valid-looking but truncated HTTP responses with no client-visible error.
- Extract prepare_backend_request() helper and create a streaming variant
that omits the total request timeout; body stream liveness is now
enforced solely by the per-chunk idle timeout
- Add 30s connect_timeout to the reqwest Client builder
- Increase CHUNK_IDLE_TIMEOUT from 30s to 120s for reasoning models
- Inject SSE error events (proxy_stream_error) before the HTTP chunked
terminator on all truncation paths so clients can detect data loss
- Wrap the streaming relay in BufWriter to reduce per-chunk TLS flush
overhead
- Bump OCSF severity for streaming truncation from Low to Medium
Closes#829
* fix(inference): drop BufWriter to preserve SSE incremental delivery
The BufWriter introduced in the previous commit buffered SSE frames until
the 16KB capacity filled or the stream ended, defeating incremental token
delivery and potentially reintroducing client-visible timeouts on healthy
streams. Revert to per-chunk write_all+flush but keep the single
format_chunk() call that coalesces framing into one write.
Also fix the streaming integration test: add a 3s mock delay that exceeds
the 1s route timeout so the test actually validates that the streaming
path omits the total request timeout. Previously the mock responded
immediately, passing regardless of timeout behavior.
The strip_headers list in send_backend_request included default_headers
names (e.g. anthropic-version), which removed the client's value. The
already_sent check then saw the original input still contained the header
and skipped injecting the route default. Neither value reached upstream.
Remove default_headers names from strip_headers so client values pass
through. The existing already_sent guard still injects defaults only when
the client omits them.
Adam Miller