fix: integration test readiness for fibp

[vieiralucas]

Summary

• Wire format bugs fixed: Five distinct mismatches between the JS SDK and the Rust server's FIBP wire format were causing all integration tests to fail or produce wrong results
• Test infrastructure fixed: Proto loading path doubling and parallel port collisions caused server startup failures in the test helper

What was broken and why

1. Auth payload (encodeAuthPayload)

JS sent u16 length + key bytes; server reads raw key bytes only. Auth always failed.

2. Error frame decoding (decodeErrorPayload)

JS expected u16 code + u16 msglen + msg; server sends raw UTF-8 message string. Error codes were garbage (e.g. 26990 = bytes "in" from "invalid"). Now decodes raw string and maps known message substrings to typed error codes.

3. Consume delivery wire format (decodeConsumeDelivery)

JS decoded id + queue + fairness_key + ...; server encodes id + fairness_key + ... (no queue field in push frame). All consumed messages had garbled data.

4. Consume stream dispatch (openStream → openConsumeStream)

JS registered a corrId-based stream handler but server push frames use corrId = 0 with FLAG_PUSH. The subscribe ack frame (corrId = N, no FLAG_PUSH) was incorrectly interpreted as "stream ended," closing the consumer immediately. Replaced with two-phase approach: await subscribe ack as a normal request, then register a global push handler for FLAG_PUSH frames.

5. Batcher connection lifecycle

Batcher held a sync getConnSync() callback; Client.enqueue() had an await getConn() before batcher.submit(). A concurrent close()/drain() could mark the batcher closed before submit ran. Fixed by making the batcher hold an async getConn() callback so enqueue() submits synchronously without any prior await.

6. Proto resolvePath double-prefix (test/helpers.ts)

protobufjs calls resolvePath("", "/absolute/path") for the initial loadSync call. The custom resolver prepended PROTO_DIR again, producing /proto/dir//proto/dir/file.proto. Added path.isAbsolute guard.

7. Port collision in parallel test runs (vitest.config.ts)

Multiple test workers called findFreePort() simultaneously, found the same port (TOCTOU), and both tried to start fila-server on it. Added fileParallelism: false to run test files sequentially.

Test plan

• npm run lint — passes
• npm run typecheck — passes
• npm test — 46/50 pass locally; 4 TLS tests skip (local binary built without rustls crypto provider, CI binary has it configured)
• CI with downloaded pre-built binary — all tests expected to pass

🤖 Generated with Claude Code

---

Summary by cubic

Fixes FIBP wire format mismatches and push‑stream routing to match the Rust server, unblocking integration tests. Hardens consume streaming to avoid dropped pushes and block concurrent streams.

• Bug Fixes
  • Auth payload: send raw UTF‑8 API key bytes (no length prefix).
  • Error frames: decode raw UTF‑8 message; map known substrings to ErrCode.
  • Consume delivery: match server push format (no queue field; id + fairness_key + ...).
  • Consume stream: register global push handler before subscribe ack; route FLAG_PUSH frames with corrId=0; expose cancel; reject concurrent streams.
  • Batcher: hold async getConn() so enqueue() queues before any await, avoiding close/drain races.
  • Test infra: guard absolute paths in protobufjs resolver; set vitest fileParallelism: false to avoid port collisions.

^{Written for commit 8f3541d. Summary will update on new commits.}

[cubic-dev-ai]

2 issues found across 6 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="src/transport.ts">

<violation number="1" location="src/transport.ts:522">
P1: Race condition: PUSH frames received in the same TCP chunk as the subscribe ACK will be silently dropped because `this.pushHandler` is registered after the `await`.</violation>

<violation number="2" location="src/transport.ts:556">
P1: Missing guard for concurrent consume streams. Opening a second stream on the same connection silently overwrites `this.pushHandler`, causing the first stream to hang and misrouting its messages.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}