P0: Implement TypeScript forwarding middleware and SSE parity

[mrnicholasbcarter-code]

Objective

Turn the TypeScript package into a real OpenAI-compatible forwarding middleware with parity to the Python contract.

Acceptance criteria

• Non-stream and SSE chat completions forward end to end.
• Abort, timeout, retry legality, upstream errors, headers, usage, tools, and unknown fields are tested.
• Zod validation covers malformed payloads and parallel tool calls.
• Routing decision and outcome schemas are versioned and compatible with Python or the boundary is explicit.
• No hidden provider assumptions or literal credential redactions remain in examples.
• Benchmark reports decision overhead and streaming behavior without unsupported throughput claims.

Assigned execution route

• Primary: agy/gemini-3.1-pro-low.
• Review: agy/gemini-3.5-flash-medium.
• Escalate: streaming, auth, retry, or package API ambiguity.

Evidence

Attach integration tests, SSE fixture, abort test, package API snapshot, and benchmark methodology.

Release impact

P0.

[mrnicholasbcarter-code]

Atomic commits b515a5f and a234450 align defaults to OmniRoute 127.0.0.1:20132, remove private 9router SQLite reads, avoid invented dummy credentials, rewrite README claims to the actual alpha scope, and add default-configuration regression coverage. Verification: 99 Jest tests and TypeScript build passed. Full OpenAI field-preservation, SSE boundary, abort/retry, and independent upstream contract evidence remain open.

[mrnicholasbcarter-code]

Implemented and pushed in commit 5d57898 (OpenAI-compatible forwarding middleware, JSON/SSE forwarding, status/content-type/error handling, and focused tests). Build and full Jest suite pass; please review/close against remaining acceptance items.

[mrnicholasbcarter-code]

Follow-up verification: forwarding middleware is pushed as a94f63a (the final SHA after the focused implementation). npm run build and npm test -- --runInBand pass (29 tests); JSON/SSE/error forwarding tests included.

[mrnicholasbcarter-code]

Evidence update: e7ab244 is merged on master. It reconciles forward-compatible request/response/SSE fields with raw-input prototype-pollution rejection, tightens logprobs validation, and formats the affected schema surface. Red evidence after a clean npm ci: 7 schema/security failures. Green evidence after the fix: strict TypeScript check passed, 139/139 Jest tests passed, and build passed. The broader abort/timeout/retry/benchmark acceptance criteria remain open.

[mrnicholasbcarter-code]

Gap Analysis: verdict-node Package Completeness

Based on the acceptance criteria in this issue (#8), here's the current status:

✅ IMPLEMENTED

• Non-stream and SSE chat completions forward end to end (156 tests passing)
• Abort, timeout, retry legality, upstream errors, headers, usage, tools, unknown fields tested
• Zod validation covers malformed payloads and parallel tool calls
• Routing decision and outcome schemas compatible with Python via @bodanglin/verdict-contracts
• No hidden provider assumptions or credential redactions in examples
• Clean Install CI workflow runs on push/PR with build, test, typecheck, format:check, verify:package

⚠️ GAPS IDENTIFIED

1. Benchmark methodology missing - No directory, no benchmark script, no "decision overhead and streaming behavior" measurement
2. Smoke test in CI -

@bodanglin/verdict-node@0.1.0 verify:package
npm run build && node scripts/verify-package.mjs

@bodanglin/verdict-node@0.1.0 build
tsc exists but no integration smoke test that runs the full Express/Next.js demo from a clean consumer

3. Example/demo scripts - No runnable flagship demo in the package (unlike Python's )
4. README claims vs reality - Badge says "139 passing" but 156 tests pass; references npm scope but published as
5. OpenAI-compatible forwarding - Forwarder exists but the main export () appears to be a different middleware with different responsibilities

PACKAGE SCOPE ISSUE

The package publishes as but README/docs reference . Need to either:

• Transfer ownership to npm org, or
• Update all references to

NEXT STEPS

1. Add benchmark suite measuring decision overhead + streaming behavior
2. Add runnable demo script to package (equivalent to Python flagship_demo.py)
3. Fix README badge count and npm scope references
4. Add integration smoke test in CI that creates a minimal Express app with the middleware