Fix: inference-parser misses /chat/completions from /v1-less proxies (opencode, litellm)

[msteinder]

Why

OpenAI-compatible proxies such as litellm (used by opencode) post to /chat/completions without the /v1 prefix. The OnRequest dispatch switch in inference-parser only matched /v1/chat/completions and /v1/completions, so these requests fell through to the default: return Continue arm — pctx.Extensions.Inference was never populated and abctl showed — for model, method, and token counts on every opencode request.

Observed before the fix:

9    14:13:56.91   out   │┌ req   —    —                    ete-litellm.ai-mode…
9    14:13:59.27   out   │└ resp  —    —    200  2.36s       ete-litellm.ai-mode…

What changed

authbridge/authlib/plugins/inferenceparser/plugin.go

• Extended the OnRequest switch case to also match /chat/completions and /completions.

authbridge/authlib/plugins/inferenceparser/plugin_test.go

• Added TestInferenceParser_VlessPath_ChatCompletions and TestInferenceParser_VlessPath_Completions.

No changes to OnResponse or OnResponseFrame — the default branch there already routes to the OpenAI parsers, which is correct.

Testing

$ cd authbridge/authlib && go test ./plugins/inferenceparser/...
ok  github.com/rossoctl/cortex/authbridge/authlib/plugins/inferenceparser

Verified end-to-end: opencode requests through authbridge-proxy --demo now show model, token counts, and method in abctl.

Related issue(s)

Fixes #822

Assisted-By: Claude (Anthropic AI) noreply@anthropic.com

Summary by CodeRabbit

• Bug Fixes
  • Added support for OpenAI-compatible /chat/completions and /completions endpoints without the /v1 prefix.
  • Requests through compatible proxies are now correctly parsed, including model details and action tracking.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 430ff944-b13b-46c3-b1e0-c94f6351ec9a

📥 Commits

Reviewing files that changed from the base of the PR and between cb8bceb and 4bcb677.

📒 Files selected for processing (2)

• authbridge/authlib/plugins/inferenceparser/plugin.go
• authbridge/authlib/plugins/inferenceparser/plugin_test.go

Included review availability: Your plan provides up to 1 included review per hour; 0 remain after this review.

---

📝 Walkthrough

Walkthrough

The inference parser now handles OpenAI-compatible /chat/completions and /completions paths without the /v1 prefix. Tests verify inference metadata, model extraction, and parser continuation.

Changes

OpenAI endpoint support

Layer / File(s)	Summary
Unversioned endpoint dispatch and validation authbridge/authlib/plugins/inferenceparser/plugin.go, authbridge/authlib/plugins/inferenceparser/plugin_test.go	The request parser accepts unversioned chat and completion paths. Tests verify parsed models, inference extensions, action state, and Continue results.

Estimated code review effort: 2 (Simple) | ~10 minutes

Merge Risk: ⚪ Minimal · up to 4bcb6

The change enables metadata extraction for two additional compatible completion paths using the existing parsing behavior, with no change to authorization, routing, or request handling. No actionable merge-blocking risk remains.

Suggested reviewers: huang195, ibrahim2595

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 25.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 4 functions across 2 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly identifies the inference-parser fix for /v1-less OpenAI-compatible proxy paths. It is concise and directly related to the primary change.
Linked Issues check	✅ Passed	The changes satisfy issue #822 by adding /chat/completions and /completions to OnRequest path matching. The added tests verify inference parsing for both paths, including model capture and action stat…
Out of Scope Changes check	✅ Passed	All code and test changes directly support issue #822. No unrelated or out-of-scope changes are present.

Full details: Linked Issues check

Explanation

The changes satisfy issue #822 by adding /chat/completions and /completions to OnRequest path matching. The added tests verify inference parsing for both paths, including model capture and action status.

• Fix all pre-merge checks with AI

✨ Finishing Touches 💡 1

📝 Generate docstrings 💡

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/inference-parser-vless-paths

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[mrsabath]

Summary

Correct, minimal, well-scoped fix. Extends the OnRequest dispatch to match the /v1-less /chat/completions and /completions variants that litellm/opencode emit, restoring model and token telemetry in abctl.

What I verified against the source (PR head 4bcb677):

• endpointPath() strips the query string via strings.Cut(path, "?"), so the new exact-match cases stay robust for /chat/completions?foo=bar.
• Response side needs no change: OnResponse gates only on anthropicMessagesPath and defaults everything else to the OpenAI parsers, but it only runs when ext was populated on the request. So the request-switch fix is what unblocks the response telemetry too, end-to-end.
• Legacy /completions (prompt-based) routed to parseOpenAIRequest still captures model + sampling params + IsAction correctly, which is exactly the telemetry intended.
• Tests assert the real regression (Extensions.Inference != nil, correct Model, IsAction) using the same pipeline.Context{Path:...} shape the code reads.

My only note is the symmetric /messages gap for Anthropic-compatible proxies (inline), which is a non-blocking follow-up.

Areas reviewed: Go (inference-parser plugin + tests), Tests, commit/PR conventions.
Commits: 1, signed-off, conventional prefix, under 72 chars. CI: all green (Go CI x3, CodeQL go/python, Bandit, Trivy, pre-commit, CodeRabbit).

Verdict: APPROVE — no must-fix issues; the single comment is a suggestion for a follow-up.

[mrsabath]

suggestion (non-blocking): same-class gap on the Anthropic side. anthropicMessagesPath is /v1/messages, so a /v1-stripping proxy (this PR's exact scenario) posting Anthropic-style traffic to /messages would still fall through to the default arm and record no inference telemetry. Out of scope for #822, but worth a follow-up (or a matching /messages case here) so the asymmetry is a deliberate choice rather than an oversight.