fix(api-proxy): fall back to next alias candidate when Copilot rejects model for endpoint by Copilot · Pull Request #6134 · github/gh-aw-firewall

Copilot · 2026-07-12T14:03:20Z

The Test Coverage Reporter workflow was failing all runs with 0 AIC because the summarization alias resolved to gpt-5.4-mini (highest semver from copilot/gpt-5*mini*), which the Copilot API permanently rejects on /chat/completions with 400 model "gpt-5.4-mini" is not accessible via the /chat/completions endpoint. The error was unrecognized and forwarded directly to the Copilot CLI.

Root cause

The gpt-5-mini alias pattern copilot/gpt-5*mini* matches both gpt-5-mini and gpt-5.4-mini. Version sorting picks gpt-5.4-mini (5.4 > 5.0) as the winner. The Copilot API lists it in the models catalogue but restricts it to /responses only.

Changes

New: endpoint-blocked detection + per-request candidate fallback

upstream-response.js: adds MODEL_ENDPOINT_BLOCKED_PATTERN (/not accessible via the .+? endpoint/i) and parseModelEndpointBlockedFromBody; buffers 400s for this case; threads onModelEndpointBlockedRetry callback through to handle400WithRetry
upstream-retry.js: new branch (b) fires onModelEndpointBlockedRetry() before the existing transient-model-not-supported branch — permanent endpoint restriction, no delay
upstream-http.js: implements onModelEndpointBlockedRetry — reads req.awfModelCandidates, finds current model in body, picks next ranked candidate, rewrites body, retries

Ranked candidates wired through resolution pipeline

model-resolver.js (_resolveAliasPatterns): returns candidates: unique (full sorted list, e.g. [gpt-5.4-mini, claude-haiku-4.5]) alongside resolvedModel
model-body-rewriter.js: propagates candidates from resolution result
model-config.js: body transform signature becomes async (body, req) — stores req.awfModelCandidates on the request for later use by retry logic
body-handler.js: passes req to bodyTransform(body, req)

Documentation

model-api-mapping.json (both containers/api-proxy/ and docs/): notes added to gpt-5.4* family documenting the Copilot /chat/completions restriction and the proxy's fallback behaviour

Fallback flow (after fix)

COPILOT_MODEL=summarization
→ alias resolves: resolvedModel=gpt-5.4-mini, candidates=[gpt-5.4-mini, claude-haiku-4.5]
→ req.awfModelCandidates = [gpt-5.4-mini, claude-haiku-4.5]
→ Copilot: 400 "gpt-5.4-mini not accessible via /chat/completions"
→ parseModelEndpointBlockedFromBody → true
→ onModelEndpointBlockedRetry(): next candidate = claude-haiku-4.5
→ body rewritten with model: claude-haiku-4.5 → retry → ✓ success

When the Copilot API rejects a resolved model with "not accessible via the /chat/completions endpoint" (e.g. gpt-5.4-mini), the api-proxy now falls back to the next ranked candidate from the alias resolution rather than returning the 400 error to the client. Root cause: The `summarization` alias resolves via the `gpt-5-mini` sub-alias which uses pattern `copilot/gpt-5*mini*`. This matches both `gpt-5-mini` and `gpt-5.4-mini` from the Copilot model list. The version sorter picks `gpt-5.4-mini` (v5.4 > v5.0) as the top candidate. But the Copilot API only supports `gpt-5.4-mini` via /responses, not via /chat/completions. The result is a 400 that previously was not recognized and was forwarded directly to the client (Copilot CLI), causing all three Test Coverage Reporter runs to fail with 0 AIC. Changes: - model-resolver.js: `_resolveAliasPatterns` now returns the full ranked `candidates` array alongside `resolvedModel` so callers have access to fallback options. - model-body-rewriter.js: propagates `candidates` from resolution result. - model-config.js: body transform now accepts (body, req) and stores `req.awfModelCandidates` for use during retry. - body-handler.js: passes `req` to `bodyTransform(body, req)`. - upstream-response.js: adds `MODEL_ENDPOINT_BLOCKED_PATTERN` matching "not accessible via the ... endpoint" errors; adds `parseModelEndpointBlockedFromBody`; threads `onModelEndpointBlockedRetry` callback through `handleUpstreamResponse` and `handle400WithRetry`. - upstream-retry.js: new branch (b) handles endpoint-blocked 400s by calling `onModelEndpointBlockedRetry()` before the transient retry. - upstream-http.js: implements `onModelEndpointBlockedRetry` — reads `req.awfModelCandidates`, finds current model in the body, picks the next ranked candidate, rewrites the body, and retries upstream. - model-api-mapping.json + docs/model-api-mapping.json: updated notes for gpt-5.4 family to document the Copilot /chat/completions restriction. - upstream-retry.test.js: adds 4 tests for endpoint-blocked fallback. - model-resolver.test.js: adds test verifying `candidates` is returned.

Copilot

Pull request overview

This PR improves the api-proxy’s Copilot model-alias handling by detecting “model not accessible via endpoint” 400s and retrying the same request with the next ranked alias candidate, preventing hard failures when the highest-semver candidate is responses-only on Copilot.

Changes:

Add endpoint-blocked (“not accessible via the … endpoint”) detection and a no-delay retry path that falls back to the next alias candidate for Copilot.
Thread ranked alias candidates through model resolution/body-rewrite so retry logic can reuse them without re-resolving.
Document the Copilot /chat/completions restriction for gpt-5.4-mini in model mapping notes.

Show a summary per file

File	Description
docs/model-api-mapping.json	Documents Copilot endpoint restriction + fallback behavior for `gpt-5.4-mini`.
containers/api-proxy/model-api-mapping.json	Mirrors the same restriction + fallback note for the runtime mapping file.
containers/api-proxy/upstream-retry.test.js	Adds unit coverage for the endpoint-blocked retry branch.
containers/api-proxy/upstream-retry.js	Adds endpoint-blocked retry branch ahead of transient model-not-supported retry.
containers/api-proxy/upstream-response.js	Buffers relevant 400 bodies and wires endpoint-blocked parsing + callback into retry handling.
containers/api-proxy/upstream-http.js	Implements request-body rewrite + resend using next ranked candidate.
containers/api-proxy/model-resolver.test.js	Verifies ranked candidates are returned to support fallback.
containers/api-proxy/model-resolver.js	Returns full ranked candidate list alongside `resolvedModel`.
containers/api-proxy/model-body-rewriter.js	Propagates `candidates` from resolver into rewrite result.
containers/api-proxy/model-config.js	Stores ranked candidates on `req` during body transform for later retry use.
containers/api-proxy/body-handler.js	Passes `req` into `bodyTransform` to allow storing per-request state.

Review details

Tip

Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Files reviewed: 11/11 changed files
Comments generated: 2
Review effort level: Low

+          // Determine which model was sent in the current body.
+          const parsed = parseBodyAsObject(body);
+          const currentModel = parsed && parsed.model;
+          if (!currentModel) return false;
+
+          const currentIdx = candidates.indexOf(currentModel);
+          if (currentIdx < 0 || currentIdx >= candidates.length - 1) return false;
+
+          const nextModel = candidates[currentIdx + 1];
+
+          // Rewrite the body with the next candidate.
+          const newParsed = parseBodyAsObject(body);
+          if (!newParsed) return false;
+          newParsed.model = nextModel;
+          const newBody = Buffer.from(JSON.stringify(newParsed), 'utf8');
+
+          // Update the candidates list so if the next model also fails we can
+          // continue falling back (by shifting the current index forward).
+          sendUpstreamRequest(requestHeaders, {
+            body: newBody, targetHost, upstreamPath, req, res, provider, requestId, startTime, span, requestBytes,
+            hasRetried,
+            modelNotSupportedRetryCount,
+          });
+          return true;
+        },


  async function transformRequestBody(body, provider, req, requestId, bodyTransform) {
    if (bodyTransform && (req.method === 'POST' || req.method === 'PUT' || req.method === 'PATCH')) {
-      const transformed = await bodyTransform(body);
+      const transformed = await bodyTransform(body, req);


Initial plan

9d825ef

Copilot AI assigned Copilot and lpcox Jul 12, 2026

Copilot started work on behalf of lpcox July 12, 2026 14:03 View session

Copilot AI linked an issue Jul 12, 2026 that may be closed by this pull request

📊 Copilot Token Usage Report2026-07-12 #6128

Open

Copilot AI changed the title ~~[WIP] Fix missing token usage fields in report~~ fix(api-proxy): fall back to next alias candidate when Copilot rejects model for endpoint Jul 12, 2026

Copilot AI requested a review from lpcox July 12, 2026 14:39

Copilot finished work on behalf of lpcox July 12, 2026 14:39

lpcox marked this pull request as ready for review July 12, 2026 15:25

Copilot AI review requested due to automatic review settings July 12, 2026 15:25

Copilot started reviewing on behalf of lpcox July 12, 2026 15:25 View session

Copilot AI reviewed Jul 12, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix(api-proxy): fall back to next alias candidate when Copilot rejects model for endpoint#6134

fix(api-proxy): fall back to next alias candidate when Copilot rejects model for endpoint#6134
lpcox with Copilot wants to merge 2 commits into
mainfrom
copilot/copilot-token-usage-report

Copilot AI commented Jul 12, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Conversation

Copilot AI commented Jul 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Root cause

Changes

New: endpoint-blocked detection + per-request candidate fallback

Ranked candidates wired through resolution pipeline

Documentation

Fallback flow (after fix)

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Review details

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Copilot AI commented Jul 12, 2026 •

edited

Loading