fix(discovery): resolve duplicate /v1 paths and redirect aborts (#5904)

Integrated into release/v3.8.44. Thanks @hamsa0x7 for diagnosing the doubled /v1 discovery path and the REDIRECT_BLOCKED probe-loop abort (#5899). De-scoped to the discovery fix (the #5903 session-affinity work is handled by #5943) and added Rule #18 regression guards.

CHANGELOG.md

@@ -12,7 +12,7 @@ _TBD_
 
 ### 🔧 Bug Fixes
 
-_TBD_
+- **fix(providers): Api Airforce model discovery no longer produces a doubled `/v1` path** — a base URL ending in `/v1/chat/completions` (e.g. `https://api.airforce/v1/chat/completions`) was only stripped of `/chat/completions`, leaving a trailing `/v1` that the endpoint builder then doubled into `…/v1/v1/models`. That 308 redirect was surfaced as `REDIRECT_BLOCKED` and aborted the whole discovery probe loop before the correct `…/v1/models` candidate. The `/v1` suffix is now stripped independently (guarding a host literally named `v1`), and a `REDIRECT_BLOCKED` on one candidate continues to the next endpoint instead of aborting. Regression guards: `tests/unit/provider-models-route.test.ts`. ([#5904](https://github.com/diegosouzapw/OmniRoute/pull/5904) — thanks [@hamsa0x7](https://github.com/hamsa0x7)). Also reported/fixed independently by [@anki1kr](https://github.com/anki1kr) in [#5920](https://github.com/diegosouzapw/OmniRoute/pull/5920).
 
 ### 📝 Maintenance
 

config/quality/file-size-baseline.json

@@ -294,7 +294,7 @@
     "_rebaseline_2026_06_27_5193_antigravity_test": "#5193 own test growth: oauth-providers-config.test.ts 870->873 (+3: antigravity projectId assertion + 50ms tick for the now fire-and-forget onboarding, matching the no-PKCE/no-openid flow).",
     "tests/unit/oauth-providers-config.test.ts": 873,
     "tests/unit/perplexity-web.test.ts": 959,
-    "tests/unit/provider-models-route.test.ts": 1628,
+    "tests/unit/provider-models-route.test.ts": 1752,
     "tests/unit/provider-validation-specialty.test.ts": 2874,
     "_rebaseline_pr4613_compatible_provider_groups": "Reconcile #4613 already-merged growth: providers-page-utils.test.ts 1004->1052 (+48, buildCompatibleProviderGroups partition unit test). Fast-gate PR->release does not run check:file-size, so this surfaced post-merge.",
     "tests/unit/providers-page-utils.test.ts": 1052,
@@ -356,6 +356,7 @@
   "_rebaseline_2026_06_17_4107_pending_reaper": "PR #4107 own growth: usageHistory.ts 854->934 (+80 = orphaned-pending-request reaper — sweepStalePendingRequests() evicts pending details older than 15min + a hard 5000 cap, plus an unref'd 5min sweep timer wired lazily into trackPendingRequest). Fixes an unbounded memory leak where abnormally-terminated requests left payload previews in pendingById forever. Cohesive with the existing pending-request bookkeeping (mirrors the normal removal path: decrement counters + cleanup buckets); not extractable.",
   "_rebaseline_2026_06_17_4116_combo_hedge_listener": "combo.ts: +9 lines from #4116 (detach per-target listener from shared hedge abort signal to fix a listener leak). Behavior-preserving cleanup; 5289 -> 5298.",
   "_rebaseline_2026_06_20_4355_gpt5x_pro_pricing": "PR #4355 own growth: pricing.ts 1581->1592 (+11 = pure-data pricing rows for openai gpt-5.5-pro + gpt-5.4-pro, closing the $0 gap that tripped the catalog pricing gate after the #4324 sweep added them to the registry; -pro mirrors its base family tier). provider-models-route.test.ts 1616->1618 (+2 = test-only alignment to the intentional opencode-go discovery behavior: owned_by stamp + T39 two-endpoint fail-path fetchCalls). Both are data/test-only; not extractable.",
+  "_rebaseline_2026_07_02_5899_airforce_v1_discovery": "PR #5904 own growth: provider-models-route.test.ts 1628->1752 (+124 = test-only Rule #18 regression guards for the Api Airforce /v1/v1/models discovery bug (#5899): (a) a baseUrl ending in /v1/chat/completions must probe .../v1/models not the doubled .../v1/v1/models, and the host-guard case http://v1; (b) a REDIRECT_BLOCKED on one candidate must continue to the next endpoint instead of aborting the probe loop. Both guards fail on the pre-fix code. Test-only additions cohesive with the existing provider-models discovery suite (shared seedConnection/callRoute harness); not separately extractable without duplicating the harness.",
   "_rebaseline_2026_06_19_4293_codex_spark_scope": "PR #4293 (isolate Codex Spark quota scope) own growth, MEASURED on the actual merged tree (release/v3.8.30 + #4293). Production: auth.ts 2219->2279 (+60) threads requestedModel into Codex quota-policy/headroom/preflight/P2C scoring so normal Codex and GPT-5.3-Codex-Spark windows are evaluated independently; chatCore.ts 5116->5125 (+9) passes the failing model scope into Codex 429 failover (markCodexScopeRateLimited) instead of a connection-wide rateLimitedUntil write; accountFallback.ts 1727->1731 (+4) scopes Codex model-lock keys to codex vs spark. Heavy parsing/display logic lives in new leaf helpers under the cap (codexQuotaScopes.ts, codexUsageQuotas.ts, codexFailover.ts). Tests: account-fallback-service 1544->1569, executor-codex 1336->1339, sse-auth 1527->1553, usage-service-hardening 1612->1633 (added Spark-scope regression coverage). Cohesive wiring at existing selection/failover lockout boundaries; not extractable.",
   "_rebaseline_2026_06_20_4447_openai_gpt41mini_o_mini_pricing": "PR #4447 own growth: pricing.ts 1592->1620 (+28 = pure-data pricing rows closing the null/$0 gap for registry-exposed OpenAI ids gpt-4.1-mini, gpt-4.1-nano, o3-mini, o4-mini that tripped the catalog pricing gate; getPricingForModel does an exact lookup, so a missing key resolves to null. Official OpenAI per-1M prices + the table's derived-field convention (reasoning=output*1.5, cache_creation=input, cached=official). Restore-green for a pre-existing release/v3.8.32 red surfaced by #4432's __RUN_ALL__ run. Cohesive data; not extractable.",
   "_rebaseline_2026_06_20_web_cookie_validator_shadow_fix": "validation.ts 4518->4522 (+4 = move the generic web-cookie validateWebCookieProvider dispatch from the TOP of validateProviderApiKey to a FALLBACK after SPECIALTY_VALIDATORS, plus a comment, so #4023's generic AUTH_007 ping no longer shadows the rich per-provider validators (grok-web #3474 IP-reputation/Cloudflare, chatgpt-web cf-mitigated, claude/gemini/copilot/qwen/t3-web). Restores provider-validation-specialty.test.ts (112/112) while keeping web-cookie-auth007 (5/5). Behavior fix at an existing dispatch boundary; not extractable.",

src/app/api/providers/[id]/models/route.ts

@@ -533,7 +533,9 @@ export async function GET(
         base = base.slice(0, -17);
       } else if (base.endsWith("/completions")) {
         base = base.slice(0, -12);
-      } else if (base.endsWith("/v1")) {
+      }
+
+      if (base.endsWith("/v1") && !base.endsWith("://v1")) {
         base = base.slice(0, -3);
       }
 
@@ -576,6 +578,11 @@ export async function GET(
           }
         } catch (err: any) {
           if (err.message === "auth_failed") break; // Don't try other endpoints if auth failed
+
+          if (err?.code === "REDIRECT_BLOCKED") {
+            continue; // Try next endpoint
+          }
+
           const status = getSafeOutboundFetchErrorStatus(err);
           if (status) {
             throw err;

tests/unit/provider-models-route.test.ts

@@ -320,6 +320,130 @@ test("provider models route discovers SiliconFlow models from configured China b
   ]);
 });
 
+test("provider models route handles local hostnames named 'v1' correctly", async () => {
+  const connection = await seedConnection("openai-compatible-local-v1", {
+    apiKey: "sk-local",
+    providerSpecificData: {
+      baseUrl: "http://v1/chat/completions",
+    },
+  });
+  const seenUrls: string[] = [];
+
+  globalThis.fetch = async (url) => {
+    seenUrls.push(String(url));
+    return Response.json({
+      data: [{ id: "local-v1-model", name: "Local v1 Model" }],
+    });
+  };
+
+  const response = await callRoute(connection.id);
+  const body = (await response.json()) as any;
+
+  assert.equal(response.status, 200);
+  assert.equal(body.source, "api");
+  assert.deepEqual(seenUrls, ["http://v1/v1/models"]);
+});
+
+test("provider models route correctly strips standard /v1 paths", async () => {
+  const connection = await seedConnection("openai-compatible-standard-v1", {
+    apiKey: "sk-standard",
+    providerSpecificData: {
+      baseUrl: "https://api.openai.com/v1",
+    },
+  });
+  const seenUrls: string[] = [];
+
+  globalThis.fetch = async (url) => {
+    seenUrls.push(String(url));
+    return Response.json({
+      data: [{ id: "standard-model", name: "Standard Model" }],
+    });
+  };
+
+  const response = await callRoute(connection.id);
+  const body = (await response.json()) as any;
+
+  assert.equal(response.status, 200);
+  assert.equal(body.source, "api");
+  assert.deepEqual(seenUrls, ["https://api.openai.com/v1/models"]);
+});
+
+test("provider models route strips /v1 when it precedes /chat/completions (#5899 no double /v1)", async () => {
+  // Regression for #5899 (Api Airforce): a baseUrl of the form
+  // "https://api.airforce/v1/chat/completions" must probe ".../v1/models" — NOT
+  // ".../v1/v1/models". The old `else if` strip chain only removed
+  // "/chat/completions", leaving a trailing "/v1" that the endpoint builder then
+  // doubled, producing a 308 redirect that aborted discovery.
+  const connection = await seedConnection("openai-compatible-airforce-v1", {
+    apiKey: "sk-airforce",
+    providerSpecificData: {
+      baseUrl: "https://api.airforce/v1/chat/completions",
+    },
+  });
+  const seenUrls: string[] = [];
+
+  globalThis.fetch = async (url) => {
+    seenUrls.push(String(url));
+    return Response.json({
+      data: [{ id: "airforce-model", name: "Airforce Model" }],
+    });
+  };
+
+  const response = await callRoute(connection.id);
+  const body = (await response.json()) as any;
+
+  assert.equal(response.status, 200);
+  assert.equal(body.source, "api");
+  // First probed endpoint must have a single /v1 — no ".../v1/v1/models".
+  assert.equal(seenUrls[0], "https://api.airforce/v1/models");
+  assert.ok(
+    !seenUrls.some((u) => u.includes("/v1/v1/")),
+    `no endpoint should contain a doubled /v1: ${JSON.stringify(seenUrls)}`
+  );
+});
+
+test("provider models route continues probing past a REDIRECT_BLOCKED endpoint (#5899)", async () => {
+  // Regression for #5899: a REDIRECT_BLOCKED error on one candidate endpoint must
+  // not abort the whole probe loop — discovery should fall through to the next
+  // endpoint instead of surfacing an empty catalog.
+  const connection = await seedConnection("openai-compatible-redirect-v1", {
+    apiKey: "sk-redirect",
+    providerSpecificData: {
+      baseUrl: "https://redirect.example",
+    },
+  });
+  const seenUrls: string[] = [];
+
+  globalThis.fetch = async (url) => {
+    const u = String(url);
+    seenUrls.push(u);
+    // First candidate ".../v1/models" answers with a real 308 redirect →
+    // safeOutboundFetch throws a SafeOutboundFetchError(REDIRECT_BLOCKED). The old
+    // code re-threw on it (status 503) and aborted the loop; the fix `continue`s.
+    if (u === "https://redirect.example/v1/models") {
+      return new Response(null, {
+        status: 308,
+        headers: { location: "https://redirect.example/models" },
+      });
+    }
+    return Response.json({
+      data: [{ id: "redirect-model", name: "Redirect Model" }],
+    });
+  };
+
+  const response = await callRoute(connection.id);
+  const body = (await response.json()) as any;
+
+  assert.equal(response.status, 200);
+  // Without the REDIRECT_BLOCKED `continue`, discovery aborted and fell back to a
+  // non-api catalog. The fix lets it reach the next endpoint and return live models.
+  assert.equal(body.source, "api");
+  assert.ok(
+    seenUrls.length >= 2,
+    `expected the loop to continue past REDIRECT_BLOCKED: ${JSON.stringify(seenUrls)}`
+  );
+});
+
 test("provider models route returns static catalog entries for providers with hardcoded models", async () => {
   const connection = await seedConnection("bailian-coding-plan", {
     apiKey: "bailian-key",