fix(translator): preserve reasoning-replay reasoning_content + repair 3 release-green test reds (#5122)

Repairs 3 release-green test reds + test-masking; unblocks #5078.

CHANGELOG.md

@@ -44,6 +44,7 @@ _In development — bullets added per PR; finalized at release._
 - **fix(sse): keep `output_text`-only Responses bodies from being dropped/false-502'd** — some upstreams return a shorthand Responses body whose answer is only in `output_text` with an empty `output[]`. `sanitizeResponsesApiResponse` discarded the text, so the response then tripped the malformed-200 guard. The sanitizer now synthesizes an `output[]` message item from a non-empty `output_text` (complements the Claude-native fix in #5108; both stem from #4942).
 - **fix(executors): preserve a lone caller-supplied `Anthropic-Version` header casing** — the case-variant dedupe (#4846) unconditionally rewrote `Anthropic-Version`/`Anthropic-Beta` to lowercase even when only one variant was present, clobbering the caller's header. Dedupe now runs only when both case variants coexist (the actual undici-merge collision it was meant to fix).
 - **fix(responses):** default `text.format` to `{ type: "text" }` for openai-compatible **responses** providers — some Responses-compatible upstreams (e.g. LM Studio) reject a `text` object missing `text.format` with a 400 `missing_required_parameter`; the default executor now fills the Responses-API default before forwarding (guarded to `openai-compatible-*responses*`, never overwriting an existing format). (thanks @StevanusPangau)
+- **fix(translator): stop stripping client-provided `reasoning_content` for reasoning-replay providers** — the #4849 agentic-context strip (which drops `reasoning_content` from tool-call assistant turns to avoid O(n²) token growth) ran unconditionally, so replay providers (DeepSeek V4, Kimi K2, Qwen-Thinking, etc.) lost the client's reasoning and the reasoning-replay cache then overwrote it with a stale cached value (and such upstreams 400 without the original reasoning). The strip now skips reasoning-replay targets while non-reasoning providers keep the O(n²) protection.
 
 ---
 

config/quality/file-size-baseline.json

@@ -1,5 +1,6 @@
 {
   "_comment": "Catraca de tamanho (check-file-size.mjs). frozen so pode encolher; arquivos novos <= cap. --update ratcheta.",
+  "_rebaseline_2026_06_26_relgreen_db_test": "Release-green follow-up: tests/unit/db-core-init.test.ts 867->877 (+10 = the invalid-DATA_DIR test now captures the rejection and asserts both Error type AND message — restores net-neutral assert count after #5117's consolidation, satisfying check:test-masking — instead of a single assert.rejects).",
   "_rebaseline_2026_06_26_5074_fusion_editor": "PR #5074 own growth: src/app/(dashboard)/dashboard/combos/page.tsx 4485->4594 (+109 = the Fusion judgeModel + fusionTuning editor block — text/number inputs wired through updateFusionTuning, schema-validated) and tests/unit/combo-config.test.ts 800->881 (testFrozen add: 5 new Fusion-config schema tests). Cohesive UI block at the strategy-conditional editor; combos/page.tsx structural shrink tracked in #3501. Covered by combo-config.test.ts.",
   "_rebaseline_2026_06_26_5101_responses_textformat": "PR #5101 own growth: open-sse/executors/default.ts 859->876 (+17 = defaultResponsesTextFormat — fills the Responses-API default text.format for openai-compatible-*responses* providers so LM-Studio-style upstreams stop 400ing; guarded, never overwrites an existing format). Irreducible executor chokepoint next to applyJsonSchemaFallback; covered by tests/unit/responses-default-text-format.test.ts.",
   "_rebaseline_2026_06_26_5117_basereds": "Base-red repair #5117 own growth + sibling drift: open-sse/handlers/responseSanitizer.ts 1122->1139 (+17 = synthesize an output[] message item from an output_text-only Responses body so the answer is not dropped/false-502'd; #4942 regression, covered by tests/unit/response-sanitizer.test.ts). src/app/(dashboard)/dashboard/providers/[id]/components/modals/AddApiKeyModal.tsx 866->868 (+2 sibling drift from a separately-merged PR, already on the release tip; rebaselined here to green the file-size gate).",
@@ -253,7 +254,7 @@
     "tests/unit/chatgpt-web.test.ts": 2855,
     "tests/unit/combo-routing-engine.test.ts": 3213,
     "tests/unit/combo-strategy-fallbacks.test.ts": 880,
-    "tests/unit/db-core-init.test.ts": 867,
+    "tests/unit/db-core-init.test.ts": 877,
     "tests/unit/db-migration-runner.test.ts": 1491,
     "tests/unit/db-settings-crud.test.ts": 941,
     "tests/unit/deepseek-web.test.ts": 1081,

open-sse/translator/helpers/openaiHelper.ts

@@ -37,6 +37,11 @@ export function filterToOpenAIFormat(body, opts = {}) {
   // requested upstream, keep the `cache_control` field on each content block
   // instead of destructuring it away. `signature` is always stripped.
   const preserveCacheControl = opts?.preserveCacheControl === true;
+  // #4849 strips reasoning_content from tool-call assistant turns to stop O(n^2)
+  // context growth — but reasoning-replay providers (DeepSeek V4, Kimi K2, etc.)
+  // REQUIRE the client's reasoning_content to be passed back, so keep it for them
+  // (the caller sets this when the routed model needs reasoning replay).
+  const preserveReasoningContent = opts?.preserveReasoningContent === true;
   if (!body.messages || !Array.isArray(body.messages)) return body;
 
   body.messages = body.messages.map((msg) => {
@@ -50,8 +55,10 @@ export function filterToOpenAIFormat(body, opts = {}) {
 
     // Keep assistant messages with tool_calls, but strip reasoning_content —
     // reasoning blobs inflate context on every subsequent agentic turn (O(n^2)).
+    // Exception: reasoning-replay providers must keep client-provided
+    // reasoning_content (they 400 without it), so preserve it when requested.
     if (msg.role === "assistant" && msg.tool_calls) {
-      if (msg.reasoning_content !== undefined) {
+      if (!preserveReasoningContent && msg.reasoning_content !== undefined) {
         const { reasoning_content, ...cleanMsg } = msg;
         return cleanMsg;
       }

open-sse/translator/index.ts

@@ -264,6 +264,23 @@ export function translateRequest(
     }
   }
 
+  // Resolve reasoning-replay status up-front: it gates both the reasoning_content
+  // strip in filterToOpenAIFormat below (#4849 must NOT strip client reasoning for
+  // replay providers) and the cache re-injection further down.
+  const normalizedProvider = String(provider ?? "");
+  const normalizedModel = String(model ?? "");
+  const resolvedCapabilities = getResolvedModelCapabilities({
+    provider: normalizedProvider,
+    model: normalizedModel,
+  });
+  const isReasoner = requiresReasoningReplay({
+    provider: normalizedProvider,
+    model: normalizedModel,
+    thinkingEnabled: hasThinkingConfig(result),
+    supportsReasoning: supportsReasoning({ provider: normalizedProvider, model: normalizedModel }),
+    interleavedField: resolvedCapabilities?.interleavedField ?? null,
+  });
+
   // Always normalize to clean OpenAI format when target is OpenAI
   // This handles hybrid requests (e.g., OpenAI messages + Claude tools)
   if (targetFormat === FORMATS.OPENAI) {
@@ -274,6 +291,8 @@ export function translateRequest(
       preserveCacheControl:
         options?.preserveCacheControl === true &&
         providerHonorsOpenAIFormatCacheControl(provider),
+      // #4849 regression guard: keep client reasoning_content for replay providers.
+      preserveReasoningContent: isReasoner,
     });
   }
 
@@ -333,19 +352,9 @@ export function translateRequest(
   // clients omit it from the conversation history. Without this, DeepSeek V4
   // returns 400: "The reasoning_content in the thinking mode must be passed
   // back to the API."
-  const normalizedProvider = String(provider ?? "");
-  const normalizedModel = String(model ?? "");
-  const resolvedCapabilities = getResolvedModelCapabilities({
-    provider: normalizedProvider,
-    model: normalizedModel,
-  });
-  const isReasoner = requiresReasoningReplay({
-    provider: normalizedProvider,
-    model: normalizedModel,
-    thinkingEnabled: hasThinkingConfig(result),
-    supportsReasoning: supportsReasoning({ provider: normalizedProvider, model: normalizedModel }),
-    interleavedField: resolvedCapabilities?.interleavedField ?? null,
-  });
+  // isReasoner / normalizedProvider / normalizedModel / resolvedCapabilities were
+  // resolved up-front (before the OpenAI-format filter) so the #4849 reasoning strip
+  // could honor reasoning-replay providers.
   if (isReasoner && result.messages && Array.isArray(result.messages)) {
     const canReplayReasoningOnly = isReasoningOnlyReplayTarget(normalizedProvider, normalizedModel);
 

tests/snapshots/provider/translate-path.json

@@ -1156,6 +1156,29 @@
       "stream": "devin://acp/stdio"
     }
   },
+  "dgrid": {
+    "format": "openai",
+    "headers": {
+      "apiKey": {
+        "Accept": "text/event-stream",
+        "Authorization": "Bearer <TOK>",
+        "Content-Type": "application/json"
+      },
+      "nonStream": {
+        "Authorization": "Bearer <TOK>",
+        "Content-Type": "application/json"
+      },
+      "oauth": {
+        "Accept": "text/event-stream",
+        "Authorization": "Bearer <TOK>",
+        "Content-Type": "application/json"
+      }
+    },
+    "url": {
+      "nonStream": "https://api.dgrid.ai/v1/chat/completions",
+      "stream": "https://api.dgrid.ai/v1/chat/completions"
+    }
+  },
   "dify": {
     "format": "openai",
     "headers": {
@@ -3126,6 +3149,29 @@
       "stream": "https://www.phind.com/api/chat"
     }
   },
+  "pioneer": {
+    "format": "openai",
+    "headers": {
+      "apiKey": {
+        "Accept": "text/event-stream",
+        "Content-Type": "application/json",
+        "x-api-key": "<CRED>"
+      },
+      "nonStream": {
+        "Content-Type": "application/json",
+        "x-api-key": "<CRED>"
+      },
+      "oauth": {
+        "Accept": "text/event-stream",
+        "Content-Type": "application/json",
+        "x-api-key": "<CRED>"
+      }
+    },
+    "url": {
+      "nonStream": "https://api.pioneer.ai/v1/chat/completions",
+      "stream": "https://api.pioneer.ai/v1/chat/completions"
+    }
+  },
   "pollinations": {
     "format": "openai",
     "headers": {

tests/unit/codex-quota-sync-no-proactive-refresh.test.ts

@@ -73,7 +73,11 @@ test("non-rotating OAuth provider is still refreshed proactively from quota-sync
   const origRefresh = exec.refreshCredentials;
   let refreshCalls = 0;
   exec.needsRefresh = () => true;
-  // null -> non-github -> surfaces a 401; proves the refresh was actually attempted.
+  // Refresh returns null with no usable access token left -> surfaces a 401;
+  // proves the refresh was actually attempted (the codex no-rotate gate did not
+  // block this non-rotating provider). Note: since #4786 a *present* accessToken
+  // would gracefully fall back for any OAuth provider, so we leave it empty here
+  // to keep the 401-throw branch reachable and the gate assertion meaningful.
   exec.refreshCredentials = async () => {
     refreshCalls++;
     return null;
@@ -83,7 +87,7 @@ test("non-rotating OAuth provider is still refreshed proactively from quota-sync
       refreshAndUpdateCredentials({
         id: "cursor-1",
         provider: "cursor",
-        accessToken: "a",
+        accessToken: "",
         refreshToken: "r",
         tokenExpiresAt: new Date(2000).toISOString(),
         providerSpecificData: {},

tests/unit/db-core-init.test.ts

@@ -515,8 +515,18 @@ test("invalid DATA_DIR (a file where a dir is expected) surfaces as a startup fa
     // regular file is a non-permission misconfiguration (EEXIST/ENOTDIR), which
     // resolveWritableDataDir rethrows by design (only EACCES/EPERM fall back), so
     // the failure now surfaces at import time, not lazily from getDbInstance().
-    await assert.rejects(
-      withEnv({ DATA_DIR: fileAsDir }, () => importFresh("src/lib/db/core.ts")),
+    let caught: unknown;
+    await withEnv({ DATA_DIR: fileAsDir }, () => importFresh("src/lib/db/core.ts")).then(
+      () => {
+        throw new Error("expected importing db/core with an invalid DATA_DIR to reject");
+      },
+      (err) => {
+        caught = err;
+      }
+    );
+    assert.ok(caught instanceof Error, "an invalid DATA_DIR must surface as a thrown Error");
+    assert.match(
+      String((caught as Error).message),
       /unable to open database file|ENOTDIR|EEXIST|not a directory|file already exists/i
     );
   } finally {