packages/opencode/src/provider/transform.ts (1)

688-701: ⚠️ Potential issue | 🔴 Critical | ⚡ Quick win

Wrap the @ai-sdk/openai switch clause in braces.

Line 688 declares a lexical variable directly under the case statement, triggering Biome's noSwitchDeclarations error rule. Add braces to block-scope the clause.

Suggested fix

-    case "@ai-sdk/openai":
-      // https://v5.ai-sdk.dev/providers/ai-sdk-providers/openai
-      const openaiEfforts = openaiReasoningEfforts(model.api.id, model.release_date)
-      if (!openaiEfforts) return {}
-      return Object.fromEntries(
-        openaiEfforts.map((effort) => [
-          effort,
-          {
-            reasoningEffort: effort,
-            reasoningSummary: "auto",
-            include: ["reasoning.encrypted_content"],
-          },
-        ]),
-      )
+    case "@ai-sdk/openai": {
+      // https://v5.ai-sdk.dev/providers/ai-sdk-providers/openai
+      const openaiEfforts = openaiReasoningEfforts(model.api.id, model.release_date)
+      if (!openaiEfforts) return {}
+      return Object.fromEntries(
+        openaiEfforts.map((effort) => [
+          effort,
+          {
+            reasoningEffort: effort,
+            reasoningSummary: "auto",
+            include: ["reasoning.encrypted_content"],
+          },
+        ]),
+      )
+    }

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@packages/opencode/src/provider/transform.ts` around lines 688 - 701, The
switch clause for case "@ai-sdk/openai" declares lexical variables
(openaiEfforts) directly under the case which triggers the noSwitchDeclarations
error; wrap the entire case body in a block (add { ... } after the case) so
openaiEfforts and any other const/let declarations are block-scoped, preserving
the existing logic that calls openaiReasoningEfforts(model.api.id,
model.release_date), guards on !openaiEfforts, and returns the
Object.fromEntries(...) mapping.

packages/opencode/src/provider/provider.ts (1)

1530-1544: ⚠️ Potential issue | 🟠 Major | ⚡ Quick win

Guard the OpenAI/Azure body rewrite against non-string payloads.

At line 1535, the code uses JSON.parse(opts.body as string) without verifying that opts.body is actually a string—it could be FormData, a stream, or another BodyInit type from the Fetch API. Similarly, line 1539 assumes every item in body.input is an object with an in check that could fail on primitives. In this generic fetch wrapper, either condition failing will throw before the request is sent, converting an otherwise valid provider call into a hard failure.

Suggested fix

           if (
             (model.api.npm === "@ai-sdk/openai" || model.api.npm === "@ai-sdk/azure") &&
-            opts.body &&
+            typeof opts.body === "string" &&
             opts.method === "POST"
           ) {
-            const body = JSON.parse(opts.body as string)
-            const keepIds = body.store === true
-            if (!keepIds && Array.isArray(body.input)) {
-              for (const item of body.input) {
-                if ("id" in item) {
-                  delete item.id
-                }
-              }
-              opts.body = JSON.stringify(body)
-            }
+            try {
+              const body = JSON.parse(opts.body) as Record<string, unknown>
+              const keepIds = body.store === true
+              if (!keepIds && Array.isArray(body.input)) {
+                for (const item of body.input) {
+                  if (item && typeof item === "object" && "id" in item) {
+                    delete (item as Record<string, unknown>).id
+                  }
+                }
+                opts.body = JSON.stringify(body)
+              }
+            } catch {
+              // leave non-JSON payloads untouched
+            }
           }

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@packages/opencode/src/provider/provider.ts` around lines 1530 - 1544, The
body-rewrite for OpenAI/Azure requests in provider.ts is unsafe because it
unconditionally JSON.parse(opts.body as string) and assumes body.input contains
objects; change the logic in the block that checks model.api.npm (the code
creating keepIds and parsing body) to first verify opts.body is a string (typeof
opts.body === "string") before JSON.parse, and only proceed if parsing yields an
object with Array.isArray(body.input); when iterating body.input ensure each
item is an object (typeof item === "object" && item !== null) before checking or
deleting the "id" property; if any guard fails, leave opts.body untouched so
non-string BodyInit types (FormData/streams/primitives) are not mutated.

packages/opencode/src/provider/transform.ts (1)

675-677: ⚠️ Potential issue | 🟡 Minor | ⚡ Quick win

Use the shared GPT-5 matcher for Azure reasoning variants.

This branch only recognizes gpt-5 and gpt-5-*, so versioned Azure IDs like gpt-5.4 never get the minimal variant. That makes Azure GPT-5.x variants diverge from the OpenAI path in the same file.

Proposed fix

-      if (id.includes("gpt-5-") || id === "gpt-5") {
+      if (GPT5_FAMILY_RE.test(id)) {
         azureEfforts.unshift("minimal")
       }

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@packages/opencode/src/provider/transform.ts` around lines 675 - 677, The
branch that adds "minimal" to azureEfforts currently checks
id.includes("gpt-5-") || id === "gpt-5" and misses versioned Azure names like
"gpt-5.4"; replace that ad-hoc check with the project's shared GPT-5 matcher
used elsewhere (e.g., the helper function named isGpt5Model or matchesGpt5) so
the condition becomes "if (sharedGpt5Matcher(id)) {
azureEfforts.unshift('minimal') }", ensuring all GPT-5 variants (including
gpt-5.x) get the minimal Azure variant.