fix: handle non-text response parts in GitHub action (#6173)

2 files changed, 186 insertions, 3 deletions

diff --git a/packages/opencode/src/cli/cmd/github.ts b/packages/opencode/src/cli/cmd/github.ts
index 748a96384..26e0fb73d 100644
--- a/packages/opencode/src/cli/cmd/github.ts
+++ b/packages/opencode/src/cli/cmd/github.ts
@@ -158,6 +158,29 @@ export function parseGitHubRemote(url: string): { owner: string; repo: string }
   return { owner: match[1], repo: match[2] }
 }
 
+/**
+ * Extracts displayable text from assistant response parts.
+ * Returns null for tool-only or reasoning-only responses (signals summary needed).
+ * Throws for truly unusable responses (empty, step-start only, etc.).
+ */
+export function extractResponseText(parts: MessageV2.Part[]): string | null {
+  // Priority 1: Look for text parts
+  const textPart = parts.findLast((p) => p.type === "text")
+  if (textPart) return textPart.text
+
+  // Priority 2: Reasoning-only - return null to signal summary needed
+  const reasoningPart = parts.findLast((p) => p.type === "reasoning")
+  if (reasoningPart) return null
+
+  // Priority 3: Tool-only - return null to signal summary needed
+  const toolParts = parts.filter((p) => p.type === "tool" && p.state.status === "completed")
+  if (toolParts.length > 0) return null
+
+  // No usable parts - throw with debug info
+  const partTypes = parts.map((p) => p.type).join(", ") || "none"
+  throw new Error(`Failed to parse response. Part types found: [${partTypes}]`)
+}
+
 export const GithubCommand = cmd({
   command: "github",
   describe: "manage GitHub agent",
@@ -890,10 +913,41 @@ export const GithubRunCommand = cmd({
           )
         }
 
-        const match = result.parts.findLast((p) => p.type === "text")
-        if (!match) throw new Error("Failed to parse the text response")
+        const text = extractResponseText(result.parts)
+        if (text) return text
+
+        // No text part (tool-only or reasoning-only) - ask agent to summarize
+        console.log("Requesting summary from agent...")
+        const summary = await SessionPrompt.prompt({
+          sessionID: session.id,
+          messageID: Identifier.ascending("message"),
+          model: {
+            providerID,
+            modelID,
+          },
+          tools: { "*": false }, // Disable all tools to force text response
+          parts: [
+            {
+              id: Identifier.ascending("part"),
+              type: "text",
+              text: "Summarize the actions (tool calls & reasoning) you did for the user in 1-2 sentences.",
+            },
+          ],
+        })
+
+        if (summary.info.role === "assistant" && summary.info.error) {
+          console.error(summary.info)
+          throw new Error(
+            `${summary.info.error.name}: ${"message" in summary.info.error ? summary.info.error.message : ""}`,
+          )
+        }
+
+        const summaryText = extractResponseText(summary.parts)
+        if (!summaryText) {
+          throw new Error("Failed to get summary from agent")
+        }
 
-        return match.text
+        return summaryText
       }
 
       async function getOidcToken() {
diff --git a/packages/opencode/test/cli/github-action.test.ts b/packages/opencode/test/cli/github-action.test.ts
new file mode 100644
index 000000000..e2d4a6932
--- /dev/null
+++ b/packages/opencode/test/cli/github-action.test.ts
@@ -0,0 +1,129 @@
+import { test, expect, describe } from "bun:test"
+import { extractResponseText } from "../../src/cli/cmd/github"
+import type { MessageV2 } from "../../src/session/message-v2"
+
+// Helper to create minimal valid parts
+function createTextPart(text: string): MessageV2.Part {
+  return {
+    id: "1",
+    sessionID: "s",
+    messageID: "m",
+    type: "text" as const,
+    text,
+  }
+}
+
+function createReasoningPart(text: string): MessageV2.Part {
+  return {
+    id: "1",
+    sessionID: "s",
+    messageID: "m",
+    type: "reasoning" as const,
+    text,
+    time: { start: 0 },
+  }
+}
+
+function createToolPart(tool: string, title: string, status: "completed" | "running" = "completed"): MessageV2.Part {
+  if (status === "completed") {
+    return {
+      id: "1",
+      sessionID: "s",
+      messageID: "m",
+      type: "tool" as const,
+      callID: "c1",
+      tool,
+      state: {
+        status: "completed",
+        input: {},
+        output: "",
+        title,
+        metadata: {},
+        time: { start: 0, end: 1 },
+      },
+    }
+  }
+  return {
+    id: "1",
+    sessionID: "s",
+    messageID: "m",
+    type: "tool" as const,
+    callID: "c1",
+    tool,
+    state: {
+      status: "running",
+      input: {},
+      time: { start: 0 },
+    },
+  }
+}
+
+function createStepStartPart(): MessageV2.Part {
+  return {
+    id: "1",
+    sessionID: "s",
+    messageID: "m",
+    type: "step-start" as const,
+  }
+}
+
+describe("extractResponseText", () => {
+  test("returns text from text part", () => {
+    const parts = [createTextPart("Hello world")]
+    expect(extractResponseText(parts)).toBe("Hello world")
+  })
+
+  test("returns last text part when multiple exist", () => {
+    const parts = [createTextPart("First"), createTextPart("Last")]
+    expect(extractResponseText(parts)).toBe("Last")
+  })
+
+  test("returns text even when tool parts follow", () => {
+    const parts = [createTextPart("I'll help with that."), createToolPart("todowrite", "3 todos")]
+    expect(extractResponseText(parts)).toBe("I'll help with that.")
+  })
+
+  test("returns null for reasoning-only response (signals summary needed)", () => {
+    const parts = [createReasoningPart("Let me think about this...")]
+    expect(extractResponseText(parts)).toBeNull()
+  })
+
+  test("returns null for tool-only response (signals summary needed)", () => {
+    // This is the exact scenario from the bug report - todowrite with no text
+    const parts = [createToolPart("todowrite", "8 todos")]
+    expect(extractResponseText(parts)).toBeNull()
+  })
+
+  test("returns null for multiple completed tools", () => {
+    const parts = [
+      createToolPart("read", "src/file.ts"),
+      createToolPart("edit", "src/file.ts"),
+      createToolPart("bash", "bun test"),
+    ]
+    expect(extractResponseText(parts)).toBeNull()
+  })
+
+  test("ignores running tool parts (throws since no completed tools)", () => {
+    const parts = [createToolPart("bash", "", "running")]
+    expect(() => extractResponseText(parts)).toThrow("Failed to parse response")
+  })
+
+  test("throws with part types on empty array", () => {
+    expect(() => extractResponseText([])).toThrow("Part types found: [none]")
+  })
+
+  test("throws with part types on unhandled parts", () => {
+    const parts = [createStepStartPart()]
+    expect(() => extractResponseText(parts)).toThrow("Part types found: [step-start]")
+  })
+
+  test("prefers text over reasoning when both present", () => {
+    const parts = [createReasoningPart("Internal thinking..."), createTextPart("Final answer")]
+    expect(extractResponseText(parts)).toBe("Final answer")
+  })
+
+  test("prefers text over tools when both present", () => {
+    const parts = [createToolPart("read", "src/file.ts"), createTextPart("Here's what I found")]
+    expect(extractResponseText(parts)).toBe("Here's what I found")
+  })
+})