feat: add structured output support via --json-schema argument (#687)

* feat: add structured output support

Add support for Agent SDK structured outputs.

New input: json_schema
Output: structured_output (JSON string)
Access: fromJSON(steps.id.outputs.structured_output).field

Docs: https://docs.claude.com/en/docs/agent-sdk/structured-outputs

* rm unused

* refactor: simplify structured outputs to use claude_args

Remove json_schema input in favor of passing --json-schema flag directly
in claude_args. This simplifies the interface by treating structured outputs
like other CLI flags (--model, --max-turns, etc.) instead of as a special
input that gets injected.

Users now specify: claude_args: '--json-schema {...}'
Instead of separate: json_schema: {...}

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* chore: remove unused json-schema util and revert version

- Remove src/utils/json-schema.ts (no longer used after refactor)
- Revert Claude Code version from 2.0.45 back to 2.0.42

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

---------

Co-authored-by: Claude <noreply@anthropic.com>

.github/workflows/test-structured-output.yml

@@ -1,16 +1,10 @@
-name: Test Structured Outputs (Optimized)
-
-# This workflow uses EXPLICIT prompts that tell Claude exactly what to return.
-# This makes tests fast, deterministic, and focuses on testing OUR code, not Claude's reasoning.
-#
-# NOTE: Disabled until Agent SDK structured outputs feature is released
-# The --json-schema flag is not yet available in public Claude Code releases
+name: Test Structured Outputs
 
 on:
-  # Disabled - uncomment when feature is released
-  # push:
-  #   branches: [main]
-  # pull_request:
+  push:
+    branches:
+      - main
+  pull_request:
   workflow_dispatch:
 
 permissions:
@@ -28,7 +22,6 @@ jobs:
         id: test
         uses: ./base-action
         with:
-          # EXPLICIT: Tell Claude exactly what to return - no reasoning needed
           prompt: |
             Run this command: echo "test"
 
@@ -37,43 +30,41 @@ jobs:
             - number_field: 42
             - boolean_true: true
             - boolean_false: false
-          json_schema: |
-            {
-              "type": "object",
-              "properties": {
-                "text_field": {"type": "string"},
-                "number_field": {"type": "number"},
-                "boolean_true": {"type": "boolean"},
-                "boolean_false": {"type": "boolean"}
-              },
-              "required": ["text_field", "number_field", "boolean_true", "boolean_false"]
-            }
           anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
-          allowed_tools: "Bash"
+          claude_args: |
+            --allowedTools Bash
+            --json-schema '{"type":"object","properties":{"text_field":{"type":"string"},"number_field":{"type":"number"},"boolean_true":{"type":"boolean"},"boolean_false":{"type":"boolean"}},"required":["text_field","number_field","boolean_true","boolean_false"]}'
 
       - name: Verify outputs
         run: |
+          # Parse the structured_output JSON
+          OUTPUT='${{ steps.test.outputs.structured_output }}'
+
           # Test string pass-through
-          if [ "${{ steps.test.outputs.text_field }}" != "hello" ]; then
-            echo "❌ String: expected 'hello', got '${{ steps.test.outputs.text_field }}'"
+          TEXT_FIELD=$(echo "$OUTPUT" | jq -r '.text_field')
+          if [ "$TEXT_FIELD" != "hello" ]; then
+            echo "❌ String: expected 'hello', got '$TEXT_FIELD'"
             exit 1
           fi
 
           # Test number → string conversion
-          if [ "${{ steps.test.outputs.number_field }}" != "42" ]; then
-            echo "❌ Number: expected '42', got '${{ steps.test.outputs.number_field }}'"
+          NUMBER_FIELD=$(echo "$OUTPUT" | jq -r '.number_field')
+          if [ "$NUMBER_FIELD" != "42" ]; then
+            echo "❌ Number: expected '42', got '$NUMBER_FIELD'"
             exit 1
           fi
 
           # Test boolean → "true" conversion
-          if [ "${{ steps.test.outputs.boolean_true }}" != "true" ]; then
-            echo "❌ Boolean true: expected 'true', got '${{ steps.test.outputs.boolean_true }}'"
+          BOOLEAN_TRUE=$(echo "$OUTPUT" | jq -r '.boolean_true')
+          if [ "$BOOLEAN_TRUE" != "true" ]; then
+            echo "❌ Boolean true: expected 'true', got '$BOOLEAN_TRUE'"
             exit 1
           fi
 
           # Test boolean → "false" conversion
-          if [ "${{ steps.test.outputs.boolean_false }}" != "false" ]; then
-            echo "❌ Boolean false: expected 'false', got '${{ steps.test.outputs.boolean_false }}'"
+          BOOLEAN_FALSE=$(echo "$OUTPUT" | jq -r '.boolean_false')
+          if [ "$BOOLEAN_FALSE" != "false" ]; then
+            echo "❌ Boolean false: expected 'false', got '$BOOLEAN_FALSE'"
             exit 1
           fi
 
@@ -90,7 +81,6 @@ jobs:
         id: test
         uses: ./base-action
         with:
-          # EXPLICIT: No file reading, no analysis
           prompt: |
             Run: echo "ready"
 
@@ -98,46 +88,38 @@ jobs:
             - items: ["apple", "banana", "cherry"]
             - config: {"key": "value", "count": 3}
             - empty_array: []
-          json_schema: |
-            {
-              "type": "object",
-              "properties": {
-                "items": {
-                  "type": "array",
-                  "items": {"type": "string"}
-                },
-                "config": {"type": "object"},
-                "empty_array": {"type": "array"}
-              },
-              "required": ["items", "config", "empty_array"]
-            }
           anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
-          allowed_tools: "Bash"
+          claude_args: |
+            --allowedTools Bash
+            --json-schema '{"type":"object","properties":{"items":{"type":"array","items":{"type":"string"}},"config":{"type":"object"},"empty_array":{"type":"array"}},"required":["items","config","empty_array"]}'
 
       - name: Verify JSON stringification
         run: |
+          # Parse the structured_output JSON
+          OUTPUT='${{ steps.test.outputs.structured_output }}'
+
           # Arrays should be JSON stringified
-          ITEMS='${{ steps.test.outputs.items }}'
-          if ! echo "$ITEMS" | jq -e '. | length == 3' > /dev/null; then
-            echo "❌ Array not properly stringified: $ITEMS"
+          if ! echo "$OUTPUT" | jq -e '.items | length == 3' > /dev/null; then
+            echo "❌ Array not properly formatted"
+            echo "$OUTPUT" | jq '.items'
             exit 1
           fi
 
           # Objects should be JSON stringified
-          CONFIG='${{ steps.test.outputs.config }}'
-          if ! echo "$CONFIG" | jq -e '.key == "value"' > /dev/null; then
-            echo "❌ Object not properly stringified: $CONFIG"
+          if ! echo "$OUTPUT" | jq -e '.config.key == "value"' > /dev/null; then
+            echo "❌ Object not properly formatted"
+            echo "$OUTPUT" | jq '.config'
             exit 1
           fi
 
           # Empty arrays should work
-          EMPTY='${{ steps.test.outputs.empty_array }}'
-          if ! echo "$EMPTY" | jq -e '. | length == 0' > /dev/null; then
-            echo "❌ Empty array not properly stringified: $EMPTY"
+          if ! echo "$OUTPUT" | jq -e '.empty_array | length == 0' > /dev/null; then
+            echo "❌ Empty array not properly formatted"
+            echo "$OUTPUT" | jq '.empty_array'
             exit 1
           fi
 
-          echo "✅ All complex types JSON stringified correctly"
+          echo "✅ All complex types handled correctly"
 
   test-edge-cases:
     name: Test Edge Cases
@@ -158,43 +140,41 @@ jobs:
             - empty_string: ""
             - negative: -5
             - decimal: 3.14
-          json_schema: |
-            {
-              "type": "object",
-              "properties": {
-                "zero": {"type": "number"},
-                "empty_string": {"type": "string"},
-                "negative": {"type": "number"},
-                "decimal": {"type": "number"}
-              },
-              "required": ["zero", "empty_string", "negative", "decimal"]
-            }
           anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
-          allowed_tools: "Bash"
+          claude_args: |
+            --allowedTools Bash
+            --json-schema '{"type":"object","properties":{"zero":{"type":"number"},"empty_string":{"type":"string"},"negative":{"type":"number"},"decimal":{"type":"number"}},"required":["zero","empty_string","negative","decimal"]}'
 
       - name: Verify edge cases
         run: |
+          # Parse the structured_output JSON
+          OUTPUT='${{ steps.test.outputs.structured_output }}'
+
           # Zero should be "0", not empty or falsy
-          if [ "${{ steps.test.outputs.zero }}" != "0" ]; then
-            echo "❌ Zero: expected '0', got '${{ steps.test.outputs.zero }}'"
+          ZERO=$(echo "$OUTPUT" | jq -r '.zero')
+          if [ "$ZERO" != "0" ]; then
+            echo "❌ Zero: expected '0', got '$ZERO'"
             exit 1
           fi
 
           # Empty string should be empty (not "null" or missing)
-          if [ "${{ steps.test.outputs.empty_string }}" != "" ]; then
-            echo "❌ Empty string: expected '', got '${{ steps.test.outputs.empty_string }}'"
+          EMPTY_STRING=$(echo "$OUTPUT" | jq -r '.empty_string')
+          if [ "$EMPTY_STRING" != "" ]; then
+            echo "❌ Empty string: expected '', got '$EMPTY_STRING'"
             exit 1
           fi
 
           # Negative numbers should work
-          if [ "${{ steps.test.outputs.negative }}" != "-5" ]; then
-            echo "❌ Negative: expected '-5', got '${{ steps.test.outputs.negative }}'"
+          NEGATIVE=$(echo "$OUTPUT" | jq -r '.negative')
+          if [ "$NEGATIVE" != "-5" ]; then
+            echo "❌ Negative: expected '-5', got '$NEGATIVE'"
             exit 1
           fi
 
           # Decimals should preserve precision
-          if [ "${{ steps.test.outputs.decimal }}" != "3.14" ]; then
-            echo "❌ Decimal: expected '3.14', got '${{ steps.test.outputs.decimal }}'"
+          DECIMAL=$(echo "$OUTPUT" | jq -r '.decimal')
+          if [ "$DECIMAL" != "3.14" ]; then
+            echo "❌ Decimal: expected '3.14', got '$DECIMAL'"
             exit 1
           fi
 
@@ -214,29 +194,27 @@ jobs:
           prompt: |
             Run: echo "test"
             Return EXACTLY: {test-result: "passed", item_count: 10}
-          json_schema: |
-            {
-              "type": "object",
-              "properties": {
-                "test-result": {"type": "string"},
-                "item_count": {"type": "number"}
-              },
-              "required": ["test-result", "item_count"]
-            }
           anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
-          allowed_tools: "Bash"
+          claude_args: |
+            --allowedTools Bash
+            --json-schema '{"type":"object","properties":{"test-result":{"type":"string"},"item_count":{"type":"number"}},"required":["test-result","item_count"]}'
 
       - name: Verify sanitized names work
         run: |
-          # Hyphens should be preserved (GitHub Actions allows them)
-          if [ "${{ steps.test.outputs.test-result }}" != "passed" ]; then
-            echo "❌ Hyphenated name failed"
+          # Parse the structured_output JSON
+          OUTPUT='${{ steps.test.outputs.structured_output }}'
+
+          # Hyphens should be preserved in the JSON
+          TEST_RESULT=$(echo "$OUTPUT" | jq -r '.["test-result"]')
+          if [ "$TEST_RESULT" != "passed" ]; then
+            echo "❌ Hyphenated name failed: expected 'passed', got '$TEST_RESULT'"
             exit 1
           fi
 
           # Underscores should work
-          if [ "${{ steps.test.outputs.item_count }}" != "10" ]; then
-            echo "❌ Underscore name failed"
+          ITEM_COUNT=$(echo "$OUTPUT" | jq -r '.item_count')
+          if [ "$ITEM_COUNT" != "10" ]; then
+            echo "❌ Underscore name failed: expected '10', got '$ITEM_COUNT'"
             exit 1
           fi
 
@@ -254,16 +232,10 @@ jobs:
         uses: ./base-action
         with:
           prompt: "Run: echo 'complete'. Return: {done: true}"
-          json_schema: |
-            {
-              "type": "object",
-              "properties": {
-                "done": {"type": "boolean"}
-              },
-              "required": ["done"]
-            }
           anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
-          allowed_tools: "Bash"
+          claude_args: |
+            --allowedTools Bash
+            --json-schema '{"type":"object","properties":{"done":{"type":"boolean"}},"required":["done"]}'
 
       - name: Verify execution file contains structured_output
         run: |

README.md

@@ -13,6 +13,7 @@ A general-purpose [Claude Code](https://claude.ai/code) action for GitHub PRs an
 - 💬 **PR/Issue Integration**: Works seamlessly with GitHub comments and PR reviews
 - 🛠️ **Flexible Tool Access**: Access to GitHub APIs and file operations (additional tools can be enabled via configuration)
 - 📋 **Progress Tracking**: Visual progress indicators with checkboxes that dynamically update as Claude completes tasks
+- 📊 **Structured Outputs**: Get validated JSON results that automatically become GitHub Action outputs for complex automations
 - 🏃 **Runs on Your Infrastructure**: The action executes entirely on your own GitHub runner (Anthropic API calls go to your chosen provider)
 - ⚙️ **Simplified Configuration**: Unified `prompt` and `claude_args` inputs provide clean, powerful configuration aligned with Claude Code SDK
 

action.yml

@@ -113,10 +113,6 @@ inputs:
     description: "Newline-separated list of Claude Code plugin marketplace Git URLs to install from (e.g., 'https://github.com/user/marketplace1.git\nhttps://github.com/user/marketplace2.git')"
     required: false
     default: ""
-  json_schema:
-    description: "JSON schema for structured output validation. When provided, Claude will return validated JSON matching this schema, and the action will automatically set GitHub Action outputs for each field."
-    required: false
-    default: ""
 
 outputs:
   execution_file:
@@ -128,6 +124,9 @@ outputs:
   github_token:
     description: "The GitHub token used by the action (Claude App token if available)"
     value: ${{ steps.prepare.outputs.github_token }}
+  structured_output:
+    description: "JSON string containing all structured output fields when --json-schema is provided in claude_args. Use fromJSON() to parse: fromJSON(steps.id.outputs.structured_output).field_name"
+    value: ${{ steps.claude-code.outputs.structured_output }}
 
 runs:
   using: "composite"
@@ -178,7 +177,6 @@ runs:
         TRACK_PROGRESS: ${{ inputs.track_progress }}
         ADDITIONAL_PERMISSIONS: ${{ inputs.additional_permissions }}
         CLAUDE_ARGS: ${{ inputs.claude_args }}
-        JSON_SCHEMA: ${{ inputs.json_schema }}
         ALL_INPUTS: ${{ toJson(inputs) }}
 
     - name: Install Base Action Dependencies
@@ -194,7 +192,7 @@ runs:
         # Install Claude Code if no custom executable is provided
         if [ -z "${{ inputs.path_to_claude_code_executable }}" ]; then
           echo "Installing Claude Code..."
-          curl -fsSL https://claude.ai/install.sh | bash -s 2.0.42
+          curl -fsSL https://claude.ai/install.sh | bash -s 2.0.45
           echo "$HOME/.local/bin" >> "$GITHUB_PATH"
         else
           echo "Using custom Claude Code executable: ${{ inputs.path_to_claude_code_executable }}"
@@ -233,7 +231,6 @@ runs:
         INPUT_SHOW_FULL_OUTPUT: ${{ inputs.show_full_output }}
         INPUT_PLUGINS: ${{ inputs.plugins }}
         INPUT_PLUGIN_MARKETPLACES: ${{ inputs.plugin_marketplaces }}
-        JSON_SCHEMA: ${{ inputs.json_schema }}
 
         # Model configuration
         GITHUB_TOKEN: ${{ steps.prepare.outputs.GITHUB_TOKEN }}

base-action/action.yml

@@ -67,10 +67,6 @@ inputs:
     description: "Newline-separated list of Claude Code plugin marketplace Git URLs to install from (e.g., 'https://github.com/user/marketplace1.git\nhttps://github.com/user/marketplace2.git')"
     required: false
     default: ""
-  json_schema:
-    description: "JSON schema for structured output validation. When provided, Claude will return validated JSON matching this schema, and the action will automatically set GitHub Action outputs for each field (e.g., access via steps.id.outputs.field_name)"
-    required: false
-    default: ""
 
 outputs:
   conclusion:
@@ -79,6 +75,9 @@ outputs:
   execution_file:
     description: "Path to the JSON file containing Claude Code execution log"
     value: ${{ steps.run_claude.outputs.execution_file }}
+  structured_output:
+    description: "JSON string containing all structured output fields when --json-schema is provided in claude_args (use fromJSON() or jq to parse)"
+    value: ${{ steps.run_claude.outputs.structured_output }}
 
 runs:
   using: "composite"
@@ -115,7 +114,7 @@ runs:
       run: |
         if [ -z "${{ inputs.path_to_claude_code_executable }}" ]; then
           echo "Installing Claude Code..."
-          curl -fsSL https://claude.ai/install.sh | bash -s 2.0.42
+          curl -fsSL https://claude.ai/install.sh | bash -s 2.0.45
         else
           echo "Using custom Claude Code executable: ${{ inputs.path_to_claude_code_executable }}"
           # Add the directory containing the custom executable to PATH

base-action/src/run-claude.ts

@@ -12,11 +12,6 @@ const PIPE_PATH = `${process.env.RUNNER_TEMP}/claude_prompt_pipe`;
 const EXECUTION_FILE = `${process.env.RUNNER_TEMP}/claude-execution-output.json`;
 const BASE_ARGS = ["--verbose", "--output-format", "stream-json"];
 
-type ExecutionMessage = {
-  type: string;
-  structured_output?: Record<string, unknown>;
-};
-
 /**
  * Sanitizes JSON output to remove sensitive information when full output is disabled
  * Returns a safe summary message or null if the message should be completely suppressed
@@ -127,91 +122,54 @@ export function prepareRunConfig(
   };
 }
 
-/**
- * Sanitizes output field names to meet GitHub Actions output naming requirements
- * GitHub outputs must be alphanumeric, hyphen, or underscore only
- */
-function sanitizeOutputName(name: string): string {
-  return name.replace(/[^a-zA-Z0-9_-]/g, "_");
-}
-
-/**
- * Converts values to string format for GitHub Actions outputs
- * GitHub outputs must always be strings
- */
-function convertToString(value: unknown): string {
-  switch (typeof value) {
-    case "string":
-      return value;
-    case "boolean":
-    case "number":
-      return String(value);
-    case "object":
-      return value === null ? "" : JSON.stringify(value);
-    case "undefined":
-      return "";
-    default:
-      // Handle Symbol, Function, etc.
-      return String(value);
-  }
-}
-
 /**
  * Parses structured_output from execution file and sets GitHub Action outputs
- * Only runs if json_schema was explicitly provided by the user
+ * Only runs if --json-schema was explicitly provided in claude_args
+ * Exported for testing
  */
-async function parseAndSetStructuredOutputs(
+export async function parseAndSetStructuredOutputs(
   executionFile: string,
 ): Promise<void> {
   try {
     const content = await readFile(executionFile, "utf-8");
-    const messages = JSON.parse(content) as ExecutionMessage[];
+    const messages = JSON.parse(content) as {
+      type: string;
+      structured_output?: Record<string, unknown>;
+    }[];
 
-    const result = messages.find(
+    // Search backwards - result is typically last or second-to-last message
+    const result = messages.findLast(
       (m) => m.type === "result" && m.structured_output,
     );
 
     if (!result?.structured_output) {
-      const error = new Error(
-        "json_schema was provided but Claude did not return structured_output. " +
-          "The schema may be invalid or Claude failed to call the StructuredOutput tool.",
+      throw new Error(
+        `--json-schema was provided but Claude did not return structured_output.\n` +
+          `Found ${messages.length} messages. Result exists: ${!!result}\n`,
       );
-      core.setFailed(error.message);
-      throw error;
     }
 
-    // Set GitHub Action output for each field
-    const entries = Object.entries(result.structured_output);
-    core.info(`Setting ${entries.length} structured output(s)`);
-
-    for (const [key, value] of entries) {
-      const sanitizedKey = sanitizeOutputName(key);
-      if (!sanitizedKey) {
-        core.warning(`Skipping invalid output key: "${key}"`);
-        continue;
-      }
-
-      const stringValue = convertToString(value);
-
-      // Truncate long values in logs for readability
-      const displayValue =
-        stringValue.length > 100
-          ? `${stringValue.slice(0, 97)}...`
-          : stringValue;
-
-      core.setOutput(sanitizedKey, stringValue);
-      core.info(`✓ ${sanitizedKey}=${displayValue}`);
-    }
+    // Set the complete structured output as a single JSON string
+    // This works around GitHub Actions limitation that composite actions can't have dynamic outputs
+    const structuredOutputJson = JSON.stringify(result.structured_output);
+    core.setOutput("structured_output", structuredOutputJson);
+    core.info(
+      `Set structured_output with ${Object.keys(result.structured_output).length} field(s)`,
+    );
   } catch (error) {
-    const errorMsg = `Failed to parse structured outputs: ${error}`;
-    core.setFailed(errorMsg);
-    throw new Error(errorMsg);
+    if (error instanceof Error) {
+      throw error; // Preserve original error and stack trace
+    }
+    throw new Error(`Failed to parse structured outputs: ${error}`);
   }
 }
 
 export async function runClaude(promptPath: string, options: ClaudeOptions) {
   const config = prepareRunConfig(promptPath, options);
 
+  // Detect if --json-schema is present in claude args
+  const hasJsonSchema = options.claudeArgs?.includes("--json-schema") ?? false;
+
   // Create a named pipe
   try {
     await unlink(PIPE_PATH);
@@ -395,13 +353,23 @@ export async function runClaude(promptPath: string, options: ClaudeOptions) {
       core.warning(`Failed to process output for execution metrics: ${e}`);
     }
 
-    core.setOutput("conclusion", "success");
     core.setOutput("execution_file", EXECUTION_FILE);
 
-    // Parse and set structured outputs only if user provided json_schema
-    if (process.env.JSON_SCHEMA) {
-      await parseAndSetStructuredOutputs(EXECUTION_FILE);
+    // Parse and set structured outputs only if user provided --json-schema in claude_args
+    if (hasJsonSchema) {
+      try {
+        await parseAndSetStructuredOutputs(EXECUTION_FILE);
+      } catch (error) {
+        const errorMessage =
+          error instanceof Error ? error.message : String(error);
+        core.setFailed(errorMessage);
+        core.setOutput("conclusion", "failure");
+        process.exit(1);
+      }
     }
+
+    // Set conclusion to success if we reached here
+    core.setOutput("conclusion", "success");
   } else {
     core.setOutput("conclusion", "failure");
 

base-action/test/structured-output.test.ts

@@ -1,15 +1,11 @@
 #!/usr/bin/env bun
 
-import { describe, test, expect, afterEach } from "bun:test";
+import { describe, test, expect, afterEach, beforeEach, spyOn } from "bun:test";
 import { writeFile, unlink } from "fs/promises";
 import { tmpdir } from "os";
 import { join } from "path";
-
-// Import the type for testing
-type ExecutionMessage = {
-  type: string;
-  structured_output?: Record<string, unknown>;
-};
+import { parseAndSetStructuredOutputs } from "../src/run-claude";
+import * as core from "@actions/core";
 
 // Mock execution file path
 const TEST_EXECUTION_FILE = join(tmpdir(), "test-execution-output.json");
@@ -19,9 +15,9 @@ async function createMockExecutionFile(
   structuredOutput?: Record<string, unknown>,
   includeResult: boolean = true,
 ): Promise<void> {
-  const messages: ExecutionMessage[] = [
-    { type: "system", subtype: "init" } as any,
-    { type: "turn", content: "test" } as any,
+  const messages: any[] = [
+    { type: "system", subtype: "init" },
+    { type: "turn", content: "test" },
   ];
 
   if (includeResult) {
@@ -30,14 +26,25 @@ async function createMockExecutionFile(
       cost_usd: 0.01,
       duration_ms: 1000,
       structured_output: structuredOutput,
-    } as any);
+    });
   }
 
   await writeFile(TEST_EXECUTION_FILE, JSON.stringify(messages));
 }
 
-describe("Structured Output - Pure Functions", () => {
+// Spy on core functions
+let setOutputSpy: any;
+let infoSpy: any;
+
+beforeEach(() => {
+  setOutputSpy = spyOn(core, "setOutput").mockImplementation(() => {});
+  infoSpy = spyOn(core, "info").mockImplementation(() => {});
+});
+
+describe("parseAndSetStructuredOutputs", () => {
   afterEach(async () => {
+    setOutputSpy?.mockRestore();
+    infoSpy?.mockRestore();
     try {
       await unlink(TEST_EXECUTION_FILE);
     } catch {
@@ -45,297 +52,107 @@ describe("Structured Output - Pure Functions", () => {
     }
   });
 
-  describe("sanitizeOutputName", () => {
-    test("should keep valid characters", () => {
-      const sanitize = (name: string) => name.replace(/[^a-zA-Z0-9_-]/g, "_");
-      expect(sanitize("valid_name-123")).toBe("valid_name-123");
+  test("should set structured_output with valid data", async () => {
+    await createMockExecutionFile({
+      is_flaky: true,
+      confidence: 0.85,
+      summary: "Test looks flaky",
     });
 
-    test("should replace invalid characters with underscores", () => {
-      const sanitize = (name: string) => name.replace(/[^a-zA-Z0-9_-]/g, "_");
-      expect(sanitize("invalid@name!")).toBe("invalid_name_");
-      expect(sanitize("has spaces")).toBe("has_spaces");
-      expect(sanitize("has.dots")).toBe("has_dots");
+    await parseAndSetStructuredOutputs(TEST_EXECUTION_FILE);
+
+    expect(setOutputSpy).toHaveBeenCalledWith(
+      "structured_output",
+      '{"is_flaky":true,"confidence":0.85,"summary":"Test looks flaky"}',
+    );
+    expect(infoSpy).toHaveBeenCalledWith(
+      "Set structured_output with 3 field(s)",
+    );
+  });
+
+  test("should handle arrays and nested objects", async () => {
+    await createMockExecutionFile({
+      items: ["a", "b", "c"],
+      config: { key: "value", nested: { deep: true } },
     });
 
-    test("should handle special characters", () => {
-      const sanitize = (name: string) => name.replace(/[^a-zA-Z0-9_-]/g, "_");
-      expect(sanitize("$field%name&")).toBe("_field_name_");
-      expect(sanitize("field[0]")).toBe("field_0_");
+    await parseAndSetStructuredOutputs(TEST_EXECUTION_FILE);
+
+    const callArgs = setOutputSpy.mock.calls[0];
+    expect(callArgs[0]).toBe("structured_output");
+    const parsed = JSON.parse(callArgs[1]);
+    expect(parsed).toEqual({
+      items: ["a", "b", "c"],
+      config: { key: "value", nested: { deep: true } },
     });
   });
 
-  describe("convertToString", () => {
-    const convertToString = (value: unknown): string => {
-      switch (typeof value) {
-        case "string":
-          return value;
-        case "boolean":
-        case "number":
-          return String(value);
-        case "object":
-          return value === null ? "" : JSON.stringify(value);
-        default:
-          return JSON.stringify(value);
-      }
-    };
-
-    test("should keep strings as-is", () => {
-      expect(convertToString("hello")).toBe("hello");
-      expect(convertToString("")).toBe("");
+  test("should handle special characters in field names", async () => {
+    await createMockExecutionFile({
+      "test-result": "passed",
+      "item.count": 10,
+      "user@email": "test",
     });
 
-    test("should convert booleans to strings", () => {
-      expect(convertToString(true)).toBe("true");
-      expect(convertToString(false)).toBe("false");
-    });
+    await parseAndSetStructuredOutputs(TEST_EXECUTION_FILE);
 
-    test("should convert numbers to strings", () => {
-      expect(convertToString(42)).toBe("42");
-      expect(convertToString(3.14)).toBe("3.14");
-      expect(convertToString(0)).toBe("0");
-    });
-
-    test("should convert null to empty string", () => {
-      expect(convertToString(null)).toBe("");
-    });
-
-    test("should JSON stringify objects", () => {
-      expect(convertToString({ foo: "bar" })).toBe('{"foo":"bar"}');
-    });
-
-    test("should JSON stringify arrays", () => {
-      expect(convertToString([1, 2, 3])).toBe("[1,2,3]");
-      expect(convertToString(["a", "b"])).toBe('["a","b"]');
-    });
-
-    test("should handle nested structures", () => {
-      const nested = { items: [{ id: 1, name: "test" }] };
-      expect(convertToString(nested)).toBe(
-        '{"items":[{"id":1,"name":"test"}]}',
-      );
-    });
+    const callArgs = setOutputSpy.mock.calls[0];
+    const parsed = JSON.parse(callArgs[1]);
+    expect(parsed["test-result"]).toBe("passed");
+    expect(parsed["item.count"]).toBe(10);
+    expect(parsed["user@email"]).toBe("test");
   });
 
-  describe("parseAndSetStructuredOutputs integration", () => {
-    test("should parse and set simple structured outputs", async () => {
-      await createMockExecutionFile({
-        is_antonly: true,
-        confidence: 0.95,
-        risk: "low",
-      });
+  test("should throw error when result exists but structured_output is undefined", async () => {
+    const messages = [
+      { type: "system", subtype: "init" },
+      { type: "result", cost_usd: 0.01, duration_ms: 1000 },
+    ];
+    await writeFile(TEST_EXECUTION_FILE, JSON.stringify(messages));
 
-      // In a real test, we'd import and call parseAndSetStructuredOutputs
-      // For now, we simulate the behavior
-      const content = await Bun.file(TEST_EXECUTION_FILE).text();
-      const messages = JSON.parse(content) as ExecutionMessage[];
-      const result = messages.find(
-        (m) => m.type === "result" && m.structured_output,
-      );
-
-      expect(result?.structured_output).toEqual({
-        is_antonly: true,
-        confidence: 0.95,
-        risk: "low",
-      });
-    });
-
-    test("should handle array outputs", async () => {
-      await createMockExecutionFile({
-        affected_areas: ["auth", "database", "api"],
-        severity: "high",
-      });
-
-      const content = await Bun.file(TEST_EXECUTION_FILE).text();
-      const messages = JSON.parse(content) as ExecutionMessage[];
-      const result = messages.find(
-        (m) => m.type === "result" && m.structured_output,
-      );
-
-      expect(result?.structured_output?.affected_areas).toEqual([
-        "auth",
-        "database",
-        "api",
-      ]);
-    });
-
-    test("should handle nested objects", async () => {
-      await createMockExecutionFile({
-        analysis: {
-          category: "test",
-          details: { count: 5, passed: true },
-        },
-      });
-
-      const content = await Bun.file(TEST_EXECUTION_FILE).text();
-      const messages = JSON.parse(content) as ExecutionMessage[];
-      const result = messages.find(
-        (m) => m.type === "result" && m.structured_output,
-      );
-
-      expect(result?.structured_output?.analysis).toEqual({
-        category: "test",
-        details: { count: 5, passed: true },
-      });
-    });
-
-    test("should handle missing structured_output", async () => {
-      await createMockExecutionFile(undefined, true);
-
-      const content = await Bun.file(TEST_EXECUTION_FILE).text();
-      const messages = JSON.parse(content) as ExecutionMessage[];
-      const result = messages.find(
-        (m) => m.type === "result" && m.structured_output,
-      );
-
-      expect(result).toBeUndefined();
-    });
-
-    test("should handle empty structured_output", async () => {
-      await createMockExecutionFile({});
-
-      const content = await Bun.file(TEST_EXECUTION_FILE).text();
-      const messages = JSON.parse(content) as ExecutionMessage[];
-      const result = messages.find(
-        (m) => m.type === "result" && m.structured_output,
-      );
-
-      expect(result?.structured_output).toEqual({});
-    });
-
-    test("should handle all supported types", async () => {
-      await createMockExecutionFile({
-        string_field: "hello",
-        number_field: 42,
-        boolean_field: true,
-        null_field: null,
-        array_field: [1, 2, 3],
-        object_field: { nested: "value" },
-      });
-
-      const content = await Bun.file(TEST_EXECUTION_FILE).text();
-      const messages = JSON.parse(content) as ExecutionMessage[];
-      const result = messages.find(
-        (m) => m.type === "result" && m.structured_output,
-      );
-
-      expect(result?.structured_output).toMatchObject({
-        string_field: "hello",
-        number_field: 42,
-        boolean_field: true,
-        null_field: null,
-        array_field: [1, 2, 3],
-        object_field: { nested: "value" },
-      });
-    });
+    await expect(
+      parseAndSetStructuredOutputs(TEST_EXECUTION_FILE),
+    ).rejects.toThrow(
+      "--json-schema was provided but Claude did not return structured_output",
+    );
   });
 
-  describe("output naming with prefix", () => {
-    test("should apply prefix correctly", () => {
-      const prefix = "CLAUDE_";
-      const key = "is_antonly";
-      const sanitizedKey = key.replace(/[^a-zA-Z0-9_-]/g, "_");
-      const outputName = prefix + sanitizedKey;
+  test("should throw error when no result message exists", async () => {
+    const messages = [
+      { type: "system", subtype: "init" },
+      { type: "turn", content: "test" },
+    ];
+    await writeFile(TEST_EXECUTION_FILE, JSON.stringify(messages));
 
-      expect(outputName).toBe("CLAUDE_is_antonly");
-    });
-
-    test("should handle empty prefix", () => {
-      const prefix = "";
-      const key = "result";
-      const sanitizedKey = key.replace(/[^a-zA-Z0-9_-]/g, "_");
-      const outputName = prefix + sanitizedKey;
-
-      expect(outputName).toBe("result");
-    });
-
-    test("should sanitize and prefix invalid keys", () => {
-      const prefix = "OUT_";
-      const key = "invalid@key!";
-      const sanitizedKey = key.replace(/[^a-zA-Z0-9_-]/g, "_");
-      const outputName = prefix + sanitizedKey;
-
-      expect(outputName).toBe("OUT_invalid_key_");
-    });
+    await expect(
+      parseAndSetStructuredOutputs(TEST_EXECUTION_FILE),
+    ).rejects.toThrow(
+      "--json-schema was provided but Claude did not return structured_output",
+    );
   });
 
-  describe("error scenarios", () => {
-    test("should handle malformed JSON", async () => {
-      await writeFile(TEST_EXECUTION_FILE, "invalid json {");
+  test("should throw error with malformed JSON", async () => {
+    await writeFile(TEST_EXECUTION_FILE, "{ invalid json");
 
-      let error: Error | undefined;
-      try {
-        const content = await Bun.file(TEST_EXECUTION_FILE).text();
-        JSON.parse(content);
-      } catch (e) {
-        error = e as Error;
-      }
-
-      expect(error).toBeDefined();
-      expect(error?.message).toContain("JSON");
-    });
-
-    test("should handle empty execution file", async () => {
-      await writeFile(TEST_EXECUTION_FILE, "[]");
-
-      const content = await Bun.file(TEST_EXECUTION_FILE).text();
-      const messages = JSON.parse(content) as ExecutionMessage[];
-      const result = messages.find(
-        (m) => m.type === "result" && m.structured_output,
-      );
-
-      expect(result).toBeUndefined();
-    });
-
-    test("should handle missing result message", async () => {
-      const messages = [
-        { type: "system", subtype: "init" },
-        { type: "turn", content: "test" },
-      ];
-      await writeFile(TEST_EXECUTION_FILE, JSON.stringify(messages));
-
-      const content = await Bun.file(TEST_EXECUTION_FILE).text();
-      const parsed = JSON.parse(content) as ExecutionMessage[];
-      const result = parsed.find(
-        (m) => m.type === "result" && m.structured_output,
-      );
-
-      expect(result).toBeUndefined();
-    });
+    await expect(
+      parseAndSetStructuredOutputs(TEST_EXECUTION_FILE),
+    ).rejects.toThrow();
   });
 
-  describe("value truncation in logs", () => {
-    test("should truncate long string values for display", () => {
-      const longValue = "a".repeat(150);
-      const displayValue =
-        longValue.length > 100 ? `${longValue.slice(0, 97)}...` : longValue;
+  test("should throw error when file does not exist", async () => {
+    await expect(
+      parseAndSetStructuredOutputs("/nonexistent/file.json"),
+    ).rejects.toThrow();
+  });
 
-      expect(displayValue).toBe("a".repeat(97) + "...");
-      expect(displayValue.length).toBe(100);
-    });
+  test("should handle empty structured_output object", async () => {
+    await createMockExecutionFile({});
 
-    test("should not truncate short values", () => {
-      const shortValue = "short";
-      const displayValue =
-        shortValue.length > 100 ? `${shortValue.slice(0, 97)}...` : shortValue;
+    await parseAndSetStructuredOutputs(TEST_EXECUTION_FILE);
 
-      expect(displayValue).toBe("short");
-    });
-
-    test("should truncate exactly 100 character values", () => {
-      const value = "a".repeat(100);
-      const displayValue =
-        value.length > 100 ? `${value.slice(0, 97)}...` : value;
-
-      expect(displayValue).toBe(value);
-    });
-
-    test("should truncate 101 character values", () => {
-      const value = "a".repeat(101);
-      const displayValue =
-        value.length > 100 ? `${value.slice(0, 97)}...` : value;
-
-      expect(displayValue).toBe("a".repeat(97) + "...");
-    });
+    expect(setOutputSpy).toHaveBeenCalledWith("structured_output", "{}");
+    expect(infoSpy).toHaveBeenCalledWith(
+      "Set structured_output with 0 field(s)",
+    );
   });
 });

docs/usage.md

@@ -185,6 +185,74 @@ For a comprehensive guide on migrating from v0.x to v1.0, including step-by-step
       Focus on the changed files in this PR.
 ```
 
+## Structured Outputs
+
+Get validated JSON results from Claude that automatically become GitHub Action outputs. This enables building complex automation workflows where Claude analyzes data and subsequent steps use the results.
+
+### Basic Example
+
+```yaml
+- name: Detect flaky tests
+  id: analyze
+  uses: anthropics/claude-code-action@v1
+  with:
+    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
+    prompt: |
+      Check the CI logs and determine if this is a flaky test.
+      Return: is_flaky (boolean), confidence (0-1), summary (string)
+    claude_args: |
+      --json-schema '{"type":"object","properties":{"is_flaky":{"type":"boolean"},"confidence":{"type":"number"},"summary":{"type":"string"}},"required":["is_flaky"]}'
+
+- name: Retry if flaky
+  if: fromJSON(steps.analyze.outputs.structured_output).is_flaky == true
+  run: gh workflow run CI
+```
+
+### How It Works
+
+1. **Define Schema**: Provide a JSON schema via `--json-schema` flag in `claude_args`
+2. **Claude Executes**: Claude uses tools to complete your task
+3. **Validated Output**: Result is validated against your schema
+4. **JSON Output**: All fields are returned in a single `structured_output` JSON string
+
+### Accessing Structured Outputs
+
+All structured output fields are available in the `structured_output` output as a JSON string:
+
+**In GitHub Actions expressions:**
+
+```yaml
+if: fromJSON(steps.analyze.outputs.structured_output).is_flaky == true
+run: |
+  CONFIDENCE=${{ fromJSON(steps.analyze.outputs.structured_output).confidence }}
+```
+
+**In bash with jq:**
+
+```yaml
+- name: Process results
+  run: |
+    OUTPUT='${{ steps.analyze.outputs.structured_output }}'
+    IS_FLAKY=$(echo "$OUTPUT" | jq -r '.is_flaky')
+    SUMMARY=$(echo "$OUTPUT" | jq -r '.summary')
+```
+
+**Note**: Due to GitHub Actions limitations, composite actions cannot expose dynamic outputs. All fields are bundled in the single `structured_output` JSON string.
+
+### Complete Example
+
+See `examples/test-failure-analysis.yml` for a working example that:
+
+- Detects flaky test failures
+- Uses confidence thresholds in conditionals
+- Auto-retries workflows
+- Comments on PRs
+
+### Documentation
+
+For complete details on JSON Schema syntax and Agent SDK structured outputs:
+https://docs.claude.com/en/docs/agent-sdk/structured-outputs
+
 ## Ways to Tag @claude
 
 These examples show how to interact with Claude using comments in PRs and issues. By default, Claude will be triggered anytime you mention `@claude`, but you can customize the exact trigger phrase using the `trigger_phrase` input in the workflow.

examples/test-failure-analysis.yml

@@ -43,38 +43,23 @@ jobs:
             - is_flaky: true if likely flaky, false if real bug
             - confidence: number 0-1 indicating confidence level
             - summary: brief one-sentence explanation
-          json_schema: |
-            {
-              "type": "object",
-              "properties": {
-                "is_flaky": {
-                  "type": "boolean",
-                  "description": "Whether this appears to be a flaky test failure"
-                },
-                "confidence": {
-                  "type": "number",
-                  "minimum": 0,
-                  "maximum": 1,
-                  "description": "Confidence level in the determination"
-                },
-                "summary": {
-                  "type": "string",
-                  "description": "One-sentence explanation of the failure"
-                }
-              },
-              "required": ["is_flaky", "confidence", "summary"]
-            }
+          claude_args: |
+            --json-schema '{"type":"object","properties":{"is_flaky":{"type":"boolean","description":"Whether this appears to be a flaky test failure"},"confidence":{"type":"number","minimum":0,"maximum":1,"description":"Confidence level in the determination"},"summary":{"type":"string","description":"One-sentence explanation of the failure"}},"required":["is_flaky","confidence","summary"]}'
 
       # Auto-retry only if flaky AND high confidence (>= 0.7)
       - name: Retry flaky tests
         if: |
-          steps.detect.outputs.is_flaky == 'true' &&
-          steps.detect.outputs.confidence >= '0.7'
+          fromJSON(steps.detect.outputs.structured_output).is_flaky == true &&
+          fromJSON(steps.detect.outputs.structured_output).confidence >= 0.7
         env:
           GH_TOKEN: ${{ github.token }}
         run: |
-          echo "🔄 Flaky test detected (confidence: ${{ steps.detect.outputs.confidence }})"
-          echo "Summary: ${{ steps.detect.outputs.summary }}"
+          OUTPUT='${{ steps.detect.outputs.structured_output }}'
+          CONFIDENCE=$(echo "$OUTPUT" | jq -r '.confidence')
+          SUMMARY=$(echo "$OUTPUT" | jq -r '.summary')
+
+          echo "🔄 Flaky test detected (confidence: $CONFIDENCE)"
+          echo "Summary: $SUMMARY"
           echo ""
           echo "Triggering automatic retry..."
 
@@ -84,10 +69,13 @@ jobs:
       # Low confidence flaky detection - skip retry
       - name: Low confidence detection
         if: |
-          steps.detect.outputs.is_flaky == 'true' &&
-          steps.detect.outputs.confidence < '0.7'
+          fromJSON(steps.detect.outputs.structured_output).is_flaky == true &&
+          fromJSON(steps.detect.outputs.structured_output).confidence < 0.7
         run: |
-          echo "⚠️ Possible flaky test but confidence too low (${{ steps.detect.outputs.confidence }})"
+          OUTPUT='${{ steps.detect.outputs.structured_output }}'
+          CONFIDENCE=$(echo "$OUTPUT" | jq -r '.confidence')
+
+          echo "⚠️ Possible flaky test but confidence too low ($CONFIDENCE)"
           echo "Not retrying automatically - manual review recommended"
 
       # Comment on PR if this was a PR build
@@ -96,16 +84,29 @@ jobs:
         env:
           GH_TOKEN: ${{ github.token }}
         run: |
+          OUTPUT='${{ steps.detect.outputs.structured_output }}'
+          IS_FLAKY=$(echo "$OUTPUT" | jq -r '.is_flaky')
+          CONFIDENCE=$(echo "$OUTPUT" | jq -r '.confidence')
+          SUMMARY=$(echo "$OUTPUT" | jq -r '.summary')
+
           pr_number=$(gh pr list --head "${{ github.event.workflow_run.head_branch }}" --json number --jq '.[0].number')
 
           if [ -n "$pr_number" ]; then
+            if [ "$IS_FLAKY" = "true" ]; then
+              TITLE="🔄 Flaky Test Detected"
+              ACTION="✅ Automatically retrying the workflow"
+            else
+              TITLE="❌ Test Failure"
+              ACTION="⚠️ This appears to be a real bug - manual intervention needed"
+            fi
+
             gh pr comment "$pr_number" --body "$(cat <<EOF
-          ## ${{ steps.detect.outputs.is_flaky == 'true' && '🔄 Flaky Test Detected' || '❌ Test Failure' }}
+          ## $TITLE
 
-          **Analysis**: ${{ steps.detect.outputs.summary }}
-          **Confidence**: ${{ steps.detect.outputs.confidence }}
+          **Analysis**: $SUMMARY
+          **Confidence**: $CONFIDENCE
 
-          ${{ steps.detect.outputs.is_flaky == 'true' && '✅ Automatically retrying the workflow' || '⚠️ This appears to be a real bug - manual intervention needed' }}
+          $ACTION
 
           [View workflow run](${{ github.event.workflow_run.html_url }})
           EOF

src/modes/agent/index.ts

@@ -149,19 +149,6 @@ export const agentMode: Mode = {
       claudeArgs = `--mcp-config '${escapedOurConfig}'`;
     }
 
-    // Add JSON schema if provided
-    const jsonSchema = process.env.JSON_SCHEMA || "";
-    if (jsonSchema) {
-      // Validate it's valid JSON
-      try {
-        JSON.parse(jsonSchema);
-      } catch (e) {
-        throw new Error(`Invalid JSON schema provided: ${e}`);
-      }
-      const escapedSchema = jsonSchema.replace(/'/g, "'\\''");
-      claudeArgs += ` --json-schema '${escapedSchema}'`;
-    }
-
     // Append user's claude_args (which may have more --mcp-config flags)
     claudeArgs = `${claudeArgs} ${userClaudeArgs}`.trim();
 

src/modes/tag/index.ts

@@ -177,19 +177,6 @@ export const tagMode: Mode = {
     // Add required tools for tag mode
     claudeArgs += ` --allowedTools "${tagModeTools.join(",")}"`;
 
-    // Add JSON schema if provided
-    const jsonSchema = process.env.JSON_SCHEMA || "";
-    if (jsonSchema) {
-      // Validate it's valid JSON
-      try {
-        JSON.parse(jsonSchema);
-      } catch (e) {
-        throw new Error(`Invalid JSON schema provided: ${e}`);
-      }
-      const escapedSchema = jsonSchema.replace(/'/g, "'\\''");
-      claudeArgs += ` --json-schema '${escapedSchema}'`;
-    }
-
     // Append user's claude_args (which may have more --mcp-config flags)
     if (userClaudeArgs) {
       claudeArgs += ` ${userClaudeArgs}`;