> ## Documentation Index
> Fetch the complete documentation index at: https://docs.ai-stats.phaseo.app/llms.txt
> Use this file to discover all available pages before exploring further.

# Run a coding-review agent with runtime tools

> Build a resumable coding-review workflow with local runtime tools, strict JSON, and human approval checkpoints.

Use this recipe when one workflow should:

* inspect code-related context with local runtime tools
* return strict structured output instead of free-form prose
* pause high-risk decisions for human approval
* checkpoint progress between model turns and tool calls

## 1. Define one narrow output contract

```ts theme={null}
type CodeReviewPlan = {
  risk: "low" | "medium" | "high";
  summary: string;
  recommendedActions: string[];
  needsApproval: boolean;
};
```

This keeps the first coding workflow inspectable in logs and easy to pass to downstream worker steps.

## 2. Define small runtime tools

```ts theme={null}
import { defineTool } from "@ai-stats/agent-sdk";

const lookupDiffSummary = defineTool({
  id: "lookup-diff-summary",
  description: "Return a short summary of the changed files in one pull request.",
  async execute(input: { pullRequestId: string }) {
    return {
      pullRequestId: input.pullRequestId,
      files: ["apps/api/src/pipeline/before/index.ts"],
      summary: "Changes touch request validation.",
    };
  },
});

const lookupFailingChecks = defineTool({
  id: "lookup-failing-checks",
  description: "Return the failing CI checks for one pull request.",
  async execute(input: { pullRequestId: string }) {
    return {
      pullRequestId: input.pullRequestId,
      failingChecks: ["web-typecheck"],
    };
  },
});
```

Keep these tools deterministic and idempotent so the checkpointed step data stays useful if a worker resumes later.

## 3. Create the agent with a human-review seam

```ts theme={null}
import { createAgent } from "@ai-stats/agent-sdk";

const codingReviewAgent = createAgent<string, CodeReviewPlan>({
  id: "coding-review-agent",
  model: "ai-stats/free",
  instructions:
    "Use the available tools to inspect the coding task, then return a strict JSON review plan.",
  tools: [lookupDiffSummary, lookupFailingChecks],
  parseOutput(text) {
    return JSON.parse(text) as CodeReviewPlan;
  },
  humanReview: ({ parsedOutput }) =>
    parsedOutput?.needsApproval
      ? {
          reason: "coding_review_approval_required",
          payload: parsedOutput,
        }
      : null,
});
```

## 4. Use the gateway-backed adapter for strict JSON

```ts theme={null}
import {
  createGatewayAgentClient,
} from "@ai-stats/agent-sdk";

const client = createGatewayAgentClient({
  clientOptions: {
    apiKey: process.env.AI_STATS_API_KEY!,
  },
  responseFormat: {
    type: "json_schema",
    name: "code_review_plan",
    schema: {
      type: "object",
      properties: {
        risk: {
          type: "string",
          enum: ["low", "medium", "high"],
        },
        summary: { type: "string" },
        recommendedActions: {
          type: "array",
          items: { type: "string" },
          minItems: 1,
        },
        needsApproval: { type: "boolean" },
      },
      required: ["risk", "summary", "recommendedActions", "needsApproval"],
      additionalProperties: false,
    },
  },
  plugins: [{ id: "response-healing", mode: "strict" }],
});
```

## 5. Run and resume the workflow

```ts theme={null}
const result = await codingReviewAgent.run({
  input: "Review pull request pr_482 and produce a remediation plan.",
  client,
  onEvent(event) {
    console.log(event.type, event.runId);
  },
});

if (result.run.status === "waiting_for_human") {
  const continued = await codingReviewAgent.continueRun({
    run: result,
    client,
    humanInput: "Approved. Finalize the plan.",
  });

  console.log(continued.output);
}
```

## 6. What to verify

After one run with tools and one run that pauses for approval, confirm:

* request detail views show the runtime tool loop clearly
* paused runs enter `waiting_for_human`
* step records checkpoint after tool execution
* schema failures or plugin-healing activity stay visible in logs

## Related guides

* [TypeScript Agent SDK](../sdk-reference/typescript/agent-sdk.mdx)
* [Build a durable agent loop](./agent-sdk-durable-loop.mdx)
* [Validate structured outputs](../guides/structured-outputs-validation.mdx)
* [Recover malformed structured JSON](./response-healing-for-structured-json.mdx)