Fixes

agents/claude.ts

@@ -5,13 +5,19 @@ import { log } from "../utils/cli.ts";
 import { addInstructions } from "./instructions.ts";
 import { agent, createAgentEnv, installFromNpmTarball } from "./shared.ts";
 
-// effort levels for Claude Code
-// Claude Code automatically selects the best model based on effort
-const claudeEffortLevels: Record<Effort, "low" | "medium" | "high"> = {
-  nothink: "low",
-  think: "medium",
-  max: "high",
-} as const;
+// Model selection based on effort level
+// Note: nothink uses Haiku for speed, think uses Sonnet for balance, max uses Opus for capability
+const claudeEffortModels: Record<Effort, string> = {
+  nothink: "claude-haiku-4-5-20250929",
+  think: "claude-sonnet-4-5-20250929",
+  max: "claude-opus-4-5-20250929",
+};
+
+// FUTURE: Consider using Anthropic's "effort" parameter (beta) with Opus 4.5 for all tasks.
+// This would allow a single model with effort levels ("low", "medium", "high") controlling
+// token spend across responses, tool calls, and thinking. Requires beta header "effort-2025-11-24".
+// See: https://platform.claude.com/docs/en/build-with-claude/effort
+// This approach could replace model selection if effort proves effective for controlling capability.
 
 export const claude = agent({
   name: "claude",
@@ -30,9 +36,9 @@ export const claude = agent({
     const prompt = addInstructions({ payload, repo });
     log.group("Full prompt", () => log.info(prompt));
 
-    // get effort level - Claude Code automatically selects the best model
-    const effortLevel = claudeEffortLevels[effort];
-    log.info(`Using effort: ${effortLevel}`);
+    // select model based on effort level
+    const model = claudeEffortModels[effort];
+    log.info(`Using model: ${model} (effort: ${effort})`);
 
     // SECURITY: For PUBLIC repos, Claude Code spawns subprocesses with full process.env, leaking API keys.
     // disable native Bash; agents use MCP bash tool which filters secrets.
@@ -62,7 +68,7 @@ export const claude = agent({
     const queryOptions: Options = {
       ...sandboxOptions,
       mcpServers,
-      effort: effortLevel,
+      model,
       pathToClaudeCodeExecutable: cliPath,
       env: createAgentEnv({ ANTHROPIC_API_KEY: apiKey }),
     };

agents/codex.ts

@@ -1,19 +1,32 @@
 import { mkdirSync, writeFileSync } from "node:fs";
 import { join } from "node:path";
 import type { McpHttpServerConfig } from "@anthropic-ai/claude-agent-sdk";
-import { Codex, type CodexOptions, type ThreadEvent } from "@openai/codex-sdk";
+import {
+  Codex,
+  type CodexOptions,
+  type ModelReasoningEffort,
+  type ThreadEvent,
+  type ThreadOptions,
+} from "@openai/codex-sdk";
 import type { Effort } from "../external.ts";
 import { log } from "../utils/cli.ts";
 import { addInstructions } from "./instructions.ts";
 import { agent, installFromNpmTarball, setupProcessAgentEnv } from "./shared.ts";
 
+// model configuration based on effort level
+const codexModel: Record<Effort, string> = {
+  nothink: "gpt-5.1-codex-mini",
+  think: "gpt-5.1-codex",
+  max: "gpt-5.1-codex-max",
+} as const;
+
 // reasoning effort configuration based on effort level
-// uses model_reasoning_effort parameter for o1 models
-const codexReasoningEffort: Record<Effort, string | undefined> = {
+// uses modelReasoningEffort parameter from ThreadOptions
+const codexReasoningEffort: Record<Effort, ModelReasoningEffort | undefined> = {
   nothink: "low",
   think: undefined, // use default
-  max: "xhigh",
-} as const;
+  max: "high",
+};
 
 interface WriteCodexConfigParams {
   tempHome: string;
@@ -88,12 +101,12 @@ export const codex = agent({
       CODEX_HOME: codexDir, // point Codex to our config directory
     });
 
-    // get reasoning effort based on effort level
+    // get model and reasoning effort based on effort level
+    const model = codexModel[effort];
     const modelReasoningEffort = codexReasoningEffort[effort];
+    log.info(`Using model: ${model}`);
     if (modelReasoningEffort) {
-      log.info(`Using model_reasoning_effort: ${modelReasoningEffort}`);
-    } else {
-      log.info(`Using default reasoning effort`);
+      log.info(`Using modelReasoningEffort: ${modelReasoningEffort}`);
     }
 
     // Configure Codex
@@ -108,22 +121,24 @@ export const codex = agent({
 
     const codex = new Codex(codexOptions);
 
-    // Build thread options with optional model_reasoning_effort
+    // Build thread options with model and optional model_reasoning_effort
     const baseThreadOptions = payload.sandbox
       ? {
+          model,
           approvalPolicy: "never" as const,
           sandboxMode: "read-only" as const,
           networkAccessEnabled: false,
         }
       : {
+          model,
           approvalPolicy: "never" as const,
           // use danger-full-access to allow git operations (workspace-write blocks .git directory writes)
           sandboxMode: "danger-full-access" as const,
           networkAccessEnabled: true,
         };
 
-    const threadOptions = modelReasoningEffort
-      ? { ...baseThreadOptions, model_reasoning_effort: modelReasoningEffort }
+    const threadOptions: ThreadOptions = modelReasoningEffort
+      ? { ...baseThreadOptions, modelReasoningEffort }
       : baseThreadOptions;
 
     const thread = codex.startThread(threadOptions);

agents/cursor.ts

@@ -1,7 +1,8 @@
 import { spawn } from "node:child_process";
-import { mkdirSync, writeFileSync } from "node:fs";
+import { existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
 import { homedir } from "node:os";
 import { join } from "node:path";
+import type { Effort } from "../external.ts";
 import { log } from "../utils/cli.ts";
 import { addInstructions } from "./instructions.ts";
 import {
@@ -11,6 +12,14 @@ import {
   installFromCurl,
 } from "./shared.ts";
 
+// effort configuration for Cursor
+// only "max" overrides the model; nothink/think use default ("auto")
+const cursorEffortModels: Record<Effort, string | null> = {
+  nothink: null, // use default (auto)
+  think: null, // use default (auto)
+  max: "opus-4.5-thinking",
+} as const;
+
 // cursor cli event types inferred from stream-json output
 interface CursorSystemEvent {
   type: "system";
@@ -91,10 +100,36 @@ export const cursor = agent({
       executableName: "cursor-agent",
     });
   },
-  run: async ({ payload, apiKey, cliPath, mcpServers, repo, effort: _effort }) => {
+  run: async ({ payload, apiKey, cliPath, mcpServers, repo, effort }) => {
     configureCursorMcpServers({ mcpServers, cliPath });
     configureCursorSandbox({ sandbox: payload.sandbox ?? false, isPublicRepo: repo.isPublic });
 
+    // determine model based on effort level
+    // respect project's .cursor/cli.json if it specifies a model
+    const projectCliConfigPath = join(process.cwd(), ".cursor", "cli.json");
+    let modelOverride: string | null = null;
+
+    if (existsSync(projectCliConfigPath)) {
+      try {
+        const projectConfig = JSON.parse(readFileSync(projectCliConfigPath, "utf-8"));
+        if (projectConfig.model) {
+          log.info(`Using model from project .cursor/cli.json: ${projectConfig.model}`);
+        } else {
+          modelOverride = cursorEffortModels[effort];
+        }
+      } catch {
+        modelOverride = cursorEffortModels[effort];
+      }
+    } else {
+      modelOverride = cursorEffortModels[effort];
+    }
+
+    if (modelOverride) {
+      log.info(`Using model: ${modelOverride} (effort: ${effort})`);
+    } else if (!existsSync(projectCliConfigPath)) {
+      log.info(`Using default model (effort: ${effort})`);
+    }
+
     // track logged model_call_ids to avoid duplicates
     // cursor emits each assistant message twice: once without model_call_id, then again with it
     const loggedModelCallIds = new Set<string>();
@@ -171,16 +206,16 @@ export const cursor = agent({
 
       // configure sandbox mode if enabled
       // in sandbox mode: remove --force flag and rely on cli-config.json sandbox settings
+      const baseArgs = ["--print", fullPrompt, "--output-format", "stream-json", "--approve-mcps"];
+
+      // add model flag if we have an override
+      if (modelOverride) {
+        baseArgs.push("--model", modelOverride);
+      }
+
       const cursorArgs = payload.sandbox
-        ? [
-            "--print",
-            fullPrompt,
-            "--output-format",
-            "stream-json",
-            "--approve-mcps",
-            // --force removed in sandbox mode to enforce safety checks
-          ]
-        : ["--print", fullPrompt, "--output-format", "stream-json", "--approve-mcps", "--force"];
+        ? baseArgs // --force removed in sandbox mode to enforce safety checks
+        : [...baseArgs, "--force"];
 
       if (payload.sandbox) {
         log.info("🔒 sandbox mode enabled: restricting to read-only operations");

agents/gemini.ts

@@ -1,6 +1,7 @@
 import { mkdirSync, readFileSync, writeFileSync } from "node:fs";
 import { homedir } from "node:os";
 import { join } from "node:path";
+import type { Effort } from "../external.ts";
 import { log } from "../utils/cli.ts";
 import { spawn } from "../utils/subprocess.ts";
 import { addInstructions } from "./instructions.ts";
@@ -11,12 +12,13 @@ import {
   installFromGithub,
 } from "./shared.ts";
 
-// model configuration based on effort level
-// auto for default, flash for fast/low-cost, pro for max capability
-const geminiModels = {
-  nothink: "gemini-2.5-flash",
-  think: "gemini-2.5-pro", // auto selection
-  max: "gemini-2.5-pro",
+// effort configuration: model + thinking level
+// thinkingLevel is set via settings.json modelConfig.generateContentConfig.thinkingConfig
+// see: https://ai.google.dev/gemini-api/docs/thinking#thinking-levels
+const geminiEffortConfig: Record<Effort, { model: string; thinkingLevel: string }> = {
+  nothink: { model: "gemini-2.5-flash", thinkingLevel: "LOW" },
+  think: { model: "gemini-2.5-flash", thinkingLevel: "HIGH" },
+  max: { model: "gemini-2.5-pro", thinkingLevel: "HIGH" },
 } as const;
 
 // gemini cli event types inferred from stream-json output (NDJSON format)
@@ -165,16 +167,16 @@ export const gemini = agent({
     });
   },
   run: async ({ payload, apiKey, mcpServers, cliPath, repo, effort }) => {
-    configureGeminiMcpServers({ mcpServers, isPublicRepo: repo.isPublic });
+    // get model and thinking level based on effort
+    const { model, thinkingLevel } = geminiEffortConfig[effort];
+    log.info(`Using model: ${model}, thinkingLevel: ${thinkingLevel}`);
+
+    configureGeminiSettings({ mcpServers, isPublicRepo: repo.isPublic, thinkingLevel });
 
     if (!apiKey) {
       throw new Error("google_api_key or gemini_api_key is required for gemini agent");
     }
 
-    // get model based on effort level
-    const model = geminiModels[effort];
-    log.info(`Using model: ${model}`);
-
     const sessionPrompt = addInstructions({ payload, repo });
     log.group("Full prompt", () => log.info(sessionPrompt));
 
@@ -290,16 +292,22 @@ export const gemini = agent({
 type ConfigureGeminiParams = {
   mcpServers: ConfigureMcpServersParams["mcpServers"];
   isPublicRepo: boolean;
+  thinkingLevel: string;
 };
 
 /**
- * Configure MCP servers for Gemini by writing to settings.json.
- * Gemini CLI uses `httpUrl` for HTTP/streamable transport, `url` for SSE transport.
- * See: https://github.com/google-gemini/gemini-cli/blob/main/docs/get-started/configuration.md
+ * Configure Gemini CLI settings by writing to settings.json.
+ * - MCP servers: uses `httpUrl` for HTTP/streamable transport
+ * - thinkingLevel: configured via modelConfig.generateContentConfig.thinkingConfig
+ * - For public repos, excludeTools disables native shell
  *
- * For public repos, also configures excludeTools to disable native shell.
+ * See: https://github.com/google-gemini/gemini-cli/blob/main/docs/get-started/configuration.md
  */
-function configureGeminiMcpServers({ mcpServers, isPublicRepo }: ConfigureGeminiParams): void {
+function configureGeminiSettings({
+  mcpServers,
+  isPublicRepo,
+  thinkingLevel,
+}: ConfigureGeminiParams): void {
   const realHome = homedir();
   const geminiConfigDir = join(realHome, ".gemini");
   const settingsPath = join(geminiConfigDir, "settings.json");
@@ -343,17 +351,26 @@ function configureGeminiMcpServers({ mcpServers, isPublicRepo }: ConfigureGemini
     log.info(`Adding MCP server '${serverName}' at ${serverConfig.url}...`);
   }
 
-  // merge with existing settings, overwriting mcpServers
-  // for public repos, exclude native shell tool to prevent secret leakage via env
+  // merge with existing settings, overwriting mcpServers and modelConfig
   const newSettings: Record<string, unknown> = {
     ...existingSettings,
     mcpServers: geminiMcpServers,
+    // configure thinking level via modelConfig
+    // see: https://ai.google.dev/api/generate-content (ThinkingConfig)
+    modelConfig: {
+      generateContentConfig: {
+        thinkingConfig: {
+          thinkingLevel,
+        },
+      },
+    },
   };
 
+  // for public repos, exclude native shell tool to prevent secret leakage via env
   if (isPublicRepo) {
     newSettings.excludeTools = ["run_shell_command"];
   }
 
   writeFileSync(settingsPath, JSON.stringify(newSettings, null, 2), "utf-8");
-  log.info(`» MCP config written to ${settingsPath}`);
+  log.info(`» Gemini settings written to ${settingsPath}`);
 }

effort.md

@@ -0,0 +1,92 @@
+# Effort Levels
+
+Pullfrog supports three effort levels that control model selection and reasoning depth:
+
+- **`nothink`** — Fast, minimal reasoning. Best for simple tasks.
+- **`think`** — Balanced (default). Good for most tasks.
+- **`max`** — Maximum capability. Best for complex tasks requiring deep reasoning.
+
+The effort level can be specified via the `effort` input in `action.yml` or in the payload's `effort` field.
+
+---
+
+## Claude Code
+
+Claude Code uses model selection based on effort level.
+
+| Effort | Model | Description |
+|--------|-------|-------------|
+| `nothink` | `claude-haiku-4-5-20250929` | Fast, efficient (Haiku) |
+| `think` | `claude-sonnet-4-5-20250929` | Balanced (Sonnet) |
+| `max` | `claude-opus-4-5-20250929` | Maximum capability (Opus) |
+
+> **Future direction:** Anthropic's beta `effort` parameter (`low`/`medium`/`high`) could replace model selection, using Opus 4.5 for all tasks with effort controlling token spend. See [Anthropic Effort Docs](https://platform.claude.com/docs/en/build-with-claude/effort).
+
+---
+
+## Codex (OpenAI)
+
+Codex uses both model selection and the `modelReasoningEffort` parameter from `ThreadOptions`.
+
+| Effort | Model | `modelReasoningEffort` | Description |
+|--------|-------|------------------------|-------------|
+| `nothink` | `gpt-5.1-codex-mini` | `"low"` | Smaller model, reduced reasoning |
+| `think` | `gpt-5.1-codex` | default | Standard model, default reasoning |
+| `max` | `gpt-5.1-codex-max` | `"high"` | Largest model, maximum reasoning |
+
+Valid values for `modelReasoningEffort`: `"minimal"` | `"low"` | `"medium"` | `"high"`
+
+Reference: [Codex Config Reference](https://developers.openai.com/codex/config-reference/)
+
+---
+
+## Gemini
+
+Gemini uses a combination of model selection and `thinkingLevel` configuration via `settings.json`.
+
+| Effort | Model | `thinkingLevel` | Description |
+|--------|-------|-----------------|-------------|
+| `nothink` | `gemini-2.5-flash` | `LOW` | Fast model, minimal thinking |
+| `think` | `gemini-2.5-flash` | `HIGH` | Fast model, deep thinking |
+| `max` | `gemini-2.5-pro` | `HIGH` | Most capable model, deep thinking |
+
+The `thinkingLevel` is configured via:
+```json
+{
+  "modelConfig": {
+    "generateContentConfig": {
+      "thinkingConfig": {
+        "thinkingLevel": "LOW"
+      }
+    }
+  }
+}
+```
+
+Reference: [Gemini Thinking Docs](https://ai.google.dev/gemini-api/docs/thinking#thinking-levels)
+
+---
+
+## Cursor
+
+Cursor uses model selection via the `--model` CLI flag. Project-level configuration in `.cursor/cli.json` takes precedence if a `model` is specified there.
+
+| Effort | Model | Description |
+|--------|-------|-------------|
+| `nothink` | `auto` (default) | Let Cursor select optimal model |
+| `think` | `auto` (default) | Let Cursor select optimal model |
+| `max` | `opus-4.5-thinking` | Claude 4.5 Opus with thinking |
+
+**Note:** If the project has `.cursor/cli.json` with a `model` field, that model is used regardless of effort level.
+
+---
+
+## OpenCode
+
+OpenCode does not currently have affordances for effort-level configuration. The effort parameter is ignored.
+
+| Effort | Behavior |
+|--------|----------|
+| `nothink` | No effect |
+| `think` | No effect |
+| `max` | No effect |

fixtures/cursor-effort.ts

@@ -0,0 +1,30 @@
+import type { Effort, Payload } from "../external.ts";
+
+/**
+ * Test fixture for Cursor effort levels.
+ * Runs all three effort levels in sequence.
+ *
+ * Run with:
+ *   AGENT_OVERRIDE=cursor pnpm play cursor-effort.ts
+ *
+ * Effort levels:
+ *   - "nothink": auto (default model)
+ *   - "think": auto (default model)
+ *   - "max": opus-4.5-thinking
+ *
+ * Note: If project has .cursor/cli.json with "model" specified,
+ * that takes precedence over effort-based model selection.
+ */
+
+const efforts: Effort[] = ["nothink", "think", "max"];
+
+export default efforts.map((effort) => ({
+  "~pullfrog": true,
+  agent: "cursor",
+  prompt: "What is 2 + 2? Reply with just the number.",
+  event: {
+    trigger: "workflow_dispatch",
+  },
+  modes: [],
+  effort,
+})) satisfies Payload[];

fixtures/gemini-effort.ts

@@ -0,0 +1,27 @@
+import type { Effort, Payload } from "../external.ts";
+
+/**
+ * Test fixture for Gemini effort levels.
+ * Runs all three effort levels in sequence.
+ *
+ * Run with:
+ *   AGENT_OVERRIDE=gemini pnpm play gemini-effort.ts
+ *
+ * Effort levels:
+ *   - "nothink": gemini-2.5-flash + LOW thinking
+ *   - "think": gemini-2.5-flash + HIGH thinking
+ *   - "max": gemini-2.5-pro + HIGH thinking
+ */
+
+const efforts: Effort[] = ["nothink", "think", "max"];
+
+export default efforts.map((effort) => ({
+  "~pullfrog": true,
+  agent: "gemini",
+  prompt: "What is 2 + 2? Reply with just the number.",
+  event: {
+    trigger: "workflow_dispatch",
+  },
+  modes: [],
+  effort,
+})) satisfies Payload[];

mcp/comment.ts

@@ -5,7 +5,12 @@ import { agentsManifest } from "../external.ts";
 import type { ToolContext } from "../main.ts";
 import { fetchWorkflowRunInfo } from "../utils/api.ts";
 import { buildPullfrogFooter, stripExistingFooter } from "../utils/buildPullfrogFooter.ts";
-import { createOctokit, getGitHubInstallationToken, parseRepoContext, type OctokitWithPlugins } from "../utils/github.ts";
+import {
+  createOctokit,
+  getGitHubInstallationToken,
+  type OctokitWithPlugins,
+  parseRepoContext,
+} from "../utils/github.ts";
 import { execute, tool } from "./shared.ts";
 
 /**
@@ -82,7 +87,11 @@ function buildImplementPlanLink(
   return `[Implement plan ➔](${apiUrl}/trigger/${owner}/${repo}/${issueNumber}?action=implement&comment_id=${commentId})`;
 }
 
-async function addFooter(body: string, payload: Payload, octokit?: OctokitWithPlugins): Promise<string> {
+async function addFooter(
+  body: string,
+  payload: Payload,
+  octokit?: OctokitWithPlugins
+): Promise<string> {
   const bodyWithoutFooter = stripExistingFooter(body);
   const footer = await buildCommentFooter({ payload, octokit });
   return `${bodyWithoutFooter}${footer}`;
@@ -494,6 +503,6 @@ export function ReplyToReviewCommentTool(ctx: ToolContext) {
         body: result.data.body,
         in_reply_to_id: result.data.in_reply_to_id,
       };
-    }),
+    }, "reply_to_review_comment"),
   });
 }

mcp/reviewComments.ts

@@ -93,95 +93,95 @@ export function GetReviewCommentsTool(ctx: ToolContext) {
       "Get all review comments and their replies for a specific pull request review. Returns line-by-line comments that were left on specific code locations, including any threaded replies.",
     parameters: GetReviewComments,
     execute: execute(async ({ pull_number, review_id }) => {
-      // fetch all review threads using graphql
-      const response = await ctx.octokit.graphql<GraphQLResponse>(REVIEW_THREADS_QUERY, {
-        owner: ctx.owner,
-        repo: ctx.name,
-        pullNumber: pull_number,
-      });
+        // fetch all review threads using graphql
+        const response = await ctx.octokit.graphql<GraphQLResponse>(REVIEW_THREADS_QUERY, {
+          owner: ctx.owner,
+          repo: ctx.name,
+          pullNumber: pull_number,
+        });
 
-      const pullRequest = response.repository?.pullRequest;
-      if (!pullRequest) {
-        return {
-          review_id,
-          pull_number,
-          comments: [],
-          count: 0,
-        };
-      }
-
-      const threadNodes = pullRequest.reviewThreads?.nodes;
-      if (!threadNodes) {
-        return {
-          review_id,
-          pull_number,
-          comments: [],
-          count: 0,
-        };
-      }
-
-      const allComments: {
-        id: number;
-        body: string;
-        path: string;
-        line: number | null;
-        side: "LEFT" | "RIGHT";
-        start_line: number | null;
-        start_side: "LEFT" | "RIGHT" | null;
-        user: string | null;
-        created_at: string;
-        updated_at: string;
-        html_url: string;
-        in_reply_to_id: number | null;
-        pull_request_review_id: number | null;
-      }[] = [];
-
-      // iterate through all threads (filter out nulls)
-      for (const thread of threadNodes) {
-        if (!thread?.comments?.nodes) continue;
-
-        // filter out null comments
-        const threadComments = thread.comments.nodes.filter(
-          (c): c is GraphQLReviewComment => c !== null
-        );
-        if (threadComments.length === 0) continue;
-
-        // find the root comment (the one with replyTo == null) to determine thread ownership
-        const rootComment = threadComments.find((c) => c.replyTo === null);
-        if (!rootComment) continue;
-
-        // check if this thread belongs to the target review using the root comment
-        const threadBelongsToReview = rootComment.pullRequestReview?.databaseId === review_id;
-        if (!threadBelongsToReview) continue;
-
-        // include all comments from this thread (original + replies)
-        // side info comes from thread level, not comment level
-        for (const comment of threadComments) {
-          allComments.push({
-            id: comment.databaseId,
-            body: comment.body,
-            path: comment.path,
-            line: comment.line,
-            start_line: comment.startLine,
-            side: thread.diffSide,
-            start_side: thread.startDiffSide,
-            user: comment.author?.login ?? null,
-            created_at: comment.createdAt,
-            updated_at: comment.updatedAt,
-            html_url: comment.url,
-            in_reply_to_id: comment.replyTo?.databaseId ?? null,
-            pull_request_review_id: comment.pullRequestReview?.databaseId ?? null,
-          });
+        const pullRequest = response.repository?.pullRequest;
+        if (!pullRequest) {
+          return {
+            review_id,
+            pull_number,
+            comments: [],
+            count: 0,
+          };
         }
-      }
 
-      return {
-        review_id,
-        pull_number,
-        comments: allComments,
-        count: allComments.length,
-      };
-    }),
+        const threadNodes = pullRequest.reviewThreads?.nodes;
+        if (!threadNodes) {
+          return {
+            review_id,
+            pull_number,
+            comments: [],
+            count: 0,
+          };
+        }
+
+        const allComments: {
+          id: number;
+          body: string;
+          path: string;
+          line: number | null;
+          side: "LEFT" | "RIGHT";
+          start_line: number | null;
+          start_side: "LEFT" | "RIGHT" | null;
+          user: string | null;
+          created_at: string;
+          updated_at: string;
+          html_url: string;
+          in_reply_to_id: number | null;
+          pull_request_review_id: number | null;
+        }[] = [];
+
+        // iterate through all threads (filter out nulls)
+        for (const thread of threadNodes) {
+          if (!thread?.comments?.nodes) continue;
+
+          // filter out null comments
+          const threadComments = thread.comments.nodes.filter(
+            (c): c is GraphQLReviewComment => c !== null
+          );
+          if (threadComments.length === 0) continue;
+
+          // find the root comment (the one with replyTo == null) to determine thread ownership
+          const rootComment = threadComments.find((c) => c.replyTo === null);
+          if (!rootComment) continue;
+
+          // check if this thread belongs to the target review using the root comment
+          const threadBelongsToReview = rootComment.pullRequestReview?.databaseId === review_id;
+          if (!threadBelongsToReview) continue;
+
+          // include all comments from this thread (original + replies)
+          // side info comes from thread level, not comment level
+          for (const comment of threadComments) {
+            allComments.push({
+              id: comment.databaseId,
+              body: comment.body,
+              path: comment.path,
+              line: comment.line,
+              start_line: comment.startLine,
+              side: thread.diffSide,
+              start_side: thread.startDiffSide,
+              user: comment.author?.login ?? null,
+              created_at: comment.createdAt,
+              updated_at: comment.updatedAt,
+              html_url: comment.url,
+              in_reply_to_id: comment.replyTo?.databaseId ?? null,
+              pull_request_review_id: comment.pullRequestReview?.databaseId ?? null,
+            });
+          }
+        }
+
+        return {
+          review_id,
+          pull_number,
+          comments: allComments,
+          count: allComments.length,
+        };
+      }),
   });
 }
 

mcp/shared.ts

@@ -2,6 +2,7 @@ import type { StandardSchemaV1 } from "@standard-schema/spec";
 import { encode as toonEncode } from "@toon-format/toon";
 import type { FastMCP, Tool } from "fastmcp";
 import type { ToolContext } from "../main.ts";
+import { formatJsonValue, log } from "../utils/cli.ts";
 
 export const tool = <const params>(toolDef: Tool<any, StandardSchemaV1<params>>) => toolDef;
 
@@ -36,13 +37,22 @@ export const handleToolError = (error: unknown): ToolResult => {
 /**
  * Helper to wrap a tool execute function with error handling.
  * Captures ctx in closure so tools don't need to handle try/catch.
+ * @param fn - the function to execute
+ * @param toolName - optional tool name for error logging
  */
-export const execute = <T>(fn: (params: T) => Promise<Record<string, any> | string>) => {
+export const execute = <T>(
+  fn: (params: T) => Promise<Record<string, any> | string>,
+  toolName?: string
+) => {
   return async (params: T): Promise<ToolResult> => {
     try {
       const result = await fn(params);
       return handleToolSuccess(result);
     } catch (error) {
+      const errorMessage = error instanceof Error ? error.message : String(error);
+      const prefix = toolName ? `[${toolName}]` : "tool";
+      log.error(`${prefix} error: ${errorMessage}`);
+      log.debug(`${prefix} params: ${formatJsonValue(params)}`);
       return handleToolError(error);
     }
   };

play.ts

@@ -25,7 +25,6 @@ export async function run(prompt: string): Promise<AgentResult> {
 
     const inputs: Inputs = {
       prompt,
-      effort: "nothink",
     };
 
     const result = await main(inputs);
@@ -187,6 +186,30 @@ Examples:
 
         if (typeof module.default === "string") {
           prompt = module.default;
+        } else if (Array.isArray(module.default)) {
+          // Array of Payloads - run each in sequence
+          const payloads = module.default;
+          log.info(`Running ${payloads.length} payloads in sequence...`);
+
+          let allSuccess = true;
+          for (let i = 0; i < payloads.length; i++) {
+            const payload = payloads[i];
+            const label = payload.effort
+              ? `[${i + 1}/${payloads.length}] effort=${payload.effort}`
+              : `[${i + 1}/${payloads.length}]`;
+            log.info(`\n${"=".repeat(60)}`);
+            log.info(`${label}`);
+            log.info(`${"=".repeat(60)}\n`);
+
+            const payloadPrompt = JSON.stringify(payload, null, 2);
+            const result = await run(payloadPrompt);
+            if (!result.success) {
+              allSuccess = false;
+              log.error(`Payload ${i + 1} failed`);
+            }
+          }
+
+          process.exit(allSuccess ? 0 : 1);
         } else if (typeof module.default === "object") {
           // Payload objects (with ~pullfrog) should be stringified
           prompt = JSON.stringify(module.default, null, 2);