Merge pull request #21 from pullfrog/effort

add effort as an input + support parsing from payload

README.md

@@ -84,6 +84,7 @@ on:
       prompt:
         description: 'Agent prompt'
         type: string
+    secrets: inherit
 
 permissions:
   id-token: write
@@ -105,14 +106,14 @@ jobs:
       - name: Run agent
         uses: pullfrog/action@v0
         with:
-          prompt: ${{ github.event.inputs.prompt }}
+          prompt: ${{ inputs.prompt }}
-
+        env:
           # feel free to comment out any you won't use
-          anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
-          openai_api_key: ${{ secrets.OPENAI_API_KEY }}
+          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
-          google_api_key: ${{ secrets.GOOGLE_API_KEY }}
+          GOOGLE_API_KEY: ${{ secrets.GOOGLE_API_KEY }}
-          gemini_api_key: ${{ secrets.GEMINI_API_KEY }}
+          GEMINI_API_KEY: ${{ secrets.GEMINI_API_KEY }}
-          cursor_api_key: ${{ secrets.CURSOR_API_KEY }}
+          CURSOR_API_KEY: ${{ secrets.CURSOR_API_KEY }}
           
 ```
 

action.yml

@@ -1,30 +1,18 @@
-name: "Pullfrog Claude Code Action"
+name: "Pullfrog Action"
-description: "Execute Claude Code with a prompt using Anthropic API"
+description: "Execute coding agents with a prompt"
 author: "Pullfrog"
 
 inputs:
   prompt:
-    description: "Prompt to send to Claude Code"
+    description: "Prompt to send to the agent"
     required: true
-    default: "Hello from Claude Code!"
+  effort:
-  anthropic_api_key:
+    description: "Effort level: nothink (fast), think (default), max (most capable)"
-    description: "Anthropic API key for Claude Code authentication"
-    required: false
-  openai_api_key:
-    description: "OpenAI API key for Codex authentication"
-    required: false
-  google_api_key:
-    description: "Google API key for Jules authentication"
-    required: false
-  gemini_api_key:
-    description: "Gemini API key for Jules authentication"
-    required: false
-  cursor_api_key:
-    description: "Cursor API key for Cursor authentication"
     required: false
+    default: "think"
 
 runs:
-  using: "node20"
+  using: "node24"
   main: "entry"
 
 branding:

agents/claude.ts

@@ -1,9 +1,24 @@
 import { type Options, query, type SDKMessage } from "@anthropic-ai/claude-agent-sdk";
+import type { Effort } from "../external.ts";
 import packageJson from "../package.json" with { type: "json" };
 import { log } from "../utils/cli.ts";
 import { addInstructions } from "./instructions.ts";
 import { agent, createAgentEnv, installFromNpmTarball } from "./shared.ts";
 
+// 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: "haiku",
+  think: "opusplan",
+  max: "opus",
+};
+
+// 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",
   install: async () => {
@@ -14,13 +29,17 @@ export const claude = agent({
       executablePath: "cli.js",
     });
   },
-  run: async ({ payload, mcpServers, apiKey, cliPath, repo }) => {
+  run: async ({ payload, mcpServers, apiKey, cliPath, repo, effort }) => {
     // Ensure API key is NOT in process.env - only pass via SDK's env option
     delete process.env.ANTHROPIC_API_KEY;
 
     const prompt = addInstructions({ payload, repo });
     log.group("Full prompt", () => log.info(prompt));
 
+    // 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.
     // for private repos, native Bash is allowed since secrets are less exposed.
@@ -46,15 +65,17 @@ export const claude = agent({
 
     // Pass secrets via SDK's env option only (not process.env)
     // This ensures secrets are only available to Claude Code subprocess, not user code
+    const queryOptions: Options = {
+      ...sandboxOptions,
+      mcpServers,
+      model,
+      pathToClaudeCodeExecutable: cliPath,
+      env: createAgentEnv({ ANTHROPIC_API_KEY: apiKey }),
+    };
+
     const queryInstance = query({
       prompt,
-      options: {
+      options: queryOptions,
-        ...sandboxOptions,
-        mcpServers,
-        // model: "claude-opus-4-5",
-        pathToClaudeCodeExecutable: cliPath,
-        env: createAgentEnv({ ANTHROPIC_API_KEY: apiKey }),
-      },
     });
 
     // Stream the results

agents/codex.ts

@@ -1,11 +1,33 @@
 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 modelReasoningEffort parameter from ThreadOptions
+const codexReasoningEffort: Record<Effort, ModelReasoningEffort | undefined> = {
+  nothink: "low",
+  think: undefined, // use default
+  max: "high",
+};
+
 interface WriteCodexConfigParams {
   tempHome: string;
   mcpServers: Record<string, McpHttpServerConfig>;
@@ -60,7 +82,7 @@ export const codex = agent({
       executablePath: "bin/codex.js",
     });
   },
-  run: async ({ payload, mcpServers, apiKey, cliPath, repo }) => {
+  run: async ({ payload, mcpServers, apiKey, cliPath, repo, effort }) => {
     const tempHome = process.env.PULLFROG_TEMP_DIR!;
 
     // create config directory for codex before setting HOME
@@ -79,6 +101,14 @@ export const codex = agent({
       CODEX_HOME: codexDir, // point Codex to our config directory
     });
 
+    // 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 modelReasoningEffort: ${modelReasoningEffort}`);
+    }
+
     // Configure Codex
     const codexOptions: CodexOptions = {
       apiKey,
@@ -90,20 +120,28 @@ export const codex = agent({
     }
 
     const codex = new Codex(codexOptions);
-    const thread = codex.startThread(
+
-      payload.sandbox
+    // Build thread options with model and optional model_reasoning_effort
-        ? {
+    const baseThreadOptions = payload.sandbox
-            approvalPolicy: "never",
+      ? {
-            sandboxMode: "read-only",
+          model,
-            networkAccessEnabled: false,
+          approvalPolicy: "never" as const,
-          }
+          sandboxMode: "read-only" as const,
-        : {
+          networkAccessEnabled: false,
-            approvalPolicy: "never",
+        }
-            // use danger-full-access to allow git operations (workspace-write blocks .git directory writes)
+      : {
-            sandboxMode: "danger-full-access",
+          model,
-            networkAccessEnabled: true,
+          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: ThreadOptions = modelReasoningEffort
+      ? { ...baseThreadOptions, modelReasoningEffort }
+      : baseThreadOptions;
+
+    const thread = codex.startThread(threadOptions);
 
     try {
       const streamedTurn = await thread.runStreamed(addInstructions({ payload, repo }));

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 }) => {
+  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
-        ? [
+        ? baseArgs // --force removed in sandbox mode to enforce safety checks
-            "--print",
+        : [...baseArgs, "--force"];
-            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"];
 
       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";
@@ -11,6 +12,15 @@ import {
   installFromGithub,
 } from "./shared.ts";
 
+// 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)
 interface GeminiInitEvent {
   type: "init";
@@ -156,8 +166,12 @@ export const gemini = agent({
       ...(githubInstallationToken && { githubInstallationToken }),
     });
   },
-  run: async ({ payload, apiKey, mcpServers, cliPath, repo }) => {
+  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");
@@ -172,6 +186,8 @@ export const gemini = agent({
     if (payload.sandbox) {
       // sandbox mode: read-only tools only
       args = [
+        "--model",
+        model,
         "--allowed-tools",
         "read_file,list_directory,search_file_content,glob,save_memory,write_todos",
         "--allowed-mcp-server-names",
@@ -183,7 +199,7 @@ export const gemini = agent({
     } else {
       // normal mode: --yolo for auto-approval
       // for public repos, shell is excluded via settings.json excludeTools
-      args = ["--yolo", "--output-format=stream-json", "-p", sessionPrompt];
+      args = ["--model", model, "--yolo", "--output-format=stream-json", "-p", sessionPrompt];
       if (repo.isPublic) {
         log.info("🔒 public repo: native shell disabled via excludeTools, using MCP bash");
       }
@@ -276,16 +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.
+ * Configure Gemini CLI settings by writing to settings.json.
- * Gemini CLI uses `httpUrl` for HTTP/streamable transport, `url` for SSE transport.
+ * - MCP servers: uses `httpUrl` for HTTP/streamable transport
- * See: https://github.com/google-gemini/gemini-cli/blob/main/docs/get-started/configuration.md
+ * - 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");
@@ -329,17 +351,26 @@ function configureGeminiMcpServers({ mcpServers, isPublicRepo }: ConfigureGemini
     log.info(`Adding MCP server '${serverName}' at ${serverConfig.url}...`);
   }
 
-  // merge with existing settings, overwriting mcpServers
+  // merge with existing settings, overwriting mcpServers and modelConfig
-  // for public repos, exclude native shell tool to prevent secret leakage via env
   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}`);
 }

agents/opencode.ts

@@ -22,7 +22,15 @@ export const opencode = agent({
       installDependencies: true,
     });
   },
-  run: async ({ payload, apiKey: _apiKey, apiKeys, mcpServers, cliPath, repo }) => {
+  run: async ({
+    payload,
+    apiKey: _apiKey,
+    apiKeys,
+    mcpServers,
+    cliPath,
+    repo,
+    effort: _effort,
+  }) => {
     // 1. configure home/config directory
     const tempHome = process.env.PULLFROG_TEMP_DIR!;
     const configDir = join(tempHome, ".config", "opencode");
@@ -60,10 +68,9 @@ export const opencode = agent({
 
     // add API keys from apiKeys object
     for (const [key, value] of Object.entries(apiKeys || {})) {
-      const upperKey = key.toUpperCase();
+      env[key.toUpperCase()] = value;
-      env[upperKey] = value;
       // also set GOOGLE_GENERATIVE_AI_API_KEY for Google provider compatibility
-      if (upperKey === "GEMINI_API_KEY") {
+      if (key === "GEMINI_API_KEY") {
         env.GOOGLE_GENERATIVE_AI_API_KEY = value;
       }
     }

agents/shared.ts

@@ -6,7 +6,13 @@ import { join } from "node:path";
 import { pipeline } from "node:stream/promises";
 import type { McpHttpServerConfig } from "@anthropic-ai/claude-agent-sdk";
 import type { show } from "@ark/util";
-import { type AgentManifest, type AgentName, agentsManifest, type Payload } from "../external.ts";
+import {
+  type AgentManifest,
+  type AgentName,
+  agentsManifest,
+  type Effort,
+  type Payload,
+} from "../external.ts";
 import { log } from "../utils/cli.ts";
 import { getGitHubInstallationToken } from "../utils/github.ts";
 
@@ -40,6 +46,7 @@ export interface AgentConfig {
   mcpServers: Record<string, McpHttpServerConfig>;
   cliPath: string;
   repo: RepoInfo;
+  effort: Effort;
 }
 
 /**

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` | `haiku` | Fast, efficient |
+| `think` | `opusplan` | Opus for planning, Sonnet for execution |
+| `max` | `opus` | Full 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 |

entry.ts

@@ -5,9 +5,7 @@
  */
 
 import * as core from "@actions/core";
-import { flatMorph } from "@ark/util";
+import { Inputs, main } from "./main.ts";
-import { agents } from "./agents/index.ts";
-import { type Inputs, main } from "./main.ts";
 import { log } from "./utils/cli.ts";
 
 async function run(): Promise<void> {
@@ -20,12 +18,10 @@ async function run(): Promise<void> {
   }
 
   try {
-    const inputs: Required<Inputs> = {
+    const inputs = Inputs.assert({
       prompt: core.getInput("prompt", { required: true }),
-      ...flatMorph(agents, (_, agent) =>
+      effort: core.getInput("effort") || "think",
-        agent.apiKeyNames.map((inputKey) => [inputKey, core.getInput(inputKey)])
+    });
-      ),
-    };
 
     const result = await main(inputs);
 

external.ts

@@ -20,22 +20,22 @@ export interface AgentManifest {
 export const agentsManifest = {
   claude: {
     displayName: "Claude Code",
-    apiKeyNames: ["anthropic_api_key"],
+    apiKeyNames: ["ANTHROPIC_API_KEY"],
     url: "https://claude.com/claude-code",
   },
   codex: {
     displayName: "Codex CLI",
-    apiKeyNames: ["openai_api_key"],
+    apiKeyNames: ["OPENAI_API_KEY"],
     url: "https://platform.openai.com/docs/guides/codex",
   },
   cursor: {
     displayName: "Cursor CLI",
-    apiKeyNames: ["cursor_api_key"],
+    apiKeyNames: ["CURSOR_API_KEY"],
     url: "https://cursor.com/",
   },
   gemini: {
     displayName: "Gemini CLI",
-    apiKeyNames: ["google_api_key", "gemini_api_key"],
+    apiKeyNames: ["GOOGLE_API_KEY", "GEMINI_API_KEY"],
     url: "https://ai.google.dev/gemini-api/docs",
   },
   opencode: {
@@ -51,6 +51,10 @@ export const AgentName = type.enumerated(...Object.keys(agentsManifest));
 
 export type AgentApiKeyName = (typeof agentsManifest)[AgentName]["apiKeyNames"][number];
 
+// effort level type - controls model selection and thinking level
+export const Effort = type.enumerated("nothink", "think", "max");
+export type Effort = typeof Effort.infer;
+
 // base interface for common payload event fields
 interface BasePayloadEvent {
   issue_number?: number;
@@ -279,6 +283,12 @@ export interface Payload extends DispatchOptions {
    */
   modes: readonly Mode[];
 
+  /**
+   * Effort level for model selection (nothink, think, max)
+   * Defaults to "think" if not specified
+   */
+  readonly effort?: Effort;
+
   /**
    * Optional IDs of the issue, PR, or comment that the agent is working on
    */

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[];

main.ts

@@ -7,8 +7,7 @@ import { encode as toonEncode } from "@toon-format/toon";
 import { type } from "arktype";
 import { type Agent, agents } from "./agents/index.ts";
 import type { AgentResult } from "./agents/shared.ts";
-import type { AgentName, Payload } from "./external.ts";
+import { type AgentName, agentsManifest, Effort, type Payload } from "./external.ts";
-import { agentsManifest } from "./external.ts";
 import { ensureProgressCommentUpdated, reportProgress } from "./mcp/comment.ts";
 import { createMcpConfigs } from "./mcp/config.ts";
 import { startMcpHttpServer } from "./mcp/server.ts";
@@ -18,29 +17,13 @@ import type { PrepResult } from "./prep/index.ts";
 import { fetchRepoSettings, fetchWorkflowRunInfo, type RepoSettings } from "./utils/api.ts";
 import { log } from "./utils/cli.ts";
 import { reportErrorToComment } from "./utils/errorReport.ts";
-import {
+import { createOctokit, parseRepoContext, setupGitHubInstallationToken } from "./utils/github.ts";
-  createOctokit,
-  parseRepoContext,
-  setupGitHubInstallationToken,
-} from "./utils/github.ts";
 import { setupGitAuth, setupGitConfig } from "./utils/setup.ts";
 import { Timer } from "./utils/timer.ts";
 
-// runtime validation using agents (needed for ArkType)
-// Note: The AgentName type is defined in external.ts, this is the runtime validator
-
-export const AgentInputKey = type.enumerated(
-  ...Object.values(agents).flatMap((agent) => agent.apiKeyNames)
-);
-export type AgentInputKey = typeof AgentInputKey.infer;
-
-const keyInputDefs = flatMorph(agents, (_, agent) =>
-  agent.apiKeyNames.map((inputKey) => [inputKey, "string | undefined?"] as const)
-);
-
 export const Inputs = type({
   prompt: "string",
-  ...keyInputDefs,
+  "effort?": Effort,
 });
 
 export type Inputs = typeof Inputs.infer;
@@ -84,7 +67,6 @@ export async function main(inputs: Inputs): Promise<MainResult> {
 
     // phase 3: resolve agent (needs repo settings)
     const agent = resolveAgent({
-      inputs,
       payload,
       repoSettings: githubSetup.repoSettings,
     });
@@ -93,7 +75,6 @@ export async function main(inputs: Inputs): Promise<MainResult> {
     // phase 4: validate API key (sync, needs agent) - fail fast before long-running operations
     const apiKeySetup = validateApiKey({
       agent,
-      inputs,
       owner: githubSetup.owner,
       name: githubSetup.name,
     });
@@ -225,24 +206,19 @@ export async function main(inputs: Inputs): Promise<MainResult> {
 }
 
 /**
- * Get agents that have matching API keys in the inputs
+ * Check if an agent has API keys available (from process.env)
  */
-/**
+function agentHasApiKeys(agent: Agent): boolean {
- * Check if an agent has API keys available (inputs or process.env for opencode)
- */
-function agentHasApiKeys(agent: Agent, inputs: Inputs): boolean {
   if (agent.name === "opencode") {
-    // check inputs first, then process.env
+    // opencode accepts any API_KEY from environment
-    const hasInputKey = Object.keys(inputs).some((key) => key.includes("api_key"));
-    if (hasInputKey) return true;
     return Object.keys(process.env).some((key) => key.includes("API_KEY") && process.env[key]);
   }
-  const inputsRecord = inputs as Record<string, string | undefined>;
+  // check if any of the agent's expected keys are in environment
-  return agent.apiKeyNames.some((inputKey) => inputsRecord[inputKey]);
+  return agent.apiKeyNames.some((envKey) => !!process.env[envKey]);
 }
 
-function getAvailableAgents(inputs: Inputs): Agent[] {
+function getAvailableAgents(): Agent[] {
-  return Object.values(agents).filter((agent) => agentHasApiKeys(agent, inputs));
+  return Object.values(agents).filter((agent) => agentHasApiKeys(agent));
 }
 
 /**
@@ -275,7 +251,7 @@ function buildMissingApiKeyError(params: { agent: Agent; owner: string; name: st
   } else {
     const inputKeys =
       params.agent.apiKeyNames.length > 0 ? params.agent.apiKeyNames : getAllPossibleKeyNames();
-    const secretNames = inputKeys.map((key) => `\`${key.toUpperCase()}\``);
+    const secretNames = inputKeys.map((key) => `\`${key}\``);
     secretNameList = inputKeys.length === 1 ? secretNames[0] : `one of ${secretNames.join(" or ")}`;
   }
 
@@ -361,16 +337,16 @@ async function initializeGitHub(token: string): Promise<GitHubSetup> {
 }
 
 function resolveAgent({
-  inputs,
   payload,
   repoSettings,
 }: {
-  inputs: Inputs;
   payload: Payload;
   repoSettings: RepoSettings;
 }): Agent {
   const agentOverride = process.env.AGENT_OVERRIDE as AgentName | undefined;
-  log.debug(`» determineAgent: agentOverride=${agentOverride}, payload.agent=${payload.agent}, repoSettings.defaultAgent=${repoSettings.defaultAgent}`);
+  log.debug(
+    `» determineAgent: agentOverride=${agentOverride}, payload.agent=${payload.agent}, repoSettings.defaultAgent=${repoSettings.defaultAgent}`
+  );
   const configuredAgentName = agentOverride || payload.agent || repoSettings.defaultAgent || null;
 
   if (configuredAgentName) {
@@ -388,13 +364,13 @@ function resolveAgent({
     }
 
     // for repo-level defaults, check if agent has matching keys before selecting
-    if (agentHasApiKeys(agent, inputs)) {
+    if (agentHasApiKeys(agent)) {
       log.info(`Selected configured agent: ${agent.name}`);
       return agent;
     }
 
     // fall through to auto-selection
-    const availableAgents = getAvailableAgents(inputs);
+    const availableAgents = getAvailableAgents();
     log.warning(
       `Repo default agent ${agent.name} has no matching API keys. Available: ${
         availableAgents.map((a) => a.name).join(", ") || "none"
@@ -402,7 +378,7 @@ function resolveAgent({
     );
   }
 
-  const availableAgents = getAvailableAgents(inputs);
+  const availableAgents = getAvailableAgents();
   if (availableAgents.length === 0) {
     throw new Error("no agents available - missing API keys");
   }
@@ -425,8 +401,13 @@ function parsePayload(inputs: Inputs): Payload {
     if (!("~pullfrog" in parsedPrompt)) {
       throw new Error();
     }
-    return parsedPrompt as Payload;
+    // internal invocation: use effort from payload, fallback to input, default to "think"
+    return {
+      ...parsedPrompt,
+      effort: parsedPrompt.effort ?? inputs.effort ?? "think",
+    } as Payload;
   } catch {
+    // external invocation: use effort from input
     return {
       "~pullfrog": true,
       agent: null,
@@ -435,6 +416,7 @@ function parsePayload(inputs: Inputs): Payload {
         trigger: "unknown",
       },
       modes,
+      effort: inputs.effort ?? "think",
     };
   }
 }
@@ -447,22 +429,22 @@ async function installAgentCli(params: { agent: Agent; token: string }): Promise
   return params.agent.install();
 }
 
-function collectApiKeys(agent: Agent, inputs: Inputs): Record<string, string> {
+function collectApiKeys(agent: Agent): Record<string, string> {
   const apiKeys: Record<string, string> = {};
-  const inputsRecord = inputs as Record<string, string | undefined>;
 
-  for (const inputKey of agent.apiKeyNames) {
+  // read API keys from environment variables
-    const value = inputsRecord[inputKey];
+  for (const envKey of agent.apiKeyNames) {
+    const value = process.env[envKey];
     if (value) {
-      apiKeys[inputKey] = value;
+      apiKeys[envKey] = value;
     }
   }
 
-  // for OpenCode: also check process.env for any API_KEY variables
+  // for OpenCode: check process.env for any API_KEY variables
   if (agent.name === "opencode" && Object.keys(apiKeys).length === 0) {
     for (const [key, value] of Object.entries(process.env)) {
       if (value && typeof value === "string" && key.includes("API_KEY")) {
-        apiKeys[key.toLowerCase()] = value;
+        apiKeys[key] = value;
       }
     }
   }
@@ -470,13 +452,8 @@ function collectApiKeys(agent: Agent, inputs: Inputs): Record<string, string> {
   return apiKeys;
 }
 
-function validateApiKey(params: {
+function validateApiKey(params: { agent: Agent; owner: string; name: string }): ApiKeySetup {
-  agent: Agent;
+  const apiKeys = collectApiKeys(params.agent);
-  inputs: Inputs;
-  owner: string;
-  name: string;
-}): ApiKeySetup {
-  const apiKeys = collectApiKeys(params.agent, params.inputs);
 
   if (Object.keys(apiKeys).length === 0) {
     return {
@@ -497,7 +474,8 @@ function validateApiKey(params: {
 }
 
 async function runAgent(ctx: AgentContext): Promise<AgentResult> {
-  log.info(`Running ${ctx.agent.name}...`);
+  const effort = ctx.payload.effort ?? "think";
+  log.info(`Running ${ctx.agent.name} with effort=${effort}...`);
   // strip context from event
   const { context: _context, ...eventWithoutContext } = ctx.payload.event;
   // format: prompt + two newlines + TOON encoded event
@@ -516,6 +494,7 @@ async function runAgent(ctx: AgentContext): Promise<AgentResult> {
       defaultBranch: ctx.repo.default_branch,
       isPublic: !ctx.repo.private,
     },
+    effort,
   });
 }
 

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
+        // fetch all review threads using graphql
-      const response = await ctx.octokit.graphql<GraphQLResponse>(REVIEW_THREADS_QUERY, {
+        const response = await ctx.octokit.graphql<GraphQLResponse>(REVIEW_THREADS_QUERY, {
-        owner: ctx.owner,
+          owner: ctx.owner,
-        repo: ctx.name,
+          repo: ctx.name,
-        pullNumber: pull_number,
+          pullNumber: pull_number,
-      });
+        });
 
-      const pullRequest = response.repository?.pullRequest;
+        const pullRequest = response.repository?.pullRequest;
-      if (!pullRequest) {
+        if (!pullRequest) {
-        return {
+          return {
-          review_id,
+            review_id,
-          pull_number,
+            pull_number,
-          comments: [],
+            comments: [],
-          count: 0,
+            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,
-          });
         }
-      }
 
-      return {
+        const threadNodes = pullRequest.reviewThreads?.nodes;
-        review_id,
+        if (!threadNodes) {
-        pull_number,
+          return {
-        comments: allComments,
+            review_id,
-        count: allComments.length,
+            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

@@ -3,10 +3,8 @@ import { existsSync, readFileSync } from "node:fs";
 import { extname, join, resolve } from "node:path";
 import { pathToFileURL } from "node:url";
 import { fromHere } from "@ark/fs";
-import { flatMorph } from "@ark/util";
 import arg from "arg";
 import { config } from "dotenv";
-import { agents } from "./agents/index.ts";
 import type { AgentResult } from "./agents/shared.ts";
 import { type Inputs, main } from "./main.ts";
 import { log } from "./utils/cli.ts";
@@ -25,26 +23,9 @@ export async function run(prompt: string): Promise<AgentResult> {
     const originalCwd = process.cwd();
     process.chdir(tempDir);
 
-    // check if prompt is a pullfrog payload and extract agent
+    const inputs: Inputs = {
-    // note: agent from payload will be used by determineAgent with highest precedence
-    // we don't need to extract it here since main() will parse the payload
-    const inputs = {
       prompt,
-      ...flatMorph(agents, (_, agent) => {
+    };
-        // for OpenCode, scan all API_KEY environment variables
-        if (agent.name === "opencode") {
-          const opencodeKeys: Array<[string, string | undefined]> = [];
-          for (const [key, value] of Object.entries(process.env)) {
-            if (value && typeof value === "string" && key.includes("API_KEY")) {
-              opencodeKeys.push([key.toLowerCase(), value]);
-            }
-          }
-          return opencodeKeys;
-        }
-        // for other agents, use apiKeyNames
-        return agent.apiKeyNames.map((inputKey) => [inputKey, process.env[inputKey.toUpperCase()]]);
-      }),
-    } as Required<Inputs>;
 
     const result = await main(inputs);
 
@@ -205,6 +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);