shockbot/agents/claude.ts

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: mini uses Haiku for speed, auto uses opusplan for balance, max uses Opus for capability
const claudeEffortModels: Record<Effort, string> = {
  mini: "haiku",
  auto: "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 () => {
    const versionRange = packageJson.dependencies["@anthropic-ai/claude-agent-sdk"] || "latest";
    return await installFromNpmTarball({
      packageName: "@anthropic-ai/claude-agent-sdk",
      version: versionRange,
      executablePath: "cli.js",
    });
  },
  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.
    const disallowedTools = repo.isPublic ? ["Bash"] : [];
    const sandboxOptions: Options = payload.sandbox
      ? {
          permissionMode: "default",
          disallowedTools: ["Bash", "WebSearch", "WebFetch", "Write"],
          async canUseTool(toolName, input, _options) {
            if (toolName.startsWith("mcp__gh_pullfrog__"))
              return { behavior: "allow", updatedInput: input, updatedPermissions: [] };
            return { behavior: "deny", message: "tool not allowed in sandbox mode" };
          },
        }
      : {
          permissionMode: "bypassPermissions" as const,
          disallowedTools,
        };

    if (payload.sandbox) {
      log.info("🔒 sandbox mode enabled: restricting to read-only operations");
    }

    // 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: queryOptions,
    });

    // Stream the results
    for await (const message of queryInstance) {
      log.debug(JSON.stringify(message, null, 2));
      const handler = messageHandlers[message.type];
      await handler(message as never);
    }

    return {
      success: true,
      output: "",
    };
  },
});

type SDKMessageType = SDKMessage["type"];

type SDKMessageHandler<type extends SDKMessageType = SDKMessageType> = (
  data: Extract<SDKMessage, { type: type }>
) => void | Promise<void>;

type SDKMessageHandlers = {
  [type in SDKMessageType]: SDKMessageHandler<type>;
};

// Track bash tool IDs to identify when bash tool results come back
const bashToolIds = new Set<string>();

const messageHandlers: SDKMessageHandlers = {
  assistant: (data) => {
    if (data.message?.content) {
      for (const content of data.message.content) {
        if (content.type === "text" && content.text?.trim()) {
          log.box(content.text.trim(), { title: "Claude" });
        } else if (content.type === "tool_use") {
          // Track bash tool IDs
          if (content.name === "bash" && content.id) {
            bashToolIds.add(content.id);
          }

          log.toolCall({
            toolName: content.name,
            input: content.input,
          });
        }
      }
    }
  },
  user: (data) => {
    if (data.message?.content) {
      for (const content of data.message.content) {
        if (content.type === "tool_result") {
          const toolUseId = (content as any).tool_use_id;
          const isBashTool = toolUseId && bashToolIds.has(toolUseId);

          if (isBashTool) {
            // Log bash output in a collapsed group
            const outputContent =
              typeof content.content === "string"
                ? content.content
                : Array.isArray(content.content)
                  ? content.content
                      .map((c: any) => (typeof c === "string" ? c : c.text || JSON.stringify(c)))
                      .join("\n")
                  : String(content.content);

            log.startGroup(`bash output`);
            if (content.is_error) {
              log.warning(outputContent);
            } else {
              log.info(outputContent);
            }
            log.endGroup();
            // Clean up the tracked ID
            bashToolIds.delete(toolUseId);
          } else if (content.is_error) {
            const errorContent =
              typeof content.content === "string" ? content.content : String(content.content);
            log.warning(`Tool error: ${errorContent}`);
          }
        }
      }
    }
  },
  result: async (data) => {
    if (data.subtype === "success") {
      const usage = data.usage;
      const inputTokens = usage?.input_tokens || 0;
      const cacheRead = usage?.cache_read_input_tokens || 0;
      const cacheWrite = usage?.cache_creation_input_tokens || 0;
      const outputTokens = usage?.output_tokens || 0;
      const totalInput = inputTokens + cacheRead + cacheWrite;

      await log.summaryTable([
        [
          { data: "Cost", header: true },
          { data: "Input", header: true },
          { data: "Cache Read", header: true },
          { data: "Cache Write", header: true },
          { data: "Output", header: true },
        ],
        [
          `$${data.total_cost_usd?.toFixed(4) || "0.0000"}`,
          String(totalInput),
          String(cacheRead),
          String(cacheWrite),
          String(outputTokens),
        ],
      ]);
    } else if (data.subtype === "error_max_turns") {
      log.error(`Max turns reached: ${JSON.stringify(data)}`);
    } else if (data.subtype === "error_during_execution") {
      log.error(`Execution error: ${JSON.stringify(data)}`);
    } else {
      log.error(`Failed: ${JSON.stringify(data)}`);
    }
  },
  system: () => {},
  stream_event: () => {},
  tool_progress: () => {},
  auth_status: () => {},
};