import { z } from 'zod/v4'
import { ALL_MODELS } from './models.js'
import { SupportedChatModel, fallbackModel } from './config.js'
// Re-export for consumers
export { ALL_MODELS, SupportedChatModel }
export type { SupportedChatModel as SupportedChatModelType }
export const TaskMode = z.enum(['review', 'debug', 'plan', 'create', 'general'])
export type TaskMode = z.infer<typeof TaskMode>
export const ConsultLlmArgs = z.object({
files: z
.array(z.string())
.optional()
.describe(
'Array of file paths to include as context. All files are added as context with file paths and code blocks.',
),
prompt: z
.string()
.describe(
'Your question or request for the consultant LLM. Ask neutral, open-ended questions without suggesting specific solutions to avoid biasing the analysis.',
),
model: SupportedChatModel.optional()
.default(fallbackModel)
.describe(
'LLM model to use. Prefer gpt-5.1-codex-max when user mentions Codex. This parameter is ignored when `web_mode` is `true`.',
),
task_mode: TaskMode.optional()
.default('general')
.describe(
'Controls the system prompt persona. Choose based on the task: "review": critical code reviewer for finding bugs, security issues, and quality problems. "debug": focused troubleshooter for root cause analysis from errors, logs, and stack traces — ignores style issues. "plan": constructive architect for exploring trade-offs and designing solutions — always includes a final recommendation. "create": generative writer for producing documentation, content, or designs. "general" (default): neutral prompt that defers to your instructions in the prompt field.',
),
web_mode: z
.boolean()
.optional()
.default(false)
.describe(
"If true, copy the formatted prompt to the clipboard instead of querying an LLM. When true, the `model` parameter is ignored. Use this to paste the prompt into browser-based LLM services. IMPORTANT: Only use this when the user specifically requests it. When true, wait for the user to provide the external LLM's response before proceeding with any implementation.",
),
thread_id: z
.string()
.optional()
.describe(
'Thread/session ID for resuming a conversation. Works with Codex CLI (gpt-*) and Gemini CLI (gemini-*) in CLI mode. Returned in the response prefix as [thread_id:xxx].',
),
git_diff: z
.object({
repo_path: z
.string()
.optional()
.describe(
'Path to git repository (defaults to current working directory)',
),
files: z
.array(z.string())
.min(1, 'At least one file is required for git diff')
.describe('Specific files to include in diff'),
base_ref: z
.string()
.optional()
.default('HEAD')
.describe(
'Git reference to compare against (e.g., "HEAD", "main", commit hash)',
),
})
.optional()
.describe(
'Generate git diff output to include as context. Shows uncommitted changes by default.',
),
})
const consultLlmInputSchema = z.toJSONSchema(ConsultLlmArgs, {
target: 'openapi-3.0',
})
export const toolSchema = {
name: 'consult_llm',
description: `Ask a more powerful AI for help with complex problems. Provide your question in the prompt field and always include relevant code files as context.
Be specific about what you want: code implementation, code review, bug analysis, architecture advice, etc.
IMPORTANT: Ask neutral, open-ended questions. Avoid suggesting specific solutions or alternatives in your prompt as this can bias the analysis. Instead of "Should I use X or Y approach?", ask "What's the best approach for this problem?" Let the consultant LLM provide unbiased recommendations.
For multi-turn conversations with CLI models (Codex and Gemini), the response includes a [thread_id:xxx] prefix. Extract this ID and pass it as the thread_id parameter in follow-up requests to maintain conversation context.`,
inputSchema: consultLlmInputSchema,
} as const
