// The tools simpleSearch offers a model, shared by the MCP server (main.ts) and the LM Studio
// plugin (index.ts), so both describe and run them the same way.

import { z } from "zod";
import { KNOWN_ENGINES } from "./config";
import { Ctx, inspectLayout, proposeRules, readWebpage, selfTest, webSearch } from "./ops";
import { annotate } from "./research";

export interface ToolSpec<S extends z.ZodRawShape = z.ZodRawShape> {
  name: string;
  description: string;
  shape: S;
  run: (args: z.infer<z.ZodObject<S>>) => Promise<string>;
}

const spec = <S extends z.ZodRawShape>(s: ToolSpec<S>): ToolSpec => s as unknown as ToolSpec;

export function toolSpecs(ctx: Ctx): ToolSpec[] {
  const engineEnum = z.enum(KNOWN_ENGINES);

  // Timing for each search and page read: start, how many other calls were running at that
  // moment, and duration with the result status. Counts and times only; never a query or address.
  let inFlight = 0;
  const timed = async (kind: string, fn: () => Promise<string>): Promise<string> => {
    const n = ++ctx.calls;
    const others = inFlight++;
    const started = Date.now();
    ctx.log(`call ${n} (${kind}) started, ${others} other call(s) running`);
    try {
      const out = await fn();
      const status = /^status: (\w+)/.exec(out)?.[1] ?? "?";
      ctx.log(`call ${n} (${kind}) finished in ${((Date.now() - started) / 1000).toFixed(1)} s: ${status}`);
      return out;
    } finally {
      inFlight--;
    }
  };

  const goalField = z
    .string()
    .optional()
    .describe("The user's question or the task this call serves, in one sentence. Pass the same goal on every call for one task.");

  return [
    spec({
      name: "web_search",
      description:
        "Search the web through the user's browser. Returns a status line and a numbered list of results with title, URL and description. Read promising results with read_webpage before relying on them.",
      shape: {
        query: z.string().describe("What to search for, in plain words."),
        goal: goalField,
        engine: engineEnum.optional().describe("Only use this engine. Leave empty to use the configured engines in order."),
      },
      run: async ({ query, goal, engine }) => {
        const r = ctx.research.resolve(goal, query);
        if (r && ctx.research.isBlocked(r)) return ctx.research.refuse(r, "search");
        return annotate(await timed("search", () => webSearch(ctx, query, engine)), r, ctx.research, "search");
      },
    }),
    spec({
      name: "read_webpage",
      description:
        "Open a URL in the user's browser and return its text. With a goal, long pages come back as the passages most relevant to it plus an outline; pass full: true for the whole text. Long text comes in parts; to continue, call again with the offset given in the previous response.",
      shape: {
        url: z.string().describe("Full URL starting with https:// or http://"),
        goal: goalField,
        full: z.boolean().optional().describe("Return the whole text in order instead of the passages most relevant to the goal."),
        offset: z.number().int().min(0).optional().describe("Character position to continue from. Leave empty for the start of the page."),
      },
      run: async ({ url, goal, full, offset }) => {
        const r = ctx.research.resolve(goal);
        if (r && ctx.research.isBlocked(r)) return ctx.research.refuse(r, "page");
        // Only a goal the model stated focuses a page; a task named after a search query is too rough.
        const out = await timed("read", () => readWebpage(ctx, url, offset ?? 0, { goal: ctx.research.focusGoal(r), full }));
        return annotate(out, r, ctx.research, "page");
      },
    }),
    spec({
      name: "search_self_test",
      description:
        "Check the browser connection and every search engine, and report what works. Use when searches keep failing or when the user asks whether search is working.",
      shape: {},
      run: () => selfTest(ctx),
    }),
    spec({
      name: "inspect_search_layout",
      description:
        "Use only after a search reported parser_failed, or when the user asks to repair a search engine. Loads a fresh results page from that engine and returns a sample of its structure for writing new parser rules.",
      shape: { engine: engineEnum },
      run: ({ engine }) => inspectLayout(ctx, engine),
    }),
    spec({
      name: "propose_parser_rules",
      description:
        "Submit replacement CSS selectors for one search engine's result parser, based on inspect_search_layout. They are tested against the sample page before being accepted.",
      shape: {
        engine: engineEnum,
        container: z.string().describe("Selector for the element that wraps one whole search result."),
        title: z.string().describe("Selector, inside the container, for the result's title."),
        link: z.string().describe("Selector, inside the container, for the link to the result."),
        snippet: z.string().optional().describe("Selector, inside the container, for the result's description."),
        exclude: z.string().optional().describe("Containers matching this selector are skipped, for example ads."),
        url_param: z.string().optional().describe("If result links go through a redirect on the engine's site, the query parameter holding the real URL."),
      },
      run: ({ engine, ...rules }) => proposeRules(ctx, engine, rules),
    }),
  ];
}

// The tools simpleSearch offers a model, shared by the MCP server (main.ts) and the LM Studio
// plugin (index.ts), so both describe and run them the same way.

import { z } from "zod";
import { KNOWN_ENGINES } from "./config";
import { Ctx, inspectLayout, proposeRules, readWebpage, selfTest, webSearch } from "./ops";
import { annotate } from "./research";

export interface ToolSpec<S extends z.ZodRawShape = z.ZodRawShape> {
  name: string;
  description: string;
  shape: S;
  run: (args: z.infer<z.ZodObject<S>>) => Promise<string>;
}

const spec = <S extends z.ZodRawShape>(s: ToolSpec<S>): ToolSpec => s as unknown as ToolSpec;

export function toolSpecs(ctx: Ctx): ToolSpec[] {
  const engineEnum = z.enum(KNOWN_ENGINES);

  // Timing for each search and page read: start, how many other calls were running at that
  // moment, and duration with the result status. Counts and times only; never a query or address.
  let inFlight = 0;
  const timed = async (kind: string, fn: () => Promise<string>): Promise<string> => {
    const n = ++ctx.calls;
    const others = inFlight++;
    const started = Date.now();
    ctx.log(`call ${n} (${kind}) started, ${others} other call(s) running`);
    try {
      const out = await fn();
      const status = /^status: (\w+)/.exec(out)?.[1] ?? "?";
      ctx.log(`call ${n} (${kind}) finished in ${((Date.now() - started) / 1000).toFixed(1)} s: ${status}`);
      return out;
    } finally {
      inFlight--;
    }
  };

  const goalField = z
    .string()
    .optional()
    .describe("The user's question or the task this call serves, in one sentence. Pass the same goal on every call for one task.");

  return [
    spec({
      name: "web_search",
      description:
        "Search the web through the user's browser. Returns a status line and a numbered list of results with title, URL and description. Read promising results with read_webpage before relying on them.",
      shape: {
        query: z.string().describe("What to search for, in plain words."),
        goal: goalField,
        engine: engineEnum.optional().describe("Only use this engine. Leave empty to use the configured engines in order."),
      },
      run: async ({ query, goal, engine }) => {
        const r = ctx.research.resolve(goal, query);
        if (r && ctx.research.isBlocked(r)) return ctx.research.refuse(r, "search");
        return annotate(await timed("search", () => webSearch(ctx, query, engine)), r, ctx.research, "search");
      },
    }),
    spec({
      name: "read_webpage",
      description:
        "Open a URL in the user's browser and return its text. With a goal, long pages come back as the passages most relevant to it plus an outline; pass full: true for the whole text. Long text comes in parts; to continue, call again with the offset given in the previous response.",
      shape: {
        url: z.string().describe("Full URL starting with https:// or http://"),
        goal: goalField,
        full: z.boolean().optional().describe("Return the whole text in order instead of the passages most relevant to the goal."),
        offset: z.number().int().min(0).optional().describe("Character position to continue from. Leave empty for the start of the page."),
      },
      run: async ({ url, goal, full, offset }) => {
        const r = ctx.research.resolve(goal);
        if (r && ctx.research.isBlocked(r)) return ctx.research.refuse(r, "page");
        // Only a goal the model stated focuses a page; a task named after a search query is too rough.
        const out = await timed("read", () => readWebpage(ctx, url, offset ?? 0, { goal: ctx.research.focusGoal(r), full }));
        return annotate(out, r, ctx.research, "page");
      },
    }),
    spec({
      name: "search_self_test",
      description:
        "Check the browser connection and every search engine, and report what works. Use when searches keep failing or when the user asks whether search is working.",
      shape: {},
      run: () => selfTest(ctx),
    }),
    spec({
      name: "inspect_search_layout",
      description:
        "Use only after a search reported parser_failed, or when the user asks to repair a search engine. Loads a fresh results page from that engine and returns a sample of its structure for writing new parser rules.",
      shape: { engine: engineEnum },
      run: ({ engine }) => inspectLayout(ctx, engine),
    }),
    spec({
      name: "propose_parser_rules",
      description:
        "Submit replacement CSS selectors for one search engine's result parser, based on inspect_search_layout. They are tested against the sample page before being accepted.",
      shape: {
        engine: engineEnum,
        container: z.string().describe("Selector for the element that wraps one whole search result."),
        title: z.string().describe("Selector, inside the container, for the result's title."),
        link: z.string().describe("Selector, inside the container, for the link to the result."),
        snippet: z.string().optional().describe("Selector, inside the container, for the result's description."),
        exclude: z.string().optional().describe("Containers matching this selector are skipped, for example ads."),
        url_param: z.string().optional().describe("If result links go through a redirect on the engine's site, the query parameter holding the real URL."),
      },
      run: ({ engine, ...rules }) => proposeRules(ctx, engine, rules),
    }),
  ];
}