// End-to-end tests of the preprocessor, commands and tools, using the SDK's real Chat and
// ChatMessage classes and a fake controller in place of LM Studio.

import * as assert from "node:assert/strict";
import * as fs from "node:fs";
import * as os from "node:os";
import * as path from "node:path";
import { test } from "node:test";

const HOME = fs.mkdtempSync(path.join(os.tmpdir(), "throughline-test-"));
process.env.THROUGHLINE_HOME = HOME;

import { loadChatState } from "../src/chat";
import { BLOCK_CLOSE, FULL_OPEN, HEADER_OPEN } from "../src/digest";
import { loadLedger, listProjects } from "../src/ledger";
import { estimateTokens, stripThroughLineBlocks } from "../src/text";
import { Session } from "./helpers";
import { firstUserText } from "../src/commands";

test("a prediction with no chat attached is left alone and gets no tools", async () => {
  const s = new Session(null);
  const out = await s.send("Hello.");
  assert.equal(out.text, "Hello.");
  assert.deepEqual(await s.tools(), []);
});

test("a chat with no project receives nothing and gets no tools", async () => {
  const s = new Session("1000000000001");
  const out = await s.send("What is the capital of Portugal?");
  assert.equal(out.text, "What is the capital of Portugal?");
  assert.deepEqual(out.statuses, []);
  assert.deepEqual(await s.tools(), []);
  assert.equal(loadChatState("1000000000001"), null, "nothing written for an ordinary chat");
});

test("a full project lifecycle in one chat, then continued in another", async () => {
  const s = new Session("1000000000002");
  await s.send("Compare the licenses and default context lengths of three inference runtimes, citing the source for each.");
  s.reply("Sure.");

  // Start without a goal: the first message of the chat becomes the goal, word for word.
  const started = await s.send("/tl start");
  assert.match(started.text, /^\[throughLine handled a command/);
  assert.match(started.text, /Started the project "compare-licenses-default-context-lengths/);
  s.reply("Started.");
  const slug = loadChatState("1000000000002")!.project!;
  const ledger = loadLedger(slug)!;
  assert.equal(ledger.goal, "Compare the licenses and default context lengths of three inference runtimes, citing the source for each.");
  assert.equal(ledger.goalSource, "user_message");
  assert.equal((await s.tools()).length, 4);

  // First message after the start carries the full state.
  const first = await s.send("Go ahead.");
  assert.ok(first.text.includes(FULL_OPEN));
  assert.ok(first.text.endsWith(BLOCK_CLOSE));
  assert.match(first.statuses[0], /full state shown/);

  // The model plans and records through its tools; results stay short.
  let r = await s.call("ledger_plan", { action: "add", steps: ["Find each license", "Find each default context length", "Draft the table"] });
  assert.match(r, /^status: ok\nchanged: steps 1 to 3 added/);
  r = await s.call("ledger_plan", { action: "update", step: 1, status: "doing" });
  assert.match(r, /now: 1\. Find each license \(doing\)/);
  r = await s.call("ledger_record", { kind: "finding", text: "Runtime A is MIT licensed.", source: "https://example.org/a/LICENSE" });
  assert.match(r, /F1 added: Runtime A is MIT licensed\. \(source: https:\/\/example\.org\/a\/LICENSE\)/);
  assert.ok(estimateTokens(r) < 100, `tool result is short (${estimateTokens(r)} tokens)`);
  r = await s.call("ledger_record", { kind: "finding", text: "The user lives in a small town and likes jazz.", source: "user" });
  assert.match(r, /^status: error\nreason: This reads like a personal fact about the user/);
  r = await s.call("ledger_record", { kind: "question", text: "Does runtime C publish a default at all?" });
  assert.match(r, /Q1 opened/);
  r = await s.call("ledger_resolve", { question: "Q1", answer: "No; it uses the model's maximum.", record_as_finding: true, source: "turn 7" });
  assert.match(r, /Q1 resolved; F2 added/);
  r = await s.call("ledger_record", { kind: "decision", text: "Cite primary sources only.", reason: "secondary sites were stale" });
  assert.match(r, /D1 added/);
  r = await s.call("ledger_plan", { action: "update", step: 1, status: "done" });
  r = await s.call("ledger_plan", { action: "update", step: 9, status: "done" });
  assert.match(r, /^status: error\nreason: There is no step 9/);
  s.reply("Licenses found.");

  // Next turn: the full state is still recent, so only the short header, with what changed.
  const second = await s.send("Continue.");
  assert.ok(second.text.includes(HEADER_OPEN));
  assert.ok(!second.text.includes(FULL_OPEN));
  assert.match(second.text, /Changed since the last header: steps 1 to 3 added; F1 added; Q1 opened; Q1 resolved; F2 added; D1 added; step 1 now done\n/);
  const header = second.text.slice(second.text.indexOf(HEADER_OPEN));
  assert.ok(estimateTokens(header) < 150, `header is small (${estimateTokens(header)} tokens)`);

  // Third turn with nothing changed: no change line at all.
  s.reply("Working.");
  const third = await s.send("Keep going.");
  assert.ok(!third.text.includes("Changed since"));

  // Once more than the verbatim window has passed since the last full state, it comes back.
  s.toolOutput(30000);
  const fourth = await s.send("And now?");
  assert.ok(fourth.text.includes(FULL_OPEN));
  const full = fourth.text.slice(fourth.text.indexOf(FULL_OPEN));
  assert.ok(estimateTokens(full) <= 600, `full state within cap (${estimateTokens(full)} tokens)`);
  assert.match(full, /F2 Does runtime C publish a default at all\? No; it uses the model's maximum\./);

  // Commands from the user change the goal and constraints; the model cannot.
  await s.send("/tl constraint Use only primary sources.");
  assert.deepEqual(loadLedger(slug)!.constraints, ["Use only primary sources."]);

  // A new chat continues the project; its first message gets the full state with everything.
  const s2 = new Session("1000000000003");
  assert.deepEqual(await s2.tools(), []);
  const opened = await s2.send(`/tl open ${slug.slice(0, 12)}`);
  assert.match(opened.text, /now continues the project/);
  assert.equal((await s2.tools()).length, 4);
  const cont = await s2.send("Where were we?");
  assert.ok(cont.text.includes(FULL_OPEN));
  assert.match(cont.text, /Constraints: Use only primary sources\./);
  assert.match(cont.text, /F1 Runtime A is MIT licensed\./);
  assert.match(cont.text, /D1 Cite primary sources only\./);
  assert.deepEqual(loadLedger(slug)!.chats.sort(), ["1000000000002", "1000000000003"]);

  // /tl off stops the second chat: nothing appended, no tools.
  await s2.send("/tl off");
  const after = await s2.send("Thanks.");
  assert.equal(after.text, "Thanks.");
  assert.deepEqual(await s2.tools(), []);

  // Every change is in the event log, and versions are kept.
  const events = fs.readFileSync(path.join(HOME, "projects", slug, "events.jsonl"), "utf8").trim().split("\n").map((l) => JSON.parse(l));
  assert.ok(events.some((e) => e.op === "goal_set" || e.op === "project_created"));
  assert.ok(events.filter((e) => e.op === "constraint_added").every((e) => e.actor === "user"));
  assert.ok(fs.readdirSync(path.join(HOME, "projects", slug, "versions")).length > 5);
  assert.match(fs.readFileSync(path.join(HOME, "projects", slug, "ledger.md"), "utf8"), /## Findings\n\n- F1: Runtime A is MIT licensed\./);

  // /tl forget needs the exact name, then moves the project to the trash and unbinds chats.
  const askFirst = await s.send("/tl forget");
  assert.match(askFirst.text, new RegExp(`send /tl forget ${slug}`));
  await s.send(`/tl forget ${slug}`);
  assert.ok(!listProjects().includes(slug));
  assert.equal(fs.readdirSync(path.join(HOME, "trash")).length, 1);
  assert.equal(loadChatState("1000000000002")!.project, undefined);
  const gone = await s.send("Hello again.");
  assert.equal(gone.text, "Hello again.");
});

test("a chat with documents gets the state once, then nothing added", async () => {
  const s = new Session("1000000000004");
  await s.send("/tl start Summarize the attached notes into a plan.");
  s.reply("Started.");
  const first = await s.send("Here they are.", { files: true });
  assert.ok(first.text.includes(FULL_OPEN), "first turn of the project in the chat still shows the full state");
  assert.ok(first.text.includes("not repeated on later messages"));
  s.reply("Read them.");
  const next = await s.send("Next part.");
  assert.equal(next.text, "Next part.", "nothing added, so rag-v1 searches with the user's own words");
  assert.match(next.statuses[0], /quiet: documents attached/);
  assert.equal((await s.tools()).length, 4, "the ledger tools stay available");
  const diag = fs.readFileSync(path.join(HOME, "diagnostics.jsonl"), "utf8").trim().split("\n").map((l) => JSON.parse(l));
  assert.ok(diag.every((d) => d.event === "preprocessor_order" && d.ragRanFirst === false && Object.keys(d).length === 3));
});

test("starting twice, unknown commands and help", async () => {
  const s = new Session("1000000000005");
  const empty = await s.send("/tl start");
  assert.match(empty.text, /No goal yet\. Set one with \/tl goal/);
  const again = await s.send("/tl start Something else");
  assert.match(again.text, /already tracking/);
  const unknown = await s.send("/tl frobnicate");
  assert.match(unknown.text, /Unknown command/);
  const help = await s.send("/tl");
  assert.match(help.text, /\/tl start \[goal\]/);
  assert.equal(loadChatState("1000000000005")!.project, "project");
  const goal = await s.send("/tl goal Write a short comparison.");
  assert.match(goal.text, /Goal set, and the project is now named "write-short-comparison"/);
  const renamed = loadChatState("1000000000005")!.project!;
  assert.equal(renamed, "write-short-comparison");
  assert.equal(loadLedger(renamed)!.goalSource, "command");
  assert.ok(!listProjects().includes("project"));
  const again2 = await s.send("/tl goal Write a longer comparison.");
  assert.match(again2.text, /Goal of "write-short-comparison" set to/, "only a placeholder name is replaced");
});

test("the goal is never taken from an earlier command or its stored reply", () => {
  assert.equal(firstUserText(["/tl list", "[throughLine handled a command from the user. Reply...]\nresult", "Plan my week."]), "Plan my week.");
  assert.equal(firstUserText(["/tl help"]), "");
});

test("text taken from history has throughLine's blocks removed", () => {
  const withHeader = `Do the thing.\n\n${HEADER_OPEN}\nProject: x | Goal: y\nNow: no plan yet.\n${BLOCK_CLOSE}`;
  assert.equal(stripThroughLineBlocks(withHeader), "Do the thing.");
  assert.equal(stripThroughLineBlocks("Next.\n\n[throughLine: project x, no plan yet. Background state for the assistant; ledger_view has the rest.]"), "Next.");
});

// End-to-end tests of the preprocessor, commands and tools, using the SDK's real Chat and
// ChatMessage classes and a fake controller in place of LM Studio.

import * as assert from "node:assert/strict";
import * as fs from "node:fs";
import * as os from "node:os";
import * as path from "node:path";
import { test } from "node:test";

const HOME = fs.mkdtempSync(path.join(os.tmpdir(), "throughline-test-"));
process.env.THROUGHLINE_HOME = HOME;

import { loadChatState } from "../src/chat";
import { BLOCK_CLOSE, FULL_OPEN, HEADER_OPEN } from "../src/digest";
import { loadLedger, listProjects } from "../src/ledger";
import { estimateTokens, stripThroughLineBlocks } from "../src/text";
import { Session } from "./helpers";
import { firstUserText } from "../src/commands";

test("a prediction with no chat attached is left alone and gets no tools", async () => {
  const s = new Session(null);
  const out = await s.send("Hello.");
  assert.equal(out.text, "Hello.");
  assert.deepEqual(await s.tools(), []);
});

test("a chat with no project receives nothing and gets no tools", async () => {
  const s = new Session("1000000000001");
  const out = await s.send("What is the capital of Portugal?");
  assert.equal(out.text, "What is the capital of Portugal?");
  assert.deepEqual(out.statuses, []);
  assert.deepEqual(await s.tools(), []);
  assert.equal(loadChatState("1000000000001"), null, "nothing written for an ordinary chat");
});

test("a full project lifecycle in one chat, then continued in another", async () => {
  const s = new Session("1000000000002");
  await s.send("Compare the licenses and default context lengths of three inference runtimes, citing the source for each.");
  s.reply("Sure.");

  // Start without a goal: the first message of the chat becomes the goal, word for word.
  const started = await s.send("/tl start");
  assert.match(started.text, /^\[throughLine handled a command/);
  assert.match(started.text, /Started the project "compare-licenses-default-context-lengths/);
  s.reply("Started.");
  const slug = loadChatState("1000000000002")!.project!;
  const ledger = loadLedger(slug)!;
  assert.equal(ledger.goal, "Compare the licenses and default context lengths of three inference runtimes, citing the source for each.");
  assert.equal(ledger.goalSource, "user_message");
  assert.equal((await s.tools()).length, 4);

  // First message after the start carries the full state.
  const first = await s.send("Go ahead.");
  assert.ok(first.text.includes(FULL_OPEN));
  assert.ok(first.text.endsWith(BLOCK_CLOSE));
  assert.match(first.statuses[0], /full state shown/);

  // The model plans and records through its tools; results stay short.
  let r = await s.call("ledger_plan", { action: "add", steps: ["Find each license", "Find each default context length", "Draft the table"] });
  assert.match(r, /^status: ok\nchanged: steps 1 to 3 added/);
  r = await s.call("ledger_plan", { action: "update", step: 1, status: "doing" });
  assert.match(r, /now: 1\. Find each license \(doing\)/);
  r = await s.call("ledger_record", { kind: "finding", text: "Runtime A is MIT licensed.", source: "https://example.org/a/LICENSE" });
  assert.match(r, /F1 added: Runtime A is MIT licensed\. \(source: https:\/\/example\.org\/a\/LICENSE\)/);
  assert.ok(estimateTokens(r) < 100, `tool result is short (${estimateTokens(r)} tokens)`);
  r = await s.call("ledger_record", { kind: "finding", text: "The user lives in a small town and likes jazz.", source: "user" });
  assert.match(r, /^status: error\nreason: This reads like a personal fact about the user/);
  r = await s.call("ledger_record", { kind: "question", text: "Does runtime C publish a default at all?" });
  assert.match(r, /Q1 opened/);
  r = await s.call("ledger_resolve", { question: "Q1", answer: "No; it uses the model's maximum.", record_as_finding: true, source: "turn 7" });
  assert.match(r, /Q1 resolved; F2 added/);
  r = await s.call("ledger_record", { kind: "decision", text: "Cite primary sources only.", reason: "secondary sites were stale" });
  assert.match(r, /D1 added/);
  r = await s.call("ledger_plan", { action: "update", step: 1, status: "done" });
  r = await s.call("ledger_plan", { action: "update", step: 9, status: "done" });
  assert.match(r, /^status: error\nreason: There is no step 9/);
  s.reply("Licenses found.");

  // Next turn: the full state is still recent, so only the short header, with what changed.
  const second = await s.send("Continue.");
  assert.ok(second.text.includes(HEADER_OPEN));
  assert.ok(!second.text.includes(FULL_OPEN));
  assert.match(second.text, /Changed since the last header: steps 1 to 3 added; F1 added; Q1 opened; Q1 resolved; F2 added; D1 added; step 1 now done\n/);
  const header = second.text.slice(second.text.indexOf(HEADER_OPEN));
  assert.ok(estimateTokens(header) < 150, `header is small (${estimateTokens(header)} tokens)`);

  // Third turn with nothing changed: no change line at all.
  s.reply("Working.");
  const third = await s.send("Keep going.");
  assert.ok(!third.text.includes("Changed since"));

  // Once more than the verbatim window has passed since the last full state, it comes back.
  s.toolOutput(30000);
  const fourth = await s.send("And now?");
  assert.ok(fourth.text.includes(FULL_OPEN));
  const full = fourth.text.slice(fourth.text.indexOf(FULL_OPEN));
  assert.ok(estimateTokens(full) <= 600, `full state within cap (${estimateTokens(full)} tokens)`);
  assert.match(full, /F2 Does runtime C publish a default at all\? No; it uses the model's maximum\./);

  // Commands from the user change the goal and constraints; the model cannot.
  await s.send("/tl constraint Use only primary sources.");
  assert.deepEqual(loadLedger(slug)!.constraints, ["Use only primary sources."]);

  // A new chat continues the project; its first message gets the full state with everything.
  const s2 = new Session("1000000000003");
  assert.deepEqual(await s2.tools(), []);
  const opened = await s2.send(`/tl open ${slug.slice(0, 12)}`);
  assert.match(opened.text, /now continues the project/);
  assert.equal((await s2.tools()).length, 4);
  const cont = await s2.send("Where were we?");
  assert.ok(cont.text.includes(FULL_OPEN));
  assert.match(cont.text, /Constraints: Use only primary sources\./);
  assert.match(cont.text, /F1 Runtime A is MIT licensed\./);
  assert.match(cont.text, /D1 Cite primary sources only\./);
  assert.deepEqual(loadLedger(slug)!.chats.sort(), ["1000000000002", "1000000000003"]);

  // /tl off stops the second chat: nothing appended, no tools.
  await s2.send("/tl off");
  const after = await s2.send("Thanks.");
  assert.equal(after.text, "Thanks.");
  assert.deepEqual(await s2.tools(), []);

  // Every change is in the event log, and versions are kept.
  const events = fs.readFileSync(path.join(HOME, "projects", slug, "events.jsonl"), "utf8").trim().split("\n").map((l) => JSON.parse(l));
  assert.ok(events.some((e) => e.op === "goal_set" || e.op === "project_created"));
  assert.ok(events.filter((e) => e.op === "constraint_added").every((e) => e.actor === "user"));
  assert.ok(fs.readdirSync(path.join(HOME, "projects", slug, "versions")).length > 5);
  assert.match(fs.readFileSync(path.join(HOME, "projects", slug, "ledger.md"), "utf8"), /## Findings\n\n- F1: Runtime A is MIT licensed\./);

  // /tl forget needs the exact name, then moves the project to the trash and unbinds chats.
  const askFirst = await s.send("/tl forget");
  assert.match(askFirst.text, new RegExp(`send /tl forget ${slug}`));
  await s.send(`/tl forget ${slug}`);
  assert.ok(!listProjects().includes(slug));
  assert.equal(fs.readdirSync(path.join(HOME, "trash")).length, 1);
  assert.equal(loadChatState("1000000000002")!.project, undefined);
  const gone = await s.send("Hello again.");
  assert.equal(gone.text, "Hello again.");
});

test("a chat with documents gets the state once, then nothing added", async () => {
  const s = new Session("1000000000004");
  await s.send("/tl start Summarize the attached notes into a plan.");
  s.reply("Started.");
  const first = await s.send("Here they are.", { files: true });
  assert.ok(first.text.includes(FULL_OPEN), "first turn of the project in the chat still shows the full state");
  assert.ok(first.text.includes("not repeated on later messages"));
  s.reply("Read them.");
  const next = await s.send("Next part.");
  assert.equal(next.text, "Next part.", "nothing added, so rag-v1 searches with the user's own words");
  assert.match(next.statuses[0], /quiet: documents attached/);
  assert.equal((await s.tools()).length, 4, "the ledger tools stay available");
  const diag = fs.readFileSync(path.join(HOME, "diagnostics.jsonl"), "utf8").trim().split("\n").map((l) => JSON.parse(l));
  assert.ok(diag.every((d) => d.event === "preprocessor_order" && d.ragRanFirst === false && Object.keys(d).length === 3));
});

test("starting twice, unknown commands and help", async () => {
  const s = new Session("1000000000005");
  const empty = await s.send("/tl start");
  assert.match(empty.text, /No goal yet\. Set one with \/tl goal/);
  const again = await s.send("/tl start Something else");
  assert.match(again.text, /already tracking/);
  const unknown = await s.send("/tl frobnicate");
  assert.match(unknown.text, /Unknown command/);
  const help = await s.send("/tl");
  assert.match(help.text, /\/tl start \[goal\]/);
  assert.equal(loadChatState("1000000000005")!.project, "project");
  const goal = await s.send("/tl goal Write a short comparison.");
  assert.match(goal.text, /Goal set, and the project is now named "write-short-comparison"/);
  const renamed = loadChatState("1000000000005")!.project!;
  assert.equal(renamed, "write-short-comparison");
  assert.equal(loadLedger(renamed)!.goalSource, "command");
  assert.ok(!listProjects().includes("project"));
  const again2 = await s.send("/tl goal Write a longer comparison.");
  assert.match(again2.text, /Goal of "write-short-comparison" set to/, "only a placeholder name is replaced");
});

test("the goal is never taken from an earlier command or its stored reply", () => {
  assert.equal(firstUserText(["/tl list", "[throughLine handled a command from the user. Reply...]\nresult", "Plan my week."]), "Plan my week.");
  assert.equal(firstUserText(["/tl help"]), "");
});

test("text taken from history has throughLine's blocks removed", () => {
  const withHeader = `Do the thing.\n\n${HEADER_OPEN}\nProject: x | Goal: y\nNow: no plan yet.\n${BLOCK_CLOSE}`;
  assert.equal(stripThroughLineBlocks(withHeader), "Do the thing.");
  assert.equal(stripThroughLineBlocks("Next.\n\n[throughLine: project x, no plan yet. Background state for the assistant; ledger_view has the rest.]"), "Next.");
});