/**
 * toknmtr agent — CLI entry point.
 *
 * Walks TRANSCRIPT_ROOT recursively for *.jsonl transcripts (including subagent
 * transcripts under `/subagents/`, which carry real Task/workflow sub-agent usage —
 * often Sonnet/Haiku — attributed to their parent session; only workflow `journal.jsonl`
 * orchestration files are skipped), parses only the bytes appended since the last
 * run (per-file cursor, see cursor.ts), and pushes the resulting events/tool calls/session
 * metadata to the server in one sweep.
 *
 * Usage:
 *   node --experimental-strip-types agent/run.ts [--once] [--backfill] [--interval=<ms>]
 *
 *   --once       Run a single sweep and exit (default behavior is a continuous loop that
 *                sweeps every --interval ms until killed).
 *   --backfill   Ignore stored cursors and reparse every transcript from byte 0 (cursors
 *                are still updated afterward, so later normal runs resume incrementally).
 *                Implies a single sweep (the point of a backfill is one full catch-up
 *                pass, not a repeating one). Safe to re-run — ingest is idempotent.
 *   --interval   Sweep interval in ms for continuous mode (default 60000).
 *
 * Env: TOKNMTR_URL, TOKNMTR_TOKEN (see push.ts). host = os.hostname().
 */
import { closeSync, openSync, readSync, readdirSync, statSync, type Dirent } from 'node:fs';
import { hostname } from 'node:os';
import { basename, join } from 'node:path';
import { getOffset, save as saveCursors, setOffset } from './cursor.ts';
import {
	parseTranscript,
	TRANSCRIPT_ROOT,
	type ParsedEvent,
	type SessionMeta,
	type ToolCall
} from './parse.ts';
import { pushBatch, pushUsageGauges } from './push.ts';
import { fetchUsageGauges } from './usage.ts';

const HOST = hostname();
const JOURNAL_NAME = 'journal.jsonl';
const DEFAULT_INTERVAL_MS = 60_000;
/** How often the subscription-% gauge scrape runs in continuous mode (it's slow + spawns a
 *  real `claude` session, so it must NOT run every sweep — mirrors the widget's 5-min cadence). */
const USAGE_INTERVAL_MS = 5 * 60_000;

interface Args {
	backfill: boolean;
	once: boolean;
	intervalMs: number;
}

function parseArgs(argv: string[]): Args {
	const backfill = argv.includes('--backfill');
	const once = argv.includes('--once');
	const intervalArg = argv.find((a) => a.startsWith('--interval='));
	const parsedInterval = intervalArg ? Number(intervalArg.slice('--interval='.length)) : NaN;
	const intervalMs =
		Number.isFinite(parsedInterval) && parsedInterval > 0 ? parsedInterval : DEFAULT_INTERVAL_MS;
	return { backfill, once, intervalMs };
}

function walk(dir: string, out: string[]): void {
	let entries: Dirent[];
	try {
		entries = readdirSync(dir, { withFileTypes: true });
	} catch {
		return; // root doesn't exist yet (fresh machine, no transcripts) — nothing to do
	}
	for (const entry of entries) {
		const full = join(dir, entry.name);
		if (entry.isDirectory()) walk(full, out);
		else if (entry.isFile() && entry.name.endsWith('.jsonl')) out.push(full);
	}
}

/**
 * All transcript files under TRANSCRIPT_ROOT. Includes subagent transcripts
 * (`.../subagents/agent-<id>.jsonl`) — real Task/workflow sub-agent usage that attributes
 * to the parent session via each line's own `sessionId`. Only workflow `journal.jsonl`
 * files are excluded: they carry orchestration bookkeeping (`type:"started"` lines with no
 * `uuid`), not message events, so they'd yield zero events anyway.
 */
function listTranscripts(): string[] {
	const out: string[] = [];
	walk(TRANSCRIPT_ROOT, out);
	return out.filter((f) => basename(f) !== JOURNAL_NAME);
}

interface NewBytes {
	text: string;
	newOffset: number;
}

/**
 * Read the byte range [offset, size) of `path` and trim it to the last complete line
 * (i.e. up to and including the last `\n`), so a file mid-write never has a partial JSON
 * line handed to the parser. The trailing partial line (if any) is left for next time —
 * newOffset stays before it.
 */
function readNewLines(path: string, offset: number, size: number): NewBytes {
	if (offset >= size) return { text: '', newOffset: offset };
	const length = size - offset;
	const buf = Buffer.alloc(length);
	const fd = openSync(path, 'r');
	try {
		readSync(fd, buf, 0, length, offset);
	} finally {
		closeSync(fd);
	}
	const lastNewline = buf.lastIndexOf(0x0a); // '\n'
	if (lastNewline === -1) return { text: '', newOffset: offset };
	return {
		text: buf.subarray(0, lastNewline + 1).toString('utf8'),
		newOffset: offset + lastNewline + 1
	};
}

interface CursorUpdate {
	path: string;
	offset: number;
	size: number;
}

async function sweep(backfill: boolean): Promise<void> {
	const files = listTranscripts();
	const events: ParsedEvent[] = [];
	const toolCalls: ToolCall[] = [];
	const sessions: SessionMeta[] = [];
	const cursorUpdates: CursorUpdate[] = [];
	let filesWithNewData = 0;

	for (const path of files) {
		let size: number;
		try {
			size = statSync(path).size;
		} catch {
			continue; // file disappeared mid-walk — skip
		}

		const offset = backfill ? 0 : getOffset(path, size);
		const { text, newOffset } = readNewLines(path, offset, size);
		if (!text) continue;

		const sessionIdFallback = basename(path, '.jsonl');
		const result = parseTranscript(HOST, sessionIdFallback, text);
		events.push(...result.events);
		toolCalls.push(...result.toolCalls);
		if (result.events.length > 0 || result.toolCalls.length > 0) sessions.push(result.session);

		cursorUpdates.push({ path, offset: newOffset, size });
		filesWithNewData++;
	}

	const pushResult = await pushBatch({ host: HOST, events, toolCalls, sessions });

	// Only commit cursors after a successful push, so a failed/unreachable-server push
	// gets retried (nothing is silently lost) on the next sweep. This runs even after a
	// --backfill sweep (which ignored the *old* cursor values when deciding where to start
	// reading) so subsequent normal runs pick up incrementally from here rather than
	// re-walking full history every time.
	for (const u of cursorUpdates) setOffset(u.path, u.offset, u.size);
	if (cursorUpdates.length > 0) saveCursors();

	console.log(
		`toknmtr agent: scanned ${files.length} transcript(s), ${filesWithNewData} with new data, ` +
			`pushed ${pushResult.received} event(s) in ${pushResult.requests} request(s) ` +
			`(${toolCalls.length} tool call(s), ${sessions.length} session(s)).`
	);
}

/**
 * Best-effort subscription-% gauge scrape + push. Never throws — a missing PTY driver, no
 * `claude` on PATH, or an unreachable server must not break the transcript sweep. Disabled
 * entirely by setting TOKNMTR_NO_USAGE=1 (e.g. on a headless container with no `claude`).
 */
async function sweepUsage(): Promise<void> {
	if (process.env.TOKNMTR_NO_USAGE === '1') return;
	try {
		const gauges = await fetchUsageGauges();
		if (!gauges) return;
		await pushUsageGauges(gauges);
		console.log(
			`toknmtr agent: pushed usage gauges (session=${gauges.session_pct ?? '–'}% ` +
				`week=${gauges.week_all_pct ?? '–'}% sonnet=${gauges.week_sonnet_pct ?? '–'}%).`
		);
	} catch (err) {
		console.error('toknmtr agent: usage gauge scrape/push failed (non-fatal):', err);
	}
}

async function main(): Promise<void> {
	const args = parseArgs(process.argv.slice(2));

	// A backfill is a pure historical-transcript catch-up: skip the (slow, live-only) gauge scrape.
	if (args.backfill) {
		await sweep(true);
		return;
	}

	if (args.once) {
		await sweep(false);
		await sweepUsage();
		return;
	}

	// Continuous mode: sweep transcripts every --interval, scrape usage gauges on a slower cadence.
	let lastUsageAt = 0;
	for (;;) {
		await sweep(false);
		if (Date.now() - lastUsageAt >= USAGE_INTERVAL_MS) {
			await sweepUsage();
			lastUsageAt = Date.now();
		}
		await new Promise((resolve) => setTimeout(resolve, args.intervalMs));
	}
}

main().catch((err: unknown) => {
	console.error('toknmtr agent failed:', err);
	process.exitCode = 1;
});