@bongos/core 1.21.174 → 1.21.176

package/scripts/gds/session-cost.js

@@ -1,430 +1,7 @@
 #!/usr/bin/env node
-// scripts/gds/session-cost.js — ADR 0028 path 2 (#560).
-//
-// Capture the cost of an interactive builder session. Reads the Claude Code
-// session transcript JSONL (which carries per-turn `message.usage` token
-// counts) AND this session's ad-hoc subagents (#563 — Agent/Task spawns at
-// <project-dir>/<session-id>/subagents/...), prices them via src/bongos/llm-pricing.js,
-// and POSTs ONE cost_log row (category=api, source=session, source_ref=<session-id>).
-// Grader subprocess transcripts are excluded (already captured via the worker
-// envelope, #486) — see priceSubagents / the ADR 0028 non-double-count invariant.
-//
-// Designed to run automatically at session end via a `SessionEnd` hook
-// (.claude/settings.json), which pipes { session_id, transcript_path, cwd,
-// reason } as JSON on stdin. Also runnable by hand:
-//   node scripts/gds/session-cost.js --transcript <path> [--session <id>]
-//   node scripts/gds/session-cost.js            # newest transcript for $PWD
-//
-// CONTRACT: best-effort, NON-FATAL, idempotent.
-//   - exits 0 on every path (a SessionEnd hook must never break session exit).
-//   - no GDS session / 401 / missing transcript / $0 cost → quiet skip.
-//   - idempotent on (source='session', source_ref=session_id): re-running on the
-//     same session is a no-op (cost_log ON CONFLICT DO NOTHING). One row per
-//     session, written once at end — the SessionEnd seam fires exactly once.
-
-const fs = require('fs');
-const os = require('os');
-const path = require('path');
-const { cliClient, loadSession, arg } = require('./cli-lib');
-const { priceUsage } = require('../../src/bongos/llm-pricing');
-const { resolveEnv } = require('../../src/instance-config');
-
-
-// Claude Code encodes a project transcript dir by replacing every non-alnum
-// char in the cwd with '-' (e.g. /Users/x/Claude-OTB → -Users-x-Claude-OTB).
-function encodeCwd(cwd) {
-  return cwd.replace(/[^a-zA-Z0-9]/g, '-');
-}
-
-// The grader (src/bongos/grader.js) spawns each worker as a `claude -p` subprocess,
-// which persists its OWN transcript in the SAME project dir as the interactive
-// session. Those are captured separately via the worker's cost envelope (#486),
-// so pricing one here would DOUBLE-COUNT grader spend (Lars review, #562). A
-// grader transcript's first user message is the worker system prompt — match it
-// to exclude it from the newest-file fallback. (The SessionEnd hook path is
-// unaffected: it passes an explicit transcript_path and never calls this.)
-// Match the SHAPE of the worker preamble, not a closed list of persona names: a
-// new worker added to modules/grading/grader-workers/ would otherwise slip past
-// this and be treated as a builder session — which now means real money, since
-// backfill-sessions.js uses this to decide what NOT to pay for. Every worker
-// opens `You are the **<Persona>** … adversarial subagent` (narc/hacker/efficiency
-// say "a specialist adversarial subagent", quality says "specialist — an
-// adversarial subagent"). Still anchored on that preamble rather than the bare
-// phrase, so a real session that merely DISCUSSES adversarial subagents in its
-// first message is not misread as one.
-const GRADER_WORKER_SIGNATURE = /You are the \*\*[^*\n]{1,40}\*\*[^\n]{0,60}?adversarial subagent/;
-
-function firstUserText(file) {
-  // Scan only until the first user message — avoids fully parsing a large
-  // interactive transcript just to classify it.
-  let raw;
-  try { raw = fs.readFileSync(file, 'utf8'); } catch (_) { return ''; }
-  for (const line of raw.split('\n')) {
-    if (!line.trim()) continue;
-    let o;
-    try { o = JSON.parse(line); } catch (_) { continue; }
-    if (o && o.message && o.message.role === 'user') {
-      const c = o.message.content;
-      return typeof c === 'string' ? c : JSON.stringify(c);
-    }
-  }
-  return '';
-}
-
-function looksLikeGraderTranscript(file) {
-  return GRADER_WORKER_SIGNATURE.test(firstUserText(file).slice(0, 400));
-}
-
-function newestTranscriptFor(cwd) {
-  const dir = path.join(os.homedir(), '.claude', 'projects', encodeCwd(cwd));
-  let files;
-  try {
-    files = fs.readdirSync(dir).filter((f) => f.endsWith('.jsonl'));
-  } catch (_) {
-    return null;
-  }
-  if (!files.length) return null;
-  const withMtime = files.map((f) => {
-    const full = path.join(dir, f);
-    return { full, mtime: fs.statSync(full).mtimeMs };
-  });
-  withMtime.sort((a, b) => b.mtime - a.mtime);
-  // Newest transcript that is NOT a grader worker subprocess (those are #486's,
-  // not ours — see GRADER_WORKER_SIGNATURE). Fall back to the plain newest only
-  // if every candidate looks like a grader (shouldn't happen in practice).
-  const picked = withMtime.find((c) => !looksLikeGraderTranscript(c.full)) || withMtime[0];
-  console.error(`session-cost: no --transcript/stdin given; using newest non-grader transcript ${path.basename(picked.full)}`);
-  return picked.full;
-}
-
-// Read piped stdin (the SessionEnd hook payload) if any; null when run on a TTY.
-function readStdin() {
-  return new Promise((resolve) => {
-    if (process.stdin.isTTY) return resolve(null);
-    let data = '';
-    process.stdin.setEncoding('utf8');
-    process.stdin.on('data', (c) => { data += c; });
-    process.stdin.on('end', () => resolve(data || null));
-    // guard: if nothing arrives promptly (non-TTY but not piped), don't hang
-    setTimeout(() => resolve(data || null), 400);
-  });
-}
-
-// --- Per-tier USD split (task 1344, System 1) --------------------------------
-// priceUsage() returns a per-tier `components` map; the ledger persists the
-// 4-way split so spend can be attributed to the tier it came from (cache_read
-// is the marathon sink). cache_write here SUMS the 5m + 1h creation tiers into
-// one figure (migration 123 has a single cache_write_usd column). Pure.
-function emptyTiers() {
-  return { fresh_input_usd: 0, cache_read_usd: 0, cache_write_usd: 0, output_usd: 0 };
-}
-// Fold one priceUsage().components object into a tier accumulator.
-function addComponents(acc, components) {
-  if (!components) return acc;
-  acc.fresh_input_usd += (components.input && components.input.usd) || 0;
-  acc.cache_read_usd += (components.cache_read && components.cache_read.usd) || 0;
-  acc.cache_write_usd += ((components.cache_write_5m && components.cache_write_5m.usd) || 0)
-                       + ((components.cache_write_1h && components.cache_write_1h.usd) || 0);
-  acc.output_usd += (components.output && components.output.usd) || 0;
-  return acc;
-}
-// Fold one tiers object into an accumulator (merging two priced subtrees).
-function addTiers(acc, t) {
-  if (!t) return acc;
-  acc.fresh_input_usd += Number(t.fresh_input_usd) || 0;
-  acc.cache_read_usd += Number(t.cache_read_usd) || 0;
-  acc.cache_write_usd += Number(t.cache_write_usd) || 0;
-  acc.output_usd += Number(t.output_usd) || 0;
-  return acc;
-}
-function roundTiers(t) {
-  return {
-    fresh_input_usd: Math.round((Number(t.fresh_input_usd) || 0) * 10000) / 10000,
-    cache_read_usd: Math.round((Number(t.cache_read_usd) || 0) * 10000) / 10000,
-    cache_write_usd: Math.round((Number(t.cache_write_usd) || 0) * 10000) / 10000,
-    output_usd: Math.round((Number(t.output_usd) || 0) * 10000) / 10000,
-  };
-}
-
-// Sum usage per model across the transcript, then price each model's total.
-//
-// task 1003217 — usage is deduped by `message.id`. Claude Code writes one
-// transcript line per content block of an assistant message (thinking / text /
-// tool_use) and repeats the SAME `message.usage` on every one of them, so
-// summing per line priced a single API call two or three times (measured 2.22x
-// inflation over 12 real transcripts). The evidence — including the cache-chain
-// identity that proves a line-group is one call — is written up above
-// buildSessionDigest in ./session-digest-build.js. A transcript whose messages
-// carry no `id` degrades to the old per-line behaviour.
-function priceTranscript(transcriptPath) {
-  const raw = fs.readFileSync(transcriptPath, 'utf8');
-  const perModel = new Map(); // model -> {input, output, read, c5, c1h}
-  const seenMessageIds = new Set();
-  for (const line of raw.split('\n')) {
-    if (!line.trim()) continue;
-    let o;
-    try { o = JSON.parse(line); } catch (_) { continue; }
-    const msg = o && o.message;
-    const usage = msg && msg.usage;
-    // Skip non-API turns: `<synthetic>` is Claude Code's marker for locally
-    // generated messages (compaction, interrupts) — no real API call, usage is
-    // a zero placeholder. Pricing them would mislabel them as an unknown model.
-    if (!usage || !msg.model || msg.model === '<synthetic>') continue;
-    // One API call, many lines — price it once (task 1003217).
-    const msgId = typeof msg.id === 'string' && msg.id ? msg.id : null;
-    if (msgId !== null) {
-      if (seenMessageIds.has(msgId)) continue;
-      seenMessageIds.add(msgId);
-    }
-    const m = perModel.get(msg.model) || { input: 0, output: 0, read: 0, c5: 0, c1h: 0 };
-    m.input += Number(usage.input_tokens) || 0;
-    m.output += Number(usage.output_tokens) || 0;
-    m.read += Number(usage.cache_read_input_tokens) || 0;
-    const cc = usage.cache_creation;
-    if (cc && (cc.ephemeral_5m_input_tokens != null || cc.ephemeral_1h_input_tokens != null)) {
-      m.c5 += Number(cc.ephemeral_5m_input_tokens) || 0;
-      m.c1h += Number(cc.ephemeral_1h_input_tokens) || 0;
-    } else {
-      m.c5 += Number(usage.cache_creation_input_tokens) || 0; // no breakdown → 5m
-    }
-    perModel.set(msg.model, m);
-  }
-  let total = 0;
-  const unknownModels = [];
-  const breakdown = [];
-  const tiers = emptyTiers();
-  for (const [model, m] of perModel) {
-    const synthUsage = {
-      input_tokens: m.input,
-      output_tokens: m.output,
-      cache_read_input_tokens: m.read,
-      cache_creation_input_tokens: m.c5 + m.c1h,
-      cache_creation: { ephemeral_5m_input_tokens: m.c5, ephemeral_1h_input_tokens: m.c1h },
-    };
-    const priced = priceUsage(model, synthUsage);
-    total += priced.usd;
-    addComponents(tiers, priced.components);
-    if (priced.unknown_model) unknownModels.push(model);
-    breakdown.push({ model, usd: priced.usd });
-  }
-  // Invariant: the four tiers sum to total (each priceUsage().usd is its own
-  // component sum). The cost ledger relies on this for the session/cron rows.
-  return { total, breakdown, unknownModels, tiers };
-}
-
-// Recursively collect *.jsonl under a dir (the session's subagents subtree).
-function listJsonlRecursive(dir) {
-  const out = [];
-  let entries;
-  try { entries = fs.readdirSync(dir, { withFileTypes: true }); } catch (_) { return out; }
-  for (const e of entries) {
-    const p = path.join(dir, e.name);
-    if (e.isDirectory()) out.push(...listJsonlRecursive(p));
-    else if (e.name.endsWith('.jsonl')) out.push(p);
-  }
-  return out;
-}
-
-// #563 (ADR 0028, Q2 gap): price this session's ad-hoc subagents. Agent/Task-tool
-// spawns persist at <project-dir>/<session-id>/subagents/...  — a sibling of the
-// main transcript FILE (<project-dir>/<session-id>.jsonl). Their token usage is
-// in no other ledger (the main transcript has 0 sidechain; #486 only covers the
-// grader). SKIP grader-signature transcripts: a grader run CAN nest inside a
-// subagents/ dir (an autonomous-builder subagent that ran a ship), and its cost
-// is already captured via the worker envelope (#486) — re-pricing it here would
-// double-count (the ADR 0028 invariant Lars flagged). Exclusion is by signature,
-// not location, precisely because the grader can appear in either place.
-function priceSubagents(sessionDir) {
-  const files = listJsonlRecursive(sessionDir);
-  let subagentTotal = 0;
-  let subagentFiles = 0;
-  let graderSkipped = 0;
-  const unknownModels = [];
-  const tiers = emptyTiers();
-  for (const f of files) {
-    if (looksLikeGraderTranscript(f)) { graderSkipped++; continue; }
-    let priced;
-    try { priced = priceTranscript(f); } catch (_) { continue; }
-    if (priced.total > 0) { subagentTotal += priced.total; subagentFiles++; }
-    addTiers(tiers, priced.tiers); // a $0 file contributes 0 → tiers stay == subagentTotal
-    unknownModels.push(...priced.unknownModels);
-  }
-  return { subagentTotal, subagentFiles, graderSkipped, unknownModels, tiers };
-}
-
-// Pure: turn a priced session (interactive transcript + subagent subtree) into
-// the cost_log row payload(s). Task 1344 splits the single legacy 'session' row
-// into TWO call-site-tagged rows — source='<base>:interactive' (source_ref =
-// sessionId) and '<base>:subagent' (source_ref = sessionId+':sub'), distinct on
-// the (source,source_ref) idempotency index — each carrying its tier split. A
-// cron-launched session sets <PREFIX>_COST_CRON_ROUTINE=<name> so base='cron:<name>'
-// instead of 'session'. The rows' amounts reconcile to the prior single-row
-// figure (priorSingle vs rowsTotal, equal within independent-rounding epsilon).
-// The routine a cron-launched session belongs to, or ''. run-routine.js writes
-// the canonical <PREFIX>_COST_CRON_ROUTINE (stepEnv); a legacy OTB_ spelling set
-// by an older runner is still read, with a deprecation warning (task 1001441).
-function cronRoutineFromEnv(env = process.env) {
-  return (resolveEnv('COST_CRON_ROUTINE', { env }) || '').trim();
-}
-
-function buildCostRows({ sessionId, inter, sub, cronRoutine }) {
-  const base = cronRoutine ? `cron:${cronRoutine}` : 'session';
-  const id8 = String(sessionId).slice(0, 8);
-  const unknownModels = [...new Set([...(inter.unknownModels || []), ...(sub.unknownModels || [])])];
-  const unkNote = unknownModels.length ? ` [unpriced model(s) priced as opus: ${unknownModels.join(',')}]` : '';
-  const round4 = (n) => Math.round((Number(n) || 0) * 10000) / 10000;
-
-  const rows = [];
-  const interAmt = round4(inter.total);
-  if (interAmt > 0) {
-    rows.push({
-      amount_usd: interAmt,
-      category: 'api',
-      source: `${base}:interactive`,
-      source_ref: sessionId,
-      description: `session ${id8} interactive — $${inter.total.toFixed(4)}${unkNote}`.slice(0, 1000),
-      tiers: roundTiers(inter.tiers),
-    });
-  }
-  const subAmt = round4(sub.subagentTotal);
-  if (subAmt > 0) {
-    const skipNote = sub.graderSkipped > 0 ? ` (skipped ${sub.graderSkipped} grader transcript(s) — in #486)` : '';
-    rows.push({
-      amount_usd: subAmt,
-      category: 'api',
-      source: `${base}:subagent`,
-      source_ref: `${sessionId}:sub`,
-      description: `session ${id8} subagent — ${sub.subagentFiles} file(s) $${sub.subagentTotal.toFixed(4)}${skipNote}${unkNote}`.slice(0, 1000),
-      tiers: roundTiers(sub.tiers),
-    });
-  }
-  const priorSingle = round4(inter.total + sub.subagentTotal);
-  // Degenerate guard: total>0 but each leg rounds to $0 → still log it as one
-  // interactive row so spend is never silently dropped.
-  if (rows.length === 0 && priorSingle > 0) {
-    rows.push({
-      amount_usd: priorSingle,
-      category: 'api',
-      source: `${base}:interactive`,
-      source_ref: sessionId,
-      description: `session ${id8} — $${priorSingle.toFixed(4)}${unkNote}`.slice(0, 1000),
-      tiers: roundTiers(addTiers(addTiers(emptyTiers(), inter.tiers), sub.tiers)),
-    });
-  }
-  const rowsTotal = round4(rows.reduce((s, r) => s + r.amount_usd, 0));
-  return { rows, base, priorSingle, rowsTotal, unknownModels };
-}
-
-async function main() {
-  // Watchdog: a SessionEnd hook must never stall session exit. If the read or
-  // the POST hangs (slow/offline network — fetch has no default timeout), bail
-  // after 6s. unref() so it never keeps the process alive on its own.
-  const watchdog = setTimeout(() => process.exit(0), 6000);
-  watchdog.unref();
-  try {
-    const stdinRaw = await readStdin();
-    let transcriptPath = arg('--transcript');
-    let sessionId = arg('--session');
-    let cwd = process.cwd();
-
-    if (stdinRaw) {
-      try {
-        const hook = JSON.parse(stdinRaw);
-        transcriptPath = transcriptPath || hook.transcript_path || null;
-        sessionId = sessionId || hook.session_id || null;
-        if (hook.cwd) cwd = hook.cwd;
-      } catch (_) { /* not JSON — ignore, fall through to defaults */ }
-    }
-
-    if (!transcriptPath) transcriptPath = newestTranscriptFor(cwd);
-    if (!transcriptPath || !fs.existsSync(transcriptPath)) {
-      // nothing to price — quiet skip
-      return;
-    }
-    if (!sessionId) sessionId = path.basename(transcriptPath, '.jsonl');
-
-    // No GDS session → nothing to attribute spend to; skip quietly.
-    const session = await loadSession();
-    if (!session || !session.token) return;
-
-    const inter = priceTranscript(transcriptPath);
-
-    // #563: also capture this session's ad-hoc subagents (sibling subtree of the
-    // main transcript file), excluding grader-signature ones (already in #486).
-    const sessionDir = path.join(path.dirname(transcriptPath), sessionId);
-    const sub = priceSubagents(sessionDir);
-
-    const total = inter.total + sub.subagentTotal;
-    if (!(total > 0)) return;
-
-    // task 1344: split into call-site-tagged, tier-carrying rows; cron sessions
-    // are tagged via <PREFIX>_COST_CRON_ROUTINE (see cronRoutineFromEnv).
-    const cronRoutine = cronRoutineFromEnv();
-    const built = buildCostRows({ sessionId, inter, sub, cronRoutine });
-    const id8 = sessionId.slice(0, 8);
-
-    // --dry-run: price + print, never POST (for verification / inspection).
-    if (process.argv.includes('--dry-run')) {
-      console.log(`session-cost (dry-run): session ${id8} → $${built.rowsTotal.toFixed(4)} across ${built.rows.length} row(s) [base=${built.base}]`);
-      for (const r of built.rows) {
-        console.log(`  ${r.source} (${r.source_ref}) $${r.amount_usd.toFixed(4)}  tiers: fresh $${r.tiers.fresh_input_usd} / cache_read $${r.tiers.cache_read_usd} / cache_write $${r.tiers.cache_write_usd} / output $${r.tiers.output_usd}`);
-      }
-      console.log(`  reconcile: prior-single $${built.priorSingle.toFixed(4)} vs rows-sum $${built.rowsTotal.toFixed(4)} (Δ $${Math.abs(built.priorSingle - built.rowsTotal).toFixed(4)})`);
-      return;
-    }
-
-    const api = await cliClient({ allowUnauthenticated: true });
-    for (const r of built.rows) {
-      const res = await api.cost.postCost({
-        body: {
-          amount_usd: r.amount_usd,
-          category: r.category,
-          source: r.source,
-          source_ref: r.source_ref,
-          description: r.description,
-          fresh_input_usd: r.tiers.fresh_input_usd,
-          cache_read_usd: r.tiers.cache_read_usd,
-          cache_write_usd: r.tiers.cache_write_usd,
-          output_usd: r.tiers.output_usd,
-        },
-      });
-
-      if (res.ok && res.data && res.data.skipped) {
-        console.error(`session-cost: already logged ${r.source} for session ${id8} (idempotent skip)`);
-      } else if (res.ok) {
-        console.error(`session-cost: logged $${r.amount_usd.toFixed(4)} for session ${id8} (source=${r.source})`);
-      } else {
-        console.error(`session-cost: skipped (non-fatal): HTTP ${res.status} for ${r.source}`);
-      }
-    }
-    return;
-  } catch (err) {
-    // Best-effort: never break session exit.
-    console.error('session-cost: skipped (non-fatal):', err && err.message ? err.message : err);
-    return;
-  }
-}
-
-// Run only when invoked directly (the SessionEnd hook / CLI). When required by a
-// test, expose the pure-ish helpers without firing the IIFE / network / exit.
-if (require.main === module) {
-  // Drain instead of process.exit(): an exit while a socket/child handle is still closing aborts node on
-  // Windows (UV_HANDLE_CLOSING, src\win\async.c) — 59 SessionStart/UserPromptSubmit runs died that way
-  // (task 1003548). The unref'd backstop only fires if a hung request keeps the loop alive.
-  main().catch(() => {}).finally(() => { process.exitCode = 0; setTimeout(() => process.exit(0), 1500).unref(); });
-}
-
-module.exports = {
-  priceTranscript,
-  priceSubagents,
-  listJsonlRecursive,
-  looksLikeGraderTranscript,
-  newestTranscriptFor,
-  // task 1344 — pure tier-split + row-building helpers (unit-tested)
-  buildCostRows,
-  cronRoutineFromEnv,
-  emptyTiers,
-  addComponents,
-  addTiers,
-  roundTiers,
-};
+// scripts/gds/session-cost.js — @bongos-forwarder (task 1004848). Generated by
+// scripts/bongos/codemod-rename-scripts-gds.js; do not edit.
+// The real script is scripts/bongos/session-cost.js. This forwarder keeps callers that still name
+// the old path working, and warns once so they get repointed. Task 1003706 removes it.
+'use strict';
+require('../bongos/forwarder.js').forward(module, 'session-cost.js');