@xhulz/nina 0.62.0 → 0.63.0

package/src/commands/sizes.mjs

@@ -0,0 +1,350 @@
+/**
+ * `nina sizes` — each step's forecast against what it landed: the `Size:` line its step file carried before it was
+ * built, beside the lines of code its commits added, the rate that measures, and the steps still to land forecast
+ * again at it.
+ *
+ * The architect forecasts each step (`Size: ~N lines in M files`) and splits past the step limit, and until this
+ * nothing compared a forecast with what landed. In the project that asked for it, two steps landed about 1.5× their
+ * forecasts before a reviewer worked the factor out by hand, and a whole architect run — 28 minutes, 375k tokens —
+ * then re-derived the rates from `git diff --numstat` to re-forecast six steps. Measured at each commit, that is a
+ * line; and as a detector (`--check`) it speaks when the newest step overran and the steps after it still carry the
+ * forecasts made before it landed, so the re-forecast is set off by the data rather than noticed.
+ *
+ * A step is its file, `.claude/plans/specs/<spec>.step-<n>.md`; its commits are those whose message names the spec and
+ * `step <n>`, as the router asks a commit to; what it landed is the lines they added outside the harness's own files.
+ *
+ *   nina sizes [--project <dir>] [--check]
+ */
+
+import { existsSync, readdirSync, readFileSync } from 'node:fs';
+import { spawnSync } from 'node:child_process';
+import { dirname, join, relative, resolve } from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { HARNESS } from '../paths.mjs';
+import { defaultVocabulary } from '../vocabulary.mjs';
+import { amber, bold, dim, pink } from '../look.mjs';
+
+const PACKAGE = dirname(dirname(dirname(fileURLToPath(import.meta.url))));
+
+/** Where the architect writes a spec's steps. */
+const SPECS = join('.claude', 'plans', 'specs');
+
+/** A landed step past its forecast by more than this sets off the detector. */
+export const OVERRUN = 1.25;
+
+/** How many of a spec's newest landed steps its rate is measured over. */
+const RATE_OVER = 5;
+
+/** Files a step's commit carries that are not the step's code: the harness's own, and lockfiles. */
+const NOT_CODE = /^(\.claude\/|\.nina\/|CLAUDE\.md$)|(^|\/)(pnpm-lock\.yaml|package-lock\.json|yarn\.lock|bun\.lockb?)$/;
+
+/**
+ * The forecast a step file states, from its `Size:` line.
+ *
+ * @param {string} text - A step file.
+ * @returns {{lines: number, files: number|null}|null}
+ */
+export function forecastIn(text) {
+  const m = /\bSize:[*_\s]*~?\s*(\d[\d,]*)\s*lines?\b([^\n]*)/i.exec(text);
+  if (!m) return null;
+  const files = /^[^.;]*?\bin\s+(\d+)\s+files?\b/i.exec(m[2]);
+  return { lines: Number(m[1].replace(/,/g, '')), files: files ? Number(files[1]) : null };
+}
+
+/** Runs git in the project, or says it could not. */
+function git(dir, args) {
+  const r = spawnSync('git', ['-C', dir, ...args], { encoding: 'utf8', maxBuffer: 256 * 1024 * 1024 });
+  return r.status === 0 ? r.stdout : null;
+}
+
+/**
+ * Many files at many revisions, read by one git process: `<rev>:<path>` → its text, or null where it did not exist.
+ * The detector runs on every turn, and a process a version came to most of its second.
+ *
+ * @param {string} dir
+ * @param {string[]} wanted
+ * @returns {Map<string, string|null>}
+ */
+function blobs(dir, wanted) {
+  const keys = [...new Set(wanted)];
+  const out = new Map(keys.map((k) => [k, null]));
+  if (keys.length === 0) return out;
+  const r = spawnSync('git', ['-C', dir, 'cat-file', '--batch'], { input: `${keys.join('\n')}\n`, maxBuffer: 256 * 1024 * 1024 });
+  if (r.status !== 0) return out;
+  const buffer = r.stdout;
+  let at = 0;
+  for (const key of keys) {
+    const end = buffer.indexOf(10, at);
+    if (end === -1) break;
+    const header = /^\S+ (\S+) (\d+)$/.exec(buffer.toString('utf8', at, end));
+    at = end + 1;
+    // `<rev>:<path> missing` carries no body.
+    if (!header) continue;
+    const size = Number(header[2]);
+    if (header[1] === 'blob') out.set(key, buffer.toString('utf8', at, at + size));
+    at += size + 1;
+  }
+  return out;
+}
+
+/** Each commit's code: the lines it added to each file that is not the harness's own, read by one git process. */
+function codeOf(dir, hashes) {
+  const out = new Map(hashes.map((h) => [h, new Map()]));
+  if (hashes.length === 0) return out;
+  for (const block of (git(dir, ['show', '--no-renames', '--numstat', '--format=%x00%H', ...hashes]) ?? '').split('\0').filter((b) => b.trim())) {
+    const [hash, ...rows] = block.split('\n');
+    const code = out.get(hash.trim());
+    if (!code) continue;
+    for (const row of rows) {
+      const [added, , path] = row.split('\t');
+      if (!path || added === '-' || NOT_CODE.test(path)) continue;
+      code.set(path, (code.get(path) ?? 0) + Number(added));
+    }
+  }
+  return out;
+}
+
+/** The commits that changed each of these files, newest first, read by one git process. */
+function changesTo(dir, paths) {
+  const out = new Map(paths.map((p) => [p, []]));
+  if (paths.length === 0) return out;
+  for (const block of (git(dir, ['log', '--no-renames', '--format=%x00%H%x09%ct', '--name-only', 'HEAD', '--', ...paths]) ?? '').split('\0').filter((b) => b.trim())) {
+    const [head, ...names] = block.split('\n');
+    const [hash, seconds] = head.split('\t');
+    for (const name of names) out.get(name.trim())?.push({ hash, t: Number(seconds) * 1000 });
+  }
+  return out;
+}
+
+/** A text as a pattern that matches it literally. */
+const literal = (text) => text.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+
+/**
+ * Whether a commit message names this step of this spec: `step 10e2` names neither `step 10e2a` nor `step 10`, and
+ * `1.6-spec-c` is not `1.6-spec-c2`.
+ */
+export function names(subject, spec, step) {
+  const whole = '(?![\\w.-]*\\w)';
+  return (
+    new RegExp(`(^|[^\\w.-])${literal(spec)}${whole}`, 'i').test(subject) && new RegExp(`\\bstep[\\s-]+${literal(step)}${whole}`, 'i').test(subject)
+  );
+}
+
+/** The step files in the working tree, as `{path, spec, step}`, in their natural order. */
+function stepFiles(dir) {
+  const root = join(dir, SPECS);
+  if (!existsSync(root)) return [];
+  const out = [];
+  for (const entry of readdirSync(root, { recursive: true })) {
+    const m = /^(.*)\.step-([^/\\]+)\.md$/.exec(String(entry).split(/[/\\]/).at(-1));
+    if (!m) continue;
+    out.push({ path: relative(dir, join(root, String(entry))).split(/[/\\]/).join('/'), spec: m[1], step: m[2] });
+  }
+  return out.sort((a, b) => a.spec.localeCompare(b.spec) || a.step.localeCompare(b.step, undefined, { numeric: true }));
+}
+
+/** The middle value. */
+const median = (xs) => {
+  const s = [...xs].sort((a, b) => a - b);
+  const m = Math.floor(s.length / 2);
+  return s.length % 2 ? s[m] : (s[m - 1] + s[m]) / 2;
+};
+
+/**
+ * Every step with a forecast, landed or not: what it was forecast before it was built, and what its commits added.
+ *
+ * @param {string} dir - The project.
+ * @returns {{steps: object[], limit: number}|null} Null outside a git repository.
+ */
+export function measure(dir) {
+  const profile = (() => {
+    try {
+      return JSON.parse(readFileSync(join(dir, HARNESS, 'profile.json'), 'utf8'));
+    } catch {
+      return {};
+    }
+  })();
+  const limit = Number(profile?.vocabulary?.STEP_LINES) || Number(defaultVocabulary(PACKAGE).STEP_LINES) || 400;
+  // It runs every turn: a project with no step file does not pay for its history.
+  const files = stepFiles(dir);
+  if (files.length === 0) return git(dir, ['rev-parse', '--verify', '--quiet', 'HEAD']) === null ? null : { steps: [], limit };
+  const log = git(dir, ['log', '--format=%H%x09%ct%x09%cI%x09%s', 'HEAD']);
+  if (log === null) return null;
+  // Times as instants: an ISO time carries its zone, and two zones compared as text order wrongly.
+  const commits = log
+    .split('\n')
+    .filter(Boolean)
+    .map((row) => {
+      const [hash, seconds, at, ...rest] = row.split('\t');
+      return { hash, t: Number(seconds) * 1000, at, subject: rest.join('\t') };
+    })
+    .reverse();
+  const order = new Map(commits.map((c, i) => [c.hash, i]));
+  const named = new Map(files.map((f) => [f.path, commits.filter((c) => names(c.subject, f.spec, f.step))]));
+  const code = codeOf(dir, [...new Set([...named.values()].flat().map((c) => c.hash))]);
+  const changes = changesTo(dir, files.map((f) => f.path));
+  const wanted = [];
+  for (const f of files) {
+    wanted.push(`HEAD:${f.path}`);
+    for (const c of changes.get(f.path)) wanted.push(`${c.hash}:${f.path}`);
+    for (const c of named.get(f.path)) wanted.push(`${c.hash}^:${f.path}`, `${c.hash}:${f.path}`);
+  }
+  const text = blobs(dir, wanted);
+  const at = (rev, path) => forecastIn(text.get(`${rev}:${path}`) ?? '');
+  // When a forecast was written: the oldest commit of the newest run, before a point in history, that carried it. An
+  // edit to the rest of the step file is not a forecast. Null where no commit is found to carry it.
+  const writtenAt = (path, before, lines) => {
+    let when = null;
+    for (const c of changes.get(path)) {
+      if ((order.get(c.hash) ?? -1) >= before) continue;
+      if (at(c.hash, path)?.lines !== lines) break;
+      when = c.t;
+    }
+    return when;
+  };
+  const steps = [];
+  for (const file of files) {
+    const now = forecastIn(readFileSync(join(dir, file.path), 'utf8'));
+    const landings = named.get(file.path).filter((c) => code.get(c.hash)?.size > 0);
+    if (landings.length === 0) {
+      if (!now) continue;
+      // Forecast again in the working tree, it is newer than every commit.
+      const since = at('HEAD', file.path)?.lines === now.lines ? writtenAt(file.path, Infinity, now.lines) : Infinity;
+      steps.push({ ...file, landed: false, forecast: now, since });
+      continue;
+    }
+    // The forecast the step was built to: its file before its first commit, or as that commit wrote it.
+    const first = landings[0];
+    const before = at(`${first.hash}^`, file.path);
+    const forecast = before ?? at(first.hash, file.path);
+    if (!forecast) continue;
+    const since = (before ? writtenAt(file.path, order.get(first.hash), forecast.lines) : null) ?? first.t;
+    const lines = new Map();
+    for (const c of landings) for (const [path, added] of code.get(c.hash)) lines.set(path, (lines.get(path) ?? 0) + added);
+    const total = [...lines.values()].reduce((a, n) => a + n, 0);
+    const last = landings.at(-1);
+    // A forecast changed while the step was built — a fix round's spec — is shown, and the step measured against the first.
+    const revised = at(last.hash, file.path);
+    steps.push({
+      ...file,
+      landed: true,
+      forecast,
+      revised: revised && revised.lines !== forecast.lines ? revised : null,
+      since,
+      lines: total,
+      files: lines.size,
+      ratio: total / forecast.lines,
+      t: last.t,
+      at: last.at,
+      hash: last.hash,
+    });
+  }
+  return { steps, limit };
+}
+
+/**
+ * The rate to forecast a step again at: the median actual over forecast of its spec's newest landed steps forecast no
+ * earlier than it was. A step forecast before the last re-forecast was measured against the old rates, which that
+ * re-forecast already corrected for, and counting it would correct for them twice.
+ *
+ * @param {object[]} steps
+ * @param {{spec: string, since?: string}} step - Its spec, and when its own forecast was written.
+ * @returns {{rate: number, over: number}|null}
+ */
+export function rateOf(steps, { spec, since }) {
+  // A forecast not known to be written when is measured by nothing, rather than by everything.
+  if (since === null || since === undefined) return null;
+  const landed = steps
+    .filter((s) => s.landed && s.spec === spec && s.since >= since)
+    .sort((a, b) => a.t - b.t)
+    .slice(-RATE_OVER);
+  return landed.length ? { rate: median(landed.map((s) => s.ratio)), over: landed.length } : null;
+}
+
+/**
+ * What the detector says: the newest landed step, when it overran, and the steps of its spec still forecast as they
+ * were when it landed.
+ *
+ * @param {string} dir
+ * @param {{steps: object[], limit: number}} measured
+ * @returns {{step: object, rate: {rate: number, over: number}, stale: object[]}|null}
+ */
+export function overrun({ steps }) {
+  const newest = steps.filter((s) => s.landed).sort((a, b) => a.t - b.t).at(-1);
+  if (!newest || newest.ratio <= OVERRUN) return null;
+  // Forecast no later than the overrun step was, so its rate measures them: forecast again since, they are not stale.
+  const stale = steps.filter((s) => !s.landed && s.spec === newest.spec && s.since !== null && s.since <= newest.since);
+  return stale.length ? { step: newest, rate: rateOf(steps, stale[0]), stale } : null;
+}
+
+const times = (x) => `×${x.toFixed(2)}`;
+
+/**
+ * `nina sizes`.
+ *
+ * @param {string[]} argv
+ * @returns {Promise<number>}
+ */
+export async function sizes(argv) {
+  const arg = (name) => (argv.includes(name) ? argv[argv.indexOf(name) + 1] : null);
+  const dir = resolve(arg('--project') ?? '.');
+  const check = argv.includes('--check');
+  if (!existsSync(join(dir, HARNESS, 'profile.json'))) {
+    console.error(`  no ${HARNESS}/profile.json under ${dir} — run it inside a project, or name one with --project\n`);
+    return check ? 0 : 1;
+  }
+  const measured = measure(dir);
+  if (measured === null) {
+    if (check) console.log('sizes: current');
+    else console.error('  not a git repository, or one with no commit yet: a step is measured by its commits\n');
+    return check ? 0 : 1;
+  }
+  const { steps, limit } = measured;
+  if (check) {
+    const found = overrun(measured);
+    if (!found) {
+      console.log('sizes: current');
+      return 0;
+    }
+    const { step, rate, stale } = found;
+    const again = stale.slice(0, 4).map((s) => {
+      const next = Math.round(s.forecast.lines * rate.rate);
+      return `${s.step} ~${s.forecast.lines} → ~${next}${next > limit ? ` (past ${limit}: split)` : ''}`;
+    });
+    console.log(
+      `step ${step.step} of ${step.spec} landed ${step.lines} lines against ~${step.forecast.lines} (${times(step.ratio)}); ` +
+        `${stale.length} step(s) forecast when it was have not been forecast again — at ${times(rate.rate)}, over ${rate.over} landed step(s) forecast the same way: ` +
+        `${again.join(', ')}${stale.length > again.length ? ', …' : ''}`,
+    );
+    return 1;
+  }
+  if (steps.length === 0) {
+    console.log(`  no step file under ${SPECS} states a \`Size:\` line yet\n`);
+    return 0;
+  }
+  for (const spec of [...new Set(steps.map((s) => s.spec))]) {
+    const mine = steps.filter((s) => s.spec === spec);
+    console.log(`\n  ${bold(pink(spec))}${dim(`  landed against the forecast it was built to; limit ${limit} lines`)}`);
+    for (const s of mine.filter((x) => x.landed).sort((a, b) => a.t - b.t)) {
+      const mark = s.ratio > OVERRUN ? amber(times(s.ratio)) : times(s.ratio);
+      const files = s.forecast.files === null ? `${s.files} file(s)` : `${s.files} file(s) of ${s.forecast.files}`;
+      const revised = s.revised ? `, revised to ~${s.revised.lines} while built` : '';
+      console.log(`    ${s.step.padEnd(10)} ~${String(s.forecast.lines).padEnd(5)} → ${String(s.lines).padEnd(5)} ${mark}  ${dim(`${files}${revised} · ${s.hash.slice(0, 7)} ${s.at.slice(0, 10)}`)}`);
+    }
+    for (const s of mine.filter((x) => !x.landed)) {
+      const rate = rateOf(steps, s);
+      if (!rate) {
+        console.log(`    ${s.step.padEnd(10)} ~${String(s.forecast.lines).padEnd(5)} ${dim('to land — forecast after every step that landed, so none measures it yet')}`);
+        continue;
+      }
+      const next = Math.round(s.forecast.lines * rate.rate);
+      console.log(
+        `    ${s.step.padEnd(10)} ~${String(s.forecast.lines).padEnd(5)} → ~${String(next).padEnd(5)} ` +
+          dim(`to land, at ${times(rate.rate)} over ${rate.over} step(s) forecast as it was`) +
+          (next > limit ? amber(`  past ${limit}: split`) : ''),
+      );
+    }
+  }
+  console.log('');
+  return 0;
+}

package/src/gate.mjs

@@ -59,7 +59,7 @@
 
 import { spawnSync } from 'node:child_process';
 import { appendFileSync, closeSync, existsSync, mkdirSync, openSync, readFileSync, readSync, readdirSync, realpathSync, statSync, unlinkSync, writeFileSync } from 'node:fs';
-import { dirname, join } from 'node:path';
+import { basename, dirname, join } from 'node:path';
 import { fileURLToPath } from 'node:url';
 import { TERMINALS, declaredTokens, parseGraph } from './graph.mjs';
 import { gateDir, layerRootFor, readProfile, slugFor } from './paths.mjs';
@@ -215,7 +215,7 @@ export function loadProject(root) {
     const spec = join(root, '.claude', 'agents', `${stage}.md`);
     if (existsSync(spec)) tokens.set(stage, declaredTokens(readFileSync(spec, 'utf8')));
   }
-  return { graph, tokens, loops: loopEdges(graph), forward: forwardEdges(graph) };
+  return { root, graph, tokens, loops: loopEdges(graph), forward: forwardEdges(graph) };
 }
 
 /**
@@ -637,7 +637,138 @@ function onPost(input, project, path, at) {
   // The name an agent was launched with, if any: a later SendMessage may reach it by that name alone.
   const name = tool !== 'SendMessage' && typeof input.tool_input?.name === 'string' ? input.tool_input.name.slice(0, 80) : null;
   append(path, { k: 'dispatch', at, role, agent, via: tool, id: input.tool_use_id ?? null, ...describedAs(input), ...(name ? { name } : {}), ...(resumed === undefined ? {} : { resumed }) });
-  return null;
+  // A stage run in the foreground has finished by now, and its report is this call's result. One launched in the
+  // background, or sent a message, has not: its report arrives later, as a prompt.
+  if (tool === 'SendMessage' || response.isAsync === true || String(response.status ?? '').startsWith('async') || input.tool_input?.run_in_background === true) return null;
+  return leftBehind(input, project, path, at, agent, 'PostToolUse');
+}
+
+/** Where a tool call names a path: a shell command, or a file tool's path. A file's contents are not read, and a long
+ * run of characters with no space — a base64 blob a stage wrote — is what made the pattern below take seconds. */
+const PATH_KEYS = ['command', 'file_path', 'path', 'notebook_path'];
+
+/** A text as a pattern that matches it literally. */
+const literal = (text) => text.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+
+/** How much of a stage's transcript is read for its commands: a run's commands are seldom near this, a stray never is. */
+const STRAY_TAIL = 32 * 1024 * 1024;
+
+/**
+ * What a stage left at the root of the session's scratchpad (Hard Rule #18): each entry its commands named there, made
+ * or changed since it was dispatched, that is not a stage's directory, `<role>-<step>-XXXX`. On the first night the rule
+ * held it was read and still missed twice — a reviewer's log, an integration-tester's pointer file — and only the
+ * orchestrator, by chance, saw either. The root is the one the stage's commands used, written out or held in a
+ * variable, so nothing here assumes where Claude Code keeps it; an entry no command of the stage named, or one it only
+ * read, is not laid at its door. A `cd` into the root and a bare name are not followed.
+ *
+ * @param {string} transcript - The stage's own transcript.
+ * @param {Set<string>} roles - The pipeline's stages, whose directories the root may hold.
+ * @param {{since?: number, project?: string}} [options] - When the stage was dispatched, in ms; and the project, inside
+ *   which a directory called `scratchpad` is the project's own, not the session's.
+ * @returns {string[]} The paths left there.
+ */
+export function strays(transcript, roles, { since = 0, project = null } = {}) {
+  let text = '';
+  try {
+    const size = statSync(transcript).size;
+    const length = Math.min(size, STRAY_TAIL);
+    const buffer = Buffer.alloc(length);
+    const fd = openSync(transcript, 'r');
+    try {
+      readSync(fd, buffer, 0, length, size - length);
+    } finally {
+      closeSync(fd);
+    }
+    for (const line of buffer.toString('utf8').split('\n')) {
+      if (!line.includes('"tool_use"') || !line.includes('scratchpad')) continue;
+      let row;
+      try {
+        row = JSON.parse(line);
+      } catch {
+        continue;
+      }
+      if (!Array.isArray(row?.message?.content)) continue;
+      for (const block of row.message.content) {
+        if (block?.type !== 'tool_use' || !block.input || typeof block.input !== 'object') continue;
+        for (const key of PATH_KEYS) if (typeof block.input[key] === 'string') text += `${block.input[key]}\n`;
+      }
+    }
+  } catch {
+    return [];
+  }
+  if (!text.includes('/scratchpad')) return [];
+  const own = new RegExp(`^(${[...roles].map(literal).join('|')})-.+-[A-Za-z0-9]{4,}$`);
+  const inside = project ? `${realpathOr(project)}/` : null;
+  const out = [];
+  for (const root of new Set([...text.matchAll(/(\/[^\s"'`;|&<>()$\\]{0,400}\/scratchpad)(?![\w.-])/g)].map((m) => m[1]))) {
+    if (inside && `${realpathOr(root)}/`.startsWith(inside)) continue;
+    let entries;
+    try {
+      entries = readdirSync(root, { withFileTypes: true });
+    } catch {
+      continue;
+    }
+    // The root as its commands wrote it, and every variable they set to it.
+    const ways = new Set([root]);
+    for (const m of text.matchAll(new RegExp(`\\b([A-Za-z_]\\w*)=["']?${literal(root)}["']?(?=[\\s;&|]|$)`, 'gm'))) ways.add(`$${m[1]}`).add(`\${${m[1]}}`);
+    const named = new RegExp(`(?:${[...ways].map(literal).join('|')})/([^\\s"'\`;|&<>()/\\\\]{1,200})`, 'g');
+    const touched = new Set([...text.matchAll(named)].map((m) => m[1]));
+    for (const entry of entries) {
+      if (entry.isDirectory() && own.test(entry.name)) continue;
+      if (!touched.has(entry.name)) continue;
+      // Read, not made: a file the orchestrator left there and a reviewer read is not the reviewer's.
+      let changed = 0;
+      try {
+        const stat = statSync(join(root, entry.name));
+        changed = Math.max(stat.mtimeMs, stat.birthtimeMs || 0);
+      } catch {
+        continue;
+      }
+      if (changed >= since) out.push(join(root, entry.name));
+    }
+  }
+  return out.sort();
+}
+
+/** A path with its links resolved, where it can be. */
+function realpathOr(path) {
+  try {
+    return realpathSync(path);
+  } catch {
+    return path;
+  }
+}
+
+/**
+ * A stage's report reached the orchestrator: say what it left at the scratchpad's root, once each, for the
+ * orchestrator to move or remove before a stage beside it overwrites or deletes it.
+ *
+ * @param {string|null} named - The agent, by its id or the name it was launched with.
+ * @returns {object|null} The hook's answer, its text for the orchestrator.
+ */
+function leftBehind(input, project, path, at, named, event) {
+  if (!named || typeof input.transcript_path !== 'string' || typeof input.session_id !== 'string') return null;
+  const entries = readLedger(path);
+  const { agent, role } = recipient(entries, project, named);
+  if (!agent || !role || !project.graph.stages.has(role)) return null;
+  // From when it went out: a foreground dispatch is recorded as it comes back, so its start is its `pre`.
+  const launched = entries.findLast((e) => e.k === 'dispatch' && e.agent === agent);
+  const pre = launched?.id ? entries.find((e) => e.k === 'pre' && e.id === launched.id) : null;
+  const transcript = join(dirname(input.transcript_path), input.session_id, 'subagents', `agent-${agent}.jsonl`);
+  const told = new Set(entries.filter((e) => e.k === 'stray').flatMap((e) => e.paths ?? []));
+  const found = strays(transcript, project.graph.stages, { since: Date.parse(pre?.at ?? launched?.at ?? '') || 0, project: project.root }).filter((p) => !told.has(p));
+  if (found.length === 0) return null;
+  append(path, { k: 'stray', at, agent, role, paths: found });
+  const names = found.map((p) => `\`${basename(p)}\``).join(', ');
+  return {
+    hookSpecificOutput: {
+      hookEventName: event,
+      additionalContext:
+        `NINA, Hard Rule #18: the ${role} whose report this is worked in ${names} at the root of the session's scratchpad ` +
+        `(${dirname(found[0])}), outside any stage's \`<role>-<step>-XXXX\` directory, where a stage beside it can overwrite ` +
+        `or delete it. Move it into the directory of the stage that made it, or remove it, and say so in that stage's next dispatch.`,
+    },
+  };
 }
 
 /** How far a round is past what holds it: its issue's cap, or its edge's ceiling. Above zero asks. */
@@ -862,7 +993,10 @@ export function handle(input, { root, now = new Date(), pkg = PACKAGE }) {
     const at = now.toISOString();
     switch (input?.hook_event_name) {
       case 'UserPromptSubmit':
-        if (typeof input.prompt === 'string' && DELIVERED.test(input.prompt.slice(0, DELIVERY_HEAD))) return null;
+        if (typeof input.prompt === 'string' && DELIVERED.test(input.prompt.slice(0, DELIVERY_HEAD))) {
+          const from = /<agent-message from="([^"]+)"/.exec(input.prompt)?.[1] ?? /<task-id>([^<\s]+)<\/task-id>/.exec(input.prompt)?.[1];
+          return leftBehind(input, project, path, at, from, 'UserPromptSubmit');
+        }
         append(path, { k: 'reset', at, why: 'prompt' });
         prune(root, now);
         return null;