session-orchestrator 5.3.0 → 5.4.0

package/scripts/lib/telemetry/pricing.mjs

@@ -1,5 +1,5 @@
 /**
- * pricing.mjs — USD price table for token-bucket cost estimation (#1244).
+ * pricing.mjs — USD price table for token-bucket cost estimation (#1244, #1470).
  *
  * The subagent ledger records four token buckets per stop record
  * (`token_input_uncached`, `token_cache_read`, `token_cache_creation`,
@@ -8,23 +8,39 @@
  * an order of magnitude on a cache-heavy run — which is the whole reason this
  * table exists rather than a single `$/token` constant.
  *
- * ## What is measured and what is derived
+ * ## Every rate is a documented figure
  *
- * `input` and `output` are the first-party Anthropic API list prices as
- * published in the bundled model table, sourced on the date `PRICING_TABLE_DATE`
- * carries (2026-09-09) — that constant is the SSOT, never a second date repeated
- * in this prose. `cache_read` and `cache_creation` are DERIVED from the
- * documented ephemeral-cache multipliers (read ≈ 0.1×, 5-minute write ≈ 1.25×
- * the input rate) except where a row carries a documented figure — Claude Fable
- * 5.1 publishes $0.25/MTok cache reads directly, which is NOT 0.1× its $10
- * input rate. Every row therefore declares its own `cache_source` so a reader
- * never has to guess which numbers were read off a price list and which were
- * multiplied out here.
+ * All four rates of every row are first-party Anthropic API list prices, read
+ * off the price page each row names in `source_url` on the date
+ * `PRICING_TABLE_DATE` carries — that constant is the SSOT, never a second date
+ * repeated in this prose. Nothing here is multiplied out. `cache_read` in
+ * particular is never derived from `input`: the ratio is not a constant (Opus
+ * 5.5 reads at 0.05× its input rate, Fable 5.1 at 0.025×, Fable 5 at 0.1×), so
+ * a multiplier would misprice exactly the cache-heavy runs this table exists
+ * for. Each row carries its own provenance fields, so a consumer quoting a
+ * dollar figure can cite where it came from without reading this file.
  *
- * A row with `verified: false` carries a PLACEHOLDER: its numbers are a
- * best-effort stand-in, never a quoted price. `costUsd()` still prices such a
- * row (an estimate is more useful than a null when it is labelled), but any
- * consumer publishing a dollar figure must surface the flag.
+ * `verified: false` would mark a PLACEHOLDER row — a best-effort stand-in,
+ * never a quoted price. No current row carries it; a consumer publishing a
+ * dollar figure must still surface the flag.
+ *
+ * ## Only exact ids resolve
+ *
+ * `priceFor()` accepts an exact table key or an explicit `MODEL_ALIASES` entry
+ * and nothing else. Until #1470 it fell back to the longest table key an id
+ * started with, which handed an unlisted generation the rates of its nearest
+ * neighbour, marked verified.
+ *
+ * ## Series break (#1470, 2026-09-29)
+ *
+ * Session reports written before this fix priced records through that
+ * fallback: `claude-opus-5-5` at `claude-opus-5` rates, `claude-sonnet-5-5` at
+ * `claude-sonnet-5` rates (identical numbers, by coincidence rather than
+ * evidence), and `claude-fable-5` / `claude-opus-4-8` not at all. Nothing
+ * rewrites that history — the metrics ledgers are append-only. Measured
+ * 2026-09-29: 0 of the 9 priced sessions.jsonl reports in this repo contained
+ * such records (the first one reached subagents.jsonl after the last priced
+ * report was written).
  *
  * ## Null means UNKNOWN MODEL, never zero
  *
@@ -39,68 +55,93 @@
  */
 
 /**
- * The date the `input`/`output` rates in PRICING_TABLE were sourced.
- * Any consumer quoting a dollar figure should quote this date beside it — a
- * price without its measurement date ages silently.
+ * The date the rates in PRICING_TABLE were read off the price page. Every row's
+ * `checked_at` equals it. Any consumer quoting a dollar figure should quote this
+ * date beside it — a price without its measurement date ages silently.
  * @type {string}
  */
-export const PRICING_TABLE_DATE = '2026-09-09';
+export const PRICING_TABLE_DATE = '2026-09-29';
 
 /**
  * @typedef {Object} PricingRow
  * @property {number} input           - USD per 1M uncached prompt tokens.
  * @property {number} cache_read      - USD per 1M tokens served from the prompt cache.
- * @property {number} cache_creation  - USD per 1M tokens written to the prompt cache.
+ * @property {number} cache_creation  - USD per 1M tokens written to the prompt cache
+ *   (the `cache_write_ttl` rate).
  * @property {number} output          - USD per 1M completion tokens.
  * @property {boolean} verified       - false ⇒ the row is a labelled placeholder, not a quoted price.
  * @property {'documented'|'derived-multiplier'} cache_source - provenance of the two cache rates.
+ *   Every current row is 'documented'; 'derived-multiplier' is kept for a future row whose cache
+ *   rates are multiplied out and must say so.
+ * @property {string} source_url      - the page all four rates were read from.
+ * @property {string} checked_at      - ISO date the rates were read; equals PRICING_TABLE_DATE.
+ * @property {'anthropic'} provider
+ * @property {'5m'} cache_write_ttl   - which cache-write rate `cache_creation` is.
+ * @property {string} unit            - 'USD per 1M tokens'.
+ * @property {string} surface         - the billing surface and speed the rates apply to.
+ * @property {'standard'} tier        - the service tier the rates apply to.
+ */
+
+/** Provenance shared by every current row — all read off one page on one date. */
+const PROVENANCE = Object.freeze({
+  verified: true,
+  cache_source: 'documented',
+  source_url: 'https://platform.claude.com/docs/en/about-claude/pricing',
+  checked_at: PRICING_TABLE_DATE,
+  provider: 'anthropic',
+  cache_write_ttl: '5m',
+  unit: 'USD per 1M tokens',
+  surface: 'Claude API (first-party), standard speed',
+  tier: 'standard',
+});
+
+/**
+ * One frozen documented row: four USD rates plus the shared provenance. The
+ * rates are spread last, so provenance can never overwrite a price.
+ * @param {{input: number, cache_read: number, cache_creation: number, output: number}} rates
+ * @returns {PricingRow}
  */
+function usd(rates) {
+  return Object.freeze({ ...PROVENANCE, ...rates });
+}
 
 /**
  * Per-model USD-per-million-token rates.
  * @type {Readonly<Record<string, PricingRow>>}
  */
 export const PRICING_TABLE = Object.freeze({
-  // Claude Fable 5.1 — $10 / $50 list; cache reads documented at $0.25/MTok
-  // (NOT the usual 0.1× multiplier), cache writes derived at 1.25× input.
-  'claude-fable-5-1': Object.freeze({
-    input: 10.0,
-    cache_read: 0.25,
-    cache_creation: 12.5,
-    output: 50.0,
-    verified: true,
-    cache_source: 'documented',
-  }),
-  'claude-opus-5': Object.freeze({
-    input: 5.0,
-    cache_read: 0.5,
-    cache_creation: 6.25,
-    output: 25.0,
-    verified: true,
-    cache_source: 'derived-multiplier',
-  }),
-  'claude-sonnet-5': Object.freeze({
-    input: 2.0,
-    cache_read: 0.2,
-    cache_creation: 2.5,
-    output: 10.0,
-    verified: true,
-    cache_source: 'derived-multiplier',
-  }),
-  'claude-haiku-4-5': Object.freeze({
-    input: 1.0,
-    cache_read: 0.1,
-    cache_creation: 1.25,
-    output: 5.0,
-    verified: true,
-    cache_source: 'derived-multiplier',
-  }),
+  // Ceiling: only the 5-minute cache-write rate is priced; a 1-hour write costs 2× input
+  // (Opus 5.5: $8/MTok) and would be under-priced here. Measured 2026-09-29 over 803,788
+  // assistant usage blocks in 10,745 subagent transcripts: `usage.cache_creation
+  // .ephemeral_1h_input_tokens > 0` on 0 blocks. Revisit when the first such block appears.
+  //
+  // Ceiling: list price only — fast mode, `inference_geo: "us"` (1.1×) and batch discounts are
+  // not modelled. Measured 2026-09-29: `usage.speed` = "fast" on 0 of the 374,331 blocks that
+  // carry the field; `inference_geo` = "not_available" on 803,479 and no other value seen.
+  // Revisit when either field shows another value.
+  //
+  // Ceiling: a model absent from this table prices as null, never at a neighbour's rate, so the
+  // rollup reports `total_cost_usd: null` and `cost_records_priced < cost_records_total` — the
+  // intended way missing evidence shows up. That ratio also drops below 1 for records no row can
+  // fix: a token-less record (transcript over the hook's size cap, without usage turns, or
+  // unreadable) and a record with a token-bearing turn that names no model (#1475). Revisit when
+  // an unpriced record names a model id absent from this table: add that model as its own
+  // evidenced row, never as a family prefix or an alias to another generation.
+  'claude-fable-5-1': usd({ input: 10.0, cache_read: 0.25, cache_creation: 12.5, output: 50.0 }),
+  'claude-fable-5': usd({ input: 10.0, cache_read: 1.0, cache_creation: 12.5, output: 50.0 }),
+  'claude-opus-5-5': usd({ input: 4.0, cache_read: 0.2, cache_creation: 5.0, output: 20.0 }),
+  'claude-opus-5': usd({ input: 5.0, cache_read: 0.5, cache_creation: 6.25, output: 25.0 }),
+  'claude-opus-4-8': usd({ input: 5.0, cache_read: 0.5, cache_creation: 6.25, output: 25.0 }),
+  'claude-sonnet-5-5': usd({ input: 2.0, cache_read: 0.2, cache_creation: 2.5, output: 10.0 }),
+  'claude-sonnet-5': usd({ input: 2.0, cache_read: 0.2, cache_creation: 2.5, output: 10.0 }),
+  'claude-haiku-4-5': usd({ input: 1.0, cache_read: 0.1, cache_creation: 1.25, output: 5.0 }),
 });
 
 /**
  * Explicit aliases for model ids that appear in transcripts but are not the
- * canonical table key — dated snapshots and the `[1m]` context-window suffix
- * the harness stamps on long-context sessions.
+ * canonical table key — a dated snapshot of the SAME model and the `[1m]`
+ * context-window suffix the harness stamps on long-context sessions. An alias
+ * never points at another generation.
  * @type {Readonly<Record<string, string>>}
  */
 export const MODEL_ALIASES = Object.freeze({
@@ -113,10 +154,19 @@ export const MODEL_ALIASES = Object.freeze({
 /**
  * Look up the rate row for a model id.
  *
- * Resolution order: exact key → explicit alias → longest matching table key
- * that the id starts with (so `claude-opus-5-20260401` and `claude-opus-5[1m]`
- * both resolve to `claude-opus-5`). Longest-prefix wins so a future
- * `claude-opus-5-1` row is never shadowed by `claude-opus-5`.
+ * Resolution: exact `PRICING_TABLE` key → explicit `MODEL_ALIASES` entry → `null`,
+ * after trimming surrounding whitespace.
+ * Nothing else resolves — no prefix, no pattern, no family fallback — so
+ * `claude-opus-5-20260401`, `claude-opus-5-99` and `claude-haiku-4-5-20251002` are
+ * all UNKNOWN even though each starts with a table key.
+ *
+ * Why no prefix: from the 4.6 generation on, Anthropic ships dateless pinned
+ * model ids (https://platform.claude.com/docs/en/about-claude/models/model-ids-and-versions,
+ * retrieved 2026-09-29). The ids that extend a current table key are therefore
+ * other generations, not snapshots of it — `claude-opus-5-5` extends
+ * `claude-opus-5` — and a prefix match priced them at another generation's
+ * rates, marked verified (#1470). A future generation gets its own evidenced
+ * row, never a family prefix.
  *
  * @param {string|null|undefined} modelId
  * @returns {PricingRow|null} the row, or null when the model is UNKNOWN
@@ -129,14 +179,10 @@ export function priceFor(modelId) {
 
   if (Object.hasOwn(PRICING_TABLE, id)) return PRICING_TABLE[id];
 
-  const aliased = MODEL_ALIASES[id];
-  if (aliased && Object.hasOwn(PRICING_TABLE, aliased)) return PRICING_TABLE[aliased];
-
-  let best = null;
-  for (const key of Object.keys(PRICING_TABLE)) {
-    if (id.startsWith(key) && (best === null || key.length > best.length)) best = key;
+  if (Object.hasOwn(MODEL_ALIASES, id) && Object.hasOwn(PRICING_TABLE, MODEL_ALIASES[id])) {
+    return PRICING_TABLE[MODEL_ALIASES[id]];
   }
-  return best === null ? null : PRICING_TABLE[best];
+  return null;
 }
 
 /**
@@ -161,9 +207,10 @@ function num(n) {
 /**
  * Estimate the USD cost of one record's four token buckets.
  *
- * Each bucket is multiplied by ITS OWN rate — a cache read is ~10× cheaper
- * than an uncached prompt token and a cache write ~1.25× more expensive, so a
- * blended rate is not an approximation of this, it is a different number.
+ * Each bucket is multiplied by ITS OWN rate — depending on the row, a cache
+ * read costs a tenth to a fortieth of an uncached prompt token and a 5-minute
+ * cache write 1.25× as much, so a blended rate is not an approximation of
+ * this, it is a different number.
  *
  * @param {object} opts
  * @param {string|null|undefined} opts.model            - model id from the transcript

package/scripts/lib/tests-src-ratio.mjs

@@ -22,8 +22,11 @@
  * Here `src` is defined by NEGATION, so a new top-level directory joins the
  * denominator the moment it is committed, with no rule edit and no re-derivation.
  *
- *   universe     `git ls-files`, filtered to CODE_EXTENSIONS
- *   numerator    tracked code under `tests/`
+ *   universe     `git ls-files`, filtered to CODE_EXTENSIONS, minus declaration
+ *                files (`*.d.ts`, `*.d.mts`, `*.d.cts` — types, not executable code)
+ *   numerator    tracked code under a top-level `tests/`, plus co-located tests
+ *                anywhere: basename `*.test.<ext>` / `*.spec.<ext>`, or any file
+ *                under a `__tests__/` directory
  *   denominator  every OTHER tracked code file
  *
  * ## The four questions, answered in code rather than left open
@@ -37,8 +40,8 @@
  *   Only git-tracked?                     YES — reproducible at a SHA, and it is
  *                                         what keeps node_modules/, coverage/ and
  *                                         untracked scratch out without an ignore list.
- *   Do `tests/fixtures/` count as test LOC? YES when they are code (`.mjs`/`.js`/
- *                                         `.cjs`). A code fixture is maintained
+ *   Do `tests/fixtures/` count as test LOC? YES when they are code (any of
+ *                                         CODE_EXTENSIONS). A code fixture is maintained
  *                                         code that exists only to serve the suite —
  *                                         that is test-corpus cost. Non-code
  *                                         fixtures (`.json`, `.jsonl`, `.md`) are
@@ -62,7 +65,9 @@
  *   <repo-root>    defaults to process.cwd()
  *   --json         emit a single JSON object on stdout, nothing else
  *   --check        exit 1 when the ratio exceeds the ceiling (consolidation
- *                  wave required); exit 0 when inside the corridor
+ *                  wave required); exit 3 when the ratio is not measurable
+ *                  (`reason: 'no-tests-found'`, #1451); exit 0 when inside
+ *                  the corridor
  *   --ceiling <n>  override the TV-003 ceiling (default 1.60)
  *   --stdin        take newline-separated paths from stdin instead of
  *                  enumerating via `git ls-files` (test seam / staged-only mode)
@@ -71,6 +76,12 @@
  *   0 — measurement completed (and, under --check, ratio is within the corridor)
  *   1 — --check only: ratio exceeds the ceiling
  *   2 — tool error (missing/unreadable root, bad argv)
+ *   3 — --check only: ratio not measurable — 0 test files counted while test-like
+ *       files ARE tracked (`reason: 'no-tests-found'`, e.g. a Python repo whose
+ *       `tests/test_*.py` sit outside CODE_EXTENSIONS). Fail-closed since #1451:
+ *       before, such a repo read `ratio 0, withinCorridor true` — "not found"
+ *       reported as "in corridor". TypeScript tests count since the 2026-10-01
+ *       owner decision on #1451; other stacks still get "not measurable".
  */
 
 import { readFileSync, existsSync, statSync } from 'node:fs';
@@ -90,23 +101,65 @@ import { isMainModule } from './is-main-module.mjs';
  * Applied symmetrically to numerator and denominator by construction — a file
  * type can never inflate one side while being invisible to the other.
  */
-export const CODE_EXTENSIONS = Object.freeze(['.mjs', '.js', '.cjs']);
+export const CODE_EXTENSIONS = Object.freeze([
+  '.mjs',
+  '.js',
+  '.cjs',
+  '.jsx',
+  '.ts',
+  '.tsx',
+  '.mts',
+  '.cts',
+]);
 
-/** Path prefix that makes a tracked code file part of the TEST corpus. */
+/**
+ * TypeScript declaration files end in a CODE_EXTENSION but carry only types —
+ * no executable line a test could exercise. Excluded from BOTH sides, so a
+ * generated `.d.ts` can never move the ratio.
+ */
+export const DECLARATION_SUFFIXES = Object.freeze(['.d.ts', '.d.mts', '.d.cts']);
+
+/** Top-level path prefix that makes a tracked code file part of the TEST corpus. */
 export const TEST_PREFIX = 'tests/';
 
+/**
+ * Co-located test files, wherever they sit: a `*.test.<ext>` / `*.spec.<ext>`
+ * basename (the vitest/jest default include shape) or a `__tests__/` segment.
+ */
+const COLOCATED_TEST_RE = /\.(test|spec)\.[^./]+$|(^|\/)__tests__\//;
+
+/**
+ * A tracked path that LOOKS like a test, counted or not: a `*.test.*` / `*.spec.*`
+ * JS/TS file, anything under a `__tests__/` segment, or any file under `tests/`
+ * regardless of extension. Used only to tell "this repo has no tests" apart from
+ * "this repo has tests the recipe cannot see" (#1451) — e.g. `tests/test_x.py`.
+ */
+const TEST_LIKE_RE = /\.(test|spec)\.[cm]?[jt]sx?$/;
+
+/**
+ * @param {string} relPath repo-relative path
+ * @returns {boolean}
+ */
+export function isTestLikePath(relPath) {
+  const p = String(relPath).replace(/\\/g, '/').replace(/^\.\//, '');
+  return TEST_LIKE_RE.test(p) || p.startsWith(TEST_PREFIX) || /(^|\/)__tests__\//.test(p);
+}
+
 /** The TV-003 ceiling. Exceeding it switches the consolidation rule on. */
 export const DEFAULT_CEILING = 1.6;
 
 /** Machine-readable schema tag for the --json envelope. */
-export const SCHEMA = 'tests-src-ratio/1';
+export const SCHEMA = 'tests-src-ratio/2';
 
 /**
  * Which bucket a repo-relative path belongs to.
  *
- * `src` is defined by NEGATION — every tracked code file that is not under
- * `tests/`. That is what makes the two buckets a partition rather than two
- * globs that can overlap or leave a gap.
+ * `src` is defined by NEGATION — every tracked code file that is not a test
+ * (under `tests/`, or co-located per COLOCATED_TEST_RE). That is what makes the
+ * two buckets a partition rather than two globs that can overlap or leave a gap.
+ *
+ * Known blind spot: a nested `pkg/tests/helper.ts` that is not itself named
+ * `*.test.*` counts as src — only the TOP-LEVEL `tests/` is a test directory.
  *
  * @param {string} relPath repo-relative path, `/`-separated
  * @returns {'test'|'src'|null} null = outside the metric entirely
@@ -114,7 +167,8 @@ export const SCHEMA = 'tests-src-ratio/1';
 export function classifyPath(relPath) {
   const p = String(relPath).replace(/\\/g, '/').replace(/^\.\//, '');
   if (!CODE_EXTENSIONS.some((ext) => p.endsWith(ext))) return null;
-  return p === TEST_PREFIX.slice(0, -1) || p.startsWith(TEST_PREFIX) ? 'test' : 'src';
+  if (DECLARATION_SUFFIXES.some((ext) => p.endsWith(ext))) return null;
+  return p.startsWith(TEST_PREFIX) || COLOCATED_TEST_RE.test(p) ? 'test' : 'src';
 }
 
 /**
@@ -154,7 +208,8 @@ export function countPhysicalLines(content) {
  * @param {(relPath: string) => string|null} opts.readFile returns content, or null when unreadable
  * @param {number} [opts.ceiling]
  * @returns {{testFiles:number,testLoc:number,srcFiles:number,srcLoc:number,ratio:number|null,
- *            ceiling:number,withinCorridor:boolean,consolidationWaveRequired:boolean,skipped:number}}
+ *            ceiling:number,withinCorridor:boolean|null,consolidationWaveRequired:boolean|null,
+ *            skipped:number,testLikeFilesSeen:number,reason:'no-tests-found'|null}}
  */
 export function measure({ files, readFile, ceiling = DEFAULT_CEILING }) {
   let testFiles = 0;
@@ -162,8 +217,10 @@ export function measure({ files, readFile, ceiling = DEFAULT_CEILING }) {
   let srcFiles = 0;
   let srcLoc = 0;
   let skipped = 0;
+  let testLikeFilesSeen = 0;
 
   for (const rel of files) {
+    if (isTestLikePath(rel)) testLikeFilesSeen++;
     const bucket = classifyPath(rel);
     if (bucket === null) continue;
     const content = readFile(rel);
@@ -183,6 +240,26 @@ export function measure({ files, readFile, ceiling = DEFAULT_CEILING }) {
 
   // A repo with no src code has an undefined ratio, not an infinite one. Saying
   // `null` keeps the consumer from reading Infinity as a corridor breach.
+  // Fail closed (#1451): a zero numerator next to tracked test-like files means
+  // the recipe cannot SEE this repo's tests (e.g. `tests/test_x.py`), not that
+  // there are none. Reporting ratio 0 / withinCorridor true there turned "not found"
+  // into "green". The verdict is unknown, so every verdict field is null.
+  if (testFiles === 0 && testLikeFilesSeen > 0) {
+    return {
+      testFiles,
+      testLoc,
+      srcFiles,
+      srcLoc,
+      ratio: null,
+      ceiling,
+      withinCorridor: null,
+      consolidationWaveRequired: null,
+      skipped,
+      testLikeFilesSeen,
+      reason: 'no-tests-found',
+    };
+  }
+
   const ratio = srcLoc === 0 ? null : Number((testLoc / srcLoc).toFixed(4));
   const withinCorridor = ratio === null ? true : ratio <= ceiling;
 
@@ -196,6 +273,8 @@ export function measure({ files, readFile, ceiling = DEFAULT_CEILING }) {
     withinCorridor,
     consolidationWaveRequired: !withinCorridor,
     skipped,
+    testLikeFilesSeen,
+    reason: null,
   };
 }
 
@@ -208,11 +287,16 @@ export function definitionOf(source) {
   return {
     source: source === 'stdin' ? 'stdin path list' : 'git ls-files (tracked files only)',
     codeExtensions: [...CODE_EXTENSIONS],
-    numerator: `tracked code files under ${TEST_PREFIX}`,
+    declarationsExcluded: [...DECLARATION_SUFFIXES],
+    numerator:
+      `tracked code files under a top-level ${TEST_PREFIX}, plus co-located tests anywhere: ` +
+      'basename *.test.<ext> / *.spec.<ext>, or any file under a __tests__/ directory',
     denominator: `every OTHER tracked code file (src defined by negation, not by a directory list)`,
     lineRule: 'physical lines; blank + comment lines counted; EOF-newline-insensitive',
     excluded:
-      'non-code extensions (.md, .json, .jsonl, .yml) on BOTH sides; untracked files (node_modules/, coverage/, scratch)',
+      `declaration files (${DECLARATION_SUFFIXES.map((x) => `*${x}`).join(', ')}: types, not ` +
+      'executable code) and non-code extensions (.md, .json, .jsonl, .yml) on BOTH sides; ' +
+      'untracked files (node_modules/, coverage/, scratch)',
   };
 }
 
@@ -322,7 +406,11 @@ const USAGE =
  * because a measurement problem must not block a session start.
  *
  * @param {{ repoRoot?: string, ceiling?: number }} [opts]
- * @returns {{ severity: 'warn', message: string, ratio: number, ceiling: number } | null}
+ * The one non-null path besides a breach is `reason: 'no-tests-found'` (#1451):
+ * a repo whose tests the recipe cannot see gets a WARN, never silence — silence
+ * is what made the unmeasurable ratio read as green fleet-wide.
+ *
+ * @returns {{ severity: 'warn', message: string, ratio: number|null, ceiling: number, reason?: string } | null}
  */
 export function checkTestsSrcRatio({ repoRoot, ceiling = DEFAULT_CEILING } = {}) {
   try {
@@ -342,6 +430,18 @@ export function checkTestsSrcRatio({ repoRoot, ceiling = DEFAULT_CEILING } = {})
     };
 
     const result = measure({ files, readFile, ceiling });
+    if (result.reason === 'no-tests-found') {
+      return {
+        severity: 'warn',
+        ratio: null,
+        ceiling: result.ceiling,
+        reason: result.reason,
+        message:
+          `⚠ tests:src ratio nicht messbar — 0 Testdateien gezählt, ` +
+          `${result.testLikeFilesSeen} test-artige Dateien getrackt (#1451). ` +
+          `Detail: node scripts/lib/tests-src-ratio.mjs --json`,
+      };
+    }
     if (result.ratio === null || result.withinCorridor) return null;
 
     const dirty = isDirty(root);
@@ -402,17 +502,20 @@ function main() {
     console.log(USAGE);
     console.log('');
     console.log('The canonical tests:src LOC measurement for TV-003 (test-value.md).');
-    console.log('  numerator    tracked code under tests/');
+    console.log('  numerator    tracked code under tests/, plus *.test.* / *.spec.* / __tests__/ anywhere');
     console.log('  denominator  every other tracked code file (src by negation)');
     console.log(`  code exts    ${CODE_EXTENSIONS.join(' ')}   (.md / .json never counted)`);
+    console.log(`  excluded     declaration files ${DECLARATION_SUFFIXES.join(' ')} on both sides`);
     console.log('  lines        physical; blanks + comments counted; EOF-newline-insensitive');
     console.log('');
     console.log('  --json         machine-readable envelope on stdout');
-    console.log('  --check        exit 1 when the ratio exceeds the ceiling');
+    console.log('  --check        exit 1 when the ratio exceeds the ceiling, 3 when not measurable');
     console.log(`  --ceiling <n>  override the TV-003 ceiling (default ${DEFAULT_CEILING})`);
     console.log('  --stdin        read newline-separated paths instead of git ls-files');
     console.log('');
-    console.log('Exit: 0 ok / within corridor · 1 (--check) ceiling exceeded · 2 tool error');
+    console.log(
+      'Exit: 0 ok / within corridor · 1 (--check) ceiling exceeded · 2 tool error · 3 (--check) not measurable',
+    );
     process.exit(0);
   }
 
@@ -460,10 +563,17 @@ function main() {
     // fail-open class is not worth re-litigating per payload size.
     writeStdoutLineSync(JSON.stringify(envelope, null, 2));
   } else {
-    const r = result.ratio === null ? 'n/a (no src code)' : result.ratio.toFixed(4);
-    const verdict = result.withinCorridor
-      ? 'within corridor'
-      : `ABOVE ceiling ${result.ceiling} — TV-003 consolidation wave required`;
+    const notMeasurable = result.reason === 'no-tests-found';
+    const r = notMeasurable
+      ? 'n/a (no tests found)'
+      : result.ratio === null
+        ? 'n/a (no src code)'
+        : result.ratio.toFixed(4);
+    const verdict = notMeasurable
+      ? `NOT MEASURABLE — ${result.testLikeFilesSeen} test-like files tracked but 0 counted (#1451)`
+      : result.withinCorridor
+        ? 'within corridor'
+        : `ABOVE ceiling ${result.ceiling} — TV-003 consolidation wave required`;
     const at = envelope.ref ? ` @ ${envelope.ref}${dirty ? '+dirty' : ''}` : '';
     writeStdoutLineSync(
       `tests:src = ${r}  (${result.testLoc} test LOC / ${result.srcLoc} src LOC` +
@@ -472,6 +582,13 @@ function main() {
     );
   }
 
+  if (args.check && result.reason === 'no-tests-found') {
+    console.error(
+      `tests:src ratio not measurable — 0 test files counted but ${result.testLikeFilesSeen} ` +
+        `test-like files tracked (no-tests-found); see #1451`,
+    );
+    process.exit(3);
+  }
   process.exit(args.check && !result.withinCorridor ? 1 : 0);
 }
 

package/scripts/lib/tmux-layout/layouts.mjs

@@ -25,6 +25,7 @@
 
 import path from 'node:path';
 import { resolveStateDir } from '../platform.mjs';
+import { shellQuote } from '../sh-quote.mjs';
 import { detectVcsCommand } from './vcs-detector.mjs';
 import { isSessionCollision } from './tmux-shell.mjs';
 
@@ -48,17 +49,6 @@ try {
 // Internal helpers
 // ---------------------------------------------------------------------------
 
-/**
- * Single-quote a shell command string for use as a tmux send-keys argument.
- * Escapes embedded single-quotes using the POSIX `'...' '\'' '...'` pattern.
- *
- * @param {string} s
- * @returns {string}  e.g. 'tail -F /path/to/STATE.md'
- */
-function shellQuote(s) {
-  return "'" + String(s).replace(/'/g, "'\\''") + "'";
-}
-
 /** Poll interval (seconds) of the agent-status pane loop. */
 const STATUS_PANE_POLL_SECONDS = 2;
 

package/scripts/lib/tmux-layout/tmux-shell.mjs

@@ -6,6 +6,8 @@
 
 import { execSync } from 'node:child_process';
 
+import { shellQuote } from '../sh-quote.mjs';
+
 const MIN_MAJOR = 3;
 const MIN_MINOR = 0;
 
@@ -69,14 +71,3 @@ export function isSessionCollision(sessionName) {
     return { exists: false, error: `tmux has-session failed: ${err.message}` };
   }
 }
-
-/**
- * Shell-quote a string for safe interpolation into a shell command.
- * Uses single-quote wrapping with embedded single-quote escaping.
- *
- * @param {string} s
- * @returns {string}
- */
-function shellQuote(s) {
-  return "'" + String(s).replace(/'/g, "'\\''") + "'";
-}

package/scripts/lib/ux-grill/reconcile.mjs

@@ -187,7 +187,7 @@ export function buildIssueBody(finding, context) {
   if (!buildIssueBodyDeprecationWarned) {
     buildIssueBodyDeprecationWarned = true;
     console.warn(
-      '[deprecated] buildIssueBody from scripts/lib/ux-grill/reconcile.mjs is deprecated since 5.3.0 ' +
+      '[deprecated] buildIssueBody from scripts/lib/ux-grill/reconcile.mjs is deprecated since the 5.3 line ' +
         'and will be removed in 6.0.0 — use buildUxGrillIssueBody instead.',
     );
   }

package/scripts/lib/validate/check-agents.mjs

@@ -327,7 +327,7 @@ if (mdFiles.length === 0) {
       if (modelVal === null || (!ALLOWED_MODEL_ALIASES.has(modelVal) && !MODEL_ID_RE.test(modelVal))) {
         const got = modelVal ?? '';
         const aliasList = [...ALLOWED_MODEL_ALIASES].join('|');
-        fail(`${agentName}: model must be ${aliasList} or a full model ID like 'claude-opus-4-7' or 'claude-sonnet-5' (got: '${got}')`);
+        fail(`${agentName}: model must be ${aliasList} or a full model ID like 'claude-opus-5-5' or 'claude-sonnet-5' (got: '${got}')`);
       }
     }
 

package/scripts/lib/validate/check-hooks-symmetry.mjs

@@ -113,6 +113,11 @@ const DOCUMENTED_ASYMMETRIES = {
         'skill-invocation-telemetry.mjs',
         'enforce-scope.mjs',
         'config-protection.mjs',
+        // pre-edit-sessions-ledger-guard (#1443): same Edit/Write/MultiEdit tool-name
+        // gate as enforce-scope/config-protection above — Codex writes files through
+        // `apply_patch`, a name this hook never matches. Wired, it would run and allow
+        // everything: a silent no-op, not enforcement. Gap registered, not faked.
+        'pre-edit-sessions-ledger-guard.mjs',
         'pre-bash-destructive-guard.mjs',
         'pre-bash-staging-fence.mjs',
         'pre-bash-memory-propose-audit.mjs',