@event4u/agent-config 16.0.0 → 16.2.0

package/src/scripts/_lib/verify_clause.ts

@@ -0,0 +1,280 @@
+/**
+ * The `verify:` clause — one grammar, one parser, two importers.
+ *
+ * THE DEFECT THIS EXISTS TO NAME. A `verify:` clause names a command and never
+ * names what the command must produce. `verify: `cat notes.md`` is a legal,
+ * green, fully-conforming step field whose oracle cannot say no: the step gets
+ * flipped on a command that was never able to fail. Measured 2026-09-29 across
+ * the 22 active roadmaps: 230 clauses, 93 naming a command, and 6 naming an
+ * expectation — so 59.6% of clauses cannot even be RUN, and 97.4% state no
+ * oracle. The absence is the norm, not an outlier.
+ *
+ * The denominator moves with archival and is not the finding. An earlier
+ * reading the same day over 27 roadmaps gave 264 / 100 / 6; five archived in
+ * between, and the count of EXPECTATIONS did not move. That is why the ratio
+ * carries and the raw totals do not.
+ *
+ * ALL SIX predate this parser, and the split is the interesting part. Three sit
+ * in an unrelated roadmap and were written by an author nobody prompted. The
+ * other three are the verify lines of the roadmap that commissioned this work —
+ * which asserted, in the same document, that ZERO clauses carried an
+ * expectation. It was carrying three of them. This grammar therefore DOCUMENTS
+ * a form already in use; it does not introduce one, and the number here is the
+ * measured one so that it carries its producing command:
+ * `./scripts-run src/scripts/roadmap_verify_share`.
+ *
+ * THE GRAMMAR. The expectation half is OPTIONAL and attaches with the arrow
+ * `exec:` evidence already carries — not a new symbol:
+ *
+ *     verify: `<cmd>` -> 0          the command must exit 0
+ *     verify: `<cmd>` -> 2          … or any declared exit code
+ *     verify: `<cmd>` -> /regex/    the command's output must match
+ *     verify: `<cmd>`               runnable, but the oracle is unstated
+ *     verify: <prose>               MANUAL — a human reads it
+ *
+ * Prose stays legal deliberately. Forbidding it would have silently invalidated
+ * the large majority of lines already in the tree, and a grammar change that
+ * reds every existing producer gets weakened until it finds nothing. What the
+ * three shapes buy instead is that they are TELLABLE APART, which is what lets
+ * `closure_scan` report the unfalsifiable share as a number rather than an
+ * impression.
+ *
+ * WHY ONE PARSER. Two importers read this grammar — the run-continuation hook
+ * (which names the next step and its verify line) and `closure_scan` (which
+ * reports the share that cannot fail). A second copy of the arrow regex is how
+ * those two drift, and a drifted grammar fails silently: an expectation simply
+ * stops being read and the clause reads as unstated. {@link VERIFY_ARROW_SOURCE}
+ * is exported so a duplicate is detectable by a test rather than by a bug
+ * report.
+ */
+
+/** What a clause says the command must produce. */
+export type Expectation =
+    | { readonly kind: 'exit'; readonly code: number }
+    | { readonly kind: 'regex'; readonly source: string };
+
+export interface VerifyClause {
+    /**
+     * The command, or `null` for the prose (MANUAL) form. A clause with no
+     * command cannot be run by anything, which is precisely what makes it
+     * unfalsifiable rather than merely unstated.
+     */
+    readonly command: string | null;
+    /** The `-> 0` / `-> /re/` half, or `null` where the clause carries none. */
+    readonly expect: Expectation | null;
+    /**
+     * The clause body, whitespace-normalised — the clause, never the step.
+     *
+     * For the prose form that is the clause's own paragraph: a detector reading
+     * this must not see text the clause did not write.
+     */
+    readonly raw: string;
+}
+
+/**
+ * The expectation half, as a source string.
+ *
+ * Exported rather than kept private so the "nothing may parse the arrow a
+ * second time" test has something to compare against. Two alternatives only:
+ * a bare exit code, or a slash-delimited regex. A bare word (`-> green`) is
+ * deliberately NOT an expectation — it would read as one while deciding
+ * nothing, which is the defect wearing the fix's clothes.
+ */
+export const VERIFY_ARROW_SOURCE = '(?:->|→)[ \\t]*(\\d+|\\/.+\\/)[ \\t]*$';
+
+/** The arrow alone, anchored at both ends of a clause tail. */
+const ARROW_RE = new RegExp(`^${VERIFY_ARROW_SOURCE}`);
+
+/** A clause body split into its command head and its trailing expectation. */
+const TAIL_RE = new RegExp(`^(.*?)[ \\t]*${VERIFY_ARROW_SOURCE}`);
+
+/** The annotation form. Machine-readable, so it wins when both are present. */
+const HTML_RE = /<!--\s*verify:\s*(.*?)\s*-->/g;
+
+/**
+ * The step's LAST match, not its first.
+ *
+ * A step that DISCUSSES the clause carries the token before the clause that
+ * counts — the roadmap introducing this grammar shows `verify: `<cmd>` -> 0`
+ * in its own prose, two lines above its own verify line. First-match-wins
+ * parsed the illustration and reported its expectation as the step's. The
+ * clause is written last by convention, after the prose, so last wins.
+ */
+function lastMatch(re: RegExp, text: string): RegExpExecArray | null {
+    re.lastIndex = 0;
+    let found: RegExpExecArray | null = null;
+    let m = re.exec(text);
+    while (m !== null) {
+        found = m;
+        if (m.index === re.lastIndex) re.lastIndex += 1;
+        m = re.exec(text);
+    }
+    return found;
+}
+
+/**
+ * The backticked form, plus at most ONE token of expectation after it.
+ *
+ * The label is an ALTERNATION with a lookahead, not two independent optional
+ * backticks. The inherited form — ``` `?verify:`? ``` — let the parser skip the
+ * trailing backtick and use the LABEL's closing backtick to open the command,
+ * so ``` `verify:` stays legal … verify: `cmd` ``` captured the whole
+ * intervening sentence as the command. Nothing failed: a garbage command is
+ * still a string, and the hook printed it into a continuation message as the
+ * step's proof. The alternation alone did not fix it — the bare branch could
+ * start one character later and borrow the same backtick — so the bare branch
+ * carries `(?!`)`: a label followed by a backtick is the delimited form, which
+ * branch one already owns.
+ *
+ * The tail is capped at a single non-space token on purpose. A caller may hand
+ * this parser a whole step BLOCK already joined into one line, so "the rest of
+ * the line" can mean "the rest of the step" — and a step that continues after
+ * its verify clause would then push the expectation away from the end anchor
+ * and lose it silently. The cost is that a regex expectation here may not
+ * contain a space; `/[1-9]/` and `/unfalsifiable-verify/` are the shapes in
+ * use, and a space-bearing expectation loses its arrow loudly (it parses as no
+ * expectation) rather than half-matching.
+ *
+ * THE CAP IS THIS FORM'S, NOT THE GRAMMAR'S — said plainly because an earlier
+ * revision of this comment implied otherwise. The annotation form is
+ * `$`-anchored through {@link VERIFY_ARROW_SOURCE} and DOES keep `-> /a b/`;
+ * this form does not. The two agree on every space-free expectation, which is
+ * every expectation in the tree. Both directions are pinned by test, so closing
+ * the gap is a deliberate act against a failing assertion rather than a silent
+ * drift — which is the failure the WHY ONE PARSER note above exists to prevent,
+ * and it applies to this module's own two readers first.
+ */
+const BACKTICKED_RE = /(?:`verify:`|verify:(?!`))\s*`([^`]+)`[ \t]*((?:->|→)[ \t]*\S+)?/g;
+
+/** Prose after the label. The MANUAL form; the body is bounded separately. */
+const PROSE_RE = /(?:`verify:`|verify:(?!`))[ \t]*/g;
+
+/**
+ * A prose clause runs to the end of its PARAGRAPH — never to the end of the step.
+ *
+ * "To end of line" is too little: a prose clause wraps, and the quantity it
+ * asserts can sit on the wrap. The rest of the step is far too much, and that
+ * was the live defect — a step's evidence block sits under its clause, so a
+ * clause that named no number inherited every number written beneath it, and
+ * `unfalsifiable-verify` reported 46 hits on quantities no clause had written.
+ *
+ * A blank line is the boundary because that is what separates a clause from the
+ * evidence recorded under it in this tree. Input with no newlines at all — a
+ * caller that already joined the block — degrades to the whole string, which is
+ * the previous behaviour and keeps such a caller working rather than silently
+ * truncating it at a boundary that is not there.
+ */
+function proseBody(rest: string): string {
+    const end = rest.search(/\n[ \t]*(?:\r?\n|$)/);
+    return (end === -1 ? rest : rest.slice(0, end)).replace(/\s+/g, ' ').trim();
+}
+
+/** Build an expectation from the already-isolated arrow value. */
+function fromValue(value: string): Expectation | null {
+    if (/^\d+$/.test(value)) return { kind: 'exit', code: Number(value) };
+    const re = /^\/(.+)\/$/.exec(value);
+    if (re === null) return null;
+    const source = re[1] as string;
+    // An expectation that cannot compile is not an oracle — it is a crash
+    // waiting for the first reader that tries to apply it.
+    try {
+        new RegExp(source);
+    } catch {
+        return null;
+    }
+    return { kind: 'regex', source };
+}
+
+/**
+ * Parse an expectation from a clause tail such as `-> 0` or `-> /re/`.
+ *
+ * @returns `null` when the tail carries no arrow, or carries one whose value is
+ * neither an exit code nor a slash-delimited regex.
+ */
+export function parseExpectation(tail: string): Expectation | null {
+    const m = ARROW_RE.exec(tail.trim());
+    return m === null ? null : fromValue(m[1] as string);
+}
+
+/** Split a command-bearing body into its command and its expectation. */
+function splitBody(body: string): { command: string; expect: Expectation | null } {
+    const m = TAIL_RE.exec(body);
+    if (m === null) return { command: body.trim(), expect: null };
+    const expect = fromValue(m[2] as string);
+    // An arrow whose value decides nothing leaves the whole body as the
+    // command, rather than silently truncating it at a `->` that was prose.
+    return expect === null ? { command: body.trim(), expect: null } : { command: (m[1] as string).trim(), expect };
+}
+
+/**
+ * Parse the `verify:` clause carried by `stepText`.
+ *
+ * @param stepText one step block — the line plus its continuation lines, which
+ * is where the tree conventionally writes the clause.
+ * @returns the clause, or `null` when the text carries no `verify:` token.
+ */
+export function parseVerifyClause(stepText: string): VerifyClause | null {
+    const html = lastMatch(HTML_RE, stepText);
+    if (html !== null) {
+        const body = (html[1] as string).trim();
+        // An EMPTY annotation is not a clause and must not mask one. Returning
+        // null here — the inherited behaviour — let `<!-- verify: -->` hide a
+        // real backticked clause further along the same step, and now that two
+        // readers consume this parser it would hide it from the share too.
+        if (body !== '') {
+            const { command, expect } = splitBody(body);
+            return { command: command === '' ? null : command, expect, raw: body };
+        }
+    }
+
+    const backticked = lastMatch(BACKTICKED_RE, stepText);
+    if (backticked !== null) {
+        const command = (backticked[1] as string).trim();
+        if (command !== '') {
+            const tail = (backticked[2] ?? '').trim();
+            return { command, expect: parseExpectation(tail), raw: `\`${command}\`${tail === '' ? '' : ` ${tail}`}` };
+        }
+    }
+
+    const prose = lastMatch(PROSE_RE, stepText);
+    if (prose === null) return null;
+    const body = proseBody(stepText.slice(prose.index + prose[0].length));
+    if (body === '') return null;
+    return { command: null, expect: null, raw: body };
+}
+
+/**
+ * The clause's command half only — `null` where the clause is prose or absent.
+ *
+ * The narrowing the run-continuation hook wants: it renders a command for the
+ * next step, and a prose clause is not one. Kept here rather than at the call
+ * site so the hook imports a decision rather than re-deriving it.
+ */
+export function commandBearing(stepText: string): VerifyClause | null {
+    const clause = parseVerifyClause(stepText);
+    if (clause === null || clause.command === null) return null;
+    return clause;
+}
+
+/** Render an expectation back to its written form, for messages and evidence. */
+export function formatExpectation(expect: Expectation): string {
+    return expect.kind === 'exit' ? `-> ${String(expect.code)}` : `-> /${expect.source}/`;
+}
+
+/**
+ * The clause as a continuation-message line, empty when there is nothing to say.
+ *
+ * The expectation rides along with the command, and that is the point rather
+ * than a formatting nicety: a re-engagement told only the command can run it,
+ * see exit 0, and flip the box regardless — the whole defect the arrow exists
+ * to close. Told the expectation too, it has an oracle.
+ *
+ * Lives beside the grammar rather than at the one call site so that one module
+ * owns how a clause is read AND how it is written back, which is also what
+ * keeps the over-cap hook from paying these lines.
+ */
+export function renderVerifyLine(clause: VerifyClause | null): string {
+    if (clause === null) return '';
+    const expect = clause.expect === null ? '' : ` ${formatExpectation(clause.expect)}`;
+    return `\n  verify: ${clause.command ?? ''}${expect}`;
+}

package/src/scripts/ai_council/config.ts

@@ -543,6 +543,7 @@ export type QuorumMinPresent = number;
 
 export type { FallbackConfig } from './fallback_config.js';
 import { buildFallback, buildSecondModel, type FallbackConfig as _FallbackConfig } from './fallback_config.js';
+import { OWNERSHIP_CLASSES, allOwnerLocked, defaultModeFor, ownershipModeError } from './ownership_classes.js';
 import {
     _f,
     _pyOct,
@@ -1182,13 +1183,7 @@ const _LOCKED_IMPACT_CLASSES: ReadonlySet<string> = new Set([
     'user_required',
 ]);
 
-const _DEFAULT_RESOLUTION_MODES: Readonly<Record<string, string>> = {
-    trivial: 'agent',
-    low_impact: 'agent',
-    medium_impact: 'council',
-    high_impact: 'user',
-    user_required: 'user',
-};
+const _ALL_OWNER_LOCKED = allOwnerLocked(_LOCKED_IMPACT_CLASSES);
 
 function _build_decision_resolution(d: Dict): DecisionResolutionConfig {
     if (!_isDict(d)) {
@@ -1207,14 +1202,14 @@ function _build_decision_resolution(d: Dict): DecisionResolutionConfig {
         );
     }
     const classes = new Map<string, DecisionResolutionEntry>();
-    for (const cls of _IMPACT_CLASSES) {
+    for (const cls of [..._IMPACT_CLASSES, ...OWNERSHIP_CLASSES]) {
         const entry_raw = _orEmpty(classes_raw[cls] ?? null);
         if (!_isDict(entry_raw)) {
             throw new CouncilConfigError(
                 `\`decision_resolution.classes.${cls}\` must be a mapping.`,
             );
         }
-        const mode = _get(entry_raw, 'mode', _DEFAULT_RESOLUTION_MODES[cls] as string);
+        const mode = _get(entry_raw, 'mode', defaultModeFor(cls));
         if (!(_isStr(mode) && _VALID_RESOLUTION_MODES.has(mode))) {
             throw new CouncilConfigError(
                 `decision_resolution.classes.${cls}.mode=${_pyRepr(mode)} not in ` +
@@ -1230,10 +1225,12 @@ function _build_decision_resolution(d: Dict): DecisionResolutionConfig {
                     `the user.`,
             );
         }
+        const ownErr = ownershipModeError(cls, mode, _pyRepr);
+        if (ownErr !== null) throw new CouncilConfigError(ownErr);
         // Iron Law: `dispatch` is not configurable for locked classes
         // (step-9 P8/P11 · U3). Any nested `dispatch` key — including
         // smuggled-in YAML anchor merges — is a hard schema error.
-        if (_LOCKED_IMPACT_CLASSES.has(cls) && 'dispatch' in entry_raw) {
+        if (_ALL_OWNER_LOCKED.has(cls) && 'dispatch' in entry_raw) {
             throw new CouncilConfigError(
                 `decision_resolution.classes.${cls}.dispatch=` +
                     `${_pyRepr(entry_raw.dispatch)}: dispatch is not ` +
@@ -1251,7 +1248,7 @@ function _build_decision_resolution(d: Dict): DecisionResolutionConfig {
         // UOTL Phase 4.1 — the optional local rung, REFUSED on a locked
         // class rather than ignored: a dropped key reads as configured.
             const second_model = buildSecondModel(
-            entry_raw as Record<string, unknown>, cls, _LOCKED_IMPACT_CLASSES, _FALLBACK_DEPS);
+            entry_raw as Record<string, unknown>, cls, _ALL_OWNER_LOCKED, _FALLBACK_DEPS);
         classes.set(cls, { mode, confidence_threshold: threshold, second_model });
     }
     const fast_path_raw = _getOr(d, 'fast_path', {});

package/src/scripts/ai_council/fallback_config.ts

@@ -107,9 +107,11 @@ export function buildSecondModel(
     if (locked.has(cls)) {
         throw d.error(
             `decision_resolution.classes.${cls}.second_model=${d.repr(entry['second_model'])}: ` +
-                `a second-model rung is not configurable for high-impact / user-required ` +
-                `decisions — those classes are LOCKED to \`user\` (Iron Law) and never resolve ` +
-                `on a model of any kind.`,
+                `a second-model rung is not configurable for an owner-routed class ` +
+                `(high-impact / user-required on the impact axis; product-owned / ` +
+                `business-owned / destructive-owned on the ownership axis) — those ` +
+                `classes are LOCKED to \`user\` (Iron Law) and never resolve on a ` +
+                `model of any kind.`,
         );
     }
     const raw = entry['second_model'];

package/src/scripts/ai_council/ownership_classes.ts

@@ -0,0 +1,118 @@
+/**
+ * The ownership axis of `decision_resolution` (ADR-268 § 10).
+ *
+ * A separate module rather than more constants in `config.ts`, for two
+ * reasons. That file is 700+ lines over the source ceiling already, so every
+ * line added there moves a ratchet; and these sets are read by the loader AND
+ * by the roadmap-side gate, which has no business importing the council's
+ * whole schema to learn eight class names.
+ *
+ * The axis changes from IMPACT to OWNERSHIP: a technical decision does not
+ * become owner-owned because it is hard or high-impact. The five impact names
+ * stay accepted for one minor so no installed `.ai-council.yml` fails to load,
+ * and they keep their own (impact-axis) Iron Law — they are deliberately NOT
+ * in this module.
+ */
+
+/** The eight ownership classes, in table order. */
+export const OWNERSHIP_CLASSES: readonly string[] = [
+    'deterministic',
+    'reversible-technical',
+    'contested-technical',
+    'critical-technical',
+    'product-owned',
+    'business-owned',
+    'destructive-owned',
+    'spend-exhaustion',
+];
+
+/**
+ * Iron Law: ONLY these three are locked to `user`. `critical-technical` is
+ * deliberately absent — locking it is the § 0 break ADR-268 names by hand
+ * ("routing a technical decision to the owner because it is hard or
+ * high-impact"), so the absence is load-bearing rather than an oversight.
+ */
+export const OWNER_LOCKED_CLASSES: ReadonlySet<string> = new Set([
+    'product-owned',
+    'business-owned',
+    'destructive-owned',
+]);
+
+/**
+ * Iron Law: `spend-exhaustion` PAUSES AND REPORTS (ADR-268 § 8). It is never
+ * routed to the owner as a question, so `mode: user` on it is a hard schema
+ * error rather than an unusual-but-legal setting.
+ */
+export const NEVER_OWNER_CLASSES: ReadonlySet<string> = new Set(['spend-exhaustion']);
+
+/**
+ * Both owner-routed sets as one, given the impact axis's own locked set.
+ *
+ * A function rather than a constant because the impact half lives in the
+ * loader: importing it here would make this module depend on the schema it is
+ * imported BY, which is the cycle the split exists to avoid.
+ */
+export function allOwnerLocked(impactLocked: ReadonlySet<string>): ReadonlySet<string> {
+    return new Set<string>([...impactLocked, ...OWNER_LOCKED_CLASSES]);
+}
+
+/**
+ * Default `mode` per class, BOTH axes — the resolver column of each table.
+ *
+ * The five impact rows moved here with the eight ownership ones so there is
+ * exactly one place a reader asks "what does this class do when nobody
+ * configured it", rather than one map per axis drifting against the other.
+ */
+export const DEFAULT_RESOLUTION_MODES: Readonly<Record<string, string>> = {
+    trivial: 'agent',
+    low_impact: 'agent',
+    medium_impact: 'council',
+    high_impact: 'user',
+    user_required: 'user',
+    deterministic: 'agent',
+    'reversible-technical': 'agent',
+    'contested-technical': 'council',
+    'critical-technical': 'council',
+    'product-owned': 'user',
+    'business-owned': 'user',
+    'destructive-owned': 'user',
+    'spend-exhaustion': 'agent',
+};
+
+/** The default `mode` for a class, or `agent` for a name neither axis knows. */
+export function defaultModeFor(cls: string): string {
+    return DEFAULT_RESOLUTION_MODES[cls] ?? 'agent';
+}
+
+/**
+ * The ownership axis's mode check, as one function.
+ *
+ * Both Iron Laws live here rather than at the call site so the loader carries
+ * three lines instead of twenty-four — and so a reader looking for "which
+ * classes are locked" finds the rule next to the sets it reads.
+ *
+ * Returns the error message, or `null` when the mode is legal. `repr` is the
+ * caller's own value formatter, so the message reads identically to every
+ * other schema error in the loader.
+ */
+export function ownershipModeError(
+    cls: string,
+    mode: string,
+    repr: (v: unknown) => string,
+): string | null {
+    if (OWNER_LOCKED_CLASSES.has(cls) && mode !== 'user') {
+        return (
+            `decision_resolution.classes.${cls}.mode=${repr(mode)}: ` +
+            `class \`${cls}\` is LOCKED to \`user\` (Iron Law) — product, business and ` +
+            `destructive decisions are owner-owned and never resolve on a model.`
+        );
+    }
+    if (NEVER_OWNER_CLASSES.has(cls) && mode === 'user') {
+        return (
+            `decision_resolution.classes.${cls}.mode=${repr(mode)}: ` +
+            `class \`${cls}\` is never owner-routed (Iron Law) — spend exhaustion ` +
+            `pauses and reports, it does not ask.`
+        );
+    }
+    return null;
+}

package/src/scripts/ai_team/team_dispatch.ts

@@ -37,6 +37,10 @@
  * shape of an Apache-2.0 upstream — attribution in the repo-root NOTICE file.
  */
 import { spawnSync } from 'node:child_process';
+// read-surface: inter-agent-reply — a teammate's reply arrives through a HOST
+// primitive this package does not own, so no call site here spells the inbound
+// fetch. Declared rather than inferred; the coverage verdict is still read from
+// the import below.
 import { sanitize_text } from '../_lib/retrieval_sanitize.js';
 import * as fs from 'node:fs';
 import * as path from 'node:path';