@event4u/agent-config 14.11.0 → 14.13.0

package/src/scripts/check_pack_size.ts

@@ -219,17 +219,71 @@ export function payloadClassCounts(
     }));
 }
 
+/**
+ * The recorded BUILT-surface figure, if the budget file carries one.
+ *
+ * Read by key SHAPE (`built_surface_measurement_<date>`) rather than by the one
+ * dated key that exists today, so re-measuring the built surface is an added
+ * record rather than an edit to this function. The newest key wins.
+ */
+export function recordedBuiltPackedMb(budget: PackSizeBudget): number | null {
+    const keys = Object.keys(budget)
+        .filter((k) => /^built_surface_measurement_/.test(k))
+        .sort();
+    const newest = keys[keys.length - 1];
+    if (newest === undefined) return null;
+    const block = (budget as unknown as Record<string, unknown>)[newest];
+    if (block === null || typeof block !== 'object') return null;
+    const built = (block as { built?: unknown }).built;
+    if (built === null || typeof built !== 'object') return null;
+    const value = (built as { packed_mb?: unknown }).packed_mb;
+    return typeof value === 'number' && Number.isFinite(value) ? value : null;
+}
+
 export function evaluate(budget: PackSizeBudget, pack: PackResult): string[] {
     const errors: string[] = [];
 
     const packedMb = pack.size / 1e6;
     const entry = budget.budgets['packed_size_mb'];
+    const built = payloadIsBuilt(pack.files);
+    const pct = budget.regression_pct;
     if (entry === undefined) {
         errors.push('packed_size_mb: missing from the budget file — gate would pass vacuously');
+    } else if (built) {
+        // A BUILT payload is not the surface `max` describes, and comparing
+        // them is the category error `pack-size-budget.json` warns about in its
+        // own `measurement_conditions`: every cap in that file, `max: 9.1`
+        // included, was measured with `--ignore-scripts` on a tree with no
+        // `dist/cli/**`. The built surface it records separately is ~2 MB
+        // larger *by construction*, so a job that happens to build first turns
+        // a green branch red for a reason that has nothing to do with its diff.
+        // Measured 2026-08-30: main 8.985 (2715 entries, unbuilt) against a
+        // branch adding six source files at 9.922 (2827 entries) — the 112-file
+        // difference is `dist/cli` + `dist/cli-delegate`, not the diff.
+        //
+        // So the built payload is judged against the recorded BUILT figure,
+        // with the same creep percentage. That keeps teeth on the surface a
+        // consumer installs instead of pretending the unbuilt cap covers it.
+        const baseline = recordedBuiltPackedMb(budget);
+        if (baseline === null) {
+            errors.push(
+                `packed_size_mb: measured ${packedMb.toFixed(3)} on a BUILT payload, and the budget `
+                    + 'file records no built-surface measurement to compare it against — the unbuilt '
+                    + 'cap does not describe this tree. Record one, or pack with --ignore-scripts.',
+            );
+        } else {
+            const ceiling = baseline * (1 + pct / 100);
+            if (packedMb > ceiling) {
+                errors.push(
+                    `packed_size_mb (BUILT surface): measured ${packedMb.toFixed(3)} regressed `
+                        + `>${String(pct)}% vs the recorded built figure ${String(baseline)} `
+                        + `(ceiling ${ceiling.toFixed(3)})`,
+                );
+            }
+        }
     } else if (packedMb > entry.max) {
         errors.push(`packed_size_mb: measured ${packedMb.toFixed(3)} exceeds budget ${entry.max}`);
     } else {
-        const pct = budget.regression_pct;
         const ceiling = entry.last_measured * (1 + pct / 100);
         if (entry.last_measured > 0 && packedMb > ceiling) {
             errors.push(
@@ -436,7 +490,10 @@ export function main(argv: readonly string[] = process.argv.slice(2)): number {
         return 1;
     }
     process.stdout.write(
-        `✅  pack size within budget — ${(pack.size / 1e6).toFixed(3)} MB packed, ` +
+        // The SURFACE is named on the green path too. A figure with no surface
+        // label is how two numbers for one lever start circulating, which is
+        // the confusion this budget file opens by warning about.
+        `✅  pack size within budget (${payloadIsBuilt(pack.files) ? 'BUILT surface, vs the recorded built figure' : 'UNBUILT surface, vs max'}) — ${(pack.size / 1e6).toFixed(3)} MB packed, ` +
             `${Object.keys(perSkill).length} skills, largest share ` +
             `${Math.max(...Object.values(perSkill).map((b) => (b / total) * 100)).toFixed(2)}%\n`,
     );

package/src/scripts/check_preamble_payload_budget.ts

@@ -20,11 +20,15 @@
  *
  * Exit codes: 0 within budget · 1 over budget · 2 misuse / unreadable budget.
  */
+import { execFileSync } from 'node:child_process';
 import * as fs from 'node:fs';
+import * as os from 'node:os';
 import * as path from 'node:path';
 
 import { assertScanned, DeadScopeError } from './_lib/scan_scope.js';
 import { censusClaudeMdHierarchy, censusRuleDir, censusSkillsCatalog } from './preamble_byte_census.js';
+import { PREFIX_STABLE_SURFACES, prefixStableRoots } from './_lib/prefix_stable_surfaces.js';
+import { attributeGrowth, buildLedger, renderAttribution } from './_lib/asset_delivery_ledger.js';
 
 const HERE = path.dirname(new URL(import.meta.url).pathname);
 const REPO_ROOT = path.resolve(HERE, '..', '..');
@@ -63,12 +67,30 @@ export function readBudget(file: string = BUDGET_FILE): Budget {
     };
 }
 
+/** Look one surface root up by id. Throws on an unknown id — a renamed surface
+ *  must fail loudly here rather than degrade this gate to measuring nothing. */
+function surfaceRoot(id: string): string {
+    const s = PREFIX_STABLE_SURFACES.find((x) => x.id === id);
+    if (s === undefined) {
+        throw new Error(
+            `prefix-stable surface '${id}' is not declared in _lib/prefix_stable_surfaces.ts — ` +
+                `the payload census cannot measure a bucket whose root it cannot resolve`,
+        );
+    }
+    return s.root;
+}
+
 /** Measure only what the repo tree determines — no `~`, no transcripts, no network. */
 export function measureDeterministicPayload(
     repoRoot: string = REPO_ROOT,
 ): Array<{ name: string; tokens: number; files: number }> {
-    const projectRules = censusRuleDir(path.join(repoRoot, 'dist', 'agent-src', 'rules'));
-    const skills = censusSkillsCatalog(path.join(repoRoot, 'dist', 'agent-src', 'skills'));
+    // Roots come from the canonical prefix-stable surface registry, never from a
+    // literal here: `check_prefix_stable_mutation` guards the same boundary, and
+    // two independent lists of one boundary is the drift shape this repository
+    // has already paid for. `surfaceRoot` throws rather than silently measuring
+    // nothing if an id is renamed out from under it.
+    const projectRules = censusRuleDir(path.join(repoRoot, surfaceRoot('project-scope-rules')));
+    const skills = censusSkillsCatalog(path.join(repoRoot, surfaceRoot('preloaded-skills-catalog')));
     // Only the PROJECT half of the CLAUDE.md hierarchy is deterministic — the
     // user file and its @-imports live on whatever machine runs this.
     const claudeMd = censusClaudeMdHierarchy(repoRoot, path.join(repoRoot, '.no-such-home'));
@@ -91,19 +113,107 @@ export function measureDeterministicPayload(
     ];
 }
 
-export function evaluate(repoRoot?: string, budgetFile?: string): BudgetVerdict {
+/**
+ * A caller-supplied ceiling, and the one direction it may go.
+ *
+ * road-to-standing-payload-truth 1.1 arms this gate in CI behind a GRACE ceiling
+ * equal to the 2026-08-24 measurement, because HEAD is 28.4 % over the design
+ * ceiling and blocking at the design number would fail every pull request from
+ * the moment the workflow lands — nobody can shed 30,566 tokens inside the PR
+ * that arms the gate.
+ *
+ * **An override may only be LOOSER than the design ceiling, and it is refused if
+ * it is tighter.** That sounds backwards for one line and is the whole point: a
+ * tighter override would let a caller silently *lower* the bar this file owns,
+ * which is the config-weakening shape in reverse — the design ceiling stays the
+ * authority, and the override is a dated, expiring concession recorded in
+ * `ci_delivery`. Tightening happens by lowering `baseline_tokens`, in the file,
+ * with a reason, where a reviewer sees it.
+ */
+export function evaluate(repoRoot?: string, budgetFile?: string, overrideCeiling?: number): BudgetVerdict {
     const budget = readBudget(budgetFile);
     const buckets = measureDeterministicPayload(repoRoot);
     const measured = buckets.reduce((sum, b) => sum + b.tokens, 0);
-    const ceiling = Math.round(budget.baseline_tokens * (1 + budget.headroom_pct / 100));
+    const design = Math.round(budget.baseline_tokens * (1 + budget.headroom_pct / 100));
+    const ceiling =
+        overrideCeiling !== undefined && Number.isFinite(overrideCeiling) && overrideCeiling > design
+            ? overrideCeiling
+            : design;
     return { measured, baseline: budget.baseline_tokens, ceiling, withinBudget: measured <= ceiling, buckets };
 }
 
+/**
+ * Per-asset attribution for a refusal, against the merge-base tree.
+ *
+ * Reads the base tree through `git worktree`-free plumbing: the ledger is built
+ * over a temporary checkout of `git merge-base HEAD origin/main`. Every failure
+ * mode returns an empty list — the caller treats attribution as an explanation,
+ * never as a precondition for refusing.
+ */
+function attributeGrowthAgainstBase(): string[] | null {
+    try {
+        const base = execFileSync('git', ['merge-base', 'HEAD', 'origin/main'], {
+            cwd: REPO_ROOT,
+            encoding: 'utf-8',
+            stdio: ['ignore', 'pipe', 'ignore'],
+        }).trim();
+        if (base.length === 0) return null;
+
+        const tmp = fs.mkdtempSync(path.join(os.tmpdir(), 'payload-base-'));
+        try {
+            for (const rel of prefixStableRoots()) {
+                // `git archive` of one subtree, extracted into the scratch root.
+                // A root absent at the base ref simply yields nothing.
+                try {
+                    const tar = execFileSync('git', ['archive', base, '--', rel], {
+                        cwd: REPO_ROOT,
+                        maxBuffer: 256 * 1024 * 1024,
+                        stdio: ['ignore', 'pipe', 'ignore'],
+                    });
+                    // `-f -` is not optional: without it BSD tar reads its
+                    // default device rather than stdin, and the extraction
+                    // silently produces nothing — which reads downstream as
+                    // "the base tree was unavailable" rather than as a bug.
+                    execFileSync('tar', ['-x', '-f', '-', '-C', tmp], {
+                        input: tar,
+                        stdio: ['pipe', 'ignore', 'ignore'],
+                    });
+                } catch {
+                    /* root absent at base — nothing to extract */
+                }
+            }
+            const [rulesRel, skillsRel] = prefixStableRoots();
+            const before = buildLedger(
+                path.join(tmp, rulesRel ?? ''),
+                path.join(tmp, skillsRel ?? ''),
+                tmp,
+            );
+            const after = buildLedger(
+                path.join(REPO_ROOT, rulesRel ?? ''),
+                path.join(REPO_ROOT, skillsRel ?? ''),
+                REPO_ROOT,
+            );
+            if (before.rows.length === 0) return null;
+            return renderAttribution(attributeGrowth(before.rows, after.rows));
+        } finally {
+            fs.rmSync(tmp, { recursive: true, force: true });
+        }
+    } catch {
+        return null;
+    }
+}
+
 export function main(argv: string[] = process.argv.slice(2)): number {
     const json = argv.includes('--format=json') || argv.includes('--json');
+    // `--ceiling <n>`: the grace ceiling the CI step reads out of
+    // `ci_delivery.grace_ceiling`. Read from the budget file there, never written
+    // in the workflow, so the number has exactly one home. A non-numeric or
+    // tighter-than-design value is IGNORED rather than honoured — see `evaluate`.
+    const ci = argv.indexOf('--ceiling');
+    const override = ci !== -1 && argv[ci + 1] !== undefined ? Number(argv[ci + 1]) : undefined;
     let verdict: BudgetVerdict;
     try {
-        verdict = evaluate();
+        verdict = evaluate(undefined, undefined, override);
     } catch (err) {
         process.stderr.write(`❌  preamble-payload budget: ${(err as Error).message}\n`);
         return 2;
@@ -118,7 +228,7 @@ export function main(argv: string[] = process.argv.slice(2)): number {
             gate: 'check_preamble_payload_budget',
             scanned: verdict.buckets.reduce((n, b) => n + b.files, 0),
             units: 'payload source file(s)',
-            roots: ['dist/agent-src/rules', 'dist/agent-src/skills', 'CLAUDE.md'],
+            roots: prefixStableRoots(),
         });
     } catch (err) {
         if (err instanceof DeadScopeError) {
@@ -143,12 +253,40 @@ export function main(argv: string[] = process.argv.slice(2)): number {
             `(baseline ${verdict.baseline}, ${sign}${delta}; ceiling ${verdict.ceiling})\n`,
     );
     if (!verdict.withinBudget) {
+        // road-to-delivered-cost-truth 2.2 — a gate names its own "no". The
+        // ceiling message alone states a fact and leaves the reader to find the
+        // cause; naming the assets and their token deltas makes the refusal
+        // actionable, which is the difference between a refusal that gets fixed
+        // and one that gets suppressed.
+        //
+        // Attribution is BEST-EFFORT and never changes the verdict: the
+        // comparison needs the merge-base tree, and a shallow clone, a detached
+        // build or a first commit legitimately has none. A gate that failed to
+        // refuse because it could not explain itself would be strictly worse
+        // than one that refuses without the explanation.
+        const attribution = attributeGrowthAgainstBase();
         process.stderr.write(
             `❌  per-spawn preamble payload grew past the ratchet: ${verdict.measured} > ${verdict.ceiling} tok.\n` +
                 `    Every rule and skill description here is re-written on EVERY subagent spawn, so growth\n` +
                 `    is paid per spawn, not once. Either shrink the addition, or raise baseline_tokens in\n` +
                 `    src/config/preamble-payload-budget.json with the reason in the same commit.\n`,
         );
+        // Three distinct states, and conflating the last two is a diagnostic
+        // defect rather than a cosmetic one: "I could not look" and "I looked
+        // and nothing changed" send a reader to completely different places.
+        if (attribution === null) {
+            process.stderr.write(
+                '\n    (no per-asset attribution: the merge-base tree could not be read here, so the\n' +
+                    '     growth cannot be traced to specific assets. The refusal stands regardless.)\n',
+            );
+        } else if (attribution.length === 0) {
+            process.stderr.write(
+                '\n    Per-asset attribution: NO standing asset changed against the merge base — this\n' +
+                    '    diff did not cause the overage, it inherited it.\n',
+            );
+        } else {
+            process.stderr.write('\n' + attribution.join('\n') + '\n');
+        }
         return 1;
     }
     process.stdout.write('✅  per-spawn preamble payload within the ratchet.\n');

package/src/scripts/check_prefix_stable_mutation.ts

@@ -0,0 +1,372 @@
+#!/usr/bin/env tsx
+/**
+ * A mid-session mutation of a declared prefix-stable surface is a violation.
+ *
+ * WHY
+ * ---
+ * `check_kernel_prefix_stability` guards the kernel prefix at AUTHORING time —
+ * it compares a committed snapshot and re-anchors in the same PR. `payload_hash_drift`
+ * measures prefix churn AFTER the fact, from an audit ledger. Neither can see the
+ * case this gate exists for: a hook, a script, or (once one exists) a resident
+ * process rewriting a prefix-stable surface DURING a session, which invalidates
+ * the prompt cache for the rest of it and makes the next call pay the cache-WRITE
+ * rate over the whole prefix.
+ *
+ * That is the runtime half, and until this gate it was prose.
+ *
+ * WHAT IT CHECKS
+ * --------------
+ * For every concern in `src/scripts/hook_manifest.yaml` bound on at least one
+ * MID-SESSION slot — any slot that is not a declared re-arm event — the gate reads
+ * the concern's script and looks for evidence that it writes inside a surface
+ * declared in `_lib/prefix_stable_surfaces.ts`. Evidence is a violation unless the
+ * concern declares `re_arm: <event>`, naming the boundary at which a rebuilt
+ * prefix is expected and paid for once.
+ *
+ * THE LIST IS LOADED, NEVER RESTATED. Both this gate and
+ * `check_preamble_payload_budget` import `PREFIX_STABLE_SURFACES`. Two lists
+ * describing one boundary is the drift shape this repository has already paid for.
+ *
+ * FAIL-CLOSED ON UNDECIDABLE (council 2026-08-28, 2/2)
+ * ----------------------------------------------------
+ * A write whose target is a runtime variable cannot be classified by reading the
+ * source. Classifying that as "fine" turns every dynamic path into an accidental
+ * exemption. So: a dynamic write target in a script that ALSO carries a literal
+ * resolving into a declared surface is reported as `undecidable` and fails. A
+ * dynamic write in a script with no such literal is not reported — the gate does
+ * not claim to know what it cannot see, and reporting every dynamic write in every
+ * hook would be a gate nobody can keep green.
+ *
+ * Exit codes: 0 clean · 1 violation(s) · 2 misuse / unreadable manifest.
+ *
+ * Usage:
+ *   ./scripts-run src/scripts/check_prefix_stable_mutation [--root <dir>] [--json] [--quiet]
+ *   ./scripts-run src/scripts/check_prefix_stable_mutation --self-test
+ */
+import { spawnSync } from 'node:child_process';
+import * as fs from 'node:fs';
+import * as os from 'node:os';
+import * as path from 'node:path';
+
+import { GateLedger } from './_lib/gate_ledger.js';
+import { runGateCli, runSelfTest, type SelfTestCase } from './_lib/gate_self_test.js';
+import { DeadScopeError, reportScanned } from './_lib/scan_scope.js';
+import { PREFIX_STABLE_SURFACES, RE_ARM_EVENTS, prefixStableRoots, surfaceFor } from './_lib/prefix_stable_surfaces.js';
+
+const HERE = path.dirname(new URL(import.meta.url).pathname);
+const REPO_ROOT = path.resolve(HERE, '..', '..');
+
+/** Write-shaped calls. A read never invalidates a cache; only a write does. */
+const WRITE_CALLS = [
+    'writeFileSync',
+    'appendFileSync',
+    'createWriteStream',
+    'renameSync',
+    'unlinkSync',
+    'rmSync',
+    'rmdirSync',
+    'cpSync',
+    'copyFileSync',
+    'truncateSync',
+    'writeFile',
+    'appendFile',
+    'outputFileSync',
+] as const;
+
+export interface Finding {
+    concern: string;
+    script: string;
+    slots: string[];
+    kind: 'literal' | 'undecidable';
+    surface: string;
+    evidence: string;
+}
+
+export interface Verdict {
+    scanned: number;
+    findings: Finding[];
+    midSessionConcerns: number;
+}
+
+interface ConcernSpec {
+    script?: unknown;
+    re_arm?: unknown;
+}
+
+/** Yaml is already a dependency of the manifest linter; reuse the same reader shape. */
+function loadManifest(root: string): { concerns: Record<string, ConcernSpec>; platforms: Record<string, Record<string, string[]>> } {
+    const file = path.join(root, 'src', 'scripts', 'hook_manifest.yaml');
+    const raw = fs.readFileSync(file, 'utf-8');
+    // Minimal, dependency-free parse of the two shapes this gate needs. The full
+    // schema is `lint_hook_manifest`'s business; duplicating a YAML dependency
+    // here would add a second parser to disagree with.
+    const out = spawnSync(
+        process.execPath,
+        ['-e', 'const y=require("js-yaml");let s="";process.stdin.on("data",d=>s+=d).on("end",()=>{const d=y.load(s);process.stdout.write(JSON.stringify({concerns:d.concerns||{},platforms:d.platforms||{}}))})'],
+        { input: raw, encoding: 'utf-8', cwd: REPO_ROOT, maxBuffer: 64 * 1024 * 1024 },
+    );
+    if (out.status !== 0 || !out.stdout) {
+        throw new Error(`hook_manifest.yaml could not be parsed: ${out.stderr || 'no output'}`);
+    }
+    return JSON.parse(out.stdout) as { concerns: Record<string, ConcernSpec>; platforms: Record<string, Record<string, string[]>> };
+}
+
+/** Every slot each concern is bound on, across every platform. */
+export function slotsByConcern(platforms: Record<string, Record<string, string[]>>): Map<string, Set<string>> {
+    const m = new Map<string, Set<string>>();
+    for (const slots of Object.values(platforms)) {
+        if (slots === null || typeof slots !== 'object') continue;
+        for (const [slot, names] of Object.entries(slots)) {
+            if (!Array.isArray(names)) continue;
+            for (const n of names) {
+                const set = m.get(n) ?? new Set<string>();
+                set.add(slot);
+                m.set(n, set);
+            }
+        }
+    }
+    return m;
+}
+
+/** Slots that are NOT re-arm boundaries — i.e. fire inside a live session. */
+export function midSessionSlots(slots: Iterable<string>): string[] {
+    return [...slots].filter((s) => !RE_ARM_EVENTS.includes(s)).sort();
+}
+
+/** All string literals in a source fragment, in order. */
+function literals(src: string): string[] {
+    const out: string[] = [];
+    const re = /'([^'\\\n]*)'|"([^"\\\n]*)"|`([^`$\\\n]*)`/g;
+    let m: RegExpExecArray | null;
+    while ((m = re.exec(src)) !== null) out.push(m[1] ?? m[2] ?? m[3] ?? '');
+    return out;
+}
+
+/** The first argument expression of each write-shaped call in `src`. */
+export function writeTargets(src: string): string[] {
+    const out: string[] = [];
+    for (const call of WRITE_CALLS) {
+        const re = new RegExp(`\\b${call}\\s*\\(`, 'g');
+        let m: RegExpExecArray | null;
+        while ((m = re.exec(src)) !== null) {
+            let depth = 1;
+            let i = m.index + m[0].length;
+            const start = i;
+            while (i < src.length && depth > 0) {
+                const ch = src[i];
+                if (ch === '(' || ch === '[' || ch === '{') depth += 1;
+                else if (ch === ')' || ch === ']' || ch === '}') depth -= 1;
+                else if (ch === ',' && depth === 1) break;
+                i += 1;
+            }
+            out.push(src.slice(start, i));
+        }
+    }
+    return out;
+}
+
+/**
+ * Classify one script. Returns findings for this concern.
+ *
+ * `literal`     — a write whose own argument carries a literal resolving into a
+ *                 declared surface. Unambiguous.
+ * `undecidable` — a write with no classifiable literal, in a file that carries a
+ *                 surface literal somewhere. Fails closed.
+ */
+export function classifyScript(source: string): Array<{ kind: 'literal' | 'undecidable'; surface: string; evidence: string }> {
+    const found: Array<{ kind: 'literal' | 'undecidable'; surface: string; evidence: string }> = [];
+    const targets = writeTargets(source);
+    if (targets.length === 0) return found;
+
+    // A file-wide literal that resolves into a surface — the fail-closed trigger.
+    const fileSurfaces = literals(source)
+        .map((l) => ({ lit: l, s: surfaceFor(l) }))
+        .filter((x) => x.s !== null);
+
+    let sawDynamic = false;
+    for (const t of targets) {
+        const lits = literals(t);
+        // `path.join('dist','agent-src','rules', …)` — join the literal segments too.
+        const joined = lits.join('/');
+        const hit = lits.map((l) => surfaceFor(l)).find((s) => s !== null) ?? surfaceFor(joined);
+        if (hit) {
+            found.push({ kind: 'literal', surface: hit.id, evidence: t.trim().slice(0, 120) });
+        } else if (lits.length === 0) {
+            sawDynamic = true;
+        }
+    }
+
+    const firstSurface = fileSurfaces[0];
+    if (found.length === 0 && sawDynamic && firstSurface !== undefined && firstSurface.s !== null) {
+        found.push({
+            kind: 'undecidable',
+            surface: firstSurface.s.id,
+            evidence: `a write target is computed at runtime and this file carries the literal '${firstSurface.lit}'`,
+        });
+    }
+    return found;
+}
+
+export function evaluate(root: string = REPO_ROOT, ledger?: GateLedger): Verdict {
+    const { concerns, platforms } = loadManifest(root);
+    const bySlot = slotsByConcern(platforms);
+    const findings: Finding[] = [];
+    let scanned = 0;
+    let midSessionConcerns = 0;
+
+    ledger?.plan(Object.keys(concerns));
+
+    for (const [name, spec] of Object.entries(concerns)) {
+        const slots = midSessionSlots(bySlot.get(name) ?? []);
+        if (slots.length === 0) {
+            ledger?.outOfScope(name, 'not_applicable_kind');
+            continue;
+        }
+        midSessionConcerns += 1;
+
+        const rel = typeof spec?.script === 'string' ? spec.script : null;
+        if (rel === null) {
+            ledger?.skip(name, 'manifest_absent');
+            continue;
+        }
+        const abs = path.join(root, rel);
+        if (!fs.existsSync(abs)) {
+            ledger?.skip(name, 'no_applicable_files');
+            continue;
+        }
+        scanned += 1;
+
+        const declaredReArm = typeof spec?.re_arm === 'string' ? spec.re_arm : null;
+        if (declaredReArm !== null && RE_ARM_EVENTS.includes(declaredReArm)) {
+            ledger?.skip(name, 'declared_exemption');
+            continue;
+        }
+
+        const hits = classifyScript(fs.readFileSync(abs, 'utf-8'));
+        for (const hit of hits) {
+            findings.push({ concern: name, script: rel, slots, kind: hit.kind, surface: hit.surface, evidence: hit.evidence });
+        }
+        if (hits.length > 0) ledger?.fail(name, `${String(hits.length)} mid-session write(s) into a prefix-stable surface`);
+        else ledger?.complete(name);
+    }
+    return { scanned, findings, midSessionConcerns };
+}
+
+// ---------------------------------------------------------------- self-test
+
+function plant(dir: string, reArm: string | null, body: string): void {
+    fs.mkdirSync(path.join(dir, 'src', 'scripts'), { recursive: true });
+    fs.writeFileSync(path.join(dir, 'src', 'scripts', 'fixture_hook.ts'), body);
+    const arm = reArm === null ? '' : `\n    re_arm: ${reArm}`;
+    fs.writeFileSync(
+        path.join(dir, 'src', 'scripts', 'hook_manifest.yaml'),
+        `schema_version: 1\nconcerns:\n  fixture:\n    script: src/scripts/fixture_hook.ts${arm}\nplatforms:\n  claude:\n    post_tool_use:\n      - fixture\n`,
+    );
+}
+
+const WRITES_SURFACE = `import * as fs from 'node:fs';\nfs.writeFileSync('dist/agent-src/rules/x.md', 'hi');\n`;
+const WRITES_ELSEWHERE = `import * as fs from 'node:fs';\nfs.writeFileSync('agents/runtime/state/x.json', 'hi');\n`;
+const DYNAMIC_WITH_SURFACE_LITERAL = `import * as fs from 'node:fs';\nconst root = 'dist/agent-src/rules';\nfs.writeFileSync(target, 'hi');\n`;
+const DYNAMIC_NO_SURFACE = `import * as fs from 'node:fs';\nfs.writeFileSync(target, 'hi');\n`;
+const READS_ONLY = `import * as fs from 'node:fs';\nfs.readFileSync('dist/agent-src/rules/x.md', 'utf-8');\n`;
+
+function selfTestCases(): SelfTestCase[] {
+    const mk = (name: string, expect: 'reject' | 'accept', reArm: string | null, body: string): SelfTestCase => ({
+        name,
+        expect,
+        run: () => {
+            const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'psm-'));
+            try {
+                plant(dir, reArm, body);
+                return runGateCli(REPO_ROOT, 'src/scripts/check_prefix_stable_mutation.ts', ['--root', dir, '--quiet'], REPO_ROOT);
+            } finally {
+                fs.rmSync(dir, { recursive: true, force: true });
+            }
+        },
+    });
+    return [
+        mk('mid-session write to a declared surface, no re-arm → reject', 'reject', null, WRITES_SURFACE),
+        mk('same write with re_arm: pre_compact → accept', 'accept', 'pre_compact', WRITES_SURFACE),
+        mk('same write with re_arm: session_start → accept', 'accept', 'session_start', WRITES_SURFACE),
+        mk('write outside every declared surface → accept', 'accept', null, WRITES_ELSEWHERE),
+        mk('dynamic target + surface literal in file → reject (fail-closed)', 'reject', null, DYNAMIC_WITH_SURFACE_LITERAL),
+        mk('dynamic target, no surface literal → accept (not claimed)', 'accept', null, DYNAMIC_NO_SURFACE),
+        mk('reads a surface but never writes → accept', 'accept', null, READS_ONLY),
+    ];
+}
+
+// ---------------------------------------------------------------------- CLI
+
+export function main(argv: string[] = process.argv.slice(2)): number {
+    if (argv.includes('--self-test')) {
+        return runSelfTest({ gate: 'check_prefix_stable_mutation', cases: selfTestCases(), minCases: 6, minRejectCases: 2 });
+    }
+    const quiet = argv.includes('--quiet');
+    const json = argv.includes('--json');
+    const ri = argv.indexOf('--root');
+    const rootArg = ri !== -1 ? argv[ri + 1] : undefined;
+    const root = rootArg !== undefined ? path.resolve(rootArg) : REPO_ROOT;
+
+    const ledger = new GateLedger('check_prefix_stable_mutation');
+    let verdict: Verdict;
+    try {
+        verdict = evaluate(root, ledger);
+    } catch (err) {
+        process.stderr.write(`❌  check_prefix_stable_mutation: ${(err as Error).message}\n`);
+        return 2;
+    }
+
+    try {
+        reportScanned({
+            gate: 'check_prefix_stable_mutation',
+            scanned: verdict.scanned,
+            units: 'mid-session hook script(s)',
+            roots: ['src/scripts/hook_manifest.yaml', ...prefixStableRoots()],
+        });
+    } catch (err) {
+        if (err instanceof DeadScopeError) {
+            process.stderr.write(`❌  check_prefix_stable_mutation: ${err.message}\n`);
+            return 2;
+        }
+        throw err;
+    }
+
+    ledger.report();
+
+    if (json) {
+        process.stdout.write(JSON.stringify(verdict, null, 2) + '\n');
+        return verdict.findings.length === 0 ? 0 : 1;
+    }
+
+    if (verdict.findings.length === 0) {
+        if (!quiet) {
+            process.stdout.write(
+                `✅  no mid-session mutation of a prefix-stable surface ` +
+                    `(${String(verdict.scanned)} script(s) across ${String(verdict.midSessionConcerns)} mid-session concern(s), ` +
+                    `${String(PREFIX_STABLE_SURFACES.length)} declared surface(s)).\n`,
+            );
+        }
+        return 0;
+    }
+
+    for (const f of verdict.findings) {
+        process.stderr.write(
+            `❌  ${f.concern} (${f.script}) fires on ${f.slots.join(', ')} and ${f.kind === 'literal' ? 'writes' : 'may write'} ` +
+                `inside prefix-stable surface '${f.surface}':\n      ${f.evidence}\n`,
+        );
+    }
+    process.stderr.write(
+        `\n    A mid-session write to a prefix-stable surface invalidates the prompt cache for the\n` +
+            `    rest of the session — the next call pays the cache-WRITE rate over the whole prefix.\n` +
+            `    Either move the write to a re-arm boundary (${RE_ARM_EVENTS.join(' / ')}), or declare\n` +
+            `    're_arm: <event>' on the concern in src/scripts/hook_manifest.yaml naming the boundary\n` +
+            `    at which the rebuilt prefix is expected and paid for once.\n`,
+    );
+    return 1;
+}
+
+if (process.env['GATE_SELF_TEST_CHILD'] !== '1' || process.argv.includes('--root')) {
+    if (process.argv[1] !== undefined && import.meta.url === `file://${process.argv[1]}`) {
+        process.exit(main());
+    }
+}