feat(ms-ai-architect): Sesjon 20 — B3 create_skill (dvelende, konstruktiv op) [skip-docs]
B3 + Spor B (krav 5+6) KOMPLETT. create_skill er speilbildet av de destruktive ops: merge/saner/retire beviser "ingen kuratert verdi TAPT"; create beviser "ingen sub-bar skill FØDT" (kvalitetsgulv). Ingen ny domene-skill shippet (svar 3) — mekanismen bevist, ikke brukt. - planCreateSkill (ren): genererer scaffold + selv-validerer mot eval.mjs sine EGNE checkers (K2/K3/K5/K6/refTall + K10) — rubrikken har én kilde. Guardrail ok iff alle deterministiske K passerer + K10<terskel + navn ledig. - TEST_DOMAIN_SPEC: kanonisk dvelende-scaffold (born compliant, K5 ratio 1.0). - runCreatePlan (uren): gated dry-run, 0 skriving til skills/. - applyApprovedAction create-gren: re-plan fersk → revalider (drift+kollisjon) → skriv scaffold + persister taksonomi-tillegg → flipp ledger. - applyCategoryAdditions (taxonomy.mjs): additiv, klobrer aldri eksisterende eier. - CLI create-verb i plan-skill-op + apply-skill-op. TDD: ny test-skill-ops-create.test.mjs (16). kb-eval 84→100. Suiter uendret: validate 239 · kb-update 139 · kb-integrity 192/192. Deterministisk eval K1–K10 for de 5 ekte: 0 regresjon. 0 ekte skills/-mutasjon. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
parent
de3a9a483b
commit
e47fc9bd59
5 changed files with 758 additions and 11 deletions
|
|
@ -12,7 +12,7 @@
|
|||
// preserved (ref-file count-invariant + set-equality) and refuse (ok=false)
|
||||
// when two reference files would clobber on the same target path.
|
||||
|
||||
import { readdirSync, readFileSync, existsSync, statSync, mkdirSync, renameSync, rmSync } from 'node:fs';
|
||||
import { readdirSync, readFileSync, writeFileSync, existsSync, statSync, mkdirSync, renameSync, rmSync } from 'node:fs';
|
||||
import { join, relative, basename, dirname, sep } from 'node:path';
|
||||
import { isDeepStrictEqual } from 'node:util';
|
||||
import {
|
||||
|
|
@ -22,11 +22,30 @@ import {
|
|||
actionKey,
|
||||
setActionStatus,
|
||||
} from '../../kb-update/lib/decisions-io.mjs';
|
||||
import { loadTaxonomy, saveTaxonomy, applyCategoryReassignments } from '../../kb-update/lib/taxonomy.mjs';
|
||||
import {
|
||||
loadTaxonomy,
|
||||
saveTaxonomy,
|
||||
applyCategoryReassignments,
|
||||
applyCategoryAdditions,
|
||||
} from '../../kb-update/lib/taxonomy.mjs';
|
||||
import {
|
||||
splitFrontmatter,
|
||||
extractDescription,
|
||||
checkK2,
|
||||
checkK3,
|
||||
checkK5K6,
|
||||
checkRefConsistency,
|
||||
} from '../eval.mjs';
|
||||
import {
|
||||
computeOverlapFromInputs,
|
||||
perSkillSiblingOverlap,
|
||||
K10_OVERLAP_THRESHOLD,
|
||||
} from './sibling-overlap.mjs';
|
||||
|
||||
export const MERGE_OPERATION = 'merge_skills';
|
||||
export const SANITIZE_OPERATION = 'sanitize_skill';
|
||||
export const RETIRE_OPERATION = 'retire_skill';
|
||||
export const CREATE_OPERATION = 'create_skill';
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// Pure core — planMergeSkills
|
||||
|
|
@ -275,6 +294,192 @@ export function planRetireSkill(skill, ctx = {}) {
|
|||
return { entry, diff, guardrail };
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// Pure core — planCreateSkill (CONSTRUCTIVE: born compliant or refused)
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
/**
|
||||
* Canonical "dvelende" scaffold spec. create_skill ships no new domain skill
|
||||
* (operator answer 3 — the mechanism is proven, not exercised on production):
|
||||
* this throwaway test-domain spec is what both CLIs and the tests scaffold to
|
||||
* demonstrate a skill that passes the deterministic rubric from birth.
|
||||
*/
|
||||
export const TEST_DOMAIN_SPEC = {
|
||||
description:
|
||||
'Use this skill when validating the create_skill scaffolding mechanism against a throwaway test domain. ' +
|
||||
'Triggers on: "create-skill self-test", "scaffold validation", "dvelende create probe".',
|
||||
categories: [
|
||||
{
|
||||
category: 'core',
|
||||
files: [
|
||||
{ file: 'overview.md', title: 'Test-domene oversikt', source: 'internal://create-skill-self-test' },
|
||||
{ file: 'mechanism.md', title: 'Scaffold-mekanikk', source: 'internal://create-skill-self-test' },
|
||||
],
|
||||
},
|
||||
{
|
||||
category: 'patterns',
|
||||
files: [
|
||||
{ file: 'routing.md', title: 'Progressiv disclosure-mønster', source: 'internal://create-skill-self-test' },
|
||||
],
|
||||
},
|
||||
],
|
||||
};
|
||||
|
||||
/** Title-case a kebab/snake skill name for the SKILL.md heading. */
|
||||
function titleCase(name) {
|
||||
return name
|
||||
.split(/[-_]/)
|
||||
.filter(Boolean)
|
||||
.map((w) => w[0].toUpperCase() + w.slice(1))
|
||||
.join(' ');
|
||||
}
|
||||
|
||||
/**
|
||||
* Generate a SKILL.md that satisfies eval.mjs's deterministic rubric: a
|
||||
* use-when/quoted description (K2), a short body (K3), a routing table that
|
||||
* cites per-folder counts (refCountConsistency) AND names every reference file
|
||||
* (K5 namedRatio + K6 named start files).
|
||||
*/
|
||||
function buildSkillMd(name, description, categories) {
|
||||
const tableRows = categories
|
||||
.map((c) => `| \`references/${c.category}/\` | ${c.files.length} |`)
|
||||
.join('\n');
|
||||
const startHere = categories
|
||||
.flatMap((c) => c.files.map((f) => `- \`references/${c.category}/${f.file}\` — ${f.title ?? f.file}`))
|
||||
.join('\n');
|
||||
return (
|
||||
`---\n` +
|
||||
`name: ${name}\n` +
|
||||
`description: ${description}\n` +
|
||||
`---\n\n` +
|
||||
`> **INSTRUKSJON:** Test-domene-skill generert av create_skill-mekanismen (dvelende). ` +
|
||||
`Bruk kunnskapsbasen i \`references/\` for detaljert veiledning.\n\n` +
|
||||
`# ${titleCase(name)}\n\n` +
|
||||
`Scaffold generert av create_skill — født compliant mot K1–K10 (deterministisk delmengde). ` +
|
||||
`Dette er et test-domene; ingen ny domene-skill shippes.\n\n` +
|
||||
`## Referansekart\n\n` +
|
||||
`| Område | Antall filer |\n` +
|
||||
`|--------|--------------|\n` +
|
||||
`${tableRows}\n\n` +
|
||||
`### Start her\n` +
|
||||
`${startHere}\n`
|
||||
);
|
||||
}
|
||||
|
||||
/** Minimal curated-shaped reference body (Source-header hygiene; content is not
|
||||
* part of the deterministic gate — eval reads SKILL.md, not ref bodies). */
|
||||
function buildRefBody(f) {
|
||||
return (
|
||||
`# ${f.title ?? f.file}\n\n` +
|
||||
`**Source:** ${f.source ?? 'internal://create-skill-self-test'}\n\n` +
|
||||
`Generert referansefil for create_skill test-domene (dvelende). ` +
|
||||
`Erstattes av kuratert innhold dersom en ekte domene-skill noen gang opprettes.\n`
|
||||
);
|
||||
}
|
||||
|
||||
/**
|
||||
* Plan a new-skill scaffold. Pure; mutates nothing; reads no disk. The
|
||||
* CONSTRUCTIVE guardrail is the mirror of the destructive ones: merge/sanitize/
|
||||
* retire prove "no curated value LOST" (preservation); create proves "no
|
||||
* sub-bar skill BORN" (quality floor). It refuses (ok=false) when the generated
|
||||
* scaffold would fail any deterministic criterion eval.mjs enforces (K2/K3/K5/
|
||||
* K6/refCountConsistency), when the new description overlaps a sibling at/above
|
||||
* the K10 threshold, or when the name already belongs to a skill. The verdict is
|
||||
* re-derived from eval.mjs's OWN checkers — the rubric has a single source.
|
||||
*
|
||||
* @param {string} name new skill directory name (kebab-case)
|
||||
* @param {{ description?: string,
|
||||
* categories?: Array<{ category: string, files: Array<{file: string, title?: string, source?: string}> }>,
|
||||
* existingSkills?: string[], existingDescriptions?: Record<string,string>,
|
||||
* promptSet?: object, threshold?: number, decided_at?: string|null, note?: string }} [ctx]
|
||||
* @returns {{ entry: object, scaffold: object, guardrail: object }}
|
||||
*/
|
||||
export function planCreateSkill(name, ctx = {}) {
|
||||
if (!name) throw new Error('planCreateSkill: name is required');
|
||||
const description = ctx.description ?? '';
|
||||
const categories = ctx.categories ?? [];
|
||||
const existingSkills = ctx.existingSkills ?? [];
|
||||
const existingDescriptions = ctx.existingDescriptions ?? {};
|
||||
const promptSet = ctx.promptSet ?? {};
|
||||
const threshold = ctx.threshold ?? K10_OVERLAP_THRESHOLD;
|
||||
|
||||
// Build the scaffold artifacts.
|
||||
const skillMdContent = buildSkillMd(name, description, categories);
|
||||
const files = categories.flatMap((c) =>
|
||||
c.files.map((f) => ({
|
||||
path: `skills/${name}/references/${c.category}/${f.file}`,
|
||||
content: buildRefBody(f),
|
||||
})),
|
||||
);
|
||||
const scaffold = {
|
||||
skillMd: { path: `skills/${name}/SKILL.md`, content: skillMdContent },
|
||||
files,
|
||||
taxonomyAdditions: categories.map((c) => ({ category: c.category, skill: name })),
|
||||
};
|
||||
|
||||
// CONSTRUCTIVE guardrail — re-run eval.mjs's deterministic checkers against the
|
||||
// generated SKILL.md (single source of truth for the rubric).
|
||||
const { frontmatter, body } = splitFrontmatter(skillMdContent);
|
||||
const parsedDescription = extractDescription(frontmatter);
|
||||
const perFolder = {};
|
||||
for (const c of categories) perFolder[c.category] = c.files.length;
|
||||
const totalRefFiles = files.length;
|
||||
|
||||
const k2 = checkK2(parsedDescription);
|
||||
const k3 = checkK3(body);
|
||||
const { K5: k5, K6: k6 } = checkK5K6(body, totalRefFiles);
|
||||
const refConsistency = checkRefConsistency(body, perFolder);
|
||||
|
||||
// K10 — sibling-scope-non-overlap of the NEW description vs the existing skills
|
||||
// (cross-skill; same overlap core as eval's attachSiblingOverlap).
|
||||
const descriptionsBySkill = { ...existingDescriptions, [name]: parsedDescription };
|
||||
const overlap = computeOverlapFromInputs(descriptionsBySkill, promptSet);
|
||||
const k10map = perSkillSiblingOverlap(overlap.pairs, { threshold });
|
||||
const k10 = k10map[name] ?? { maxCombined: 0, worstSibling: null, pass: true, threshold };
|
||||
|
||||
const nameAvailable = !existingSkills.includes(name);
|
||||
|
||||
const failed = [];
|
||||
if (!nameAvailable) failed.push('name');
|
||||
if (!k2.pass) failed.push('K2');
|
||||
if (!k3.pass) failed.push('K3');
|
||||
if (!k5.pass) failed.push('K5');
|
||||
if (!k6.pass) failed.push('K6');
|
||||
if (!refConsistency.consistent) failed.push('refConsistency');
|
||||
if (!k10.pass) failed.push('K10');
|
||||
|
||||
const guardrail = {
|
||||
method:
|
||||
'constructive (mirror of the destructive guardrails): no sub-bar skill is born. The scaffold must pass ' +
|
||||
"eval.mjs's deterministic rubric from birth (K2/K3/K5/K6/refCountConsistency), must not overlap a sibling " +
|
||||
'at/above the K10 threshold, and the name must be free. ok iff failed === [].',
|
||||
nameAvailable,
|
||||
k2,
|
||||
k3,
|
||||
k5,
|
||||
k6,
|
||||
refConsistency,
|
||||
k10,
|
||||
failed,
|
||||
ok: failed.length === 0,
|
||||
};
|
||||
|
||||
const entry = {
|
||||
operation_type: CREATE_OPERATION,
|
||||
status: 'pending',
|
||||
decided_at: ctx.decided_at ?? null, // never Date.now() in a pure lib
|
||||
targets: { name },
|
||||
guardrail,
|
||||
note:
|
||||
ctx.note ??
|
||||
`Dry-run: create ${name} — scaffold ${
|
||||
guardrail.ok ? 'passes K1–K10 (deterministic) — born compliant' : 'FAILS rubric (' + failed.join(',') + ')'
|
||||
}. Applies nothing.`,
|
||||
};
|
||||
|
||||
return { entry, scaffold, guardrail };
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// Impure shell — read-only disk load + gated ledger write
|
||||
// ---------------------------------------------------------------------------
|
||||
|
|
@ -398,6 +603,35 @@ export function runRetirePlan(skill, opts) {
|
|||
return result;
|
||||
}
|
||||
|
||||
/**
|
||||
* Drive a create dry-run: read existing skill NAMES read-only (name-collision
|
||||
* check), plan the scaffold, and (only with write:true) record the PENDING entry
|
||||
* in decisions.json. Writes NOTHING under skills/ — the constructive apply is a
|
||||
* separate approved step. The scaffold content comes from the caller (the CLI's
|
||||
* TEST_DOMAIN_SPEC); K10 siblings come from the caller's existing descriptions.
|
||||
* @param {string} name
|
||||
* @param {{ skillsDir: string, dataDir: string, write?: boolean, decided_at?: string|null,
|
||||
* description?: string, categories?: object[],
|
||||
* existingDescriptions?: Record<string,string>, promptSet?: object, threshold?: number }} opts
|
||||
* @returns {{ entry: object, scaffold: object, guardrail: object }}
|
||||
*/
|
||||
export function runCreatePlan(name, opts) {
|
||||
const {
|
||||
skillsDir, dataDir, write = false, decided_at = null,
|
||||
description, categories, existingDescriptions, promptSet, threshold,
|
||||
} = opts;
|
||||
const existingSkills = existsSync(skillsDir)
|
||||
? readdirSync(skillsDir, { withFileTypes: true }).filter((e) => e.isDirectory()).map((e) => e.name)
|
||||
: [];
|
||||
const result = planCreateSkill(name, {
|
||||
description, categories, existingSkills, existingDescriptions, promptSet, threshold, decided_at,
|
||||
});
|
||||
if (write) {
|
||||
saveDecisions(recordAction(loadDecisions(dataDir), result.entry), dataDir);
|
||||
}
|
||||
return result;
|
||||
}
|
||||
|
||||
// ===========================================================================
|
||||
// Apply-path (Sesjon 18–19) — the DESTRUCTIVE frontier. An operator-approved
|
||||
// ledger entry is executed into a real skills/ mutation. Single-skill ops
|
||||
|
|
@ -469,6 +703,7 @@ export function applyApprovedAction(approvedEntry, opts) {
|
|||
const {
|
||||
pluginRoot, skillsDir, agentsDir, dataDir, taxonomyDataDir,
|
||||
apply = false, decided_at = null, categorySkill,
|
||||
createSpec, existingDescriptions, promptSet,
|
||||
} = opts;
|
||||
const op = approvedEntry?.operation_type;
|
||||
const target = approvedEntry?.targets ?? {};
|
||||
|
|
@ -481,8 +716,17 @@ export function applyApprovedAction(approvedEntry, opts) {
|
|||
fresh = runRetirePlan(target.skill, { skillsDir, dataDir, write: false, categorySkill });
|
||||
} else if (op === MERGE_OPERATION) {
|
||||
fresh = runMergePlan(target.absorber, target.absorbed, { skillsDir, dataDir, write: false, categorySkill });
|
||||
} else if (op === CREATE_OPERATION) {
|
||||
// create has no prior on-disk state to re-derive: the scaffold spec is
|
||||
// re-supplied by the caller; the disk-dependent part is name-availability.
|
||||
const spec = createSpec ?? {};
|
||||
fresh = runCreatePlan(target.name, {
|
||||
skillsDir, dataDir, write: false, decided_at: null,
|
||||
description: spec.description, categories: spec.categories,
|
||||
existingDescriptions, promptSet,
|
||||
});
|
||||
} else {
|
||||
throw new Error(`applyApprovedAction: unsupported operation_type "${op}" — applies sanitize/retire/merge only (create_skill staged later)`);
|
||||
throw new Error(`applyApprovedAction: unsupported operation_type "${op}" — applies sanitize/retire/merge/create only`);
|
||||
}
|
||||
|
||||
// 2. Idempotent revalidation against fresh disk — refuse on any failure.
|
||||
|
|
@ -491,6 +735,38 @@ export function applyApprovedAction(approvedEntry, opts) {
|
|||
return { applied: false, preview: !apply, revalidate, plan: fresh };
|
||||
}
|
||||
|
||||
// 3a-create — CONSTRUCTIVE: write the born-compliant scaffold + persist
|
||||
// taxonomy ownership additively. No archival (nothing pre-existed); the
|
||||
// guardrail's nameAvailable made real is that skills/<name> is created fresh.
|
||||
if (op === CREATE_OPERATION) {
|
||||
const allFiles = [fresh.scaffold.skillMd, ...fresh.scaffold.files];
|
||||
for (const f of allFiles) {
|
||||
const abs = join(pluginRoot, f.path);
|
||||
mkdirSync(dirname(abs), { recursive: true });
|
||||
writeFileSync(abs, f.content);
|
||||
}
|
||||
// Persist taxonomy ownership for the new categories (additive — never
|
||||
// clobbers a sibling's owned category; that would be a merge, not a create).
|
||||
const taxDir = taxonomyDataDir ?? dataDir;
|
||||
const additions = fresh.scaffold.taxonomyAdditions;
|
||||
if (additions.length) {
|
||||
saveTaxonomy(applyCategoryAdditions(loadTaxonomy(taxDir), additions), taxDir);
|
||||
}
|
||||
// Flip the ledger entry approved -> applied (gated write; audit record kept).
|
||||
saveDecisions(setActionStatus(loadDecisions(dataDir), actionKey(approvedEntry), 'applied', decided_at), dataDir);
|
||||
return {
|
||||
applied: true,
|
||||
revalidate,
|
||||
report: {
|
||||
op, target,
|
||||
filesWritten: allFiles.length,
|
||||
taxonomyAdditions: additions.length,
|
||||
createdDir: `skills/${target.name}`,
|
||||
ledgerStatus: 'applied',
|
||||
},
|
||||
};
|
||||
}
|
||||
|
||||
// 3a. merge — curated RELOCATION (not archival) + taxonomy persist + shell archive.
|
||||
if (op === MERGE_OPERATION) {
|
||||
const { absorber, absorbed } = target;
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue