clementine-agent 1.18.135 → 1.18.136

package/dist/agent/self-improve.js

@@ -31,7 +31,7 @@ const DEFAULT_CONFIG = {
     // 'prompt-override' writes markdown to ~/.clementine/prompt-overrides/.
     areas: [
         'soul', 'cron', 'workflow', 'memory', 'agent', 'communication', 'goal',
-        'advisor-rule', 'prompt-override',
+        'advisor-rule', 'prompt-override', 'skill',
     ],
     autoApply: true,
     sourceMode: 'skip',
@@ -150,6 +150,13 @@ function classifyRisk(area) {
         case 'workflow': return 'low';
         case 'advisor-rule': return 'low'; // YAML files, hot-reloaded, easily deleted
         case 'prompt-override': return 'low'; // Markdown files, hot-reloaded, easily deleted
+        // 1.18.136 — 'skill' is the new first-class self-improve target.
+        // Skills are user-facing recipes that fire across many tasks; a
+        // bad change can cascade into every cron that pins them. Medium
+        // tier so changes always surface for owner approval, never auto-
+        // apply. Frontmatter + Anthropic spec are validated before any
+        // proposal can even reach the queue (see validateProposal).
+        case 'skill': return 'medium';
         case 'soul': return 'medium'; // Core personality — needs approval
         case 'communication': return 'medium'; // Global operating instructions
         case 'memory': return 'medium'; // Memory config
@@ -1140,7 +1147,11 @@ export class SelfImproveLoop {
             `User rules at priority 100+ override engine builtins of the same id.\n` +
             `- For "prompt-override": target = "global", "agent:<slug>", or "job:<jobName>" (e.g. "job:daily-summary"). ` +
             `Use when a job/agent needs more standing guidance — markdown that gets prepended to its prompt. ` +
-            `The proposedChange is the markdown body (optionally with gray-matter frontmatter for priority/position).\n\n` +
+            `The proposedChange is the markdown body (optionally with gray-matter frontmatter for priority/position).\n` +
+            `- For "skill": target = the exact skill slug (matches Anthropic regex ^[a-z0-9][a-z0-9-]{0,63}$, e.g. "morning-briefing"). ` +
+            `Use when a skill is producing low-quality output, has a high failure rate, OR is stale (>90 days unused but still pinned to a cron). ` +
+            `Surface candidates: skills with low clementine.useCount + recent run failures, OR skills whose linked cron jobs appear in the failing-jobs list above. ` +
+            `The proposedChange must be the FULL SKILL.md (frontmatter + body): preserve the "name" field exactly, keep "description" ≤1024 chars and free of XML tags, body changes only — Clementine will preserve the clementine.useCount/createdAt/version block on save so do NOT include or alter that block.\n\n` +
             `Return your answer as a JSON object matching the schema: { "results": [ ... ] }. Up to 3 items. If absolutely nothing actionable today, return { "results": [] }.`;
         const analysisResult = await this.assistant.runPlanStep('si-analyze', analysisPrompt, {
             tier: 2,
@@ -1200,7 +1211,11 @@ export class SelfImproveLoop {
             `- Generate a SPECIFIC, MINIMAL change (not a full rewrite)\n` +
             `- Explain WHY this change should improve the metric\n` +
             `- IMPORTANT: "proposedChange" must be the COMPLETE updated file content (not just the diff), because it will replace the entire file\n` +
-            `- For source code changes: preserve all imports, exports, and function signatures. Only modify implementation details.\n\n` +
+            `- For source code changes: preserve all imports, exports, and function signatures. Only modify implementation details.\n` +
+            (selected.area === 'skill'
+                ? `- For skill changes: keep the frontmatter "name" and "description" intact (description ≤1024 chars, no XML tags). Modify the BODY (procedure section after the frontmatter). Do NOT include or alter the "clementine:" frontmatter block — Clementine preserves useCount/createdAt/version automatically on save.\n`
+                : '') +
+            `\n` +
             `Output ONLY a JSON object (no markdown, no explanation):\n` +
             `{ "area": "${selected.area}", "target": "${selected.target}", "hypothesis": "what will improve and why", "proposedChange": "the complete updated file content with your minimal change applied" }`;
         const result = await this.assistant.runPlanStep('si-hypothesize', proposalPrompt, {
@@ -1217,6 +1232,20 @@ export class SelfImproveLoop {
                 return existsSync(SOUL_FILE) ? readFileSync(SOUL_FILE, 'utf-8') : '';
             case 'cron':
                 return existsSync(CRON_FILE) ? readFileSync(CRON_FILE, 'utf-8') : '';
+            case 'skill': {
+                // 1.18.136 — read SKILL.md from the folder-form path. Fall back
+                // to the flat-form (legacy) path if no folder. The hypothesizer
+                // proposes against the BODY only — frontmatter is preserved on
+                // write by validateProposal + writeSkill (see resolveTargetPath
+                // / applyApprovedChange below).
+                const folderEntry = path.join(VAULT_DIR, '00-System', 'skills', target, 'SKILL.md');
+                const flatEntry = path.join(VAULT_DIR, '00-System', 'skills', target + '.md');
+                if (existsSync(folderEntry))
+                    return readFileSync(folderEntry, 'utf-8');
+                if (existsSync(flatEntry))
+                    return readFileSync(flatEntry, 'utf-8');
+                return '';
+            }
             case 'workflow': {
                 const wfFile = path.join(WORKFLOWS_DIR, target.endsWith('.md') ? target : `${target}.md`);
                 return existsSync(wfFile) ? readFileSync(wfFile, 'utf-8') : '';
@@ -1389,6 +1418,70 @@ export class SelfImproveLoop {
                 return `Cannot apply change — identity drift too high (similarity: ${drift.similarity.toFixed(3)}, threshold: ${DRIFT_SIMILARITY_THRESHOLD})`;
             }
         }
+        // Skill area: merge proposed frontmatter with existing clementine.*
+        // lifecycle block so an LLM-authored proposal doesn't clobber
+        // useCount / createdAt / version. The proposal supplies the new
+        // name/description/body; the lifecycle metadata stays Clementine's
+        // responsibility. This sits between validation and the generic
+        // writeFileSync below so every other area still writes raw.
+        if (pending.area === 'skill') {
+            try {
+                const proposed = matter(pending.proposedChange);
+                const existing = pending.before ? matter(pending.before) : null;
+                const proposedFm = (proposed.data ?? {});
+                const existingFm = (existing?.data ?? {});
+                const existingExt = (existingFm.clementine ?? {});
+                const proposedExt = (proposedFm.clementine ?? {});
+                const now = new Date().toISOString();
+                const mergedExt = {
+                    ...existingExt,
+                    // Allow proposal to update tools.allow / triggers / source if it
+                    // explicitly sets them, but preserve lifecycle counters from
+                    // the existing record.
+                    ...proposedExt,
+                    useCount: existingExt.useCount ?? 0,
+                    createdAt: existingExt.createdAt ?? now,
+                    updatedAt: now,
+                    version: typeof existingExt.version === 'number' ? existingExt.version + 1 : 1,
+                    lastSelfImproveExperimentId: experimentId,
+                };
+                const mergedFm = {
+                    name: proposedFm.name ?? existingFm.name,
+                    description: proposedFm.description ?? existingFm.description,
+                    ...(proposedFm.title || existingFm.title ? { title: proposedFm.title ?? existingFm.title } : {}),
+                    clementine: mergedExt,
+                };
+                const body = proposed.content.endsWith('\n') ? proposed.content : proposed.content + '\n';
+                const merged = matter.stringify(body, mergedFm);
+                mkdirSync(path.dirname(targetPath), { recursive: true });
+                writeFileSync(targetPath, merged);
+                // Record version + finish the standard apply flow below.
+                this.recordVersion(experimentId, pending.area, pending.target, pending.hypothesis, pending.before);
+                logger.info({ id: experimentId, area: pending.area, target: pending.target }, 'Applied approved skill change (lifecycle preserved)');
+                this.updateExperimentStatus(experimentId, 'approved');
+                try {
+                    unlinkSync(pendingFile);
+                }
+                catch { /* ignore */ }
+                const stateAfter = this.loadState();
+                stateAfter.pendingApprovals = Math.max(0, stateAfter.pendingApprovals - 1);
+                this.saveState(stateAfter);
+                try {
+                    appendFileSync(IMPACT_CHECKS_FILE, JSON.stringify({
+                        experimentId,
+                        area: pending.area,
+                        target: pending.target,
+                        appliedAt: new Date().toISOString(),
+                        checkAfterMs: 24 * 60 * 60 * 1000,
+                    }) + '\n');
+                }
+                catch { /* ignore impact-check schedule errors */ }
+                return `Applied skill change to ${path.relative(VAULT_DIR, targetPath)} (useCount preserved, version bumped)`;
+            }
+            catch (err) {
+                return `Cannot apply skill change — frontmatter merge failed: ${err instanceof Error ? err.message : String(err)}`;
+            }
+        }
         // Write the change (non-source areas)
         mkdirSync(path.dirname(targetPath), { recursive: true });
         writeFileSync(targetPath, pending.proposedChange);
@@ -2009,6 +2102,21 @@ export class SelfImproveLoop {
                 return SOUL_FILE;
             case 'cron':
                 return CRON_FILE;
+            case 'skill': {
+                // 1.18.136 — skill changes route through writeSkill (not raw
+                // writeFileSync) so frontmatter validation runs and the existing
+                // useCount/lastUsed/createdAt are preserved. The intercept lives
+                // in applyApprovedChange below; this still returns the canonical
+                // path so the version recorder + impact-check writer have
+                // something to anchor on.
+                const folderEntry = path.join(VAULT_DIR, '00-System', 'skills', target, 'SKILL.md');
+                const flatEntry = path.join(VAULT_DIR, '00-System', 'skills', target + '.md');
+                if (existsSync(folderEntry))
+                    return folderEntry;
+                if (existsSync(flatEntry))
+                    return flatEntry;
+                return null;
+            }
             case 'workflow': {
                 const name = target.endsWith('.md') ? target : `${target}.md`;
                 return path.join(WORKFLOWS_DIR, name);
@@ -2117,6 +2225,52 @@ export function validateProposal(area, target, proposedChange) {
         // a misbehaving LLM proposal doesn't even get cached.
         return { valid: false, error: 'source area is deprecated; propose advisor-rule or prompt-override instead' };
     }
+    if (area === 'skill') {
+        // 1.18.136 — skill body validation. The proposedChange is the FULL
+        // SKILL.md (frontmatter + body). Anthropic spec rules: name must
+        // match ^[a-z0-9][a-z0-9-]{0,63}$, description ≤1024 chars, body
+        // present + non-empty, no XML tags in description, no Anthropic-
+        // reserved words in name. Reuses the centralized validator that
+        // dashboard + MCP + auto-extract all share.
+        if (!/^[a-z0-9][a-z0-9-]{0,63}$/.test(target)) {
+            return { valid: false, error: `skill target must be a valid slug (got "${target}")` };
+        }
+        let parsed;
+        try {
+            parsed = matter(proposedChange);
+        }
+        catch (err) {
+            return { valid: false, error: `skill SKILL.md frontmatter parse error: ${err instanceof Error ? err.message : String(err)}` };
+        }
+        const fm = parsed.data;
+        const name = typeof fm.name === 'string' ? fm.name : '';
+        const description = typeof fm.description === 'string' ? fm.description : '';
+        const body = parsed.content || '';
+        if (!name || !/^[a-z0-9][a-z0-9-]{0,63}$/.test(name)) {
+            return { valid: false, error: 'skill frontmatter "name" missing or invalid slug' };
+        }
+        if (name !== target) {
+            return { valid: false, error: `skill frontmatter "name" (${name}) must match target slug (${target})` };
+        }
+        if (/\b(anthropic|claude)\b/i.test(name)) {
+            return { valid: false, error: 'skill name uses a reserved word (anthropic/claude)' };
+        }
+        if (!description.trim()) {
+            return { valid: false, error: 'skill frontmatter "description" required' };
+        }
+        if (description.length > 1024) {
+            return { valid: false, error: `skill description ${description.length} chars exceeds Anthropic spec ceiling of 1024` };
+        }
+        if (/<\w+/.test(description)) {
+            return { valid: false, error: 'skill description must not contain XML tags' };
+        }
+        if (!body.trim()) {
+            return { valid: false, error: 'skill body (procedure) required' };
+        }
+        // Body line count is a SOFT limit (warning, not error) — keeps
+        // validation strict but doesn't block proposals that legitimately
+        // need more lines.
+    }
     if (area === 'advisor-rule') {
         // Must parse as YAML and have schemaVersion: 1, id matching target, when[], then[].
         if (!/^[a-z0-9-]+$/.test(target)) {

package/dist/types.d.ts

@@ -957,7 +957,7 @@ export interface SelfImproveExperiment {
     startedAt: string;
     finishedAt: string;
     durationMs: number;
-    area: 'soul' | 'cron' | 'workflow' | 'memory' | 'agent' | 'source' | 'communication' | 'goal' | 'advisor-rule' | 'prompt-override';
+    area: 'soul' | 'cron' | 'workflow' | 'memory' | 'agent' | 'source' | 'communication' | 'goal' | 'advisor-rule' | 'prompt-override' | 'skill';
     target: string;
     hypothesis: string;
     proposedChange: string;
@@ -1014,7 +1014,7 @@ export interface SelfImproveConfig {
      */
     surfaceThreshold?: number;
     plateauLimit: number;
-    areas: ('soul' | 'cron' | 'workflow' | 'memory' | 'agent' | 'source' | 'communication' | 'goal' | 'advisor-rule' | 'prompt-override')[];
+    areas: ('soul' | 'cron' | 'workflow' | 'memory' | 'agent' | 'source' | 'communication' | 'goal' | 'advisor-rule' | 'prompt-override' | 'skill')[];
     /** Enable tiered auto-apply: low-risk changes apply without approval. Default: false. */
     autoApply?: boolean;
     /** Target a specific agent slug (for per-agent improvement cycles). */

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "clementine-agent",
-  "version": "1.18.135",
+  "version": "1.18.136",
   "description": "Clementine — Personal AI Assistant (TypeScript)",
   "type": "module",
   "main": "dist/index.js",