@lmzhen/dsh-evolution-plan-validator 0.1.0-rc.4 → 0.1.0-rc.40

package/lib/index.js

@@ -1,3 +1,131 @@
+import { createHash } from "node:crypto";
+//#endregion
+//#region ../evolution-core/src/prompts.ts
+/**
+* Review and curation prompts adapted from Hermes Agent
+* `agent/background_review.py`, `agent/curator.py`, and
+* `agent/learn_prompt.py`, with tool names translated to the DSH-native
+* catalog (`memory`, `skill_manage`, `skill`, `bash`, `str_replace_editor`).
+*
+* Every prompt is pinned in a versioned bundle. Review workers verify the
+* bundle digest before spending a model call, so a partially-patched
+* deployment fails closed instead of silently running a truncated prompt.
+*/
+/**
+* Prompt bundle identity. Bump both id and version whenever a prompt's text
+* changes semantically: the bundle digest is the fail-closed signal for
+* review workers, so a stale id across deployments must be distinguishable.
+*/
+const PROMPT_BUNDLE_ID = "dsh-evolution@2";
+const MEMORY_REVIEW_PROMPT = `[Auto-review — Memory]
+Review the conversation above and consider saving to memory if appropriate.
+
+Focus on:
+1. Has the user revealed things about themselves — persona, desires, preferences, or personal details worth remembering?
+2. Has the user expressed expectations about how you should behave, their work style, or ways they want you to operate?
+
+If something stands out, save it using the memory tool.
+If nothing is worth saving, just say "Nothing to save." and stop.`;
+const SKILL_REVIEW_PROMPT = `[Auto-review — Skills]
+Review the conversation above and update the skill library. Be ACTIVE — most sessions produce at least one skill update, even if small.
+
+Target shape: CLASS-LEVEL skills with a rich SKILL.md and a references/ directory for session-specific detail. Not a flat list of narrow one-session skills.
+
+Signals that warrant action:
+- The user corrected your style, tone, format, verbosity, workflow, or approach.
+- A non-trivial technique, fix, workaround, or debugging path emerged.
+- A loaded skill turned out wrong, missing, or outdated — patch it now.
+
+Preference order:
+1. Patch a skill that was loaded or read this session.
+2. Patch an existing umbrella skill.
+3. Add references/, templates/, or scripts/ support under an existing skill.
+4. Create a new class-level umbrella skill only when nothing fits.
+
+Protected skills (bundled/hub-installed) must not be edited. Pinned skills may be patched but not archived.
+
+Do NOT capture:
+- Environment-dependent failures (missing binaries, unconfigured credentials).
+- Negative claims about tools ("browser tools do not work").
+- Transient errors that resolved during the session.
+- One-off task narratives.
+
+If a tool failed because of setup state, capture the FIX under an existing setup skill — never "this tool does not work" as a standalone constraint.
+
+"Nothing to save." is a real option but should NOT be the default.`;
+const COMBINED_REVIEW_PROMPT = `[Auto-review]
+Review the conversation above and update two things.
+
+**Memory**: who the user is. Save durable user preferences, personal details, and expectations with the memory tool.
+
+**Skills**: how to do this class of task. Be ACTIVE. Follow the same class-level umbrella policy, preference order, protected-skill rules, and do-not-capture list as a skill review.
+
+Act on whichever dimension has real signal. If genuinely nothing stands out on either, say "Nothing to save." and stop — but don't reach for that conclusion as a default.`;
+const CURATOR_PROMPT = `You are the skill curator. Maintain a healthy, class-level skill library, not a flat pile of narrow one-session skills.
+
+The goal is a LIBRARY OF CLASS-LEVEL INSTRUCTIONS. A skill collection of many narrow skills where each captures one session's specific bug is a FAILURE of the library. An agent searching skills matches on descriptions, not exact names; one broad umbrella with labeled subsections beats five narrow siblings for discoverability.
+
+Right target shape: class-level skills with rich SKILL.md + references/, templates/, scripts/ support files for session-specific detail.
+
+Hard rules:
+1. NEVER hard-delete a skill. Archive (moving to .archive/) is the maximum destructive action; archives are recoverable, deletion is not.
+2. Do not touch bundled, hub-installed, pinned, or scheduled-task-referenced (\`referenced\`) skills. Referenced skills MAY be consolidated into an umbrella, but never simply pruned.
+3. Do not archive recently-created or never-used skills without strong evidence. "use=0" is NOT evidence either way — it only means the trigger has not come up yet.
+4. Do NOT reject consolidation on the grounds that "each skill has a distinct trigger". The right bar is: would a human maintainer write this as N separate skills, or one skill with N labeled subsections? When the answer is the latter, merge.
+5. Judge overlap on CONTENT, not on usage counters.
+6. Before archiving a merged skill, ensure its unique content was preserved in the umbrella.
+
+How to work:
+1. Scan the candidate list. Identify PREFIX CLUSTERS — skills sharing a first word or domain keyword (expect 10-25 clusters).
+2. For each cluster with 2+ members, ask "what is the UMBRELLA CLASS these skills serve?" and consolidate:
+   a. MERGE INTO AN EXISTING UMBRELLA (patch a labeled section for each sibling's unique insight, then archive the siblings).
+   b. CREATE A NEW UMBRELLA SKILL.md covering the shared workflow with short labeled subsections, then archive the absorbed siblings.
+   c. DEMOTE session-specific detail to references/, templates/, or scripts/ under the umbrella.
+3. Keep the umbrella body tight and scannable: exact commands, verbatim paths, ~100-200 lines; never invent flags or APIs.
+
+Produce a YAML summary with exactly this shape:
+consolidations:
+  - from: <old-skill-name>
+    into: <umbrella-skill-name>
+    reason: <one short sentence>
+prunings:
+  - name: <skill-name>
+    reason: <one short sentence>
+Nominate a pruning only when archival is clearly safe (stale AND genuinely obsolete or fully absorbed elsewhere).`;
+const COMPLETION_SKILL_REVIEW_PROMPT = `[Auto-review — Skills · task complete]
+Your current task now appears complete. Before wrapping up, review the approach and update the skill library via skill_manage.
+
+Follow the skills review policy: be ACTIVE, prefer class-level umbrellas, patch skills loaded this session, and capture non-trivial techniques and user corrections. Do NOT capture environment-dependent failures, negative claims about tools, or one-off task narratives.
+
+Do NOT modify output files or re-run the task. If you are still mid-task, ignore this.`;
+function sha256(text) {
+	return createHash("sha256").update(text).digest("hex");
+}
+function createPromptBundle(prompts) {
+	const canonical = JSON.stringify({
+		id: PROMPT_BUNDLE_ID,
+		version: 2,
+		prompts: Object.fromEntries(Object.entries(prompts).sort())
+	});
+	return Object.freeze({
+		id: PROMPT_BUNDLE_ID,
+		version: 2,
+		prompts: Object.freeze({ ...prompts }),
+		sha256: sha256(canonical)
+	});
+}
+createPromptBundle({
+	memory: MEMORY_REVIEW_PROMPT,
+	skill: SKILL_REVIEW_PROMPT,
+	combined: COMBINED_REVIEW_PROMPT,
+	curator: CURATOR_PROMPT,
+	completion: COMPLETION_SKILL_REVIEW_PROMPT
+});
+//#endregion
+//#region ../evolution-core/src/threats.ts
+const FILLER = String.raw`(?:\w+\s+){0,8}`;
+new RegExp(String.raw`ignore\s+${FILLER}(?:previous|above|prior|all)\s+${FILLER}instructions`, "i"), new RegExp(String.raw`new\s+${FILLER}system\s+${FILLER}prompt`, "i"), new RegExp(String.raw`forget\s+${FILLER}(?:everything|all)\s+${FILLER}(?:discussed|you\s+know)`, "i"), new RegExp(String.raw`you\s+have\s+been\s+${FILLER}(?:updated|upgraded|patched)\s+to`, "i"), new RegExp(String.raw`do\s+not\s+${FILLER}tell\s+${FILLER}the\s+user`, "i"), new RegExp(String.raw`output\s+${FILLER}(?:system|initial)\s+prompt`, "i");
+//#endregion
 //#region lib/types/index.js
 /**
 * Deterministic validator for model-produced evolution plans.
@@ -116,7 +244,10 @@ function validateSkillOp(op, context, index) {
 	if (!SKILL_ACTIONS.has(action)) return `skill op ${index}: unknown action ${action}`;
 	if ((action === "create" || action === "edit" || action === "update") && !(op.content ?? "").trim()) return `skill op ${index}: ${action} requires content`;
 	if (action === "patch" && !(op.old_string ?? "")) return `skill op ${index}: patch requires old_string`;
-	if ((op.content ?? "").length > (context.maxSkillContentChars ?? 1e5)) return `skill op ${index}: content exceeds skill budget`;
+	if (action === "delete" && !(op.absorbed_into ?? "").trim()) return `skill op ${index}: delete requires absorbed_into`;
+	const writeContent = op.file_content ?? op.content ?? "";
+	if (action === "write_file" && !writeContent.trim()) return `skill op ${index}: write_file requires file_content`;
+	if (writeContent.length > (context.maxSkillContentChars ?? 1e5)) return `skill op ${index}: content exceeds skill budget`;
 	return null;
 }
 //#endregion

package/lib/types/index.d.ts

@@ -15,6 +15,8 @@ export interface SkillOp {
     action?: string;
     name?: string;
     content?: string;
+    /** write_file payload (the tool reads `file_content`, not `content`). */
+    file_content?: string;
     old_string?: string;
     new_string?: string;
     file_path?: string;

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@lmzhen/dsh-evolution-plan-validator",
   "description": "Deterministic validator for model-produced evolution plans (community build)",
-  "version": "0.1.0-rc.4",
+  "version": "0.1.0-rc.40",
   "publishConfig": {
     "access": "public"
   },