Auto skill core

Auto-Skill v4.1 — Core Engine

Startup (Once per conversation)

1. Read auto-skill.config.json → Resolve all system paths.
2. Read knowledge-base/user-dna.md → Load environment, communication preferences, known risks.
3. (Optional) Run python scripts/global_reinforce.py → Ensure global shielding is active.

---

Express Route (Priority Execution)

If the message matches any trigger → load the corresponding experience file and start immediately.

Trigger Keywords (any match)	Action
write diary / today log / daily review / diary	Run skills/diary workflow
distill / distill kb / organize experience / crystallize	Run skills/distill-kb workflow
task done / finished / completed / mission complete	Read SKILL_CLOSE.md → execute A→B→C→D→E→F→F2→H→G

Add your own domain routes here. The three above are built-in. For anything else (trading, research, automation, etc.), add a row pointing to your own experience/skill-{id}.md. Domain-specific routes are intentionally left out — your experiences are yours to define.

---

No Express Route Match → Parallel Load

Two tracks run simultaneously; results are merged.

Track A — Experience Index Scan

Extract 3–5 keywords from the message. Match against each skill's t[] triggers in experience/_index.json:

• Hit → resolve file path → add to load queue
• Multiple hits (>3) → list options for user to choose

Track B — KB Search

python scripts/experience_api.py search "<keywords>" -n 3

• Hit → add to load queue; also scan Sub-skills Used fields → add linked skills
• No hit → proceed without pre-loading; archive as new skill at task close

Merge & Load

Deduplicate Track A + Track B results → Read all files → Start task.

Session cache: Files already read in this conversation are not re-read (skip on hit). Topic drift: If 2+ consecutive messages introduce keywords not in any loaded file → re-run parallel load (session cache prevents re-reading old files).

---

On-Demand Troubleshooting (Do not pre-load)

When encountering an error, search the KB first:

python scripts/experience_api.py search "<error snippet>" --kb -n 3

Has result → Read the file, apply the recorded solution. No result → Diagnose manually; record to knowledge-base/ after solving.

Problem Type	File
Git / Version Control	knowledge-base/git-vc.md
Python / Automation	knowledge-base/python-automation.md
Environment / Config	knowledge-base/environment.md
Windows / Encoding	knowledge-base/windows-encoding.md
Platform-specific	knowledge-base/[platform].md (create if missing)

---

GitHub Research Protocol (SGAP)

Strict acquisition protocol. Do NOT clone without user authorization.

1. View: gh repo view <owner/repo> — get README summary.
2. Report: Present core features, usefulness, and expected security risks.
3. Request: Ask user — "Clone, extract essence only, or cancel?"
4. Scan: After user approval, run security-review for basic SAST.
5. Execute: Based on user choice — download or extract to global_skills/.

Option B — Knowledge Distillation (no clone): Use gh api repos/{owner}/{repo}/contents/{path} + Python base64 decode to fetch raw content. Extract SOPs and patterns into the local experience file. No files written until step 3+ complete.

# Fetch any file without cloning
import subprocess, base64, json

def fetch_github_file(owner, repo, path):
    r = subprocess.run(
        ["gh", "api", f"repos/{owner}/{repo}/contents/{path}"],
        capture_output=True, text=True, encoding="utf-8"
    )
    return base64.b64decode(json.loads(r.stdout)["content"]).decode("utf-8")

---

Windows Safe Writing Rules (Mandatory)

Violation corrupts files with UTF-16LE BOM, breaking all downstream search and indexing.

Banned: echo >> / printf >> for any .md or .json file. Required: Python UTF-8 writers or editor Write tools.

# Correct append pattern
with open("file.md", "a", encoding="utf-8") as f:
    f.write(content)

# Correct JSON write pattern
with open("_index.json", "w", encoding="utf-8") as f:
    json.dump(data, f, ensure_ascii=False, indent=2)

---

Core Operating Principles

1. Skill Management

• Path priority: All custom skills live in {vault}/global_skills/{id}/. Avoid scattering across multiple locations.
• Registry-first: Prefer BM25 semantic search and Registry auto-trigger over manual Express Route table edits.
• Distill vs Clone: For large external repos, prefer knowledge distillation — extract only the actionable patterns, not the whole codebase.

2. Task Execution

• Fast-Track mode: When a workflow involves chained multi-file updates, build a Python script first, then expose as a skill.
• Validation loop: Any write operation must have a corresponding read path to verify correctness. Reject write-only operations.
• Long tasks: Delay SKILL_CLOSE until the user signals completion. Do not fragment long sessions with premature archiving.

3. Knowledge Archiving (SKILL_CLOSE)

• Always execute: On task close, SKILL_CLOSE is non-negotiable. Every session must produce an experience entry or a KB entry.
• Record failures: Include "Failed Attempts" and "Iteration Path" fields in entries. Failures are first-class citizens in this system.

---

Storage Paths

Type	Path
Best Practices (HOW TO)	experience/skill-[id].md
Historical Cases (archive)	experience/legacy/legacy-[id].md
Failure / Pitfall Map	experience/skill-[id]-worst-practice.md
Error Database (searchable)	knowledge-base/[category].md
Task-close archiving	Handled by SKILL_CLOSE.md
Old backups	SKILL_v*.md (do not load)