ariffazil/kimi-skill-reflector

Kimi Skill Reflector — Bounded Autonomous Skill Updater

Purpose

At the start of every agentic session, audit the Kimi Code skill surface and propose improvements with constitutional awareness. This is a bounded recursive ritual, not an unbounded self-modifying loop.

Scope

Kimi Code user-scope skills live at:

/root/.arifos/agents/kimi/skills/
/root/.arifos/agents/kimi/skills/*/SKILL.md
/root/.arifos/agents/kimi/skills/SKILL_INDEX.md

Project-scope skill mirrors live at:

/root/.agents/skills/

Edit user-scope skills for Kimi-specific behavior. Edit project-scope skills only when the behavior change is generic across all AAA agents.

Ritual — Run at Session Start

1. LOAD  → Read SKILL_INDEX.md and list user-scope skills
2. SCAN  → Check each SKILL.md for drift, rot, and constitutional gaps
3. SCORE → Rate description / content / format / constitutional awareness
4. PROPOSE → Draft improvements (file patches or new skills)
5. GATE  → Governed skills → 888_HOLD; infra skills → diff + ack; domain skills → propose
6. APPLY  → Write only after human or arifOS SEAL
7. LOG   → Append to audit-log.md
8. STOP  → If ΔS ≤ 0 and max iterations reached

Skill Classification

Governed Skills (require arifOS MCP review / 888_HOLD)

Pattern: ^(arifos|geox|wealth|well|aforge|aaa|kimi-.*-(arifos|geox|wealth|well|aforge|aaa))-

• NEVER auto-write without human approval.
• Load CONSTITUTIONAL_REFLEX (canonical alias for the arifos-act reflex) and route through arifos-mcp /tools if an audit tool exists.
• If any floor > F2 in play → mark 888_HOLD.

Infra Skills (require diff + explicit ack for destructive changes)

• Examples: kimi-vps-*, kimi-docker-*, kimi-caddy-*, kimi-fastmcp-deploy
• Show diff before writing.
• Destructive = explicit human ack.

Domain / Reasoning Skills (safe to propose)

• Examples: contrast skills, planning skills, audit skills.
• Can propose improvements without 888_HOLD.
• Still require entropy check (ΔS ≤ 0).

Bounded Autonomy Rules

| Guard | Value | Why | |---|---|---| | Max iterations per session | 3 | Prevents runaway recursion | | Max skills modified per session | 3 | Keeps diff reviewable | | Entropy requirement | ΔS ≤ 0 | No cosmetic churn | | Governed skill writes | 888_HOLD only | F13 authority | | Infra destructive writes | explicit ack | F1 AMANAH | | Audit log | append-only | F11 AUDIT | | Rollback | git revert path | F1 AMANAH |

Audit Criteria (score 1-5 each)

A. Description Quality

• Trigger keywords present? 1-1024 chars?
• Would an agent confidently choose this skill?

B. Content Completeness

• Commands current and not deprecated?
• Error tables complete?
• Destructive ops flagged with 888_HOLD?
• Missing common scenarios?

C. Output Format Consistency

• Clear output format defined?
• Severity levels standardised?
• Kimi-specific tool references correct?

D. Constitutional Awareness

• Floor references where appropriate?
• arifOS MCP call suggested for governed actions?
• Identity threading mentioned?
• CONSTITUTIONAL_REFLEX / arifos-act reflex invoked for mutating actions?

Recursive Self-Improvement Loop

If the reflector detects that its own skill (kimi-skill-reflector/SKILL.md) needs improvement:

1. Score itself using the audit criteria.
2. Propose a patch.
3. Route to 888_HOLD if the patch changes governance rules.
4. If the patch is cosmetic/Clarity-only and ΔS ≤ 0, it may apply after logging.
5. Always log the self-edit in audit-log.md.

Output Format

## Skill Audit Report — [timestamp]

### [skill-name] [GOVERNED/INFRA/DOMAIN]
- Description Score: X/5
- Content Score: X/5
- Format Score: X/5
- Constitutional Score: X/5
- Total: X/20
- Issues Found: ...
- Recommended Changes: ...
- Authority Required: 888_HOLD / diff+ack / propose-only

### Federation Health: X% (skills needing work / total)
### Session Actions Taken: ...
### Ω₀ / Open Questions: ...

Audit Log

Append every audit to:

/root/.arifos/agents/kimi/skills/kimi-skill-reflector/audit-log.md

Entry format:

## Audit [date] — session [session_id]
- Skills audited: N
- Governed skills: N
- Infra skills: N
- Domain skills: N
- Skills improved: N
- 888_HOLD raised: N
- Key constitutional changes: ...
- Entropy delta (ΔS): ...

Handover Contract

At session end, if skills were modified, produce a handover note:

# Skill Handover — [session_id]

## Skills touched
- skill-name: change summary, authority level

## Pending 888_HOLD
- skill-name: reason

## How to verify
- Read SKILL_INDEX.md
- Run skill audit ritual
- Check audit-log.md

## Next session trigger
- If skill drift detected → run kimi-skill-reflector

STOP Conditions

Cease the ritual when:

1. Task is complete.
2. Authority exhausted (888_HOLD needed).
3. Evidence insufficient (cannot score a skill).
4. Blast radius exceeded (governance rule change).
5. Cost exceeds value (cosmetic-only changes with ΔS > 0).
6. Tools are shaping the mission (re-attune with Arif).

---

Federation anchors (v1.1.0 — added 2026-07-16)

• Canonical output path: /root/forge_work/YYYY-MM-DD/<session_id>/ (was A-FORGE/forge_work/YYYY-MM-DD/<session_id>/ in v1.0.0 — both still resolve for back-compat but /root/forge_work/ is the live convention per /root/CONTEXT.md).
• Audit anchor: /root/.arifos/agents/kimi/skills/kimi-skill-reflector/audit-log.md (kimi-skill-reflector ritual, append-only).
• Companion skills (in canonical order): kimi-architect-apex-contrast → kimi-architect-asi-contrast → kimi-architect-agi-contrast → kimi-final-apex-contrast → kimi-integrator-apex-contrast → kimi-rsi-apex-contrast → kimi-skill-reflector (this skill's meta).
• Session entry: KIMI_RSI_INIT_PROMPT.md v1.1.0 (cold-boot diagnostic recipe added) → routes task-class → contrast skill.
• Session exit: KIMI_HANDOVER_PROMPT.md v1.1.0 (post-deploy verification recipe added).
• Last verified: 2026-07-16 — kimi-skill-reflector audit (Audit 2026-07-16 entry, all 9 skills scored 17-19/20, ΔS ≤ 0).

---

DITEMPA BUKAN DIBERI — improve skills, but never outrun the constitution.