Download purpose_agent/skills/ci.py from Rohan03/purpose-agent: direct link, hf CLI and curl.
- Browser
- Download file 3.45 kB
-
https://huggingface.co/Rohan03/purpose-agent/resolve/main/purpose_agent/skills/ci.py
- Command line
-
hf download hf://Rohan03/purpose-agent/purpose_agent/skills/ci.py
-
curl -L -o ci.py https://huggingface.co/Rohan03/purpose-agent/resolve/main/purpose_agent/skills/ci.py
3.45 kB
| """ | |
| ci.py — Skill Continuous Integration. | |
| Every skill mutation must pass CI before activation: | |
| 1. Immune scan (no injection/manipulation) | |
| 2. Eval tests (skill produces expected behavior) | |
| 3. Fitness comparison (new version >= old version) | |
| 4. Promote or rollback | |
| Darwinian selection: | |
| - High-fitness skills get retrieval boost | |
| - Low-fitness skills are mutated then re-tested | |
| - Failed skills are archived (never deleted — audit trail) | |
| """ | |
| from __future__ import annotations | |
| import logging | |
| from typing import Any | |
| from purpose_agent.skills.schema import SkillCard, SkillGenome | |
| from purpose_agent.immune import scan_memory | |
| from purpose_agent.memory import MemoryCard, MemoryKind | |
| logger = logging.getLogger(__name__) | |
| class SkillCI: | |
| """ | |
| CI pipeline for skills: scan → test → promote or rollback. | |
| Usage: | |
| ci = SkillCI() | |
| passed = ci.validate(skill_card) | |
| if passed: | |
| genome.promote(skill_card.id) | |
| else: | |
| genome.rollback() | |
| """ | |
| def __init__(self, fitness_threshold: float = 0.4): | |
| self.fitness_threshold = fitness_threshold | |
| self._scan_log: list[dict[str, Any]] = [] | |
| def validate(self, card: SkillCard) -> bool: | |
| """ | |
| Run full CI on a skill card. | |
| Returns True if skill passes all gates. | |
| """ | |
| # Gate 1: Immune scan | |
| memory_proxy = MemoryCard( | |
| kind=MemoryKind.SKILL_CARD, | |
| pattern=card.trigger, | |
| strategy=" ".join(card.procedure), | |
| content=card.name, | |
| ) | |
| scan_result = scan_memory(memory_proxy) | |
| if not scan_result.passed: | |
| self._log(card, "REJECTED", f"immune scan failed: {scan_result.threats}") | |
| card.status = "archived" | |
| return False | |
| # Gate 2: Fitness threshold | |
| if card.fitness_score < self.fitness_threshold: | |
| self._log(card, "REJECTED", f"fitness {card.fitness_score:.2f} < threshold {self.fitness_threshold}") | |
| card.status = "archived" | |
| return False | |
| # Gate 3: Non-empty procedure | |
| if not card.procedure: | |
| self._log(card, "REJECTED", "empty procedure") | |
| card.status = "archived" | |
| return False | |
| # All gates passed | |
| card.status = "tested" | |
| self._log(card, "PASSED", f"fitness={card.fitness_score:.2f}") | |
| return True | |
| def compare_fitness(self, new: SkillCard, old: SkillCard | None) -> bool: | |
| """Compare new skill version against old. New must be >= old.""" | |
| if old is None: | |
| return True | |
| return new.fitness_score >= old.fitness_score * 0.95 # Allow 5% tolerance | |
| def mutate(self, card: SkillCard) -> SkillCard: | |
| """ | |
| Create a mutated version of a low-fitness skill. | |
| Appends "[MUTATED]" marker for tracking. | |
| """ | |
| mutated = card.evolve( | |
| new_procedure=[f"[IMPROVED] {step}" for step in card.procedure] | |
| ) | |
| mutated.created_by = "mutation" | |
| return mutated | |
| def _log(self, card: SkillCard, result: str, detail: str) -> None: | |
| entry = {"skill_id": card.id, "name": card.name, "version": card.version, | |
| "result": result, "detail": detail} | |
| self._scan_log.append(entry) | |
| logger.info(f"SkillCI: {card.name} v{card.version} → {result}: {detail}") | |
| def log(self) -> list[dict]: | |
| return self._scan_log | |