Remix

中文说明

Remix is a universal artifact reconstruction and synthesis tool. It analyzes, compares, and rebuilds artifacts — skills, protocols, modules, features, products, and compound bundles.

Remix is a standalone tool. It works on its own with zero external dependencies beyond Python and jsonschema. When you want self-evolution or governance integration, install the optional skill-se-kit plugin.

Quick Start

pip install .

Analyze sources (scores only, no build):

remix analyze \
  --brief '{"target_profile":"skill","target_job":"evaluate sources"}' \
  --sources '[{"kind":"file","path":"./source.md"}]'

Compare sources (rankings + strategy options):

remix compare \
  --brief '{"target_profile":"skill","target_job":"pick the best source"}' \
  --sources '[{"kind":"file","path":"./a.md"},{"kind":"file","path":"./b.md"}]'

Run the full pipeline (analyze → compare → build → verify):

remix run \
  --brief '{"target_profile":"skill","target_job":"build a code review skill"}' \
  --sources '[{"kind":"file","path":"./my-skill.md"}]'

List available target profiles:

remix profiles

Core Workflow

1. Intake — collect a brief (what you want) and sources (what you have)
2. Normalize — convert diverse source formats into canonical representations
3. Analyze — score each source across configurable dimensions (0–5 scale)
4. Compare — rank sources, apply hard gates, find complementary pairs
5. Synthesize — generate 2–3 strategy options (conservative, balanced, forward-port)
6. Build — materialize the selected output artifact
7. Verify — run profile-specific checks
8. Audit & Handoff — produce provenance trail and release metadata

Target Profiles

Profile	Outputs	Use Case
skill	manifest.json, SKILL.md, tests.md	Agent skills
protocol	schema.json, examples, compatibility matrix	Interop contracts
module	package layout, source, tests	Reusable code packages
feature	spec, rollout plan, acceptance criteria	Product features
product	PRD, roadmap, capability map	Product definitions
compound	Recursive bundle of above	Multi-artifact systems

Configurable Scoring

Scoring dimensions and weights are configurable per-run via the brief:

{
  "target_profile": "skill",
  "target_job": "...",
  "scoring_overrides": {
    "task_fit": { "weight": 1.5 },
    "testability": { "weight": 0.5 },
    "custom_dimension": { "weight": 1.0, "score": 4.2 }
  }
}

Each target profile ships with sensible default weights. Override only what you need.

Extension Points

Extension	Purpose	Required?
Analyzer	Replace heuristic scoring with LLM-backed or custom analysis	No (heuristic default)
Validator	Custom manifest/proposal validation	No (null default)
EvolutionBackend	Self-evolution experience recording	No (null default)

All are Python Protocol classes — implement and inject at runtime:

from remix import RemixRuntime

runtime = RemixRuntime(
    analyzer=my_llm_analyzer,
    validator=my_validator,
    evolution_backend=my_backend,
)

Optional: Skill-SE-Kit Integration

If you want Remix to record experience and evolve over time:

pip install ".[evolution]"

from remix import from_skill_runtime

runtime = from_skill_runtime(
    skill_root="/path/to/skill",
    protocol_root="/path/to/protocol",
)

This is optional. Remix works fully without it.

Repository Layout

remix/
  SKILL.md           — skill description for agent discovery
  manifest.json      — machine-readable skill metadata
  README.md
  README.zh-CN.md
  pyproject.toml
  src/remix/         — core implementation
  tests/             — test suite

Related Projects

• Skill-SE-Kit: optional self-evolution plugin
• Agent Skill Governor: governance layer that can invoke Remix

Installation Notes

• pip >= 22.0 is recommended. Newer pip versions correctly parse pyproject.toml metadata out of the box.
• A setup.cfg file is included for backward compatibility with older pip versions (e.g., pip 21.x shipped with Python 3.9). Older pip may fail to read [project] metadata from pyproject.toml, causing the package to install as UNKNOWN. The setup.cfg file provides the same metadata in the legacy format so that pip install . works correctly regardless of pip version.