repopilot-skillforge

RepoPilot Skillforge is a local CLI that scans a repository and writes practical AI-agent onboarding files: generated agent guidance, review and release skills, detected command maps, and a stale-file doctor.

Problem

AI coding agents move faster when a repository tells them how it is built, tested, released, and reviewed. Most repositories keep that knowledge scattered across manifests, scripts, CI files, and human memory. RepoPilot Skillforge turns the first scan of a repo into a repeatable set of files an agent can actually use.

Why now

Agent workflows are becoming part of everyday development, but each repo still needs local context. A lightweight generator gives teams a clear starting point without requiring a hosted service, paid API, or manual archaeology before every task.

Quickstart

uv sync --all-extras
uv run repopilot scan tests/fixtures/python
uv run repopilot generate tests/fixtures/python
uv run repopilot doctor tests/fixtures/python

To try the reproducible demo:

bash scripts/demo.sh

The demo writes generated output under demo-output/python and creates media/demo.png.

Screenshots

What it generates

• AGENTS.generated.md: repository guidance for coding agents
• skills/review/SKILL.md: a generated review workflow
• skills/release/SKILL.md: a generated release workflow
• commands.md: detected setup, test, check, build, and release commands
• analysis.json: structured analysis with ecosystem, test frameworks, entrypoints, risky files, confidence, warnings, and source fingerprint

Architecture

The code is intentionally small:

• src/repopilot/analysis.py detects manifests, commands, test frameworks, entrypoints, risky files, and source fingerprints.
• src/repopilot/models.py defines Pydantic models for stable JSON output.
• src/repopilot/generator.py renders package templates into generated repository files.
• src/repopilot/cli.py exposes scan, generate, and doctor through Typer.
• tests/fixtures and tests/golden keep ecosystem detection reproducible.

Supported ecosystems

The MVP supports Python (pyproject.toml), Node (package.json), and Rust (Cargo.toml). Unknown repositories still receive safe fallback output with low confidence instead of invented commands.

Development

bash scripts/bootstrap.sh
bash scripts/check.sh
bash scripts/test.sh
bash scripts/build.sh
bash scripts/demo.sh
bash scripts/build_release.sh

Roadmap

• Add Go, Java, and .NET manifest detection.
• Read CI workflow files as command evidence.
• Generate richer skill variants for security, dependency upgrades, and release notes.
• Add confidence explanations for each inferred command.

FAQ

Does this call any paid API?

No. The CLI analyzes local files only.

Does it modify application source code?

No. generate writes the generated guidance files listed above and creates parent directories as needed.

How does doctor decide files are stale?

It stores a source fingerprint in analysis.json and recomputes that fingerprint from manifests, lockfiles, and source files on later runs.

What happens when detection confidence is low?

RepoPilot writes conservative output, leaves command fields empty when needed, and records warnings in analysis.json.