CAM Codebase Enhancer

Use this skill when you need an AI to actually improve and verify a codebase — not just talk about it. CAM is the only Claw variant with a validation-first architecture that rejects hallucinated success.

When to Use This Skill

• Legacy codebase modernization (e.g., Python 2 to 3.12+, jQuery to React)
• Monorepo refactoring across multiple packages
• Test suite generation and coverage improvement
• Dependency upgrades with breaking change resolution
• Research repository to production conversion
• Cross-repo pattern mining and knowledge transfer
• Any task where you need proof the agent's changes actually work

Core Capabilities

CAM operates as an 8-step workflow, each step producing verifiable artifacts:

1. Inspect — cam evaluate <repo> analyzes a repository and produces structured findings
2. Mine — cam mine <dir> extracts reusable patterns from external repos into persistent knowledge (claw.db). Achieves 100% repair rate on malformed LLM JSON.
3. Ideate — cam ideate <dir> generates novel app or improvement concepts from stored knowledge + repo inputs
4. Spec — cam create <repo> generates an explicit, reviewable creation spec before execution
5. Execute — cam create <repo> --execute writes real code changes, verified by workspace diff (not agent narration). Includes inner correction loop (up to 3 retries with workspace restore + violation feedback).
6. Validate — cam validate --spec-file <spec> checks the result against the saved spec, acceptance rules, and structured metric expectations (coverage, test count)
7. Benchmark — cam forge-benchmark measures output quality with a deterministic harness
8. Self-Enhance — cam self-enhance start periodically improves CAM's own source code via a 7-gate validation pipeline.

Invocation Examples

Evaluate a repo before touching it

cam evaluate /path/to/repo --mode quick
cam doctor audit --limit 10

Mine patterns from external repos

cam mine /path/to/source-repos --target /path/to/target --max-repos 5 --depth 2

Create and execute improvements

cam create /path/to/target \
  --repo-mode fixed \
  --request "Modernize auth module to use JWT, add tests" \
  --check "pytest -q" \
  --execute \
  --namespace-safe-retry

Validate the result

cam validate --spec-file data/create_specs/<spec>.json

Export knowledge for standalone consumption

cam forge-export --out knowledge.jsonl --max-methodologies 200

CAM-PULSE: Autonomous X-powered discovery

# Check configuration and API key
cam pulse preflight

# One-shot scan: search X for new GitHub repos, filter for novelty, assimilate
cam pulse scan --keywords "AI agent framework,new repo github.com" --from-date 2026-03-21

# Dry run (scan + filter only, no cloning/mining)
cam pulse scan --dry-run

# Start perpetual polling daemon
cam pulse daemon --interval 15

# View discovery stats
cam pulse status

# List recent discoveries
cam pulse discoveries --limit 20

# Scan history
cam pulse scans

# Daily report
cam pulse report --date 2026-03-21

HuggingFace model repo mining

# Mine a HuggingFace model repo (tiered: micro/standard/large)
cam pulse ingest https://huggingface.co/microsoft/phi-3-mini-4k-instruct

Repo freshness monitoring

# Check all tracked repos for staleness
cam pulse freshness --verbose

# Re-mine repos with significant changes
cam pulse refresh --all

# Re-mine a specific repo
cam pulse refresh https://github.com/bytedance/deer-flow

Guided interactive mode (no flags to memorize)

cam chat

Safety Notes

• Validation-first: CAM checks actual workspace diffs — if no files changed, execution is marked as failed
• Namespace guards: In fixed mode, CAM rejects changes that introduce new top-level source namespaces
• Preflight gates: Risky or ambiguous tasks trigger automatic preflight with must-clarify questions
• Budget caps: Per-agent and per-day cost limits enforced via claw.toml
• Honest failure: CAM documents its limits explicitly and does not overstate benchmark results

Prerequisites

• Python 3.12+
• git
• API keys for configured LLM providers (set via cam setup or environment variables)
• Install: pip install -e ".[dev]"

References

• README — Full project overview, proven capabilities, honest limits
• Command Guide — Complete CLI reference
• Operator Cheatsheet — Quick reference card
• Proven Capabilities — Evidence-backed claims
• Project Charter — Anti-drift expectations