finereli

code-aware

Give your coding agent architectural memory.

code-aware processes git history to build a mental model of your codebase — not just what exists, but why it was built that way and how it evolved. The output is a set of developer-notes-style markdown files that any coding agent can read before making changes.

Works with Claude Code, Codex, Copilot, Cursor, and any tool that supports the Agent Skills standard.

Install

Claude Code (marketplace)

In Claude Code, run:

/plugin marketplace add finereli/code-aware
/plugin install code-aware@finereli

That's it. The skill is now available across all your projects.

Claude Code (local)

For development or if you prefer a local install:

git clone https://github.com/finereli/code-aware.git ~/code-aware
cd ~/code-aware && npm install
claude --plugin-dir ~/code-aware

Other agents (Codex, Copilot, Cursor)

git clone https://github.com/finereli/code-aware.git ~/code-aware
cd ~/code-aware && npm install
ln -s ~/code-aware/skills/code-aware ~/.claude/skills/code-aware

Your agent will discover code-aware automatically and use it when relevant.

First Scan

Ask your agent to scan the codebase — code-aware automatically uses whatever LLM is available in your environment (Haiku on Claude Code, GPT-4.1-mini on Codex).

Or run manually:

cd /path/to/your/project
~/code-aware/skills/code-aware/scripts/scan.sh

The default quick scan processes the last 50 commits (~2 minutes). For full history:

~/code-aware/skills/code-aware/scripts/scan.sh . --full

Output lands in .code-aware/:

• index.md — list of all models with descriptions
• models/*.md — one file per architectural component
• INSIGHTS.md — six opinionated observations about the codebase

What It Produces

Models are adaptive — no hardcoded categories. The system samples observations from your actual codebase and discovers the right boundaries. A Flutter app gets different models than a Python CLI or a Cloudflare Workers API.

Each model file reads like a senior developer's personal notes:

# Contact Management

> CRUD operations, referral tracking, follow-up scheduling, and archiving

## Follow-up Scheduling
- Follow-ups include both `followup_date` and `followup_type` fields.
  Both must remain synchronized — partial updates caused 500 errors previously.
- UI refactored from two-step modal to single dialog...

## Important Invariants
- Entries with `removed_at` set are excluded from active listings and statistics.
- Only entries with neither timestamp set are considered "active and pending."
...

Real file names, real function names, actual gotchas and invariants. Not a database dump — useful context.

The Feedback Loop

code-aware creates a virtuous cycle: the more context your agent has about the codebase, the better its commit messages become. Better commit messages feed richer information back into the models. The system improves its own input quality over time.

Keeping Models Current

After committing new work, sync to process only new commits:

~/code-aware/skills/code-aware/scripts/sync.sh

Check staleness without updating:

~/code-aware/skills/code-aware/scripts/status.sh

Or just ask your agent — the skill knows when to check and update.

Configuration

code-aware auto-detects your LLM provider from environment variables:

Provider	Key	Default model	Default insights model
OpenAI	OPENAI_API_KEY	gpt-4.1-mini	gpt-4.1
Anthropic	ANTHROPIC_API_KEY	claude-haiku-4	claude-sonnet-4

Override with:

Variable	Description
CODE_AWARE_MODEL	Override the model for observation extraction and synthesis
CODE_AWARE_INSIGHTS_MODEL	Override the model for codebase insights

How It Works

1. Load — reads git log with diffs
2. Observe — LLM extracts architectural observations from commit batches (focusing on why and how, not just what)
3. Discover — samples observations and proposes natural model boundaries
4. Assign — each observation gets assigned to a model
5. Summarize — pyramidal compression: tier-0 summaries per model, then higher tiers compress 10:1
6. Consolidate — reviews all models, merges overlapping ones, renames mismatches
7. Synthesize — generates narrative developer notes per model with temporal bucketing (Recent, This Week, This Month, etc.)
8. Insights — six opinionated questions answered from the full model set
9. Export — writes markdown files to .code-aware/

License

MIT