invocator 0.1.0

Extract synthesized engineering knowledge from a GitHub repo into a learnings/ folder.

invocator

Summon the accumulated wisdom of a GitHub repository into a learnings/ grimoire. ✨

invocator is a CLI that points at any GitHub repository, walks its PRs, issues, commits, and code-review conversations, and synthesizes that history into a folder of categorized markdown files — rules, prevenções, patterns, decisions, glossary — so the tacit knowledge living in a project's timeline becomes navigable documentation.

The three spells

Spell	What it does
invocator forge key	Bind a pact with Anthropic — store and validate your API key
invocator scry cost --repo owner/name	Gaze ahead: cost / wall time / item counts before the ritual
invocator extract wisdom --repo owner/name	The full ritual — produces ./learnings/

Prerequisites

• Python 3.11+
• gh CLI installed and authenticated (gh auth login) — invocator never calls api.github.com directly
• An Anthropic API key

Install

pipx install invocator        # recommended
# or
uv tool install invocator
# or, from source
git clone https://github.com/Gab-r-x/invocator.git && cd invocator && pip install -e .

Quickstart

invocator forge key
invocator scry cost --repo OleveCo/CourseGPTBackend
invocator extract wisdom --repo OleveCo/CourseGPTBackend

After the ritual completes, ./learnings/ will contain six markdown files synthesized from the repo's history.

Output structure

File	Contents
rules.md	Conventions and imperatives ("always X", "never Y") mined from review comments
prevencoes.md	Bug patterns, regressions, and pitfalls the team has already paid for
patterns.md	Refactors and architectural patterns that recur in the codebase
decisions.md	ADR-style decisions extracted from PRs and issues
glossary.md	Domain terms and named concepts that recur ≥3 times
INDEX.md	Run metadata: repo, timestamp, model, cost, per-file bullet counts

How it works

1. summon — gh api --paginate pulls PRs, issues, commits, and review comments into a JSONL cache with per-resource watermarks (re-runs are incremental).
2. transmute — heuristics (conventional commits, review-cue regexes, labels, ADR detection, glossary mining) classify each item into one of five categories. SHA1 dedupe + top-K cap per category.
3. inscribe — one Anthropic call per category, with explicit prompt caching, writes learnings/<category>.md. A SHA256 corpus hash means re-runs without deltas are free (no LLM call). INDEX.md is rendered deterministically.

Cost

Roughly $1–3 USD for medium-sized repos with claude-sonnet-4-6. Always preview with invocator scry cost --repo X first. Re-runs without changes are free thanks to the hash cache.

Troubleshooting

Symptom	Fix
gh: command not found	Install gh: https://cli.github.com/
GH_NOT_AUTHENTICATED	Run gh auth login
NO_API_KEY	Run invocator forge key
CORPUS_TOO_LARGE	Cap inputs: --top-k 100
Want to re-fetch from scratch	invocator extract wisdom --repo X --force-refetch

Contributing

This project follows a strict phased-todo workflow. Before contributing, read:

• CLAUDE.md — workflow and project rules
• docs/STANDARDS.md — the non-negotiable engineering rules
• docs/todos/README.md — operational contract for the executor → progress → tests pipeline

License

MIT — see LICENSE.