antoine

Codebase lookup and indexing for agent skills.

The name encodes the goal: antoine = "N to 1" (an-to-one). Collapse the N ad-hoc tool calls an agent would otherwise make against a codebase (ls + cat + grep + git log + git show + …) into one call to a purpose-built kata verb (or its antoine alias — see pyproject.toml for the console-script wiring; kata-cli is the PyPI distribution name, not a command name) that returns the same information as structured data. Every verb antoine ships is a bet that some recurring N-call pattern has a 1-call replacement that is cheaper, more reliable, and easier to delegate to a subagent.

What's here

This repo manages three intertwined things — the CLI that ships the 1-call verbs, the evaluation harness that tells us whether the verbs are actually worth the bet, and the recorded results from past rounds.

• antoine/ — the package that will eventually expose the lookup verbs. Greenfield: learn / explain / whoami are honest placeholder stubs. See CLAUDE.md for build / test / architecture details.

• kata-cli — alt-published PyPI distribution carrying the same wheel content as antoine-cli. Installing either exposes the same pair of console scripts — antoine and kata (see pyproject.toml). kata-cli is the distribution label users pip install, not a command they run. The dual-publish loop is defined in .github/workflows/; see CHANGELOG.md entries for v0.7.0 / v0.7.1 for the history of how the distribution names were wired up. (A third name, code-lens-cli, was published from this repo through v0.9.2; from v0.10.0 onward it lives in its own repo — see "Results of this loop" below.)

• experiments/scripts_eval/ — the A/B-test harness for the repo-map skill (env-var-gated hooks, three-layer scoring, 5-repo round-1 corpus, multi-arm rider design with banned / directed / organic modes, multi-pair LLM-as-judge). The eval rounds are the validation gate before the learn / explain / overview / doctor verb redesign lands. See experiments/scripts_eval/README.md and experiments/scripts_eval/RUNBOOK.md.

• docs/eval-rounds/ — write-ups from completed evaluation rounds (round 01, smoke 02, round 02 so far). These are the empirical record behind the directives in CLAUDE.md — including the round-2 finding that subagents build their plans from the prompt body before consulting the skills catalog, which is why the dispatching table lives in the parent agent's instructions rather than in skill descriptions.

Results of this loop

antoine ships the tool. The first published catalog of results from running its capture/reduce/assess loop is code-lens-cli — four 1-call verbs (classify / recent / grep / profile) that antoine maintainers identified as recurring N-call patterns and packaged into a sibling distribution. Install with uv tool install code-lens-cli. Most agents will install both.

The migration history (antoine 0.10.0 → code-lens-cli 0.10.0, 2026-05-17) is the first concrete proof that the loop produces shippable artifacts. Future cells (2–8) of the loop are designed to make catalogs like this one routine rather than artisanal.