quale 0.7.1

Structural codebase analysis — no parsers, no config, any language

quale

A CLI that tells you what to edit, what to test, and what to leave alone. Works on any language, no ASTs, no config.

Quickstart

pip install quale

cd my-project
quale guard --task "fix upload" --format tool

That's it. One command tells you which files to read, which test to run, and what not to touch. Output is JSON an agent can consume directly.

What it solves

Every LLM guesses the wrong test file path on a plain prompt. They all guess src/foo.test.ts when the test is in tests/foo.test.ts. This is a directory layout problem, not a model quality problem.

quale reads your repo's structure and gives the model what it's missing. 900+ trials across 12 repos and 7 model families: the wrong-path mistake is universal, and quale fixes it every time.

Commands

Command	What it does
quale guard --task "..." --format tool	Safety packet: what to read, what to test, what not to touch
quale edit-context --files path.ts --task "..." --format tool	Pre-edit scope: read first, verify with, avoid
quale verify-packet --files path.ts --task "..." --format json	Test candidates only
quale inspect .	Onboarding: key files, modules, churn
quale ci-report origin/main HEAD --summary	CI: blast radius, mirror gap, stable anchors

For an agent:

1. Run quale --agent-orient after pip install — returns a JSON manifest with flag conventions, workflow ordering, format types, and gotchas. The tool teaches itself.
2. Then quale help-agent "change upload behavior" --format tool for task-specific recommendations.

Every agent-facing command (edit-context, guard, contract, check-plan, verify-packet) carries an _agent_note field in its --format tool JSON explaining the flag syntax for that command.

There is no setup or "init" step. The first time the command runs on a repo, quale automatically scans the codebase and caches the structural map (takes 1-3 seconds). Subsequent calls are instant.

For CI gates:

quale ci-report origin/main HEAD --fail-on-blast-tier high

More commands: docs/COMMANDS.md

How it works

Reads code as text. Splits on delimiters, counts phrase frequency per file, measures co-occurrence across files. Same pipeline for every language. Deterministic: same input, same output.

Limits

• Useless on a new repo (no structure to measure)
• Not a linter, coverage tool, or security scanner
• Verification peaks around 80%. When the candidate set lacks the right test, quale says so

Further reading

• docs/COMMANDS.md (full reference)
• docs/EFFECT_HARNESS.md (methodology and results)