quale 0.7.0.dev0

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

quale — structural codebase analysis

Reads code as text, no ASTs, no config. Answers structural questions about any codebase.

Useless on day zero. After ~1 commit, signal builds.

What it does

Every model we tested guesses the wrong test file on a plain prompt. With quale, every model picks the right one on the first try. 900+ trials, 12 repos, 7 model families.

The blind spot is structural, not semantic: no model knows your directory layout because that's not in training data.

Quickstart

quale edit-context --files src/spool.ts --task "change upload" --format tool
quale verify-packet --files src/spool.ts --task "change upload" --format tool
quale guard --task "change upload"

For an agent

Put this in AGENTS.md or an instruction file:

Before editing, run:
quale edit-context --files $FILE --task "$TASK" --format tool

The tool emits structured JSON. The agent reads it, picks the right test file, and stays in scope. Measured: 75% verify hit, zero sprawl (vs 10-20% verify, 0.4-0.65 sprawl without).

For a developer

quale edit-context --files src/spool.ts --task "..."
quale edit-context --diff HEAD~1
quale ci-report origin/main HEAD --summary
quale inspect .

For CI

quale ci-report origin/main HEAD --fail-on-blast-tier high
quale ci-report origin/main HEAD --fail-on-mirror-gap 0.70

Exit codes: 0 = pass, 1 = error, 2 = blast gate, 3 = stable anchor gate.

Cross-model verification

Model	Baseline	With quale
Qwen 235B	wrong path	correct
Gemma 4 31B	wrong dir	correct
Nemotron 30B	wrong + 277 tok	right + 112 tok
Mistral 24B	wrong dir	correct
Claude Opus 4	src/spool.test.ts	tests/spool.test.ts
Gemma 4B (local CPU)	blank JSON	correct

Every model makes the same src/foo.test.ts error. quale corrects it.

What it is not

Linter, coverage tool, dead-code detector, security policy. All output is report-only. Verification accuracy peaks around 80% -- when the candidate set lacks the right test file, quale says so rather than guessing.

How it works

Three primitives, same pipeline for every language:

1. Segmenter: split file content on delimiters
2. Vocabulary: collect unique phrases per file
3. Index: measure co-occurrence, overlap, and history

Deterministic. Same input, same output, every time.

Further reading

• docs/COMMANDS.md -- full command reference
• docs/EFFECT_HARNESS.md -- harness methodology and results