Skip to main content

A rule-based command-output optimization and context-compression layer that reduces unnecessary LLM context while preserving important information.

Project description

Quor

A rule-based command-output optimization and context-compression layer that reduces unnecessary LLM context while preserving important information.

Status: v0.1.0 released — available on PyPI. All 10 implementation phases complete (605 tests passing, ruff + mypy clean, verified on Python 3.11 and 3.14 across multiple machines).


What is Quor?

AI coding assistants spend a large share of their context window on raw command output — passing test runs, unchanged git status lines, repeated warnings, verbose build logs. That's prompt size the assistant pays for on every turn, and it crowds out the signal that actually matters: the failure, the diff, the one line that changed.

Quor sits between your shell and the assistant's context window. It runs your command exactly as it would run anyway, then applies a deterministic filtering pipeline that removes low-signal output while preserving everything that indicates success, failure, or change. The result is a smaller, higher-signal prompt — not a different one.

git status          →  quor git status
                           ↓
              [deterministic filtering pipeline]
                           ↓
        optimized output → AI assistant context

Key properties:

  • Deterministic preprocessing. Same input always produces the same output. No LLM calls, no ML models, no non-determinism in the filtering path itself.
  • Token reduction and context optimization. Removes low-signal output (unchanged lines, repeated noise, verbose logs) so more of the assistant's context budget is spent on content that matters.
  • Fail-open. Every layer degrades to the original, unfiltered output on error rather than risking data loss. A broken filter, a plugin crash, or a timeout never blocks your command or hides information — it just means nothing gets removed that turn.
  • Transparent, rule-based processing. Every filtering decision is traceable: quor explain "pytest tests/" shows exactly what each stage removed and why. Filters are plain TOML you can read, edit, and version-control.
  • Plugin extensible. Third-party stages and lifecycle plugins register via standard Python entry-points — no core changes required to add support for a new tool or add custom telemetry/policy logic.
  • Non-compiler architecture. Quor doesn't parse, execute, or semantically understand the command it runs — it runs the real tool, captures the real output, and applies text-level rules. There's no static analysis and nothing that changes what a command actually does or returns.
  • Windows-native, pip-installable. pip install quor — no Rust toolchain, no Homebrew, no compilation step.

Quor does not change what a command does, what it's allowed to access, or what it returns to you on your own terminal — it only changes what gets forwarded into the assistant's context window, deterministically and transparently.


How Quor Works

  1. The command executes normally. Quor runs the real command (git status, pytest, etc.) exactly as it would run without Quor — same process, same exit code, same side effects.
  2. The output is captured. Quor reads the command's stdout before it reaches the AI assistant.
  3. A deterministic filtering pipeline runs. Rule-based stages (configurable, auditable, no ML) mark each line as keep, remove, or protect.
  4. Important information is preserved. Failures, diffs, errors, and anything matching a "protect" rule are never removed, no matter what else a filter is configured to strip.
  5. The optimized output is sent to the AI assistant. A smaller, higher-signal version of the same output — never a summarized or reworded one.

Performance & Token Reduction

How Quor reduces context

Quor's pipeline annotates every output line with one of three decisions — KEEP, COMPRESS, or PROTECT — and the final render drops every COMPRESS line. Reduction comes entirely from removing lines that carry no new information for the assistant, never from summarizing, rewriting, or truncating the meaning of the lines that remain.

Algorithms responsible for compression:

  • remove_ansi — strips lines that are pure terminal escape/color codes once stripped of formatting.
  • strip_lines — removes lines matching configured noise patterns (e.g. PASSED lines, dot-progress output), while preserve_patterns marks matching lines PROTECT unconditionally.
  • deduplicate_consecutive — collapses runs of identical adjacent lines to the first occurrence.
  • group_repeated — collapses N repetitions of a matched pattern into the first instance plus a (×N) count.
  • max_tokens — truncates beyond a configured budget using a head, tail, or both strategy, only after the above stages have already removed redundant content.

Why reduction varies by content

Compression ratio depends entirely on how repetitive or verbose the actual output is — there is no fixed "Quor compression rate." A clean pytest run with 500 passing tests and one failure compresses aggressively (500 near-identical PASSED lines carry no new information). A git diff against a single changed file compresses very little, because nearly every line is already signal.

Compresses well: long runs of passing test output, unchanged git status entries, repeated build warnings, verbose dependency-resolution logs, ANSI-heavy terminal formatting.

Intentionally preserved, never compressed: failures, tracebacks, AssertionError and similar exception text, diff hunks, anything matching a filter's preserve_patterns, and any line already marked PROTECT by an earlier stage — no later stage can downgrade a PROTECT decision.

Why determinism matters here

Because every stage is a rule (pattern match, dedup, count, or budget) and never a model call, the same input always produces the same output, and quor explain <command> can show the exact stage-by-stage reasoning behind every decision. This is what makes the behavior auditable and predictable in a way a summarization-based approach cannot be: nothing is ever paraphrased, and nothing is removed without a rule you can inspect (and override) in a TOML file.

Benchmark framework

Token savings should be measured as:

reduction % = 1 - (output_tokens / input_tokens)

using Quor's own char / 4 token estimator (quor/tracking/db.py::count_tokens, documented as a ±20% approximation — the same figure quor gain reports), against a representative sample of real command output for each scenario, not synthetic or cherry-picked examples. A credible benchmark run should capture, per scenario: the command, raw output token count, filtered output token count, which filter/stages fired, and whether the output was a passing run, a failing run, or a mixed run — since pass/fail mix is the single biggest driver of reduction percentage for test-runner output.

Scenario Input Tokens Output Tokens Reduction % Notes
pytest — large suite, all passing To be measured To be measured To be measured
pytest — large suite, one failure To be measured To be measured To be measured
git status — many unchanged files To be measured To be measured To be measured
git diff — single small file changed To be measured To be measured To be measured
ruff check — clean To be measured To be measured To be measured
ruff check — several violations To be measured To be measured To be measured

No measurements have been recorded yet. quor gain will report real, per-project figures once Quor is in use; this table will be populated from that data (or a dedicated benchmark script) rather than estimated.


Installation

pip install quor
quor init --claude

That's it — quor and qr are both installed as commands, and quor init --claude wires up the Claude Code hook. Requires Python 3.11+.


Quick Start

# 1. Install
pip install quor

# 2. Wire up the Claude Code hook
quor init --claude

# 3. Confirm everything is healthy
quor doctor

# 4. See what Quor would do to a real command, without running anything for real
quor explain "pytest tests/"

# 5. After using Claude Code for a while, check your savings
quor gain

Expected output for step 3 (quor doctor) on a healthy install:

✓ Python ≥ 3.11 — 3.11
✓ Dependency 'typer'
✓ Dependency 'pydantic'
✓ Dependency 'orjson'
✓ Dependency 'platformdirs'
✓ Dependency 'regex'
✓ Dependency 'rich'
✓ Hook script installed — C:\Users\<you>\AppData\Local\quor\quor\hooks\claude-hook.ps1
✓ Hook responds correctly
✓ No conflicting PreToolUse hooks
✓ Tracking DB readable/writable
✓ Built-in filter tests pass
✓ Mode: optimize
✓ Plugin discovery — no third-party plugins installed

If any line shows , quor doctor prints what to run to fix it (usually quor init --claude).


Troubleshooting

'pip' is not recognized / 'python' is not recognized

The Python installer's "Add python.exe to PATH" option is easy to miss, especially on a per-user install. Two fixes:

  • Fastest, no setup needed: use the py launcher or full module invocation instead of the bare command:

    py -m pip install quor
    py -m quor doctor
    
  • Permanent fix: add your Python install's base folder and its Scripts subfolder to your User PATH (Windows key → "Edit environment variables for your account" → select Path → Edit → New). This does not require admin rights. Close and reopen your terminal afterward — existing windows won't see the change.

Multiple Python versions installed, unsure which one Quor uses

Use the py launcher to be explicit:

py --list          # show all installed versions
py -3.11 -m pip install quor
py -3.11 -m quor doctor

pip.exe / quor.exe "Access is denied" on a locked-down corporate machine

Some corporate endpoint-protection policies block execution of newly created .exe wrapper scripts even though the underlying Python interpreter is allowed to run. Work around it by always going through the interpreter instead of the wrapper:

python -m pip install quor
python -m quor doctor

Installing into a virtual environment on Windows and getting strange import errors

If your venv's path is deeply nested (e.g. under a long temp directory), Windows' classic 260-character path limit can silently truncate files during install. Create the venv somewhere short (e.g. C:\myvenv) and reinstall.

quor doctor shows a red for "Hook script installed"

Run quor init --claude — this is expected on any install where the hook hasn't been wired up yet, not a bug.

quor doctor shows a red for "No conflicting PreToolUse hooks"

Another tool already has a PreToolUse hook registered in Claude Code's settings.json. Run quor init --claude again to review and resolve the conflict; Quor will not silently overwrite another tool's hook.

Checking which version of Quor is installed

There's no quor --version flag yet — use:

pip show quor

Commands (V1)

Command Description
quor init --claude Install the Claude Code hook
quor explain <cmd> Show stage-by-stage trace of what Quor removed
quor gain Show cumulative token savings (±20%)
quor verify Run all inline filter tests
quor validate [file] Validate a filter config file
quor doctor Health check — hook responding? Tests passing?
quor schema Output the filter file JSON Schema to stdout

Both quor and qr are registered as entry points.


Roadmap: Observability (Planned)

The following are planned, not yet implemented. They're listed here so it's clear what to expect next, not as claims about current behavior:

  • Compression statistics — richer per-run breakdowns than quor gain's current cumulative view (per-stage contribution, per-filter history).
  • Estimated token savings inline — surfacing the ±20% estimate at the point of use, not only via a separate quor gain invocation.
  • Before/after preview — a way to see the unfiltered and filtered output side by side without needing to run quor explain separately.
  • Dry-run mode — run the filtering pipeline and report what would be removed without actually altering the output sent to the assistant.
  • Verbose diagnostics — an opt-in detailed trace mode for debugging filter behavior, beyond what quor explain already provides.

See docs/final/ROADMAP.md for the full version-by-version plan.


Development Status

Phase Description Status
0 Repository setup Complete
1 ContentMask primitive + pipeline engine Complete
2 Built-in compression stages Complete
3 Filter config + registry Complete
4 Command rewriter + classifier Complete
5 Claude Code hook adapter Complete
6 SQLite + JSONL tracking Complete
7 CLI commands Complete
8 Plugin infrastructure Complete
9 Plugin discovery & loading Complete
10 Packaging & release Completev0.1.0 on PyPI

605 tests passing, ruff + mypy clean on quor/ and tests/, verified on Python 3.11, 3.13, and 3.14. See docs/final/PROJECT_STATUS.md for the current snapshot, docs/final/IMPLEMENTATION_PLAN.md for the full roadmap, and CHANGELOG.md for the v0.1.0 release notes.

The operating-mode system (AUDIT / OPTIMIZE / SIMULATE) is intentionally display-only in this release — quor doctor and quor gain show the configured mode, but the dispatcher doesn't yet branch on it. This is a scoped, documented roadmap item, not a bug; see docs/final/PROJECT_STATUS.md for details.


Development Setup

git clone https://github.com/priyanshup/Quor.git
cd Quor
pip install -e ".[dev]"
pip install -e ./tests/fixtures/test_plugin
pytest tests/

The second install step is required for the plugin-discovery tests — see CONTRIBUTING.md for why it's a separate step rather than a pyproject.toml dev dependency.

Requires Python 3.11+. Windows is the primary development and CI target. Pure Python — no uv or other non-pip tooling required.


Contributing

See CONTRIBUTING.md for the full contributor guide, CODE_OF_CONDUCT.md for community standards, and SECURITY.md to report a vulnerability.


License

Apache 2.0 — see LICENSE for details.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

quor-0.1.1.tar.gz (311.7 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

quor-0.1.1-py3-none-any.whl (79.9 kB view details)

Uploaded Python 3

File details

Details for the file quor-0.1.1.tar.gz.

File metadata

  • Download URL: quor-0.1.1.tar.gz
  • Upload date:
  • Size: 311.7 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.1.0 CPython/3.13.12

File hashes

Hashes for quor-0.1.1.tar.gz
Algorithm Hash digest
SHA256 b53806497cbf97c88446ccc3bcc3c93daad36e0adf015970250a0317956d63b6
MD5 964a77ad2b960dee43bc13aab666c092
BLAKE2b-256 da42bac4f753822fb406e45cdf8198635d3b1815720dd9e4a6b687d808256422

See more details on using hashes here.

File details

Details for the file quor-0.1.1-py3-none-any.whl.

File metadata

  • Download URL: quor-0.1.1-py3-none-any.whl
  • Upload date:
  • Size: 79.9 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.1.0 CPython/3.13.12

File hashes

Hashes for quor-0.1.1-py3-none-any.whl
Algorithm Hash digest
SHA256 baca340560f38c1daaa1349a5f4be77a6740712cb8908bf4e6a485c62498f072
MD5 1225953e4a7d0e7702c14b5fb708a19d
BLAKE2b-256 fe75ce22f099991c22bcb1377ebd5467775b60fdcff87f67774c209c4cee9532

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page