Squeez

Squeeze out the juice, leave the pulp behind.

• Tool output pruner for LLM coding agents
• Pipe any tool output (pytest, grep, git log, npm build, kubectl, ...) through squeez with a task description, get back only the relevant lines
• Two models, same CLI: a generative Qwen 3.5 2B (0.80 F1, 92% compression) or a smaller extractive ModernBERT alternative
• CLI pipe, Python library, or vLLM server

Existing context pruning tools (SWE-Pruner, Zilliz Semantic Highlight, Provence) are built for source code or document paragraphs. They don't handle the mixed, unstructured format of tool output (stack traces interleaved with passing tests, grep matches with context lines, build logs with timestamps). Squeez is trained on 27 types of tool output from real SWE-bench workflows and synthetic multi-ecosystem observations.

pip install squeez
python -m pytest tests/ -v 2>&1 | squeez "find the test failure related to authentication"

Example

Task: "Find the test failure related to authentication"

$ python -m pytest tests/ -v
======================== test session starts ========================
platform linux -- Python 3.12.1, pytest-8.1.1
collected 23 items

tests/test_auth.py::test_login_valid PASSED
tests/test_auth.py::test_login_invalid PASSED
tests/test_auth.py::test_token_refresh FAILED
tests/test_auth.py::test_logout PASSED
tests/test_users.py::test_create_user PASSED
tests/test_users.py::test_delete_user PASSED
tests/test_users.py::test_list_users PASSED
tests/test_middleware.py::test_csrf_check PASSED
tests/test_middleware.py::test_rate_limit PASSED
tests/test_middleware.py::test_cors_headers PASSED

======================= FAILURES ================================
_____ test_token_refresh ________________________________________

    def test_token_refresh(self):
        token = self.client.get_token(expired=True)
>       refreshed = self.client.refresh(token)
E       AuthenticationError: Token refresh window expired
E       Expected: new token within 30m window
E       Got: rejection after 15m (timeout changed?)

tests/test_auth.py:47: AuthenticationError
================ short test summary info ========================
FAILED tests/test_auth.py::test_token_refresh
================== 1 failed, 9 passed ==========================

tests/test_auth.py::test_token_refresh FAILED

    def test_token_refresh(self):
        token = self.client.get_token(expired=True)
>       refreshed = self.client.refresh(token)
E       AuthenticationError: Token refresh window expired
E       Expected: new token within 30m window
E       Got: rejection after 15m (timeout changed?)

$ git log --oneline -25 | squeez "find the commit that changed the authentication timeout"

u6v7w8x Change auth timeout from 30m to 1h

$ npm run build 2>&1 | squeez "find the TypeScript error"

src/components/Auth.tsx(34,5): error TS2345: Argument of type 'string' is
  not assignable to parameter of type 'AuthToken'.

$ kubectl describe pod api-server-7d4b | squeez "why is the pod failing"

    State:          Waiting
      Reason:       CrashLoopBackOff
    Last State:     Terminated
      Reason:       Error
      Exit Code:    1
  Warning  BackOff  3m (x5)  kubelet  Back-off restarting failed container

Results

Evaluated on 618 manually curated held-out examples spanning 27 tool types:

Squeez-2B (2B params) outperforms the 18x larger Qwen 3.5 35B A3B by 11 recall points at the same compression level.

Quick start

With vLLM (recommended)

pip install vllm
vllm serve KRLabsOrg/squeez-2b --dtype bfloat16 --max-model-len 16384

# Use from squeez CLI
pip install squeez
export SQUEEZ_SERVER_URL=http://localhost:8000/v1
cat output.txt | squeez "find the bug"

vLLM keeps the model warm in memory with batched inference and high throughput.

Local inference (no server)

pip install squeez

cat output.txt | squeez "Find the failing traceback block"
squeez "Fix the CSRF bug" --input-file output.txt

Note: Local mode loads the model on every call. Fine for one-off use, but for repeated calls (e.g. an agent piping every tool through squeez), use vLLM.

Any OpenAI-compatible API

Works with Groq, Together, or any OpenAI-compatible server. Set the URL, model name, and API key:

export SQUEEZ_SERVER_URL=https://api.groq.com/openai/v1
export SQUEEZ_SERVER_MODEL=squeez
export SQUEEZ_API_KEY=gsk_...

Python API

from squeez.inference.extractor import ToolOutputExtractor

# Default: loads KRLabsOrg/squeez-2b locally
extractor = ToolOutputExtractor()

# Or connect to a server
extractor = ToolOutputExtractor(base_url="http://localhost:8000/v1")

filtered = extractor.extract(
    task="Find the referer validation block",
    tool_output=raw_output,
)

Use with Claude Code

Add to your CLAUDE.md:

Always when you invoke a shell command, pipe it through `squeez` and tell exactly what you want to know.

Examples:
- `bun test 2>&1 | squeez "did the tests pass?"`
- `git log --oneline -50 | squeez "find the commit that broke CSRF"`
- `cat src/auth/middleware.py | squeez "find the referer validation logic"`

Do NOT use squeez when:
- You need exact, uncompressed output (e.g. writing a patch)
- The command is interactive

Works with other coding agents (Codex CLI, OpenCode, etc.) via their equivalent instruction files.

Advanced

# squeez.yaml
server_url: "http://localhost:8000/v1"
# local_model_path: "./output/squeez_qwen"  # for local inference instead
# backend: null  # auto-detect; or "transformers", "vllm", "encoder"

export SQUEEZ_LOCAL_MODEL=KRLabsOrg/verbatim-rag-modern-bert-v2

pytest -q 2>&1 | squeez "find the failing test"
git log --oneline -50 | squeez "find the auth commit"

# Download dataset
python scripts/download_data.py

# Train generative model (Qwen 3.5 2B + LoRA)
squeez train --train-file data/train.jsonl --eval-file data/dev.jsonl

# Train token encoder
python -m squeez.encoder.train \
    --classifier-type token \
    --train-file data/encoder_train.jsonl \
    --eval-file data/encoder_dev.jsonl \
    --base-model answerdotai/ModernBERT-base \
    --output-dir output/squeez_encoder

# Evaluate
squeez eval --extractor-model output/squeez_qwen --eval-file data/test.jsonl

python scripts/build_full_dataset.py \
    --output-dir data/v3 \
    --teacher-model openai/gpt-oss-120b \
    --teacher-base-url http://localhost:8000/v1

Citation

@misc{kovács2026squeeztaskconditionedtooloutputpruning,
      title={Squeez: Task-Conditioned Tool-Output Pruning for Coding Agents}, 
      author={Ádám Kovács},
      year={2026},
      eprint={2604.04979},
      archivePrefix={arXiv},
      primaryClass={cs.SE},
      url={https://arxiv.org/abs/2604.04979}, 
}

License

Apache 2.0