🌮 TACO

A Self-Evolving Framework for Efficient Terminal Agents via Observational Context Compression

Terminal agents keep feeding raw shell output back into their own context, and that noise accumulates quadratically across multi-turn tasks — drowning out real error signals and inflating token cost.

TACO is a plug-and-play, self-evolving observational-context compression framework. Instead of hard-coded truncation, it discovers, repairs and reuses compression rules online, and keeps a global rule pool that lets new tasks bootstrap from knowledge accumulated on earlier ones.

On TerminalBench it gives +1%–4% across strong backbones (MiniMax-M2.5, DeepSeek-V3.2, Qwen3-Coder-480B, Qwen3-14B) and transfers to SWE-Bench Lite, DevEval, CRUST-Bench and CompileBench. See the paper for full results.

---

Quick start

TACO ships as terminus-2 inside the Harbor evaluation framework.

pip install -e .

harbor run \
  -d terminal-bench@2.0 \
  -a terminus-2 \
  -m openai/gpt-4o-mini \
  -n 4 \
  -o results/taco_example \
  --ak enable_compress=True \
  --ak compress_base_url="<COMPRESSION_LLM_URL>" \
  --ak compress_api_key="<COMPRESSION_LLM_KEY>" \
  --ak compress_model_name="<COMPRESSION_LLM_MODEL>" \
  --ak enable_self_evo=True \
  --ak max_turns=200

A ready-to-edit template lives at scripts/run_taco_example.sh.

Parameters

All flags are passed on the CLI as --ak <name>=<value>.

Flag	Default	Description
enable_compress	False	Master switch. Every flag below is a no-op while this is False.
compress_base_url	""	Base URL of the self-evo planner LLM (any OpenAI-compatible endpoint).
compress_api_key	""	Bearer token for the planner LLM.
compress_model_name	""	Model name served by the planner LLM.
enable_self_evo	False	Enable the online rule planner / evolver.
freeze_rules	False	Freeze the rule pool — no LLM plan, no evolution (ablation / reproducible runs).
disable_global_evo	False	Ignore the global rule pool and start each task from built-in seed rules (ablation).
uncovered_threshold	3000	Character length above which an uncovered output is flagged for new-rule proposal.
max_turns	1_000_000	Per-task turn cap, e.g. --ak max_turns=200.
model_info	None	JSON forwarded to LiteLLM, e.g. '{"max_input_tokens": 132000, "max_output_tokens": 32768}'.

Common configurations

# Full TACO (recommended)
--ak enable_compress=True --ak enable_self_evo=True \
--ak compress_base_url=... --ak compress_api_key=... --ak compress_model_name=...

# Ablation: freeze the rule pool, no evolution
--ak enable_compress=True --ak enable_self_evo=True --ak freeze_rules=True

# Ablation: local-only evolution (ignore the global pool)
--ak enable_compress=True --ak enable_self_evo=True --ak disable_global_evo=True

# Control (vanilla terminus-2)
--ak enable_compress=False

Citation

@misc{ren2026selfevolvingframeworkefficientterminal,
      title={A Self-Evolving Framework for Efficient Terminal Agents via Observational Context Compression},
      author={Jincheng Ren and Siwei Wu and Yizhi Li and Kang Zhu and Shu Xu and Boyu Feng and Ruibin Yuan and Wei Zhang and Riza Batista-Navarro and Jian Yang and Chenghua Lin},
      year={2026},
      eprint={2604.19572},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.19572},
}