01Acquire
Information source acquisition
Collect, filter, and understand reusable agent skills, then organize them into a scenario–skill repository.
71K+ valid skillsScenario extraction
arXiv · 2608.18580
Fine-grained Agentic Construction of Executable Tasks
FACET
Preserving Source Intent and Executable State in Terminal Task Synthesis
University of Science and Technology of China · Shanghai AI Laboratory · Fudan University
stokou@mail.ustc.edu.cn · * Corresponding author
$ facet build --grounded
01scenario.reconstruct()recover intent + dependencies
02environment.realize()build → inspect → repair
03artifacts.align(state)instruction · solution · verifier
✓
Executable task acceptedshared state · dense tests · clean validation
6,078Validated in study
22.77Tests per task
1.2KTraining trajectories
+8.24Best absolute gain
The central idea
Build the world first.
Then write the task.
A terminal task couples an instruction, environment, solution, and verifier. One mismatch can invalidate the whole task.
FACET preserves source knowledge and grounds every artifact in the same realized environment.
Pipeline
From scattered skills to verified tasks.
Three stages. One continuous information path.
02Reconstruct
Scenario & reference building
Recover goals, dependencies, intermediate states, tools, and I/O contracts before building aligned instruction and solution references.
Five-dimensional viewIntent preserved
03Ground
Environment-driven construction
Realize the environment first, ground every artifact in the same container state, and repair only the component that fails.
Shared executable stateTargeted repair
Shared grounding interface
One state.
Every artifact aligned.
Environment changes are not hidden in a textual handoff. Files, paths, ports, packages, schemas, and fixtures are visible to each generator in sequence.
- 01Instruction grounded in the realized initial state
- 02Solution executed against the same environment
- 03Verifier observes both initial and final states
Instruction
Solution
Verifier
realizede₀ → eₜexecutable state
Terminal-Bench 2.1
Consistent gains across scales.
Fine-tuning with only 1.2K successful trajectories.
Qwen3.5-4B+7.12
Base17.60
FACET24.72
Qwen3.5-9B+8.24
Base27.34
FACET35.58
Qwen3.5-27B+6.75
Base40.82
FACET47.57
The 27B model reaches 47.57—within 1.49 points of Qwen3.5-397B under the same evaluation setting, at roughly 1/15 the model size.
Open releases
Data and models are available.
6,020 public-release tasks and three fine-tuned Qwen3.5 checkpoints.
83/100
Generation analysis
Sequential grounding yields more valid tasks.
The forward order—environment, instruction, solution, then solution-aware verifier—reaches an 83% final yield, compared with 63% for reverse and 65% for joint generation.
Forward83%
Joint65%
Reverse63%
Appendix highlights
More data. Less narration.
Compact comparisons from the paper and appendix.
A.1
Terminal-agent dataset comparison
Same Terminus-2 scaffold. P@k in percent.
| Dataset | Trajectory | Task | ||||
|---|---|---|---|---|---|---|
| # Traj. | Turns | # Tasks | Tests | P@1 | P@3 | |
| Nemotron-Terminal | 5K | 6.12 | 15K | 6.18 | 40.67 | 48.00 |
| Endless-Terminals | 200 | 4.53 | 2,492 | 5.51 | 83.00 | 87.00 |
| Terminal-Lego | 32K | 5.77 | 15K | 16.60 | 47.00 | 49.00 |
| TerminalWorld | 200 | 11.94 | 1,530 | 3.98 | 57.00 | 82.00 |
| Tmax | 500 | 11.14 | 15K | 3.29 | 80.00 | 86.00 |
| FACET (ours) | 1.2K | 11.86 | 6,078 | 22.77 | 27.00 | 35.00 |
A.2
Terminal-Bench 2.1
| Model | Size | Base | FACET | Δ |
|---|---|---|---|---|
| Qwen3.5-4B | 4B | 17.60 | 24.72 | +7.12 |
| Qwen3.5-9B | 9B | 27.34 | 35.58 | +8.24 |
| Qwen3.5-27B | 27B | 40.82 | 47.57 | +6.75 |
| Qwen3.5-397B-A17B | 397B | 49.06 | — | reference |
A.3
Generation order
| Scheme | Validated | Initial | Final yield |
|---|---|---|---|
| Forward | 99 | 46.5% | 83/100 |
| Joint | 96 | 37.5% | 65/100 |
| Reverse | 91 | 24.2% | 63/100 |
A.4
Construction funnel
| Stage | Count | Retention |
|---|---|---|
| Scenario–skill seeds | 7,852 | — |
| Successful environments | 7,504 | 95.70% |
| First-pass valid tasks | 2,856 | 38.35% |
| Final validated tasks | 6,078 | 81.63% |
A.5
Successful trajectories
| Statistic | Value |
|---|---|
| Observation commands | 84.7% |
| Observation-only first turn | 95.6% |
| Action → observation | 53.1% |
| Top ten commands | 88.4% |
Citation
Build on FACET.
FACET is available as arXiv:2608.18580.
@misc{shi2026facet,
title = {{FACET}: Preserving Source Intent and Executable
State in Terminal Task Synthesis},
author = {Kou Shi and Zun Wang and Qisheng Su and
Shiting Huang and Ziao Zhang and Zhen Fang and
Qingnan Ren and Jin Liu and Yu Zeng and
Yiming Zhao and Lin Chen and Zehui Chen and Feng Zhao},
year = {2026},
eprint = {2608.18580},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
url = {https://arxiv.org/abs/2608.18580}
}