# OneAgent

**Repository Path**: WindNegev/one-agent

## Basic Information

- **Project Name**: OneAgent
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-09-20
- **Last Updated**: 2026-09-20

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# one-agent

`one-agent` studies adaptive delegation and context routing for repository-level software-engineering agents.

> Can a main agent learn when to solve a GitHub issue itself, when to delegate or parallelize work, what repository context each real worker needs, and how to verify and integrate patches under a matched compute budget?

The project uses real repositories, executable environments, real model rollouts, and test-based rewards. It does not fabricate worker outputs or inject synthetic failures. Small subsets are used only for engineering calibration; scientific claims require frozen held-out tasks and official evaluators.

## Research scope (v2)

The main agent receives an issue, repository state, available tools/workers, and a resource budget. It chooses among actions such as:

- solve directly;
- request read-only localization or impact analysis;
- delegate one implementation task;
- run independent analyses or candidate patches in parallel;
- request tests or review;
- accept, revise, reject, or merge a worker patch.

Workers run in isolated environments or Git worktrees, use real shell/code/test tools, and return evidence, patches, and test outcomes. The main policy never receives the gold patch or hidden evaluator results.

Out of scope for the first study: long-term user memory, automatic skill writing, general desktop automation, heterogeneous commercial-model routing, and artificial fault injection.

## Data and evaluation

- **Training/development:** SWE-Gym (2,438 real Python repository tasks; current Lite dataset has 230 tasks).
- **Primary held-out evaluation:** SWE-bench Verified (500 expert-validated GitHub issues).
- **Later transfer evaluation:** selected Terminal-Bench tasks after the repository-level result is established.
- **Model:** Qwen3-4B-Instruct-2507 is the initial trainable target. A larger frozen model will be used as a capability ceiling.
- **Inference:** vLLM for serving; an established SWE agent scaffold is used before introducing custom orchestration.
- **Training:** no RL until direct and prompted delegation baselines reveal a stable, test-verifiable policy gap. Verl remains the planned training stack.

COMPASS results are retained as historical Phase-0 diagnostics of tool use and constraint preservation, not as the headline evidence for the software-engineering claim.

## Required comparisons

1. Established single-agent scaffold with all tools.
2. Fixed-role delegation with the same model and matched total token/tool budget.
3. Prompted adaptive delegation and context routing.
4. Oracle routing/localization available only as a diagnostic upper bound.
5. Learned routing policy, only after the preregistered training gate is passed.

Report task resolution, regression-test preservation, valid-patch rate, total tokens, tool calls, GPU time, wall-clock latency, redundant work, merge conflicts, and delegation frequency separately.

## Repository map

- `docs/SWE_RESEARCH_PLAN.md` — v2 research question, data policy, baselines, and training gate.
- `docs/NATIVE_SWE_CALIBRATION.md` — pre-Docker real-agent calibration and executor failure analysis.
- `docs/ENVIRONMENT.md` — remote environment and benchmark runtime constraints.
- `docs/RESEARCH_ALIGNMENT.md` — superseded v1 COMPASS-oriented formulation, retained for provenance.
- `docs/PHASE0_RESULTS.md` — real-model COMPASS diagnostic results.
- `docs/RESEARCH_PROTOCOL.md` — reproducibility and anti-leakage rules.
- `scripts/preflight.py` — remote environment checks.

## Remote environment

Experiments run under `/home/machunyue/code/agenticrl/one-agent` with the `one-agent` conda environment. Benchmark source checkouts and large artifacts live under `/data4/machunyue`, not on the nearly full root filesystem.

Current blocker: the `machunyue` account is not a member of the `docker` group. SWE-bench evaluation cannot start until Docker daemon access is granted and a new login session is created. Hugging Face data access works through `https://hf-mirror.com`, with caches redirected to `/data4/machunyue/hf-cache`.

## Current status

- [x] Real Qwen3-4B inference and tool calling validated on the remote machine.
- [x] COMPASS diagnostics recorded; limitations identified.
- [x] Primary domain changed to repository-level software engineering.
- [x] Official SWE-Gym and SWE-bench source revisions cloned under `/data4/machunyue/benchmarks`.
- [x] Run a real native SWE-Gym task with Qwen3-4B; three protocols showed that the checkpoint is not yet a viable coding worker.
- [ ] Obtain Docker daemon access for `machunyue`.
- [ ] Validate one official gold patch end-to-end.
- [ ] Select a viable coding-specialized executor, then run a frozen single-agent development baseline.
- [ ] Implement and compare real adaptive delegation.
- [ ] Decide whether the evidence justifies SFT or RL.