Qwen3-repo

Architecture-aware repo-to-context scaffold for Qwen3 and Qwen3.6. Ingests a GitHub repository into a dependency-ordered context pack, runs an agentic coding loop via any OpenAI-compatible endpoint, and includes NL2Repo / SWE-bench evaluation runners.

Why ordering matters

Qwen3.6's hybrid architecture (Gated DeltaNet + Gated Attention) processes three out of four layers with linear attention and a fixed-size recurrent state. Placing definitions before their dependents should help the recurrent state accumulate core types and interfaces before call sites reference them.

Standard Qwen3 models use full attention and can look back to any position, so ordering matters less -- but dependency-aware packing still avoids wasting context on low-value files and keeps related code together.

Qwen's NL2Repo evaluations (score: 36.2, Qwen's published result; this scaffold provides an alternative implementation) were run via Claude Code (temp=1.0, top_p=0.95, max_turns=900). This scaffold provides an open-source alternative with architecture-aware context formatting.

Installation

pip install qwen3-repo

Quick start

Ingest a repository

from qwen3_repo import ingest_repo

# Works with any supported Qwen3 or Qwen3.6 model
context_pack, files, budget = ingest_repo(
    "https://github.com/user/repo",
    model="Qwen3-32B",
)
print(f"{len(files)} files, ~{budget.pack_budget:,} token budget")

Rank files by importance

from qwen3_repo import rank_files
from qwen3_repo.ingester import discover_files
from pathlib import Path

files = discover_files(Path("/path/to/repo"))
ranked = rank_files(files)
for r in ranked[:10]:
    print(f"{r.score:.2f}  {r.path}")

Detect vision encoder needs

from qwen3_repo import detect_vision_needs
from pathlib import Path

# Vision encoder is only relevant for Qwen3.6 (Qwen3 has no vision)
result = detect_vision_needs(Path("/path/to/repo"))
print(result["recommendation"])
# "Consider --language-model-only. No high-relevance visual assets detected. Frees ~6 GB KV cache (~100K-150K additional context tokens)."

Run the agentic scaffold

# Start a vLLM server first:
# vllm serve Qwen/Qwen3-32B --port 8000

python -m qwen3_repo.scaffold \
    --repo-path /path/to/repo \
    --task "Fix the failing test in test_auth.py" \
    --max-turns 900 \
    --temperature 1.0 \
    --top-p 0.95

# Or with SGLang:
# python -m sglang.launch_server --model Qwen/Qwen3-32B --port 30000

python -m qwen3_repo.scaffold \
    --backend sglang \
    --repo-path /path/to/repo \
    --task "Fix the failing test in test_auth.py"

Run NL2Repo evaluation

python -m qwen3_repo.eval.nl2repo \
    --tasks nl2repo_tasks.json \
    --api-url http://localhost:8000/v1 \
    --model Qwen/Qwen3.6-27B \
    --output-dir nl2repo_results

Compare Claude Code vs qwen3-repo

python -m qwen3_repo.bench_compare \
    --tasks comparison_tasks.json \
    --repo-path /path/to/repo \
    --markdown

Context ordering strategy

1. Role-based grouping: CONFIG -> TYPE_DEF -> CORE_LIB -> UTILITY -> FEATURE -> TEST -> DOC -> BUILD
2. Dependency-aware ordering: Topological sort within each group (Kahn's algorithm, importance as tiebreaker)
3. Budget trimming: Lowest-importance files dropped first; tests and docs trimmed before core code

Importance scoring

The ingestion pipeline (ingest_repo) uses these signals to order files within each role group:

The standalone rank_files() utility uses a separate scoring system with six signals (centrality, recency, coverage, role weight, structural depth, size efficiency) and customizable weights. See ranker.py for details.

Vision encoder decision (Qwen3.6 only)

Qwen3 models have no vision encoder. For Qwen3.6:

Supported models

Qwen3 (standard transformer, full attention)

Qwen3.6 (hybrid Gated DeltaNet + Gated Attention)

License

Apache 2.0