.png)
3 weeks ago
2
An open-source model that produces fast, high-precision code context.
This is an exploration inspired by SWE-grep.
- Collect actions (grep/glob/read) policies either from usage logs or open datasets
- Optimize by removing redundant actions or parallelisation
- Train model on optimized action policy
- Release model as a single file, MCP tool
Install uv:
curl -LsSf https://astral.sh/uv/install.sh | sh
uv sync
Train the model with uv run -m src.train, output:
epoch 1 path-loss 4.2273 tool-loss 1.4921
epoch 2 path-loss 2.6636 tool-loss 1.1331
epoch 3 path-loss 1.9355 tool-loss 1.0876
epoch 4 path-loss 1.5844 tool-loss 0.9886
epoch 5 path-loss 1.4470 tool-loss 0.9531
epoch 6 path-loss 1.3959 tool-loss 0.9435
Predict the best action with uv run main.py, output:
How does the deploy script decide between blue/green targets?
predicted: read:scripts/deploy.py
top tools: read (0.86), summarize (0.08), glob (0.03)
top paths: scripts/deploy.py (0.52), deploy/rollouts/blue_green.yaml (0.36), docs/metrics/rollup.md (0.01)
Where is the feature flag `modal_new_footer` evaluated before render?
predicted: grep:app/components/ModalFooter.tsx
top tools: grep (0.65), read (0.25), glob (0.05)
top paths: app/components/ModalFooter.tsx (0.63), app/features/modal/useModalFooter.ts (0.26), src/payments/webhooks/retry.go (0.01)
Add this to your Codex's config.toml:
model_provider = "openai-responses-proxied"
[model_providers.openai-responses-proxied]
name = "OpenAI using Responses with Proxy"
base_url = "http://127.0.0.1:8080/v1"
env_key = "OPENAI_API_KEY"
wire_api = "responses"
Start proxy server:
uv run src/openai_forwarder.py --host 127.0.0.1 --port 8080
Use Codex per usual and you should seen openai_forwarder.log.jsonl populated.
The data in datasets/ are synthetically generated.
- example_supervised.jsonl — 31 queries drawn from realistic engineering scenarios. Each record stores repository metadata, commits, natural-language queries, the turn/parallel budgets, latency target, and multiple ground-truth spans annotated with the tool responsible (read, grep, glob, summarize) plus line ranges and reference answers.
- example_trajectory.jsonl — Trajectory rollouts aligned to the same query IDs, logging every tool invocation (command, arguments, timestamps, observations), the final selected tool/path, and reward metrics (weighted-F1, latency, composite score).
Together these files support both supervised evaluation and replay-style reinforcement learning while sharing a single underlying corpus.
If there is enough interest and contributions from the community, we might be able to turn this into a real thing!
Gaps:
- Tools
- Determine whether grep, glob, read, and summarize are the right set of tools.
- Parallel tool usage.
- Dataset: Replace the synthetic dataset with a corpus of real repository queries (bug reports, tickets, doc requests) labeled with repo/commit identifiers, ground-truth files, and line ranges.
- Training: possibly from existing agentic workflows.
- Evaluation: potentially requires human labeling.
- Model architecture
Star the repo and start submitting issues or PRs!
