co-agent-coder 0.1.0__py3-none-any.whl

co_agent_coder-0.1.0.dist-info/METADATA

@@ -0,0 +1,396 @@
+Metadata-Version: 2.4
+Name: co-agent-coder
+Version: 0.1.0
+Summary: A private AI coding agent for the free Google Colab T4: local 7B model, git-backed memory, reviewable branches.
+Author: Lyle Okoth
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://github.com/lyle-mlengineer/coagent
+Project-URL: Documentation, https://github.com/lyle-mlengineer/coagent/blob/main/README.md
+Project-URL: Source, https://github.com/lyle-mlengineer/coagent
+Project-URL: Issues, https://github.com/lyle-mlengineer/coagent/issues
+Project-URL: Changelog, https://github.com/lyle-mlengineer/coagent/blob/main/CHANGELOG.md
+Keywords: coding-agent,colab,llm,transformers,git,qwen
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: Operating System :: POSIX :: Linux
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Software Development :: Code Generators
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: tomli>=2.0.1; python_version < "3.11"
+Provides-Extra: llm
+Requires-Dist: torch<2.7,>=2.4; extra == "llm"
+Requires-Dist: transformers<4.49,>=4.45; extra == "llm"
+Requires-Dist: accelerate<1.2,>=0.34; extra == "llm"
+Requires-Dist: bitsandbytes<0.47,>=0.43; extra == "llm"
+Requires-Dist: safetensors>=0.4.3; extra == "llm"
+Provides-Extra: ui
+Requires-Dist: ipywidgets>=8.1; extra == "ui"
+Requires-Dist: ipython>=8.18; extra == "ui"
+Requires-Dist: nbformat>=5.10; extra == "ui"
+Requires-Dist: nbconvert>=7.16; extra == "ui"
+Provides-Extra: gradio
+Requires-Dist: gradio<6,>=4.44; extra == "gradio"
+Provides-Extra: context
+Requires-Dist: tree-sitter<0.25,>=0.23; extra == "context"
+Requires-Dist: tree-sitter-python>=0.23; extra == "context"
+Requires-Dist: tree-sitter-javascript>=0.23; extra == "context"
+Requires-Dist: json5>=0.9.25; extra == "context"
+Provides-Extra: dev
+Requires-Dist: pytest>=8.2; extra == "dev"
+Requires-Dist: pytest-cov>=5.0; extra == "dev"
+Requires-Dist: ruff>=0.6; extra == "dev"
+Requires-Dist: nbformat>=5.10; extra == "dev"
+Requires-Dist: nbconvert>=7.16; extra == "dev"
+Requires-Dist: ipykernel>=6.29; extra == "dev"
+Provides-Extra: colab
+Requires-Dist: coagent[context,llm,ui]; extra == "colab"
+Provides-Extra: all
+Requires-Dist: coagent[context,dev,gradio,llm,ui]; extra == "all"
+Dynamic: license-file
+
+# coAgent
+
+<p align="center">
+  <img src="./assets/cover.png" alt="coAgent — a private coding agent for the free Google Colab T4 that commits every checkpoint back to git" width="100%" />
+</p>
+
+<p align="center">
+  <img src="https://img.shields.io/badge/Status-design%20v1.0-FFC24B?style=for-the-badge&logoColor=black" alt="Design v1.0"/>
+  <img src="https://img.shields.io/badge/Target-free%20Colab%20T4-F9AB00?style=for-the-badge&logo=googlecolab&logoColor=white" alt="Free Colab T4"/>
+  <img src="https://img.shields.io/badge/Model-7B%20NF4%20%C2%B7%2012k-7EF0C0?style=for-the-badge&logoColor=black" alt="7B NF4, 12k context"/>
+  <img src="https://img.shields.io/badge/Python-3.11-3776AB?style=for-the-badge&logo=python&logoColor=white" alt="Python 3.11"/>
+  <a href="https://github.com/lyle-mlengineer/coagent-web"><img src="https://img.shields.io/badge/Web-coagent--web-E05785?style=for-the-badge&logo=github&logoColor=white" alt="Web front door"/></a>
+  <a href="https://colab.research.google.com/github/lyle-mlengineer/coagent/blob/main/notebooks/bootstrap.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open in Colab"/></a>
+  <img src="https://img.shields.io/badge/License-Apache%202.0-D22128?style=for-the-badge&logo=apache&logoColor=white" alt="Apache 2.0"/>
+</p>
+
+**[Open in Colab](https://colab.research.google.com/github/lyle-mlengineer/coagent/blob/main/notebooks/bootstrap.ipynb) · [Design](./DESIGN.md) · [PRD](./PRD.md) · [Web front door](https://github.com/lyle-mlengineer/coagent-web)**
+
+> **Design stage.** This repository holds the full architecture (`DESIGN.md`) and product requirements (`PRD.md`); implementation starts at v0.1 in the build plan below. Every interface shown here — the cell, the commands, the config — is the target v0.1 surface, not a shipped package yet.
+
+### Run it
+
+Open the bootstrap notebook in Colab — **Runtime → Change runtime type → T4 GPU** — and run the cells. The whole install story is three lines:
+
+```python
+from coagent import coAgent
+agent = coAgent(repo="https://github.com/you/project.git")
+agent.chat("add a chunked CSV loader and test that dtypes are preserved")
+```
+
+The notebook checks your hardware before loading weights and refuses with a reason if the runtime cannot host the model. Everything after that — the plan, each edit, the test result, the commit — appears in the cell, and every checkpoint lands on a `coagent/<session>` branch you can review.
+
+**coAgent is an AI coding agent that lives inside a Google Colab notebook.** It clones your repo into the free T4, plans inside your own files, edits with a locally loaded 7B model, runs your tests on CPU, and commits every checkpoint back to git — on a branch you can review. The runtime is disposable; the repository is the memory.
+
+No API keys. No per-token cost. Nothing sent to a third-party model API — your code only travels to the repo you already own.
+
+The public landing page, waitlist, and GitHub sign-in live in [coagent-web](https://github.com/lyle-mlengineer/coagent-web). This repo is the agent itself.
+
+## What it does
+
+**Cloud coding agents assume a credit card and that you will send them your code.** Colab gives students, researchers, and independent engineers a free GPU but none of the workflow: no durable disk, no inbound ports, and a runtime that disappears without warning. coAgent is designed backwards from that constraint.
+
+- 🔒 **Private by construction** — the 7B model runs in your Colab VM. Your code is never sent to a third-party API, and there is no account beyond GitHub.
+- 💸 **Zero keys, zero marginal cost** — 7B at NF4 fits the free T4 (4.2–4.6 GB of weights, ≤10 GB peak VRAM), so exploration is never metered.
+- 🧠 **Git is the state store** — one commit per verified step, pushed or bundled off the VM before the next turn. A dead runtime costs at most one step.
+- ✅ **Tests are the critic** — a 7B model cannot hold a system in its head, so deterministic checks decide what counts as done: parse, lint, then your suite.
+- 🧯 **Honest about failure** — a VRAM OOM ladder, step/wall-clock/token budgets, loop detection, and a written report instead of a silent stop.
+- 📓 **Notebook-native** — plots render inline, artifacts persist to the repo or Drive with a manifest, and the UI works without hosting anything.
+
+## Who it's for
+
+| If you are… | coAgent gives you… |
+| --- | --- |
+| **A solo ML engineer** *(primary)* | Repo changes and test runs without leaving Colab or paying per token — with your code staying private. |
+| **A data analyst** | A cell edited, executed, and a figure rendered and saved with a manifest, not just source files. |
+| **A student or learner** | A real coding agent at zero marginal cost, on hardware you already have. |
+| **Not for:** teams needing concurrent users, monorepos over ~50k LOC, or fully autonomous long-running agents. | Those needs would compromise the T4-only, single-user design. |
+
+## How it works
+
+```mermaid
+flowchart TB
+  subgraph U["You · browser tab"]
+    NB["Notebook cell · chat, plan approval, artifacts"]
+  end
+  subgraph V["Colab VM · disposable"]
+    K["IPython kernel"]
+    A["agent loop · survey → plan → act → verify"]
+    M["Qwen2.5-Coder-7B · NF4 · 12k ctx"]
+    T["tool runtime · subprocess · git"]
+    W["/content/repo"]
+  end
+  subgraph D["Durable"]
+    GH["GitHub · branch coagent/2026-11-12-a1b2c3"]
+    DR["Drive · bundles, artifacts, HF cache"]
+  end
+  NB --> K --> A
+  A <--> M
+  A --> T --> W
+  W -->|"checkpoint: commit + push"| GH
+  W -.->|"bundle when push is unavailable"| DR
+  V -.->|"eviction"| D
+```
+
+The loop is split into five phases with **different prompts and different tool sets**, because a small model does one job well:
+
+| Phase | Sees | Can do | Sampling |
+| --- | --- | --- | --- |
+| **SURVEY** | repo map, task, plan | read tools only | T=0.2 |
+| **PLAN** | task + retrieved files | nothing — writes a plan file | T=0.7 |
+| **ACT** | plan, excerpts, last 3 observations | one write or exec | T=0.2 |
+| **VERIFY** | diff, test output | run tests/lint only | deterministic |
+| **REPORT** | plan, log | nothing | T=0.5 |
+
+**PLAN cannot write.** A model that can act while planning will start editing before it has read the file. For anything beyond a single-file change, the plan is shown and must be approved before ACT — the human supplies intent, the model executes.
+
+### A worked turn
+
+```text
+user:  add a chunked CSV loader to src/loader.py and test dtype preservation
+
+SURVEY  read src/loader.py, src/frame.py; repo map shows Frame.from_rows exists
+PLAN    .coagent/plans/chunked-csv.md written        [approved]
+ACT     edit_file src/loader.py  (+18 -3)            [syntax ok]
+VERIFY  pytest tests/test_loader.py -q → 1 failed    [assert dtype]
+ACT     edit_file src/loader.py:31                   [fix cast]
+VERIFY  pytest -q → 12 passed
+CHECKPOINT commit a1b2c3d "coagent: add chunked CSV loader" (Coagent-Step: 7) + push
+REPORT  branch coagent/2026-11-12-a1b2c3 ready; 1 file, +18 -3, 12 tests green
+```
+
+## Designed backwards from a free T4
+
+Every design decision is a response to a Colab constraint.
+
+| Colab reality | What coAgent does about it |
+| --- | --- |
+| **Ephemeral VM**, ~12 h cap, ~90 min idle disconnect | Commit per verified step, sync off the VM before the next turn, and resume from the last checkpoint on any new VM. |
+| **One free T4**, 16 GB, compute capability 7.5 | 7B at NF4, 12,288-token hard cap, fp16 + SDPA. Model resident ~6.5–8 GB, leaving ≥7 GB for your own work. |
+| **~25 tokens/s decode** | Repo map instead of exploration, diffs instead of rewrites, tests instead of self-critique. Turns-to-done is the metric. |
+| **2 vCPU, ~12.7 GB RAM** | No vector database. BM25 over function-sized chunks plus a symbol graph. |
+| **No inbound ports** | The UI lives in the notebook; no daemon, no tunnel, no hosted service. |
+| **Root-ish shell, no containers** | Approval matrix, command denylist, path jail on reads *and* writes. The denylist is a seatbelt, not a sandbox — the real control is the human click. |
+
+Preflight prints the GPU, VRAM, RAM, cache state, and chosen context budget before loading weights, then refuses anything that is not a T4 (compute capability 7.5, ≥14 GB VRAM) unless `--allow-unsupported` is passed. If the 7B cannot fit after the OOM ladder, the fallback is the 3B at 8k — never a larger model, never CPU-only, never a paid tier.
+
+## Safety and trust
+
+| Action | Default |
+| --- | --- |
+| read / grep / repo map, inside the jail | allow |
+| read outside the jail | deny unless explicitly allowed |
+| write inside the repo, on the session branch | allow |
+| run tests, lint, syntax check | allow |
+| arbitrary shell command, network, install | ask |
+| delete / move / chmod, write outside the repo | ask |
+| `git push` to `coagent/*` on a configured remote | allow at checkpoint |
+| push to `main`, `--force`, `reset --hard` | deny |
+
+- **Branch isolation** — coAgent works on `coagent/<date>-<id>` and never pushes to `main`. `git log --author=coagent` is the audit trail.
+- **Secrets** — read from Colab `userdata` into memory, injected only into the git child process, never into the repo or a log. Clone through a `GIT_ASKPASS` helper instead of putting a token in the remote URL.
+- **Redaction** — GitHub, Hugging Face, OpenAI, AWS, Google, and JWT-shaped tokens are scrubbed before any log or display write.
+- **Prompt injection** — observations are wrapped as data, side effects require approval, and the agent branch cannot reach `main`. The second layer is the real guarantee; the docs say so plainly.
+
+## How to use it
+
+The target flow is one Colab notebook, one repo URL, and a task. Everything below ships with v0.1.
+
+1. **Open the bootstrap notebook in Colab** and select a T4 runtime. Preflight reports the hardware and the context budget before a single weight is loaded.
+2. **Add a fine-grained GitHub PAT** to Colab secrets as `GH_TOKEN`, scoped to the one repo. It never enters the notebook source.
+3. **Run the cell.** Install, construct, chat — at most three lines of user code. Cold start targets ≤10 minutes; the weights are cached locally for the next session.
+4. **Describe the task.** coAgent surveys, proposes a plan, and waits for your approval before it writes anything.
+5. **Review the branch.** Every verified step is a commit with `Coagent-Session`, `Coagent-Step`, and `Coagent-Verify` trailers; `coagent push` opens the PR path.
+6. **Recover.** If the runtime dies, run `coagent resume` on a fresh VM and it replays from the last checkpoint, losing at most one step.
+
+```python
+!pip -q install co-agent-coder        # or: git+https://github.com/lyle-mlengineer/coagent
+
+from coagent import coAgent
+
+agent = coAgent(repo="https://github.com/you/my-project.git")
+agent.chat("add a chunked CSV loader and test dtype preservation")
+```
+
+### Commands
+
+| Command | What happens |
+| --- | --- |
+| `coagent chat` | Opens the notebook chat loop in the current kernel. |
+| `coagent run "<task>"` | One-shot: survey → plan → act → verify → checkpoint → report. |
+| `coagent resume` | Rebuilds session state from branch + plan + log on any VM. |
+| `coagent doctor` | Versions, GPU, cache, git auth, and a one-line verdict. |
+| `coagent report` | Session summary generated from `.coagent/sessions/*.jsonl`. |
+| `coagent eval` | Runs the task suite and reports turns-to-done. |
+| `coagent push` | Pushes `coagent/*` and optionally opens a PR. |
+
+## Configuration
+
+Per-repo settings are committed so a session is reproducible on any machine:
+
+| Key | Default | Notes |
+| --- | --- | --- |
+| `model.name` | `Qwen/Qwen2.5-Coder-7B-Instruct` | 3B is the validated fallback. |
+| `model.quant` | `nf4` | bitsandbytes NF4; fp16 compute on T4. |
+| `model.max_context` | `12288` | Hard cap, counted with the model's tokenizer. |
+| `verify.commands` | `["ruff check .", "pytest -q"]` | CPU-only by default. |
+| `verify.gpu_commands` | `[]` | Explicit opt-in; the model yields VRAM first. |
+| `verify.timeout_s` | `300` | Per command; process group is killed. |
+| `verify.require_tests` | `true` | A feature without a test is not done. |
+| `durability.mode` | `push` | `push` to git, or `bundle` to Drive when offline. |
+| `durability.remote` | `origin` | Push target; `main` is denied. |
+| `safety.deny` | force-push, `main` | Enforced in code, not in the prompt. |
+| `safety.require_approval` | network, install, delete, outside-repo writes | Default flips to deny when non-interactive. |
+
+Bootstrap sets the git identity, installs the credential helper, and warns — never silently edits — when `.gitignore` is missing `.env` or `.coagent/secrets*`.
+
+```toml
+# .coagent/config.toml
+[model]
+name = "Qwen/Qwen2.5-Coder-7B-Instruct"
+quant = "nf4"
+max_context = 12288
+
+[verify]
+commands = ["ruff check .", "pytest -q"]   # CPU-only by default
+gpu_commands = []                          # run once, after the model yields VRAM
+timeout_s = 300
+require_tests = true
+
+[durability]
+mode = "push"                              # push | bundle (Drive)
+remote = "origin"
+
+[safety]
+deny = ["git push --force", "git push origin main"]
+```
+
+## Tech stack
+
+<p align="center">
+  <img src="https://img.shields.io/badge/Python%203.11-3776AB?style=flat-square&logo=python&logoColor=white" alt="Python 3.11"/>
+  <img src="https://img.shields.io/badge/PyTorch-EE4C2C?style=flat-square&logo=pytorch&logoColor=white" alt="PyTorch"/>
+  <img src="https://img.shields.io/badge/Transformers-FFD21E?style=flat-square&logo=huggingface&logoColor=black" alt="Transformers"/>
+  <img src="https://img.shields.io/badge/bitsandbytes%20NF4-4B8BBE?style=flat-square" alt="bitsandbytes"/>
+  <img src="https://img.shields.io/badge/Qwen2.5--Coder-615CED?style=flat-square" alt="Qwen2.5-Coder"/>
+</p>
+<p align="center">
+  <img src="https://img.shields.io/badge/tree--sitter-2F6F44?style=flat-square" alt="tree-sitter"/>
+  <img src="https://img.shields.io/badge/outlines-111111?style=flat-square" alt="outlines"/>
+  <img src="https://img.shields.io/badge/nbformat-F37626?style=flat-square&logo=jupyter&logoColor=white" alt="nbformat"/>
+  <img src="https://img.shields.io/badge/pytest-0A9EDC?style=flat-square&logo=pytest&logoColor=white" alt="pytest"/>
+  <img src="https://img.shields.io/badge/Ruff-D7FF64?style=flat-square&logoColor=black" alt="Ruff"/>
+</p>
+<p align="center">
+  <img src="https://img.shields.io/badge/Google%20Colab-F9AB00?style=flat-square&logo=googlecolab&logoColor=white" alt="Google Colab"/>
+  <img src="https://img.shields.io/badge/Git-F05032?style=flat-square&logo=git&logoColor=white" alt="Git"/>
+  <img src="https://img.shields.io/badge/Drive-4285F4?style=flat-square&logo=googledrive&logoColor=white" alt="Google Drive"/>
+  <img src="https://img.shields.io/badge/Apache%202.0-D22128?style=flat-square&logo=apache&logoColor=white" alt="Apache 2.0"/>
+</p>
+
+- **Runtime** — PyTorch, Hugging Face Transformers, bitsandbytes NF4, accelerate. fp16 + SDPA on Turing; bf16 and FlashAttention-2 are never assumed.
+- **Context and editing** — tree-sitter repo map, BM25 retrieval, JSONL session logs, and SEARCH/REPLACE edits with a unique-match requirement. `outlines` is the optional constrained-decoding path.
+- **Verification** — pytest and Ruff as the reference commands, run CPU-only with `CUDA_VISIBLE_DEVICES=""`; `nbformat` + `nbconvert` for notebooks.
+- **Delivery** — `pip install` from git and one bootstrap notebook. No daemon, no tunnel, no client framework.
+
+## Project structure
+
+```text
+coagent/
+├─ DESIGN.md                 architecture and rationale — the why
+├─ PRD.md                    product requirements, personas, acceptance tests
+├─ assets/                   cover art used by this README
+├─ notebooks/                bootstrap notebook                     (planned)
+├─ src/coagent/
+│  ├─ cli.py                 chat | run | resume | doctor | report | eval | push
+│  ├─ agent/                 loop, prompts, budget, state
+│  ├─ llm/                   transformers backend, registry, grammar
+│  ├─ context/               repo map, retrieval, compression, tokens
+│  ├─ tools/                 fs, shell, python, git, notebooks, artifacts
+│  ├─ gitx/                  session, commit, durable sync, resume, lock
+│  ├─ ui/                    notebook render, optional Gradio
+│  ├─ safety/                policy, denylist, redaction
+│  └─ evals/                 harness + task suite
+└─ tests/
+```
+
+## Build plan
+
+| Release | Week | Scope | Exit criterion |
+| --- | --- | --- | --- |
+| **v0.1** | 1 | loop, six tools, transformers backend, notebook chat | fixes a failing test in a toy repo, end to end |
+| **v0.2** | 2 | repo map, SEARCH/REPLACE reliability, verify loop, durable checkpoints, resume | a killed runtime resumes losing ≤1 step |
+| **v0.3** | 3 | plan approval, artifacts, approvals UI, Drive cache, OOM ladder | 5-task eval ≥60%; injection and rejected-plan tests pass |
+| **v1.0** | 4 | constrained decoding, evals, docs, demo, T4 matrix | 20-task eval ≥70%; all acceptance tests green |
+
+> **Definition of shipped:** a stranger with a free Colab account and a repo URL can install, run a task, review the branch, and recover from a disconnect using only this README.
+
+## What it deliberately isn't
+
+| Option | Why not |
+| --- | --- |
+| LangChain / LangGraph | Hidden prompt layers and token overhead make a weak model harder to debug. The loop is ~600 lines of explicit Python. |
+| Whole-file rewrites | Catastrophic diffs above ~200 lines. SEARCH/REPLACE, with a unique-match requirement, is safer. |
+| Vector database over the repo | A CPU and memory tax beside the GPU; BM25 plus a symbol graph covers a repo of this size. |
+| Multi-agent fan-out | One GPU serializes inference; extra agents are just turns with less shared context. |
+| Hosted API (OpenAI, Anthropic) | Defeats the privacy and zero-cost guarantees. The `llm/` interface exists, but v1 ships local-only. |
+| Pro GPUs, TPUs, CPU-only, multi-GPU | The product is defined by the free tier. Preflight refuses unsupported hardware rather than pretending. |
+| `share=True` Gradio / ngrok | Tunnel flakiness, rate limits, and a public exposure surface. Colab's port proxy is enough. |
+
+## How to contribute
+
+**This is a design-led repo.** The documents are the product until the code exists, and they are expected to change with evidence.
+
+1. Open an issue before a large PR. Reference the section (`DESIGN.md §4.5`) or requirement (`PRD FR-BOOT-2`) you are changing.
+2. If a decision changes, update **both** `DESIGN.md` (rationale) and `PRD.md` (requirements) in the same PR. A design change without a rationale is a regression.
+3. Keep the constraints honest: free T4 only, 12,288-token cap, ≤10 GB peak VRAM, no hidden framework.
+4. Implementation PRs: add or update a task in `src/coagent/evals/tasks/`, run the suite CPU-only, and report turns-to-done rather than "it feels better".
+5. Conventional commits (`feat:`, `fix:`, `docs:`, `test:`) and a PR body that explains the **why**.
+
+Good first contributions: a DESIGN/PRD consistency pass, an eval task with a deterministic pass criterion, a failure-mode repro from §10, or a threat-model review of §9.
+
+Please **do not** open a public issue for a security problem. Email [lyleokoth.mlengineer@gmail.com](mailto:lyleokoth.mlengineer@gmail.com) with steps to reproduce and I will respond privately.
+
+## Author
+
+<p align="center">
+  <a href="https://lyle-mlengineer.github.io"><img src="https://img.shields.io/badge/Portfolio-Live-E05785?style=for-the-badge&logo=googlechrome&logoColor=white" alt="Portfolio"/></a>
+  <a href="https://www.linkedin.com/in/lyle-okoth/"><img src="https://img.shields.io/badge/LinkedIn-0A66C2?style=for-the-badge&logo=linkedin&logoColor=white" alt="LinkedIn"/></a>
+  <a href="https://medium.com/@lyle-okoth"><img src="https://img.shields.io/badge/Medium-12100E?style=for-the-badge&logo=medium&logoColor=white" alt="Medium"/></a>
+  <a href="https://x.com/LyleOkoth"><img src="https://img.shields.io/badge/X-000000?style=for-the-badge&logo=x&logoColor=white" alt="X"/></a>
+  <a href="mailto:lyleokoth.mlengineer@gmail.com"><img src="https://img.shields.io/badge/Email-D14836?style=for-the-badge&logo=gmail&logoColor=white" alt="Email"/></a>
+</p>
+
+|  |  |
+| --- | --- |
+| **Name** | Lyle Okoth |
+| **Email** | [lyleokoth.mlengineer@gmail.com](mailto:lyleokoth.mlengineer@gmail.com) |
+| **GitHub** | [@lyle-mlengineer](https://github.com/lyle-mlengineer) |
+| **Portfolio** | [lyle-mlengineer.github.io](https://lyle-mlengineer.github.io) |
+
+Machine learning engineer working on speech and voice AI by day, and building coAgent to make a free Colab GPU behave like a dependable teammate the rest of the time.
+
+## Attribution
+
+- **Model** — [Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct) by Alibaba, with [Qwen2.5-Coder-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-3B-Instruct) as the validated fallback.
+- **Inference** — [Hugging Face Transformers](https://huggingface.co/docs/transformers), [bitsandbytes](https://github.com/bitsandbytes-foundation/bitsandbytes) for NF4, and [outlines](https://github.com/dottxt-ai/outlines) as an optional constrained-decoding path.
+- **Context** — [tree-sitter](https://tree-sitter.github.io/tree-sitter/) for the repo map, BM25 for retrieval, and [nbformat](https://nbformat.readthedocs.io/) for notebook handling.
+- **Editing format** — SEARCH/REPLACE blocks follow the format proven by [Aider](https://aider.chat/); the multi-phase loop and tool protocol are coAgent's own.
+- **Quality** — [pytest](https://pytest.org/) and [ruff](https://docs.astral.sh/ruff/) are the reference verifier commands in the eval suite.
+- **Platform** — [Google Colab](https://colab.research.google.com/) free tier; notebooks and Drive are the runtime and the artifact store.
+- **Artwork** — the cover was composed for this project by Lyle Okoth; badges come from [shields.io](https://shields.io/) and brand marks from [Simple Icons](https://simpleicons.org/).
+- **License** — released under the [Apache License 2.0](./LICENSE).
+
+---
+
+<p align="center">
+  <i>The runtime is cattle. The repo is the pet. Every step leaves a commit.</i><br/>
+  <a href="https://lyle-mlengineer.github.io">lyle-mlengineer.github.io</a> · <a href="mailto:lyleokoth.mlengineer@gmail.com">lyleokoth.mlengineer@gmail.com</a>
+</p>

co_agent_coder-0.1.0.dist-info/RECORD

@@ -0,0 +1,51 @@
+co_agent_coder-0.1.0.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+coagent/__init__.py,sha256=_b7mtofLHlkkCXPDvzRcYDdMUhyhSgJFVvOeQ9kB8LQ,3022
+coagent/__main__.py,sha256=EEOfn5gpJDQLJJC9H8-HFjez5tXMGS34M2S-B69weXo,141
+coagent/cli.py,sha256=JKhXBTyiYL6hFqAMCBPyTKjtlzfpFWklFPR4xIiYxO4,13670
+coagent/config.py,sha256=0g0D9FkF0sXh4aO5_0AFjj-UGGCKP6a8yvx0kdWWGAY,10093
+coagent/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+coagent/agent/__init__.py,sha256=cNEFSCsmlF0yfASE0sg88wwYgsvzdZaMIKgjumTrvZ8,687
+coagent/agent/budget.py,sha256=iY02BcCQT3TrJfsh9H6i6ze4er8GuLnhVPXl0j0l6o4,7408
+coagent/agent/loop.py,sha256=drU4VIvTJhM_6g3eEwfjL0y5WvSZ2AIPYhqaHigDpXA,45517
+coagent/agent/prompts.py,sha256=6SYPbo6R6Sw3R4D2p-o1Z49n7chJPUwm8Mbp4KEu-FQ,7133
+coagent/agent/state.py,sha256=oBdlCifRM0gRys2A0wyzfyimgqxXUlRqbpjd09nfn0s,6036
+coagent/context/__init__.py,sha256=aXiSZxI0xKM8iSCM8vzZcrNRlLzXcU-tfJn5GLqGsQE,684
+coagent/context/artifacts.py,sha256=tMnDjT4vMmKQKEd2XxP--FYBuZApRKqvaj5NbxMrlGA,16127
+coagent/context/repo_map.py,sha256=d6W0AMXrINy0hEhpHzegDbrlsq275NL1roiOmCwtpig,23158
+coagent/context/retrieval.py,sha256=eHs6UVTb9drJyGmKXnNbqkGcZRv2aL0myjiYsorkcIs,13693
+coagent/context/tokens.py,sha256=LseaowPXv8C0l6voBSjJIdpqOM69L1V50KQhOV_-QhE,3485
+coagent/evals/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+coagent/gitx/__init__.py,sha256=pp90SV0dr0Fv6wClRr78RCPTndUm0tQp9LgjrgaO3zs,1201
+coagent/gitx/commit.py,sha256=cvwTco_XUZz0lXtB8KocwpYHn5C1Ij5Zl3mUxVltxuE,3107
+coagent/gitx/durable.py,sha256=7UHXneYTwTdXmtnwthUJH1Z59niFuziMH0w3SjlXF6g,4756
+coagent/gitx/git.py,sha256=a-_Di5spgfhWlVCxgrzZgMHah29HeT_JInloPbw8dw8,14166
+coagent/gitx/lock.py,sha256=eetFXKSfihDgHVXIpG9o67l-74xHzSVJpbIN1EYVRyU,5235
+coagent/gitx/log.py,sha256=9NKrOOCNzpMwKqhzXeoNPKmJ1n9fV4mCJURjv8pjVjU,2695
+coagent/gitx/resume.py,sha256=-MfRyjxjXKTpg563w5X2RbDm0I8fyBCS7g6sxcAYnfA,7709
+coagent/gitx/session.py,sha256=buo04E1BeSGQiO2azBaNYYMdu1VjY2UyxAhpLtUabCc,7887
+coagent/llm/__init__.py,sha256=mmE_Ogsz1C5ODZ9aKeFyi4pDZcb8WAQKdT4_ZM4onqA,1866
+coagent/llm/base.py,sha256=QSbofGBFWQ9KOYie5VW0PGEN17kiOnUZbFQMyqYvvGg,7475
+coagent/llm/grammar.py,sha256=7XLs5wHVVb9uEN7hr4HUHay3R7-kJ4kkRJqoTxNMFv0,10652
+coagent/llm/registry.py,sha256=ykMgpV_yoEQGdplIpXOCNuK0n7JlAJ0Xib7KZNo7Z4U,4286
+coagent/llm/transformers_backend.py,sha256=6JxETIVWbs7SyoyGZCaxHo15GN8Nr-xVXyvYhvDWLew,11062
+coagent/safety/__init__.py,sha256=0EthXjEoNes1efsfl6lynWnfcfQrxpR8uG--AYcPL9c,572
+coagent/safety/denylist.py,sha256=--YH_ERy-afNGRM8saZ-f6Y5iM7WeJ_GYRrWis9irSk,5750
+coagent/safety/policy.py,sha256=K4WH-6bSkOL-kgR_Q9VQ6GHlNJwpgfJG5giFzEhOM9E,7664
+coagent/safety/redact.py,sha256=MMVub3MT1SPvYkHF5LIXpyMGNRsjAjyeCeubZRcDuGQ,5312
+coagent/templates/config.toml,sha256=eFvc6dHnKGucqDxBLntBhzWn_cxlvjcMC_LvICBhdl4,3485
+coagent/tools/__init__.py,sha256=MvLH2FV3Le2toUGalFMfPfSn6N7koriT4hjCSguzKMk,5139
+coagent/tools/artifacts.py,sha256=14TLC8lWp-P0wXjpSrN-HaTrIgqMxgDSGIzPhNSC_b4,11839
+coagent/tools/base.py,sha256=FUDuLtbywYOyCY_UaXKtOVWZz658-BDWzU5sVe9mgnM,7953
+coagent/tools/fs.py,sha256=uCxmsvR9z6R4_eU7ZigGdHidrKl6iSyneatJvN2egRU,20026
+coagent/tools/notebooks.py,sha256=XHa7CCfOMo4UB_j2NCbXrSPbdILIzhiidLdWA3MjljE,34308
+coagent/tools/pyt.py,sha256=xTllc6bPMIn2EcNjQykMjn80A5bXCJKHl_09-BbaLiY,8852
+coagent/tools/shell.py,sha256=Juvqd5kv9jRfLJBngQ7lU7C4I7KnQyn0UWLFcrX4EfE,20186
+coagent/ui/__init__.py,sha256=fxd_Al3t220lTmS-7mwvNKEVvrWf6gCiuWcg2SxRU7I,1323
+coagent/ui/gradio_app.py,sha256=Od16fWGVLcgvOcqA668o4VnFJZcxdT5CawJ1SH6R4rM,13916
+coagent/ui/notebook.py,sha256=GPOmZgnsq3M6uKs6zpfAlP_WXa4VoXQH447DzaPkXdY,12535
+coagent/ui/render.py,sha256=S6zUfghey8L9C0yCra3MpFPjaUGW4ykKonZ0t3-Krjc,15379
+co_agent_coder-0.1.0.dist-info/METADATA,sha256=qYz_UydmxPhWny7CHMjA1rzL0XUm-U9mWP_qP-qT1GE,25515
+co_agent_coder-0.1.0.dist-info/WHEEL,sha256=mOdXR9C3wMVov7cSY8PY-XCzo3SmmWCHvGQr5csJoao,91
+co_agent_coder-0.1.0.dist-info/entry_points.txt,sha256=laA86SHKnwDygrhPCy9ngNY6sa0UgMnfxZVsmhwgiM4,133
+co_agent_coder-0.1.0.dist-info/top_level.txt,sha256=TBagh83PxN48jm4UttI1iaEcHpKrqsv3V5kJGbC0-mk,8
+co_agent_coder-0.1.0.dist-info/RECORD,,

co_agent_coder-0.1.0.dist-info/WHEEL

@@ -0,0 +1,5 @@
+Wheel-Version: 1.0
+Generator: setuptools (84.2.1)
+Root-Is-Purelib: true
+Tag: py3-none-any
+

co_agent_coder-0.1.0.dist-info/entry_points.txt

@@ -0,0 +1,5 @@
+[coagent.backends]
+transformers = coagent.llm.transformers_backend:TransformersBackend
+
+[console_scripts]
+coagent = coagent.cli:main