coder-agent

Turn issues into working code. In your terminal.

watch a run · docs · results · try it

A terminal coding agent that takes a task in plain English, reads a target repository, plans a change, edits files, runs the tests, and iterates on failures until they pass. Point it at an issue URL and it opens the pull request.

Built to learn the agentic-AI stack end to end, on a zero-cost setup: every model call goes to a free tier, every embedding is computed locally, and every file or shell action goes through a sandbox jail.

git clone https://github.com/Cha-Imaa/coder-agent.git
cd coder-agent
uv sync --extra dev
cp .env.example .env   # fill in GROQ_API_KEY and GOOGLE_API_KEY (both free tiers)

uv run coder run path/to/repo "make the failing tests pass"          # one task, approve each edit
uv run coder chat path/to/repo                                        # several tasks on one thread
uv run coder fix-issue https://github.com/OWNER/REPO/issues/N --pr    # issue in, pull request out
uv run coder index path/to/repo --query "where are durations parsed"  # build and query the index
uv run coder stats                                                    # tokens and fallbacks per run

coder run --help lists the flags: --yes to skip approvals, --sandbox docker, --resume, --model provider:model, --max-iterations, --test-command, --tokens.

One run is one pass through a LangGraph state graph: prepare detects the test command, retrieve_context pulls the relevant chunks, plan writes a short numbered plan, act reads and edits through tool calls, run_tests decides, and reflect feeds the failing output back into the next plan. The model never touches the filesystem itself: every read, edit and command is a tool call served by an MCP server, and every one that writes or executes stops at the approve node first unless --yes was given.

The MCP servers are separate processes talking JSON-RPC over stdio. The file server resolves every path inside the repository and hands run_command to the sandbox — the local jail, or a fresh Docker container with --sandbox docker; the graph sees the same tools either way.

Full walkthrough — what an agent is, why a state graph rather than a while loop, tools as a protocol, sandboxing, retrieval: docs/system_design_coding_agent.md.

Twelve in-house tasks with hidden tests across five categories, plus a thirty-problem HumanEval slice. Every number comes from a script in the repository and a results file under evals/results/.

Every task was solved in the first plan/act/test cycle. The suite is small, so 100% says the loop works, not that the agent is done — a --agent noop run that touches nothing fails all twelve.

Four retrieval arms on the same suite land within three tasks and 15% of tokens of each other, while the same task varies far more between arms than the arms do between themselves: on repositories with three to five files, retrieval mode makes no measurable difference.

K2 Think is faster per task and spends more than twice the tokens getting there: more tool-calling steps, and each step resends the conversation.

About 97% of the tokens go to act reading files and running tools — 29k against under 1k for plan. That is the number codebase retrieval is meant to move.

The results page has the per-category tables, the recall@k numbers and the failure analysis.

uv run python evals/run_evals.py      # all 12 tasks, ~356k tokens
uv run python evals/retrieval_eval.py # recall@k per retrieval mode
uv run pytest                         # the test suite, no API key needed

src/coder_agent/
  cli.py            Typer commands: run, chat, fix-issue, index, stats
  agent.py          builds the graph, MCP clients and checkpointer for one session
  graph/            state, nodes, routing, approval interrupt, context trimming
  mcp_server/       server.py (files + shell, jailed) and github.py (issues, PRs)
  sandbox/          local jail and Docker runner behind one interface
  rag/              loader, tree-sitter chunker, embeddings, Chroma index, hybrid retriever, reranker
  llm.py            provider factory, retry with backoff, fallback chain
  telemetry/        SQLite ledger of tokens, latency, iterations, outcome per run
evals/              the 12-task suite, the HumanEval slice, results and entry-point scripts
docs/               system design, results, the try-it walkthrough, figures, the player
tests/              pytest, no network and no API key needed