| name | coding-agent |
| description | Plan a coding task, generate and execute Python, and loop on reviewer feedback until the code passes -- a plan/code/judge fix-loop with governed execution. |
| subgraph | graph.yaml |
coding-agent
Plan → code → review fix-loop. The code node generates a
self-contained Python script and runs it through std.python_exec@1
(capability-gated tools:std:exec, default-deny); a judge node
reviews the code plus its execution result against the rubric. On
fail, the reviewer's rationale is re-injected into the brief and the
loop retries; on pass, the graph halts.
State
- Input:
task -- what to build.
- Outputs:
code (the approved script), run_result
(stdout/stderr/exit_code envelope), verdict, score, rationale.
Use
- Standalone:
stargraph run graph.yaml (grant tools:std:exec; set
the LM via --lm-model/--lm-url or per-node config.model).
- As a subgraph:
kind: subgraph, spec: <path to graph.yaml>;
config.max_steps bounds the fix-loop.
Tuning
Edit graph.yaml: the review rubric, the brief template, or the
code node's timeout_s.
