agent-execution

name	agent-execution
description	Use for Agents when implementing a plan. Prioritize `just` tasks, allowlisted native commands, and end-to-end completion before final reporting.

Agent Execution Discipline

Follow these rules for every implementation task.

1. Tool And Command Discipline

• Prefer native tools first.
• For shell commands, use allowlisted/native commands and keep them minimal.
• Prefer just tasks over direct command variants when a justfile target exists.
• For experiments, run and poll only through just run-experiment, just poll, just poll-raw, and just status.
• Do not run background experiments with direct uv run python ....

2. Approval Avoidance

• Avoid command shapes that commonly trigger approvals when an equivalent allowlisted command exists.
  • Especially for edits, these kinds of approval requirements should Not be triggered.
    • cat, heredoc (<<), redirection (>), uv, rm, uv
• Prefer single, direct commands over complex chained shell constructs.
• Reuse existing project scripts/tasks instead of ad-hoc command composition.

3. Execution Persistence

• Start by using the ToDo tool and follow that until completion.
• Continue until all plan steps are implemented, validated, and summarized.
• Do not pause for “check-in” messages.
• Unblock yourself using best practices and documentation.
• Note down any special actions, decisions etc required during implementation in notes.md.

4. Completion Contract

• End only after finishing all steps in the plan.
• Validate outcomes with the project’s standard commands (prefer just test, just lint, or other relevant just targets).
• Make sure to test out whatever feature was built end to end in a non-interactive manner.
• Provide the final response only after implementation and validation are done, including:
• What changed.
• What commands were run.
• What remains (only if genuinely blocked).

5. Implementation Standards

• No Fallbacks (Fail Fast): Never write defensive fallback logic (e.g., catching generic exceptions, guessing default values, or silently handling missing dependencies/configurations).
• Explicit Contracts: If an expected state, key, or configuration is missing, fail immediately and explicitly with a descriptive error (e.g., ValueError, KeyError). Code must loudly reject invalid or ambiguous conditions instead of guessing.