code-planning

name	code-planning
description	Build an ordered executable code plan for controlled experiment enablement
argument-hint
allowed-tools	Bash(*), Read, Edit, Write, Glob, Grep, Agent
license	MIT

Code Planning

Mission

Translate the idea into an ordered executable code plan that supports:

1. All-components science conditions where every canonical idea component is enabled
2. Component-disabled science conditions using the same runner interface with one or more components disabled

CRITICAL: Code must be SELF-CONTAINED in project/

• All code must be COMPLETE and INDEPENDENT in project/.
• The code must NOT depend on repos/ for core functionality.
• Everything needed to run experiments must be in project/.
• Repos are for REFERENCE ONLY - implement actual experiment code in project/.

Protocol

• Write the managed artifact code.plan at agent_reports/code/plan/latest.json with ordered stages, project target paths, expected outputs, and verification commands.
• Map each idea.json component to code implementation node(s).
• For each component node, document the component-disable mechanism (how to turn that component off through the unified runner).
• Sequence stages so each worker can finish with a clean, runnable project/ state before the next stage starts.
• Generate the managed code.blueprint artifact before delegating implementation.
• End the plan with a mandatory final_integration_smoke node that uses the real prepared dataset and the real API/model path when required.
• The real benchmark data from dataset_candidate/ MUST be used in experiments, not synthetic data.
• Keep planner outputs, worker reports, reviewer reports, hook reports, phase summaries, and controlled artifacts under the structured agent_reports/code/ layout.
• Merge worker outputs into agent_reports/code/summary.md, optional agent_reports/code/usage.md, agent_reports/code/artifacts/integration_readiness.json, and the reviewer-approved phase report agent_reports/code/phase.json.

Report Requirements

The runtime owns reports. Do not write report files with tools. Each step keeps:

1. agent_reports/code/worker/<step>/attempts/NNN.json and latest.json
2. agent_reports/code/review/<step>/attempts/NNN.json and latest.json
3. agent_reports/code/hook/<step>/attempts/NNN.json and latest.json

Phase-level reports:

• agent_reports/code/plan/latest.json - Ordered step plan
• agent_reports/code/plan/planner_report.json - Planner summary
• agent_reports/code/summary.md - Human-readable code status
• agent_reports/code/usage.md - How to run experiments (optional)
• agent_reports/code/artifacts/integration_readiness.json - Readiness status
• agent_reports/code/phase.json - Final phase verdict

Hard Rules

• Do not claim a condition is enabled unless a runnable command exists.
• Do not rely on parallel code execution; the runtime executes the ordered stages list and gates each stage before moving on.
• final_integration_smoke must not be satisfied by imports, mocks, or dry-run-only evidence.
• Every idea.json component must have a corresponding component-disable mechanism in the plan.
• The plan must include runnable entrypoints/configuration for all-components and component-disabled science conditions.
• Every step must finish through the worker/review/hook prefinish gate.