OpenPlan Operate

Use this after $openplan-core when the agent is about to do work, delegate work, or run unattended.

Session Init

1. Read the user's latest request.
2. Identify local instructions in this order when present:

• system/developer instructions;
• repository rules such as AGENTS.md, CLAUDE.md, RULES.md, ENTRYPOINTS.md;
• task-specific specs, plans, or review targets.

1. Load only files needed for the next action.
2. Detect whether the project has two surfaces:

• goal/intent/design/docs surface;
• implementation/code/execution surface.

1. Identify configurable roots when present:

• docs root such as .openplan/, docs/, repo root, separate docs repo, or user-provided path;
• code root when separate from docs;
• temporary output root for disposable briefs.

1. Classify the mode:

• align;
• research;
• design;
• implement;
• test;
• review;
• operate/dispatch;
• stop and report blocked.

1. State any assumption that affects the route, including assumed docs root.

Behavior Constraints

• Understanding-only requests do not edit files.
• Review-only requests do not fix files.
• Auto/YOLO mode still must preserve human-owned decisions.
• Do not switch branches, delete untracked state, clean working directories, or perform destructive recovery unless explicitly asked.
• Prefer existing project infrastructure and language features.
• Avoid hardcoded roles, paths, permissions, or backend names when the project is config-driven.
• Avoid hardcoded documentation roots. Use the user-provided docs root or existing project convention before proposing .openplan/, docs/, or repo root.
• Keep comments and explanations minimal unless they clarify non-obvious logic.
• If a required artifact is missing, report the missing artifact and why it matters.

Subagent Constraints

Use subagents only when delegation is explicitly requested or platform policy allows it for the task.

Subagents must receive:

• one concrete objective;
• minimal file/path context;
• read/write scope or read-only status;
• expected output contract;
• whether they may edit files;
• whether they may run commands;
• how to report blockers.

Subagents must not:

• ask the human directly when the parent is the interaction boundary;
• make acceptance, bypass, merge, scope, or architecture decisions;
• review their own authored work;
• expand scope without reporting;
• hide uncertainty;
• return ungrounded approval.

Result Contract

Every delegated or automated operation should return:

Objective:
Scope:
Actions taken:
Evidence:
Findings or output:
Open questions:
Human decisions required:
Residual risk:

References

• references/init-protocol.md: detailed initialization checklist.
• references/subagent-protocol.md: delegation prompt and result contract.
• assets/claude-code-command.md: Claude Code command/prompt seed.

Use $openplan-handoff before ending a long session, delegating complex work, or crossing from one agent platform to another.

Use $openplan-research when the selected route depends on evidence that is not already in loaded durable context.