PlanWright Control Plane

PlanWright control plane

PlanWright coordinates autonomous coding work over MCP. Humans declare objectives (desired state); agents claim and execute them; every action is written to a signed, hash-chained audit ledger.

When to use

• More than one agent session is running and they risk colliding.
• The user wants coarse objectives, not story points or subtasks.
• The user needs auditable evidence of what an agent did and who approved it.
• The user mentions "Planwright", "objectives", "control plane", or "audit trail".

How to work with it (via the planwright MCP tools)

1. Anchor to the repo first with an explicit project ID (slug lookups fail until anchored). This is the #1 gotcha. Use planwright_set_repo with the repo name (e.g., owner/repo) or project ID.

2. Read open objectives using planwright_list_objectives to see what's available in each lane (backlog, scheduled, in_progress, acceptance, done).

3. Claim one using planwright_claim_objective — claiming locks it so parallel agents skip it. Only claim objectives in the scheduled lane.

4. Decompose and plan — break down the objective into steps, then post your plan using planwright_append_plan.

5. Execute — implement the changes in the codebase.

6. Check in — record your work:

   • Use planwright_record_diff to log commits/diffs
   • Use planwright_submit_test_run to record test results
   • Use planwright_append_note for progress updates

7. Request acceptance — when done, call planwright_request_acceptance to move the objective to the acceptance lane for human review.

8. Take the next objective — repeat from step 2.

Common gotchas

• Must call planwright_set_repo first — all other tools fail without an anchored project.
• Bugs go in the bug lane, not the planning surface — use the bug tools, not objective tools.
• The audit ledger records every change automatically — surface the relevant entry when the user asks "who/what/when".

Tool categories

Session & context: planwright_set_repo, planwright_list_workspaces, planwright_list_projects, planwright_get_board_url

Objectives: planwright_list_objectives, planwright_get_objective, planwright_create_objective, planwright_update_objective, planwright_schedule_objective, planwright_claim_objective, planwright_check_alignment

Agent workflow: planwright_append_plan, planwright_record_diff, planwright_submit_test_run, planwright_request_acceptance, planwright_append_note

Initiatives: planwright_list_initiatives, planwright_create_initiative, planwright_update_initiative, planwright_delete_initiative

Bugs: planwright_list_my_bugs

Context files: planwright_list_context_files, planwright_get_context_file, planwright_push_context_file

Example conversation

User: "I have three Claude Code sessions working on different features. How do I make sure they don't step on each other?"

Agent: I'll set up PlanWright as your control plane. First, let me anchor to your repo and see the current objectives.

[Calls planwright_set_repo with the repo name] [Calls planwright_list_objectives to show available work]

Each session should claim a different objective from the scheduled lane. Claiming locks it, so other agents will skip it and move to the next available objective. When an agent finishes, it requests acceptance and takes another.