Plan Mode & Mission Ledger

Navin's Plan Mode is the orchestrator that decides how the agent plans, delegates, verifies, and recovers. It is inspired by Magentic-One's Task Ledger / Progress Ledger loop and by Deep Agents' harness ideas (structured plan, sub-agents, skills, MCP), implemented natively on Navin's board and AgentLoop - without depending on LangGraph or the Deep Agents package.

Levels

Intent	Command	Behavior
Plan only	`/blueprint`	Goal analysis + Task Ledger + board tasks. No code execution.
Plan then build	Build button → `/forge`	Execute the plan with Progress Ledger per step.
Autopilot	`/cruise`	Plan, execute, test, detect stalls, local replan until done or paused.
Long mission	`/mission`	Multi-turn durable goal + ledger + checkpoints + resume + final report.

Composer UI modes stay plan / agent. Tools stay Navin's (board, spawn, verify, test_run, files, shell, MCP, …).

Orchestrator protocol

On every step (/forge, /cruise, /mission):

Read ledger (board ledger_get) + board next.
Choose one ready step (or spawn independent ready steps when appropriate).
Act with tools.
Verify according to step validation (test / lint / verify / manual / none).
Write ledger_progress with evidence.
Decide: next | retry | ledger_replan (local) | ledger_pause.
Never mark done without evidence when validation requires tests.

If missing_info blocks progress → pause and ask the human. Do not invent facts.

Data model

Persisted at .navin/board/mission.json next to board.json.

Task Ledger: goal, status, version, constraints, facts, missing_info, acceptance_criteria, steps[], history[] (why the plan changed).

Progress Ledger: current_step_id, evidence, stall_count, loop_detected, replan_needed, budgets (max_retries, max_replans, max_stalls, optional token/cost), next_actor.

Board tasks carry optional acceptance, validation, retry_count, agent, evidence. Marking done with validation in test|lint|verify is rejected without evidence.

Runtime

Each turn, Runtime Context injects a short board digest plus mission ledger lines (goal, version, stall, pause, protocol reminder). Budgets and stall/loop detection can auto-pause the mission.

Pause / resume / manual edit

board ledger_pause / ledger_resume
Human edits via ledger_update bump version and append history (actor=human)
Chat plan panel shows goal, vN, stall badge, last change reason, and Build for /blueprint handoff

Sub-agents

spawn briefs should copy the step's acceptance + validation. The parent merges evidence before done / ledger_progress.

Equivalence (inspiration → Navin)

Capability	Navin primitive
Structured todos / plan	Board tasks + mission ledger
Sub-agents	`spawn` + project agents
Skills / MCP	Existing skills + MCP tools
Durable state / resume	`mission.json` + `/checkpoint` + `/goal`
Controlled loops / replan	Progress Ledger + `ledger_replan`
HITL	Build panel, pause, `next_actor=human`

Examples

/blueprint Add GitHub OAuth without breaking Google Auth
# review plan panel → Build

/forge Implement the approved plan

/cruise Ship dark mode end to end with tests

/mission Migrate auth to Supabase providers across the monorepo

Board Autonomy - consent, isolated branches, PR on done, issues sync
Modes
Commands
Skills