Project, repos, instructions, prior work, constraints, capabilities, conflicts.
ContextSnapshot
Agent OS · Mission Control · Issue #14
From incomplete idea to autonomous execution
One guided operator flow for defining the work, gathering evidence and references, creating and critiquing the plan, approving authority, and running safely.
01
Define the work
Preserve Brian's raw intent, compile live context, then lock outcomes instead of jumping from prompt to tasks.
Problem, audience, requested solution, attachments, assumptions, why now.
WorkRequest
Goals, expected outcomes, non-goals, appetite, success signals, closeout proof.
OutcomeContract
02
Discover and ground
The operator approves what to learn, where to look, how much to spend, and what counts as enough.
Questions, source classes, freshness, scouts, budgets, stopping rules.
DiscoveryPlan
Cited findings, repo SHAs, experiments, conflicts, confidence, open gaps.
EvidenceBundle
Functional, technical, product, design, asset, and anti-references.
ReferenceCollection
03
Create the ideal plan
Turn evidence into a selected direction, testable requirements, solution design, task graph, and independent critique.
Credible alternatives, tradeoffs, cost, reversibility, recommendation.
OptionSet
Behavior, quality, edge cases, negative requirements, acceptance, traceability.
RequirementSet
User flow, architecture, trust boundaries, recovery, migration, proof.
SolutionDesign
Task DAG, dependencies, owned surfaces, routing, estimates, checkpoints.
PlanCandidate
Completeness, feasibility, simplicity, UX, security, cost, proof, autonomy.
PlanReview
04
Approve autonomous authority
A versioned execution contract—not a prompt—defines exactly what Run authorizes.
Resolve evidence, coverage, authority, budget, verification, rollback, and stop-gate failures.
ReadinessDecision
Simulate, run selected branches, run to checkpoint, or run all approved work.
ExecutionContract
05
Run, verify, and learn
Healthy work stays quiet; the operator can steer, pause, interrupt, constrain, switch, or take over throughout.
Native sessions, action trust, capability leases, receipts, checkpoints, bounded repair.
MissionRun
Material changes show impact and require reapproval; safe local choices remain autonomous.
PlanDelta
Outcome proof, independent review, artifacts, cost, residual risk, routing feedback.
Closeout + Evidence
Mission Control · one operator surfaceConversation, mission brief, discovery workspace, plan workspace, live run, attention, evidence.
↓
Guided inception and plan studioIntent → outcomes → research → references → options → requirements → design → critique → ExecutionContract.
↓ approved contract
Durable control planeMission DAG, event journal, context compiler, budgets, schedules, runtime/model router, evaluations.
↓ native session selection
Sessionful runtime gatewayCodex, Claude, xAI/Composer, OpenCode, Goose, Cline, OpenHands, Hermes, OpenClaw, and future adapters.
↓ proposed effect
Action trust kernelActionIntent → deterministic policy → approval/capability lease → constrained executor → ExecutionReceipt.
↓
Constrained workspaces and mediated connectorsWorktrees, containers, remote hosts, GitHub, browser, deploy, messaging, databases, and SaaS.
Runtime and model are separate choices. Every consequential effect is governed independently of the model that proposed it.