local-first · v0.1

Universal Planner Runtime

Phase XII · V9 — the permanent planning layer. Given V8's immutable DecisionContext (goal/priority/action candidates, a risk assessment, a confidence model), it produces a deterministic, APPROVABLE execution plan: stages, execution order, explicit dependencies, approval requests, a rollback plan, and estimated requirements. Every step cites its originating cognition. Fail-closed: missing cognition→UNKNOWN, capability→UNPLANNABLE, provider→UNAVAILABLE, approval→WAITING_APPROVAL, contradiction→PLAN_INVALID. It performs the planning V8 forbids itself — but executes nothing, invokes no provider, mutates no memory, changes no cognition. Read-only.

3 / 5
Planned / blocked
11 / 11
Stages / steps
6 (4)
Approvals (required)
5
Rollback steps
Deterministic, approvable, unexecuted plans. V9 consumes V8's DecisionContext read-only and orders its goal/priority/action candidates into stages and steps. Every step cites the V8 action candidate it originates from and carries the underlying memory evidence, so nothing is planned beyond the reasoning. Risks and assumptions are explicit; dependencies are explicit and acyclic; high/critical-risk steps require approval and irreversible steps are flagged (never auto-rolled-back). The plan is only ever a proposal — execution stays behind the Broker (→ V5). Missing inputs fail closed to an explicit status, never an invented plan.

The Universal Planner Runtime is the permanent planning layer of GIOS: every future simulator, predictor, autonomous agent, and orchestration engine inherits this deterministic planning layer instead of implementing planning independently. It consumes V8's DecisionContext (the goal / priority / action candidates, risk assessment, and confidence model produced by the reasoning layer) and transforms it into an ordered, approvable execution plan — stages derived from the prioritized goals, a deterministic execution order, explicit stage and step dependencies, approval requests for the steps that mitigate high or critical risk, a rollback plan over the reversible steps, and estimated requirements expressed as classes and counts rather than fabricated numbers. Every step cites the V8 action candidate it originates from and carries the underlying memory evidence, so the plan is traceable all the way back through cognition and memory to the observed execution outcomes, and nothing is planned beyond the available reasoning. It fails closed: missing cognition yields UNKNOWN, a missing capability yields UNPLANNABLE, a missing provider yields UNAVAILABLE, an ungranted approval yields WAITING_APPROVAL, and contradictory constraints yield PLAN_INVALID — never an invented plan. It prepares an immutable handoff for a future Simulation Runtime (V10). It never executes, never invokes a provider, never mutates memory, never changes cognition, and never performs a side effect — execution stays behind the Broker (→ V5) — and it inherits the A0 Intelligence Kernel lifecycle.