# Interactive Planning Workflow

**Status:** Approved target-state baseline  
**Activation:** I2.

## Target State

The Planner leads a structured planning conversation while updating durable Planning state continuously.

```text
User input
  -> resolve current topic
  -> update Decisions/Findings/Risks/Artifacts/Deferrals
  -> evaluate current gate
  -> identify highest-value next unresolved item
  -> recommend/infer where evidence is sufficient
  -> ask at most one consequential owner question when genuinely required
  -> repeat until Ready to launch
```


## Planner Model Invocation Contract

Every substantive conversational Planner turn in I2 is backed by an actual model invocation through AWP's canonical Account/Model/Provider and Model Gateway seams. A deterministic state machine may validate and apply structured outcomes, but it cannot substitute for the Planner Agent interview.

Effective routing is explicit and visible:

```text
PlanningSession override
 -> Project planning.interviewerProviderId / planning.interviewerAccountId / planning.interviewerModel / planning.interviewerReasoningEffort
 -> System defaults for the same four routing dimensions
 -> configuration required
```

Planning never silently borrows the Factory execution model/account. Provider, account, model, and reasoning effort are explicit selections. Model names and supported effort levels come from the selected account/provider model catalog rather than a hard-coded or free-text-only UI. Each model-produced turn retains immutable provider/account/model/reasoning-effort provenance and the PlanningSession revision it reasoned over. Model/provider failure leaves the current structured state intact, records a visible failed/retryable turn state where applicable, and allows retry or explicit model/account change without rewriting history.

The model proposes conversational interpretation and structured effects. Application validation owns the mutation: stale revisions, invalid item transitions, OwnerRequired/PolicyRequired authority, deferral gates and launch policy remain fail-closed.

## Planner Turn Contract

1. Resolve/respond to current user input.
2. Update structured Planning state/artifacts.
3. Surface new Finding/Risk/OpenQuestion.
4. Decide whether current area is sufficiently resolved.
5. Select highest-value next unresolved planning item.
6. Lead into it with recommendation, decision brief or one consequential question.

Never default to passive `What would you like to discuss next?` when Plan state already knows.

## Agenda Priority

```text
blocking dependencies
high-impact unresolved Decisions
risks that can invalidate downstream work
missing information required by next stage
time-sensitive external setup/research
low-cost choices unlocking multiple downstream artifacts
```

Expose `Now / Next / Later`. User may reorder/park without switching to Manual mode.

## Simple Mode

- Ask OwnerRequired/PolicyRequired choices.
- Infer DelegableExpert choices from evidence/defaults and record them visibly.
- Show technical plan building in Plan Index/context summaries; do not present every technical item as chat choice.
- Escalate low-confidence consequential technical choices.
- Ask owner-level consequences instead of implementation vocabulary.

## Expert Mode

- Technical recommendation becomes conversational review item.
- Recommendation-first with 2–4 credible alternatives where useful.
- `Accept recommendation` is primary action.
- User can reject one dimension without reopening unrelated accepted decisions.

## Recommendation Contract

Show confidence, basis, consequences and what would change recommendation. Weak evidence yields no fake strong recommendation.

## Profile / Requirement Changes

Changing PlanningProfile previews added/removed/kept requirements. Accepted Decisions/artifacts are preserved; no-longer-required items become retained/not-required, not silently deleted.

## Existing Project Conflict

When repo/docs/user/recommendation conflict, show each truth source and require/record resolution only when consequential to next stage.

## Research

Planner may recommend conditional validation. User can accept, adapt or explicitly skip with bounded risk. Research does not create a mandatory stage for every Plan.

## Readiness

Planner stops interrogating an area when required items are satisfied and remaining uncertainty is accepted/bounded/deferred/irrelevant. It automatically advances.

## Launch Handoff

At Ready to launch, show concise PlanRevision, outcome/scope/non-goals, delivery/execution, connections/approvals/gaps/risks/preflight and choices `Start now | Schedule | Park`, followed by one Confirm.

## Failure / Recovery

Autosave every material change/draft. Browser refresh, session compaction, agent/model change, auth redirect and provider outage must resume exact active item/agenda from structured state.

## Increment Realization

I2 activates full workflow. I3 deepens generic Approval/autonomy; I5/I6 consume already-planned quality/delivery outputs.

## Acceptance

During long planning, user can always answer what was asked, what is happening, what was decided, why, what remains and what happens next without reconstructing chat history.