AI Native Tech Plan

AI-native Technical Planning

Write the complete requested Target State unless the user explicitly asks for a prototype or partial target. A plan is an executable contract, not a calendar or staffing forecast.

Preconditions

Read AGENTS.md, .planning/devflow/STATE.md, the active OpenSpec change, and relevant source. Add a Skill Routing Ledger. If unstable platform assumptions, current documentation, plugin/runtime behavior, or local capability evidence affects the design, run capability-research. Open Questions remain means the artifact is draft, not final; resolve them before committing to the plan.

Capability Routing

Resolve decision-resolution, implementation-planning, architecture-guidance, and when domain language is in scope domain-language-modeling through the static DevFlow contract in scripts/workflow_methodology.py. Use grilling for unresolved decisions; DevFlow owns implementation planning; architecture work may use codebase-design, while domain concepts and invariants explicitly add domain-modeling. Record evidence and triggered skills in the Skill Routing Ledger.

Required Shape

1. Target State — final external behavior and internal boundaries.
2. Scope / Non-Goals — explicit write and compatibility boundaries.
3. Architecture Decisions — choices, evidence, and rejected alternatives.
4. Completion Contract — binary acceptance criteria.
5. Critical Path — dependency-ordered required behavior protected from incidental work.
6. Incidental Finding Budget — normally one bounded RED/GREEN cycle inside the approved contract and write set.
7. Escalation Triggers (Authority Delta) — separate technical repair from concrete missing permission/risk; bind standing milestones to exact effects and targets. For model execution, define Standing Goal Execution Authority with exact task/provider/model/existing-auth credential policy/cost policy/serial identity. State that a one-use attempt receipt is technical evidence rather than one-use permission; record actual monetary cost without per-call confirmation, and route non-blocking related work to DEFER_AND_CONTINUE.
8. Capability Slices — dependency-ordered, production-complete slices; each includes implementation, validation, and cleanup.
9. Execution Ledger — owner, write set, status, evidence, and human gate.
10. Continuation Policy — default auto-until-terminal, the next-item selection rule, the six continuation outcomes, and every genuine Human Gate. A phase label alone is not a gate.
11. Generated Artifact Strategy — state whether disposable output exists. When it does, define pre-creation registration, a task/run isolated root or constrained adjacent scope, owner exit, retention, receipt, exact AUTO_CLEAN, repair, and Human Gate behavior.
12. Validation Commands — focused, broad, runtime, docs, and release checks.
13. Project Refresh Impact — for DevFlow changes, record disposition, inspected surfaces, schema decision, contract revision/digest, migration/fixture coverage, and source/release/cache proof; otherwise record why it is not applicable.
14. Project-Directed Implementation Readiness — when the project selects an external implementation provider, record the provider policy, target profile, exact capabilities, consumer/change bindings, evidence contract, limitations, and override policy. Mark unresolved readiness non-executable; do not add provider discovery, installation, activation, or fallback.
15. Risks / Rollback — stop conditions and reversible actions.
16. Goal Mode Prompt and Continue Prompt when recovery risk warrants them.
17. Review Checklist and Final Verification.

Required behavior does not belong in MVP, Future Work, or a later delivery phase. Incidental hardening does not enter the Critical Path unless it affects safe completion; classify and record it through the Incidental Finding Lifecycle instead of silently expanding the plan.

Workflow Gates

Behavior/API/data/integration/migration/compatibility/error-handling changes require OpenSpec before implementation. Apply the Goal Suitability Gate through define-goal when the user requests a goal or the work is long-running, multi-slice, migration/release oriented, cross-context, delegation-backed, or likely to lose its definition of done.

Add a SubAgent Strategy for independent Capability Slices. Record authorization state, disjoint write sets, Agent Task Contract path, main-agent-owned artifacts, and fallback. No worker starts before that contract is valid.

If a readiness Requirement applies, every product-mutation slice and mutating delegation depends on a current Ready receipt plus its ordinary authority. A direction review or draft requirement may precede that receipt; it cannot make the plan execution-ready.

For disposable output, choose one registration-only Generated Artifact Strategy. Prefer a task/run-specific isolated root. A contract must be sealed before creation; a post-creation contract cannot legitimize residue. Record the retention policy and require a fresh AUTO_CLEAN plan plus terminal cleanup receipt before completion.

Canonical Artifact Mapping

Methodology notes are review inputs. Promote approved decisions into OpenSpec or the DevFlow Execution Ledger before they satisfy a gate; canonical files win conflicts.

Output Resources

When creating a durable task ledger, copy and adapt assets/task-ledger-template.md; do not invent a second planning source. Read references/goal-prompt-template.md only when the Goal Suitability Gate says the plan needs Goal Mode or a recovery Continue Prompt.

Completion

The plan is complete when every required behavior maps to a slice, criterion, validation command, owner, rollback path, and ledger item, with no unresolved Open Questions. Each non-terminal item also names its automatic next action so the orchestrator can continue without routine phase confirmation.