octaviantocan Workflow: Plan
Phase 1 locks intent with the user. Phases 2-6 break the work into steps, dispatch one sub-agent per step, compose an overall plan, critic-review it, and save.
References: references/writing-plans.md (format conventions), references/step-planner-prompt.md (Phase 3 sub-agent prompt), references/critic-prompt.md (Phase 5 critic prompt).
Process
┌──────────────────────────────────────────────────────────────────┐
│ CENTRAL PLAN AGENT │
│ │
│ Phase 1 Brainstorm with user → .context/plans/brainstorm.md │
│ Phase 2 Gather context, break into steps, decide ordering │
│ Phase 3 Dispatch one step planner per step (parallel/seq) │
│ └─▶ .context/plans/step-N-<name>.md │
│ Phase 4 Compose overall plan from step files │
│ Phase 5 Critic sub-agent → cross-step coherence │
│ Phase 6 Address feedback, save to docs/plans/ │
└──────────────────────────────────────────────────────────────────┘
Modes. Warm handoff: if .context/plans/brainstorm.md already exists, read it and skip Phase 1. Cold start: run Phase 1 to produce it. Either way, Phases 2-6 work identically.
Phase 1: Brainstorm
The main agent talks to the user directly during the brainstorm. Do not delegate the user-intent interview; delegate only repo exploration that can run in parallel without deciding product intent.
Hard gate. Do not write code, scaffold projects, or invoke implementation skills until the brainstorm is approved. Every project, regardless of perceived simplicity — "simple" projects are where unexamined assumptions cause the most wasted work. The brainstorm can be short for trivial changes, but it must exist and be approved.
Explore first. Read the relevant codebase before asking anything. Identify existing patterns, files likely to change, related features, architectural constraints. Use Explore sub-agents for broader searches.
Ask clarifying questions. Interview until you reach shared understanding on scope, constraints, acceptance criteria, edge cases, and design.
- 3-5 questions per batch; batch unrelated questions together; only serialize when one answer shapes the next.
- Provide your recommended answer conversationally ("I'd lean toward X because Y — sound right?") so the user has something to react to.
- Keep probing across rounds — stop when all decisions are resolved, not at a question count.
- Scope check first: if the request is multiple independent subsystems, decompose into sub-projects, each with its own plan cycle.
Save to .context/plans/brainstorm.md (gitignored). Present the brainstorm in the active agent UI, iterate on feedback, repeat until approved. Use this template:
# [Feature Name]
## What we're building
One paragraph — problem, what we're solving, why.
## Architectural context
Where in the stack this lives; packages and layers involved.
Include an ASCII diagram (box-drawing characters) for moderate or
complex changes. Skip or one-sentence it for trivial changes.
## Approach A: [Name]
Core idea. Optional ASCII diagram.
**Key files:** Create `path/new.ts` — purpose. Modify `path/existing.ts` — what.
**Trade-offs:** [pros/cons]
## Approach B: [Name] (only if there is a real tradeoff)
## Recommendation
Which approach and why; or, single-approach, key design decisions.
## Out of scope
Present multiple approaches only when there is a genuine tradeoff. Otherwise, one approach.
When to grill. If the domain has dense terminology, conflicting concept names, or unresolved invariants, run workflow-grill-with-docs against the approved brainstorm before Phase 2 — it sharpens terminology and may capture ADRs you'd otherwise re-litigate later.
Phase 2: Gather Context and Determine Steps
Read .context/plans/brainstorm.md, relevant skill files (domain-effect, paw, domain-cli, extension-boundaries, skill-gen, etc.), AGENTS.md, and the key source files the brainstorm names.
Break the work into steps. Granularity is your call — simple work → fewer larger steps, complex work → more smaller steps. Each step is one coherent unit a sub-agent can execute. Decide ordering: independent steps in parallel, dependent steps sequential so later sub-agents can read earlier step files.
Phase 3: Dispatch Step Planner Sub-Agents
For each step, dispatch a planning sub-agent using the runtime's available subagent tool and references/step-planner-prompt.md. Each sub-agent writes .context/plans/step-N-<name>.md with exact paths, full code blocks, verification commands, commit messages, and a decomposition hint. Follow the model/tool rules in AGENTS.md when the runtime exposes those choices.
Phase 4: Compose Overall Plan
Read all step-N-*.md files and compose the overall plan per the "Overall Plan Format" in references/writing-plans.md: goal, architecture (2-3 sentences), overview, packages & files affected (grouped by package, all steps), execution order with dependencies, and step summaries linking each detail file.
Phase 5: Critic Pass
Dispatch a single critic sub-agent using references/critic-prompt.md. It checks cross-step coherence — conflicts, missing dependencies, ordering, gaps vs brainstorm, convention violations.
Serious issues (dependency errors, missing requirements) → re-dispatch affected step planners. Minor issues → fix directly. Cap at 3 iterations; surface remaining issues to the user.
Phase 6: Finalize and Save
Show the user a one-line summary (step count, parallel groups, critic issues resolved) and the clean plan. On approval, save to docs/plans/YYYY-MM-DD-<feature-name>.md, commit docs: add <feature-name> implementation plan. Step detail files stay in .context/plans/ (gitignored).
Hand off: "Plan complete and saved. Ready for execution."