development-lifecycle

name	development-lifecycle
description	Use when frontend/React/TypeScript/UI work. Guides understand -> plan -> grill -> TDD -> /go (4b->5). Alias: /work.

Development Lifecycle

Auto-detect phase, guide process.

If the user says "plow ahead", "do not stop", "use your best judgment", "keep going until done", or similar, use /plow-ahead as the autonomy contract: make reversible assumptions, continue through routine ambiguity, and stop only for true blockers.

Ponytail default

Ponytail before implementation: run /deslop full (write mode) during planning and keep it active through /go: delete, stdlib, native, installed dependency, one-line before custom code.

Phases

0. Worktree (automatic, silent if already isolated)

Feature branch? Skip. On main/master/develop? Spawn worktree before code. [ETHOS: Worktree Isolation]

scripts/mux-worktree.sh <type>/<branch-name>

Claude invoke silent when phase 1 start on default branch. User never run.

1. Understand

• Explore | clarify one-at-a-time | new->2-3 approaches+tradeoffs | bug->failing test->root cause
• Claude-hosted: background agents may explore alternatives, prior art, and edge cases in parallel.
• Native Codex: explore inline unless the user explicitly requests agents or invokes /swarm; skill activation alone is not consent.
• Mixed patterns area? Refactor to single pattern FIRST before add features
• GATE: no impl code until approach approved.

2. Plan

• Every step: exact file paths, exact code, expected output. No placeholders.
• Non-trivial feature -> /resilience-review for edge cases/error handling/polish or skip reason.
• Bite-sized tasks (2-5 min each)
• UI work: use /prototype for 2-3 runnable UI variations, review with user, pick best. See REFERENCE.md.
• 5+ tasks -> stacked PRs (one per logical group)

2b. Grill

GATE: no impl until plan survive grilling.

• Auto-invoke /grilling | grill until every branch resolved | update CONTEXT.md + ADRs inline
• Update plan with changes | get explicit user confirmation
• Skip only if: trivial bug fix AND <3 tasks AND no architectural decisions

3. Implement (TDD)

• Thinker/executor split (Claude-hosted default): the plan that survived grill is the spec. Fable-5/Opus-4.8 stays orchestrator; execution may go to Sonnet or /codex GPT-5.6 Sol. The orchestrator reviews the diff; below the bar -> rerun smarter.
• Native Codex: implement the approved plan inline. Do not recursively invoke /codex or auto-route to /swarm; use native agents only after explicit user consent.
• After plan survives grill, Claude may use /swarm when independent lanes safely accelerate work. In Codex, /swarm runs only when the user invokes it or explicitly asks for parallel agents.
• RED: failing test first | GREEN: minimal code to pass | defensive gaps -> RED tests
• Code is liability: reuse-first (standard library, native platform, already-installed dependency, one-line), then keep only product value, defense, or test confidence.
• Test deletion guard: verify test+assertion count not decrease after GREEN. AI may weaken tests -> reject and redo.
• REFACTOR while green | no setTimeout hacks | run --detectAsyncLeaks

4-6. Ship -- /go

Impl done -> run /go to ship. Handle all:

• 4. Verify -- types + lint + tests + browser smoke
• 4b. Review / Refine -- self-reviewer + adversarial-reviewer; ensure Resilience Review evidence for risky features
• 5. Ship -- /simplify -> /deslop -> /commit-push-pr -> code-reviewer agent
• 5b. Iterate -- monitor CI -> /resolve-pr-feedback -> AI self-review: up to 3 rounds, early-exit on clean; human review: address ALL (hook-enforced)
• 6. Compound -- codify lessons as .claude/rules/

See /go skill full details. See REFERENCE.md phase-specific checklists.

Phase Selection

Full flowchart REFERENCE.md#phase-flowchart.

User says	Phases
"Build a new feature"	1->2->2b->3->/go
"Fix this bug"	1(reproduce)->3(TDD)->/go
"Refactor this module"	1->2->2b->3->/go
"Write tests for X"	3 only
"Ship it" / "Create a PR"	/go only
"Quick question"	Just answer