gdd-explore

name	gdd-explore
description	Stage 2 of 5 - unified exploration merging inventory grep + design interview. Probes 6 connections, scans the codebase, conducts the AskUserQuestion interview, and writes .design/DESIGN.md + DESIGN-DEBT.md + DESIGN-CONTEXT.md. Use after /gdd:brief to map the existing system and lock decisions before planning. Activates for requests involving researching design direction, gathering references, or exploring visual options.
argument-hint	[--skip-interview] [--skip-scan] [--incremental] [--full]
tools	Read, Write, Bash, Grep, Glob, Task, AskUserQuestion, mcp__gdd_state__get, mcp__gdd_state__transition_stage, mcp__gdd_state__probe_connections, mcp__gdd_state__update_progress, mcp__gdd_state__set_status, mcp__gdd_state__add_blocker, mcp__gdd_state__checkpoint, mcp__gdd_state__add_decision

Get Design Done - Explore

Role: You are the Explore stage. Stage 2 of 5 in the get-design-done pipeline.

Purpose: Unified exploration merging the former scan (inventory grep) and discover (context interview) stages. Produces .design/DESIGN.md, .design/DESIGN-DEBT.md, and .design/DESIGN-CONTEXT.md.

Full procedure detail: ./explore-procedure.md.

---

Stage entry

All STATE.md persistence goes through gdd-state MCP tools - no direct edits. Plain design docs (DESIGN.md / DESIGN-DEBT.md / DESIGN-CONTEXT.md) use Write.

1. mcp__gdd_state__transition_stage with to: "explore". On gate failure: print blockers from error.context.blockers verbatim, do not advance.
2. mcp__gdd_state__get (no args) -> snapshot state for downstream steps.

---

Step 1 - Connection probe

Probe six connections, then batch-write results via ONE mcp__gdd_state__probe_connections call (unspecified connections keep their existing value):

• A - Figma (variant-agnostic): ToolSearch + regex parse of get_metadata / use_figma prefixes -> tiebreaker selection (both-sets > reads-only, figma > others, non-figma-desktop, alphabetical) -> live {prefix}get_metadata call -> available / unavailable / not_configured (with prefix= + writes=).
• B - Refero: ToolSearch presence sufficient.
• C - 21st.dev: ToolSearch mcp__21st presence.
• D - Magic Patterns: ToolSearch mcp__magic_patterns presence.
• E - paper.design: ToolSearch mcp__paper + live get_selection call.
• F - pencil.dev: find . -name "*.pen" file-presence.

Full probe specs + commit-results JSON shape: ./explore-procedure.md §Step 1.

Step 1.5 - 21st.dev Prior-Art Check (when 21st-dev: available)

For each greenfield component in scope: 21st_magic_component_search(component_name, limit: 3). Fit >= 80% -> add <prior-art> block to DESIGN.md recommending adoption; fit < 80% -> note as reference, build custom. If svgl_get_brand_logo available and brand assets are in scope, call per logo and save SVGs to .design/assets/. Skip entirely if no greenfield components in scope. Detail: ./explore-procedure.md §Step 1.5.

---

Step 2 - Inventory scan (unless --skip-scan)

Map pre-check: if .design/map/ exists with all 5 files (tokens.md, components.md, visual-hierarchy.md, a11y.md, motion.md) fresher than src/, consume them and skip the grep pass. Otherwise grep and, after Step 4, suggest /gdd:map for the next cycle.

Parallelism decision: read .design/config.json + reference/parallelism-rules.md. Record verdict via mcp__gdd_state__set_status ("explore_parallel" / "explore_serial"). Parallel -> multiple Task() in one response; serial -> sequential.

Incremental batching (Phase 53, default): --incremental (default; --full opts out) runs the change classifier first via the fingerprint store, groups the DesignContext graph into Louvain community batches, and dispatches mappers per the verdict (SKIP=0, PARTIAL=affected batches only, FULL=all). Engine: scripts/lib/explore-parallel-runner + scripts/lib/mappers/incremental-discover.cjs; dispatch concurrency comes from concurrency-tuner.cjs. Detail: ./explore-procedure.md §Incremental Batching.

Run the canonical scan grep/glob inventory (POSIX ERE, preserving PLAT-01/02): component detection (Glob **/*.{tsx,jsx,vue,svelte}), color extraction (hex / rgb / hsl / Tailwind arbitrary), typography scan (font-family / Tailwind font-* / text-*), motion scan (transition / animate- / @keyframes / framer-motion), token detection (tailwind.config / CSS custom properties / token JSON), layout detection (ordered fallback src/ -> app/ -> pages/ -> lib/ -> unknown). Write .design/DESIGN.md + .design/DESIGN-DEBT.md. Then mcp__gdd_state__update_progress for scan progress. Detail: ./explore-procedure.md §Step 2.

Step 2.x - i18n readiness probe (informational, per D-04): check package.json deps against {react-intl, next-intl, i18next, vue-i18n, formatjs, lingui} -> framework-managed; else grep Intl.(DateTimeFormat|NumberFormat|...) in src/ -> partial; else none. Emit single line Localization readiness: <state> in the report. NO gate, NO blocking - surface signal only (D-07). Detail: ./explore-procedure.md §Step 2.x.

Step 2.5 - Detect prior sketches and project-local conventions

• Sketches: list .design/sketches/* slugs, group by WINNER.md present (completed) vs absent (pending). Record via mcp__gdd_state__set_status: "explore_sketches_present". Include in DESIGN.md "Prior Explorations" section.
• Project-local skills: read ./.claude/skills/design-*-conventions.md -> include in DESIGN-CONTEXT.md <project_conventions> (these override defaults).
• Global skills: ~/.claude/gdd/global-skills/*.md (other than README) -> prepend to <project_conventions> under <global_conventions> sub-block. Project-local D-XX wins on conflict.

Step 2.6 - Graph review (gate then reviewer)

Runs only when Step 2's mapper batch + synthesizer produced .design/context-graph.json. Skip the whole step if the file is absent (pre-Phase-52 project, or --skip-scan).

1. Gate first (cheap Haiku check). Spawn design-context-reviewer-gate via Task() with the classifier signal from Step 2's incremental batching pass: change_pct (fingerprint classifier pct), classifier_action (SKIP / PARTIAL_UPDATE / ARCHITECTURE_UPDATE / FULL_UPDATE), and graph_path: ".design/context-graph.json". Gate emits a single-line {spawn, rationale} JSON decision then ## GATE COMPLETE.
2. If spawn: false: record mcp__gdd_state__set_status: "explore_graph_review_skipped", log the gate rationale (e.g., "project change 3% < 5% threshold"), set telemetry lazy_skipped: true. Done - proceed to Step 3.
3. If spawn: true: spawn design-context-reviewer via Task() with <required_reading> pointing at .design/STATE.md, .design/context-graph.json, reference/design-context-schema.md, reference/design-context-tag-vocab.md. The reviewer runs scripts/validate-design-context.cjs --json as the deterministic floor (checks 1/2/3/5) then layers the semantic checks (4/6/7/8/9) and returns the 9-check verdict inline (APPROVED or REJECT).
4. Capture verdict (read-only contract). The reviewer does NOT write .design/DESIGN-CONTEXT-REVIEW.md - it is reads-only: true, writes: []. WARN findings surface through scripts/lib/health-mirror#getHealthChecks as status: 'warn'; a hard-reject maps to status: 'fail'. Record the overall verdict via mcp__gdd_state__set_status: "explore_graph_review_approved" (or "_rejected").
5. On REJECT: record mcp__gdd_state__add_blocker with the reviewer's blocking-issues list verbatim. Do not advance to Step 3 until the synthesizer or the implicated mapper is re-run. On APPROVED (with or without WARNs): proceed.

---

Step 3 - Design interview (unless --skip-interview)

Run inline - NEVER spawn design-discussant as a subagent. AskUserQuestion only renders the native picker from the top-level skill context; spawning degrades to plain markdown (broken in Claude Desktop).

• 3.a Pre-load context: state.decisions snapshot -> BRIEF.md -> DESIGN.md -> DESIGN-CONTEXT.md <gray_areas> -> project conventions. If figma: available, call {prefix}get_variable_defs and draft tentative D-XX entries.
• 3.b Identify question set: skip areas already covered by D-XX or project convention. Default coverage: cycle goal, audience, brand direction (if no tokens), color/typography/spacing primitives (if undetected), motion preferences, gray areas from DESIGN-CONTEXT.md.
• 3.c Ask one at a time: AskUserQuestion with 4 concrete options + "Other" / "Skip". Reject generic answers ("modern", "clean") with one follow-up.
• 3.d Record after each answer: mcp__gdd_state__add_decision (atomic, auto-assigns D-NN); append a quality-classified JSON line to .design/learnings/question-quality.jsonl.
• 3.e Produce DESIGN-CONTEXT.md: summarize locked decisions, remaining gray areas, Figma tentatives. Frontmatter status: complete.

Full interview protocol + JSON line schema: ./explore-procedure.md §Step 3.

---

Step 4 - Close out explore

• If the synthesizer / mapper batch ran in Step 2: mcp__gdd_state__update_progress with task_progress: "<done>/<total>", status: "explore_mappers_done".
• If Step 2.6 ran the graph review: ensure the verdict status is recorded (explore_graph_review_approved / _rejected / _skipped); on _rejected do NOT checkpoint until re-run.
• mcp__gdd_state__checkpoint - bumps last_checkpoint.
• Stage advance to plan happens at the next stage's entry (plan owns its own transition_stage); do not edit frontmatter directly.

After Writing

Print: "=== Explore complete ===\nSaved: .design/DESIGN.md, .design/DESIGN-DEBT.md, .design/DESIGN-CONTEXT.md\nNext: @get-design-done plan".

Do NOT transition to plan (or invoke `/gdd:plan`) until BOTH `.design/DESIGN.md` AND `.design/DESIGN-CONTEXT.md` are committed AND the user has approved them. If this project uses a custom `.design` location, read the artifact paths from `.design/STATE.md` rather than assuming the default.

Rationalizations - Thought to Reality

The shortcut excuses an agent reaches for during explore, and the drift each one introduces:

Thought	Reality
"I already know this codebase, I can skip the inventory scan."	An unscanned codebase hides the tokens/components you'll duplicate - the grep pass exists to stop you reinventing what's there.
"The six connection probes are noise, I'll assume Figma is off."	A skipped probe means a wrong connection assumption silently breaks the design stage's tool dispatch.
"--skip-interview is fine, the brief covered it."	The interview locks the gray areas the brief left fuzzy; skipping it ships undecided D-XX into planning.
"I'll batch all the interview questions to save round-trips."	Batched questions overwhelm the user and smuggle in coupled assumptions - one-at-a-time keeps each decision clean.
"DESIGN-DEBT.md is optional, the scan was clean enough."	Unrecorded debt resurfaces as an unexplained constraint three stages later with no provenance.
"Prior sketches and project conventions don't apply this cycle."	Ignored conventions get overridden by defaults, producing inconsistency the audit will flag against the rest of the system.

EXPLORE COMPLETE