/supergraph:analyze

Analyze first, implement never. No code until approach is approved.

Announce: "🔍 /supergraph:analyze — framing problem, checking graph risk..."

When

• Ambiguous requirements
• Multiple valid approaches
• Task spans modules
• Graph shows high blast-risk (hub/bridge nodes)
• User says "how should I..." or "what's the best way..."

Skip for: typo fixes, config changes, clear mechanical edits.

Workflow

0. Read CONTEXT.md (if exists):

cat CONTEXT.md 2>/dev/null | head -60

Use existing domain vocabulary in all analysis — never invent new terms for concepts already named.

1. Frame the problem:

• Goal: [one sentence]
• Known constraints
• Open questions

1b. Score ambiguity (before grilling):

Signal	+1 if...
Ambiguous scope	touches multiple modules without naming one
No explicit path	no file/package/feature named
Multiple intents	could be bug fix, feature, refactor, or question
First interaction	no established context this session

Score	Action
0–1	Skip grilling — proceed with best available info
2	Show 1-line routing summary, wait for confirm
3–4	Grill: ask focused questions with recommended options

Auto-skip grilling entirely if: user said "go"/"just do it", trivial fix (<15 lines, 1 file), explicit mode command, or active plan already exists.

Grilling rules: ONE question at a time, offer recommended answer, max 3 questions. Stop when goal + constraints are clear enough that approach won't reverse on new info.

2. Check graph risk: Reuse /supergraph:scan context (CBM_PROJECT, codebase-memory-mcp, see references/codebase-memory-contract.md). If targets identified, call detect_changes, search_graph, trace_path, get_architecture, recipes hubs/bridges/cross-boundary. Empty = evidence. >20 files STOP.

2b. Serena (optional): See serena/SKILL.md:Setup. If scan not run, call initial_instructions first, then find_referencing_symbols/find_implementations per target. Skip if unavailable.

3. Propose 2-3 approaches + persona debate: For each approach: pros, cons, risk level, effort. Prefer minimal viable.

Then run 5 quick persona checks on the recommended approach:

Persona	Question
Architect	Does this fit the architecture? New coupling?
Security	What can be abused? Auth/data boundaries respected?
Performance	Latency impact? N+1 queries? Memory leaks?
UX	Error states handled? Intuitive?
Devil's Advocate	Simpler alternative? Which assumption could be wrong?

Emit verdict: GO (no blockers) / CAUTION (manageable risks, note mitigations) / STOP (critical issue — redesign needed before planning).

STOP triggers: auth bypass with no mitigation, fundamental design incompatibility, N+1 with no workaround, false core assumption.

4. Ask focused questions (one at a time): Only if the answer changes direction.

5. Recommend and hand off: Present recommendation. Once approved, summarize decisions into an analysis block in the plan file or prompt context:

## Analysis Decisions
- Approach: [chosen] | Why: [reason]
- Alternatives considered: [list] | Risks: [list]

Update CONTEXT.md if analysis crystallized new domain terms:

printf '\n## <term>\n[definition]\n' >> CONTEXT.md

→ invoke /supergraph:plan

Rules

• No implementation during analyze
• Don't over-analyze for hypothetical futures
• Ask ONE question at a time during grill — never dump multiple questions
• Always end with: "Shall I create the plan?" → invoke /supergraph:plan