mikeparcewski/smaht-synthesize

Context Synthesis

You received this skill because the hook determined this prompt needs deep context assembly.

Arguments (from hook)

Args are passed as a JSON string. Parse with json.loads(args):

• prompt: the user's original prompt
• complexity: float 0–1 (hook scoring)
• risk: bool — high-risk keywords detected
• turns: recent session turns summary (condensed, may be absent on first turn)
• context_briefing: pre-assembled adapter context from the orchestrator (present on SLOW-path only, absent on FAST-path or when orchestrator failed). When present, skip cold exploration in Step 1 and use this as the starting context.

Budget (based on complexity + risk)

| complexity | risk | rounds | parallel agents | |-----------|------|--------|----------------| | < 0.5 | no | 1 | 2 | | 0.5–0.7 | no | 1 | 3 | | > 0.7 | any | 2 | 4 | | any | yes | 2 | 3 |

Step 0: Scope Validation

Before running the full synthesis loop, validate that this prompt actually warrants agentic synthesis. Parse complexity, risk, and (if present) context_briefing from the skill args (all provided as JSON fields).

If context_briefing is present in args, its content can be used to inform all 3 checks below — treat it as verified adapter evidence (native task state, brain index hits, domain events) rather than re-deriving from scratch.

Run these 3 quick checks against the user's prompt:

1. Novel / complex intent — Is the prompt asking for something genuinely new or complex (not a clarification, follow-up, or simple yes/no)? Short rephrasing of the prior turn, confirmations ("is this right?", "looks good?"), and single-word continuations ("yes", "proceed") do NOT pass this check.
2. Cross-domain or multi-step reasoning — Does answering require synthesizing knowledge across multiple systems, domains, or layers? Single-file questions or direct factual lookups do NOT pass this check.
3. Risk flag — Is risk true (destructive/production-affecting keywords detected by the hook)?

Decision:

• 0 of 3 pass → Exit early. Output exactly:

  [Scope Check: LOW — routing to fast path]
  

  Then stop. Do NOT produce a context briefing or proceed to Step 1.

• 1 or more pass → Proceed to Step 1 below.

Step 1: Facilitator — What do I need to know?

If context_briefing is present in args: Skip the cold exploration phase. Use the provided briefing as your starting context — it already contains adapter output (brain index hits, native task state, domain events). Proceed directly to Step 3 to assess whether the briefing is sufficient or if targeted follow-up is needed.

If context_briefing is absent: Read the user's prompt and recent turns. Identify 3–5 specific questions:

• What concept, file, or system is the user asking about?
• What past decisions are relevant?
• What recent changes (events) might affect the answer?
• What constraints apply?

Do NOT search yet. Just list the questions.

Step 2: Fan-out — Parallel exploration

Spawn parallel Agent calls (one per search question). Each agent should:

Brain wiki search (synthesized knowledge — try this first):

find ~/.wicked-brain/wiki -name "*.md" 2>/dev/null | head -20
# Read any wiki article whose name matches the topic

Brain FTS search (raw chunks — for specifics):

curl -s -X POST http://localhost:4242/api \
  -H "Content-Type: application/json" \
  -d '{"action":"search","params":{"query":"QUERY_TERMS","limit":5}}'

If curl fails (brain server not running), fall back to Grep/Glob on the project files directly.

Recent events (what changed recently):

curl -s -X POST http://localhost:4242/api \
  -H "Content-Type: application/json" \
  -d '{"action":"search","params":{"query":"QUERY_TERMS event created","limit":3}}'

Each agent returns: what it found (or "nothing relevant — brain unavailable, searched files directly").

Step 3: Facilitator — Is this sufficient?

Review all agent outputs. Ask:

• Do I have specific, factual answers to the user's questions?
• Is anything critical missing?

If missing AND rounds remaining: identify the gap, spawn 1–2 targeted follow-up agents. If sufficient OR no rounds left: proceed to synthesis.

Step 4: Synthesize — Complete turn directive

Output ONLY the following block (200–300 words). This replaces Claude's own context assembly.

CONTEXT BRIEFING [smaht-synthesized | complexity={X} | risk={Y}]

**The user is asking**: {one sentence — actual intent, not a restatement}

**What is true** (verified from project knowledge):
- {specific fact 1 with source reference}
- {specific fact 2 with source reference}
- {specific fact 3 with source reference}
[add up to 2 more only if directly relevant]

**Active constraints**: {from session turns — user-stated rules that apply}

**Recommended approach**: {1–2 sentences — what to do and how, based on what was found}

**What was NOT found**: {if any critical info was missing — be explicit}

After outputting the briefing, tell Claude: "Proceed with this context. Answer the original prompt: {prompt}"