analyze

Tier-aware discipline — apply at session start and before every major step:

1. Read widest → narrowest before duplicating effort:
   • T3 (cross-project): mcp__plugin_conexus_nexus__nx_answer(...) for verb-shape questions; mcp__plugin_conexus_nexus__search(...) for keyword lookup.
   • T2 (project): mcp__plugin_conexus_nexus__memory_search(query="<topic>", project="<repo>").
   • T1 (siblings, this session): mcp__plugin_conexus_nexus__scratch(action="search", query="<topic>").
2. Reuse plans before dispatching multiple agents: mcp__plugin_conexus_nexus__plan_search(query="<task>", limit=3).
3. Write back at end — findings not stored are findings lost. Pick the tier that matches the audience:
   • mcp__plugin_conexus_nexus__scratch(action="put", ..., tags="<topic>") for sibling agents downstream THIS session (T1, narrowest scope, cheapest write).
   • mcp__plugin_conexus_nexus__memory_put(...) for project-scoped decisions, future sessions same project (T2).
   • mcp__plugin_conexus_nexus__store_put(...) for permanent cross-project knowledge, future sessions everywhere (T3).
   • mcp__plugin_conexus_nexus__plan_save(...) for multi-agent pipeline outcomes (so future callers hit plan-match).

analyze

You MUST call nx_answer for cross-corpus synthesis or ranking. Direct search returns unstructured chunks; analytical questions need the search → extract → rank → generate composition that the analyze plans provide. This is the one verb where bundling gives the biggest wins: a 3-op chain collapses from ~45s of per-step spawns to ~15s in a single claude -p call.

The call

mcp__plugin_conexus_nexus__nx_answer(
    question=<caller's phrasing>,
    dimensions={"verb": "analyze"},
    context=<area, criterion, limit — as JSON string if needed>,
)

One tool call. nx_answer handles match → run → record, including the operator-bundle optimization. Plan-miss falls through to an inline claude -p planner.

Required bindings

• area — the subject area being analysed.
• criterion — the axis used to rank candidates (e.g. "recency", "citation count", "coverage breadth").

Optional bindings

• limit — per-corpus result cap (default_bindings → 12).

Typical intent shapes

• "compare how X is handled across ml-systems and networking"
• "survey approaches to Y"
• "rank options for Z by cost"

When direct search is fine

If the question is a single-corpus lookup — e.g. "find the chunks that mention algorithm X" — use mcp__plugin_conexus_nexus__search. Analyze earns its latency cost when the question requires multi-corpus alignment, ranking by a semantic criterion, or structured extraction before synthesis.

Anti-patterns (do not do any of these)

• Calling search directly for a cross-corpus synthesis question. You get top-K chunks with no composition, no cross-corpus alignment, no ranking. If the question requires ranking or comparing across multiple collections, you need nx_answer's full DAG.
• Calling plan_match directly instead of nx_answer. You lose the record step, the inline-planner fallback, and use_count telemetry.
• Using analyze when research would suffice. research is single-concept; analyze implies cross-corpus / cross-approach synthesis. If the caller only wants to understand one thing, use /conexus:research.
• Using bare criterion. "important" is not a criterion; pick an axis the ranker can actually sort by.

See /conexus:plan-first and docs/plan-authoring-guide.md.