intercraft

improve priority P2 effort M wave 2

What it does: A content-only Claude Code plugin (skill + 14 reference docs + 1 review agent + 1 audit command; no hooks, no MCP server per .claude-plugin/plugin.json) that codifies “agent-native architecture” — 8 design principles (action parity, tools-as-primitives, context injection, shared workspace, CRUD completeness, UI integration, capability discovery, prompt-native features) for building apps where agents are first-class citizens. The agent-native-architecture SKILL.md uses an intake→routing pattern to load one of 14 reference docs on demand; /agent-native-audit fans out 8 parallel sub-agents (one per principle) and compiles a scored X/Y compliance report; agent-native-reviewer.md applies the principles to PRs. Positioned as the architecture-review complement to Clavain’s code-craft review.

Rationale: Well-structured and genuinely distinct in the roster — it is the only plugin that audits whether an application treats agents as first-class citizens (an architecture concern), not code quality (interflux/intercheck) or model routing. Its packaging is exactly what 2026 SOTA endorses: a thin skill with progressive disclosure (intake → routed reference load) and an orchestrator-worker fan-out audit. Core thesis (parity, primitives, prompt-defined features) is correct and MORE relevant in 2026, so this is not modernize/deprecate. But the content was authored on the late-2024/early-2025 “Claude Code SDK / coding-agent-is-general-agent” framing and has not absorbed the frontier shifts most relevant to its own subject. Chiefly: its mcp-tool-design.md teaches dynamic capability discovery but is silent on progressive tool disclosure / RAG-MCP — the dominant 2026 lesson that eagerly registering many tool schemas is now an anti-pattern (143K/200K window consumed before first query; Tool Search Tool ~85% reduction, Code Mode ~99.9%). For a plugin whose subject is agent tool-surface design, that omission is substantive. Secondary gaps: tool-design guidance argues against constraints (z.string over z.enum, “artificial capability limits are an anti-pattern”) with no counterweight for tool-poisoning / lethal-trifecta / MCPTox >60% attack rates; the audit emits single-shot LLM-judged X/Y scores (the “unstructured NL critique” / agent-as-judge pattern SOTA shows is unreliable) with no variance or tool-grounded signal; self-modification.md predates the experience-distillation/skill-library line and the memory-pollution caution. All are content refreshes, not rearchitecture.

SOTA gap: Progressive tool disclosure / RAG-MCP as a first-class agent-native design principle. The “MCP context-bloat backlash” and “RAG-MCP retrieval-augmented tool selection” trends establish that how a tool catalog is surfaced (lazy/indexed vs eagerly loaded) now dominates agent effectiveness — a 9th principle (“tool-surface economy: expose a small static surface, defer/index the rest”) belongs alongside the existing 8. Secondary: tool-grounded/reliability-science auditing (variance over k trials, executable signals) instead of single-shot LLM-judged scores, per the “Beyond pass@1” and “verification grounded in executable signals” trends.

Redundancy: Low/none for its core purpose — uniquely covers application-level agent-parity architecture. Mild conceptual adjacency: its review agent + audit command overlap the general review machinery (interflux/flux-drive multi-agent review, intercheck quality, clavain code-review-discipline) in mechanism, but not in subject (those judge code craft; intercraft judges agent-native architecture). The /agent-native-audit fan-out could be implemented ON TOP of flux-drive’s existing parallel-dispatch + scored-triage engine rather than reimplementing parallel sub-agent orchestration in command prose, which would reduce mechanism duplication while keeping the unique rubric.

Actions:

Add a 9th principle + reference doc on tool-surface economy / progressive disclosure: when to defer-load tool schemas, RAG-MCP-style indexed tool selection, Code Mode / Tool Search Tool, and ‘treat the tool catalog as an indexed corpus’ — and add the matching audit sub-agent so /agent-native-audit scores it. Refresh mcp-tool-design.md to cite this (drives the ‘improve’).
Add a security counterweight to mcp-tool-design.md and the skill anti-patterns: the existing ‘don’t over-constrain / z.string over z.enum / artificial limits are an anti-pattern’ advice needs a lethal-trifecta / tool-poisoning caveat (MCPTox, prompt-injection in tool descriptions, controlled-blast-radius framing) so the plugin doesn’t advise removing guardrails that 2026 security requires.
Strengthen the audit from single-shot LLM-judged X/Y to reliability-aware: prefer tool-grounded evidence (grep for actual tool/route definitions, not vibe), and note k-trial/variance or at minimum require each sub-agent to cite file:line evidence per claim to reduce agent-as-judge hallucination.
Refresh self-modification.md to reference the experience-distillation / skill-library frontier (trajectories→named reusable principles) and add the human-auditable / memory-pollution caution; optionally rebase /agent-native-audit’s fan-out on interflux flux-drive to reuse the ecosystem’s parallel-dispatch engine instead of bespoke orchestration.