intermonk

keep priority P3 effort S wave 3

What it does: A single-skill plugin (skills/dialectic, no MCP/hooks/commands) implementing a 7-phase Hegelian dialectic process. The orchestrator runs an elenctic interview, writes a neutral context_briefing.md, then spawns two ‘Electric Monk’ subagents (parallel, run_in_background) that each FULLY BELIEVE an opposing position at full conviction (anti-hedging is a hard functional requirement). Monks return text the orchestrator persists to disk; the orchestrator alone (never delegated) performs determinate negation + Boydian decomposition (Phase 4) and synthesis/Aufhebung (Phase 5), then validates via fresh monk agents + a required hostile-auditor agent (Phase 6), then recurses (Phase 7) with a per-topic dialectic_queue. Theory grounded in Rao (artificial belief system), Hegel (determinate negation), Boyd (shatter-then-recombine). Artifacts written to dialectics/[topic]/. Adopted MIT from Kyle Mathews’ hegelian-dialectic-skill.

Rationale: Mature, deliberate, and already aligned with multiple SOTA patterns. It implements the orchestrator-worker topology with genuinely isolated subagent context windows that return distilled text to a lead — the dominant production pattern (Anthropic deep-research, HALO arXiv:2505.13516) — and applies it where it is justified (two truly independent, parallelizable belief positions, not default fan-out). It uses progressive disclosure correctly: a compact SKILL, XML-tagged phase sections for selective re-reading on recursion, and references/ loaded on demand — exactly the context-engineering discipline (Anthropic write/select/compress/isolate; ~98.7% tool-token-style savings ethos applied to a skill). It has an explicit token budget table, an absolute-path output contract, a ‘monk output is testimony not evidence’ guard against confusion/poisoning, and a hostile-auditor verification gate. The prose is unusually high-craft and the structural tests + bump-version workflow show real maintenance. This is a feature-complete, well-documented capability with no architectural rot; nothing here warrants modernize/deprecate. The improvement opportunities are additive, not structural.

SOTA gap: Verification is purely free-form NL critique by LLM agents (Phase 6 monk re-evaluation + hostile auditor), with the validation monks judging a synthesis derived from their own essays. Two converging SOTA findings apply: (1) ‘generic unstructured NL critiques are largely ineffective for refinement; structured, tool-grounded feedback (tests, executors, retrieval) drives improvement’ (ReVeal arXiv:2506.11442; Code-to-Courtroom arXiv:2510.24367) — for empirical/engineering/risk domains the synthesis’s testable claims could be checked against executable/retrieval signals rather than only prose; (2) Agent-as-a-Judge self-preference bias and ‘Gaming the Judge’ (Jan 2026) — a monk grading a synthesis built from its own argument is the exact bias the literature warns to design around (mask author identity / use a distinct judge). Secondary gap: no experience-distillation loop — every dialectic starts cold and the dialectic_queue stays a per-topic file; the EvolveR/ExpeL/skill-library paradigm (arXiv:2510.16079) suggests graduating recurring sublations/contradiction patterns into a reusable named-principle library.

Redundancy: Modest, not merge-worthy. interlens exposes get_dialectic_triads + synthesize_solution (thesis/antithesis/synthesis as lens primitives) — but that is a one-shot lens lookup, not a multi-round belief-committed process. clavain /debate (structured Claude↔Codex debate), interpeer (cross-AI second opinion), and interflux/flux-explore (multi-agent adversarial review across semantic distance) all do adversarial multi-perspective reasoning, but each is single-pass or review-oriented. intermonk’s differentiator is genuine: deep, human-in-the-loop, recursive Hegelian synthesis with full-conviction belief subagents and an Aufhebung that transforms the question. No consolidation warranted; it owns a distinct decision-support niche.

Actions:

Reconcile version drift (plugin.json says 0.1.1, AGENTS.md/CLAUDE.md say 0.1.0) via bump-version.sh — trivial nit.
Add a structured-verification option to Phase 6 for empirical/engineering/risk domains: when the synthesis makes testable claims, dispatch a tool/retrieval-grounded check (web/code search or a quick executable test) rather than relying solely on free-form LLM critique — cites ReVeal/Code-to-Courtroom SOTA.
Mitigate Agent-as-a-Judge self-preference bias: have the hostile auditor (or a distinct judge agent) carry primary validation weight and mask author identity in the monk-validation prompts, so a monk is not effectively grading its own derived synthesis (cites Agent-as-a-Judge / Gaming-the-Judge).
Optional follow-on (separate bead): add a distillation step that graduates recurring contradictions/sublations from dialectic_queue.md into a reusable principle entry in interknow/intermem, turning per-topic queues into a compounding library (cites EvolveR/skill-library SOTA).