epyc-root

Process research URLs (papers, blogs, repos) through a structured intake pipeline. Extracts claims, deduplicates against a persistent index, cross-references existing work, expands literature, updates active handoffs, proposes handoff stubs, and produces a structured report. Use when ingesting new research material into the EPYC compendium.

Index the EPYC orchestrator x_* request overrides exposed through Hermes command skills.

Control the orchestrator stack lifecycle from Hermes via a thin wrapper over the existing stack launcher.

Lint, query, and maintain the project knowledge base. Use when auditing KB health, searching for compiled knowledge, or checking governance hygiene. Do not use when ingesting new research (use research-intake).

Use when reviewing a diff, commit, PR, dependency change, agent/tool change, or codebase surface for security risk. Runs a two-pass STRIDE + OWASP Top 10 + OWASP LLM + supply-chain review with exploit-path-gated findings and P0-P3 severity.

Cap or restore the escalation tier for this conversation. Controls how far the orchestrator can escalate from frontdoor to specialist models.

Disable REPL code execution for this conversation. Forces direct text responses only — no Python/shell execution.

Force a specific orchestrator role or model for this conversation. Bypasses frontdoor routing to use a deterministic current role.

Maintain governance accounting for all repository-relevant `CLAUDE.md` files and keep matrix documentation in sync. Use when adding/updating CLAUDE policy files, changing governance scope, or auditing which CLAUDE files are in vs out of migration scope.

Semantic search over the project's compiled markdown knowledge base (wiki/, handoffs/active/, handoffs/completed/, research/, progress/, docs/chapters/). Use BEFORE grep/find when investigating a topic, looking up prior decisions, or searching for cross-cutting context. ColBERT-backed; returns ranked chunks with file path, heading breadcrumb, and line range.

Compress the prose of project agent files (`agents/*.md`, `agents/shared/*.md`, `CLAUDE.md`) at authoring time using a strict project rider that preserves directive polarity (RFC 2119), procedural ordering, and structural blocks. Use when generating compressed variants of an agent file (mild / medium / aggressive levels) for per-model deployment.

Design, refactor, and validate project agent files using a thin-map architecture (shared policy plus lean role overlays). Use when creating/updating `agents/*.md`, splitting monolithic agent instructions, enforcing role schema, or migrating long operational guidance into docs.