epicsagas

Ring 3 evolution engine. Analyzes session observations, generates and improves evolved skills, shows metrics dashboard. Subcommands: status, history, rollback, reset. Use for post-session review and skill improvement.

State-persisted autonomous pipeline: spec → go → audit → eval → ship → evolve in one command. Auto-detects direct/council/interactive mode. Crash-recoverable via PIPELINE-*.json. Hands-off until PR.

Core router. Always active. Auto-invokes matching skill before every response. Runs confusion protocol on high-risk ambiguity.

Loop-breaking self-diagnosis. Use when 3+ consecutive failures occur, circular retries persist, or context overwhelms the session.

Audit phase. Parallel review: code quality + security + tests. Semantic dedup of cross-mode findings. Outputs PASS/WARN/FAIL per dimension. Validates spec coverage.

Conventional Commits 1.0 generator. Stages relevant files, infers type(scope): description, never uses --no-verify.

Context window manager. Snapshots session state, compacts conversation, and resumes via hook. Use when context exceeds 70% or a large new task begins.

4-voice parallel deliberation (Architect · Skeptic · Pragmatist · Critic) for architecture, tech selection, or design decisions with no clear answer. Anti-anchoring: each voice gets independent context. Records decision in harness-mem.

Generate interior illustrations for book chapters. Analyzes scenes, creates style-consistent illustration prompts, and produces illustration plan with placement metadata.

Phase 5: Publishing. Convert to EPUB/PDF/MOBI/TXT/Markdown, cover design, metadata, title candidates, KDP checklist, marketing plan.

AI-native publishing system: autonomous multi-phase workflows from ideation to export. Fiction, non-fiction, and technical books. Trigger: /velith or book-related requests.

Phase 3: Drafting. Plan-Then-Execute chapter generation with parallel writing and continuity validation.

Phase 4: Editing. 5-stage pipeline — editorial assessment, developmental edit, line edit, copy edit, proofread.

Phase 1: Ideation. Concept validation, market gap analysis, competitive matrix, final concept selection.

Phase 0: Onboarding. Initialize new book project — define genre, audience, language, scope, generate PRD.md and STYLE.md.

Phase 2: Outlining. Apply genre-specific structures, generate per-chapter specs, cross-reference map, structural validation.

Git workspace manager — sync repos, bump versions, project health dashboard. Works in submodule monorepos, flat multi-repo, and single projects. Triggers: sync repos, pull all, bump version, git tags, overview, dashboard, project status.

Uses SkillOpt (text-space optimizer) to train and evolve natural-language skill documents for frozen LLM agents. Generates benchmark data from real tasks, builds environment adapters, runs rollout-reflect-aggregate-update training loops, and produces an optimized skill document. Triggers on: 'optimize skill', 'SkillOpt', 'skill training', 'improve this skill', 'skill experiment'.

Audits and guides OSS project release readiness across the full lifecycle: community standards, README quality, launch strategy, distribution pipeline, and governance/security posture. Activates when reviewing, planning, or shipping an open-source project — at any stage from initial setup to public launch.

Recommends and structures learning from a field completely outside the user's expertise. Generates a weekly reading plan with summary template for cross-pollination of ideas. Triggers on: 'learn something new', 'asymmetric learning', 'weekly learning', 'outside my field', 'cross-pollination'.

Business-first risk evaluation before writing code. Forces quantification of how a technical decision affects revenue, users, or market position within 30 days. Triggers on: 'biz risk', 'business value', 'is this worth it', 'prioritize', 'ROI'. No code without a business case.

Generates adversarial counter-arguments against the user's strongest technical convictions. Forces intellectual humility by constructing the best possible case against their position. Triggers on: 'devil's advocate', 'challenge my thinking', 'counter-argument', 'am I wrong', 'convince me otherwise'.

Tests that tools and products work outside the user's optimized development environment. Generates a compatibility checklist for lower-spec or general-purpose environments. Triggers on: 'env test', 'compatibility check', 'works on other machines', 'low spec test', 'production ready'.

Forces 5-level root-cause questioning on any technical decision or justification. Reveals whether the real motive is business value or technical satisfaction. Triggers on: 'five whys', 'why', 'root cause', 'justify this', 'why this approach'. No surface-level answers accepted.

Analyzes or compares 2+ papers or methods. Triggers on "compare these papers", "analyze this method", "which is better?".

Verifies every Markdown link citation in `_papers/` and `_ideas/` notes resolves to the correct paper and that the attributed claim actually matches what the paper says. Gate skill — run before finalizing literature_review/go/check outputs.

Adversarial self-review gate that re-reads a literature review, comparison, or idea list before it's presented as final, hunting for overclaiming, cherry-picking, and gaps.

Documentation generator for public APIs, functions, and modules. Use when exports, signatures, or new modules are added or changed.

Evidence-hierarchy enforcer for research and analysis. Rates every claim by its evidence tier (primary > peer-reviewed > secondary > anecdotal) and refuses to overstate. Use when synthesizing findings or answering factual questions.

Reads and summarizes ONE specific paper. Triggers on "read this paper", "summarize X", or when the user gives a specific paper title/link.

Searches online for ML/AI papers relevant to a topic. Triggers on "find papers about X", "recent research on X", "any papers on this topic?".

Reproducibility checker for computational research and analyses. Demands pinned seeds, versions, and data hashes so a result can be re-run. Use when producing any number, figure, or benchmark a decision will rest on.

Collects market sizing, trend, and competitor data from the web and tags every claim with an evidence tier (primary > peer-reviewed/analyst > secondary press > anecdotal). Runs after `goal` has fixed the scope. Triggers on "시장 규모 알아봐", "경쟁사 조사", "이 시장 트렌드는?".

Safety gate. Adversarially re-reads FINDINGS/COMPARISON/GTM-PLAN before they're presented as final, hunting for overclaiming, unsourced numbers, and cherry-picked competitor data. Run before declaring any deliverable done.

Turns `analyze` findings into a competitor comparison matrix and a positioning read — where the user's product/idea fits relative to alternatives, and which gap is defensible. Triggers on "비교표 만들어줘", "경쟁사 매트릭스", "우리 포지셔닝은?".

Turns the positioning read from `decide` into a concrete GTM strategy document — ICP, messaging, channel mix, pricing, and a 90-day launch plan with owners and metrics. Triggers on "GTM 전략 짜줘", "출시 계획", "채널 전략".

Scopes a market research / competitive analysis / GTM engagement before any data collection starts. Defines the market, the decision the research must serve, and what "done" looks like. Triggers on "시장조사 시작", "새 GTM 프로젝트", "경쟁사 분석해줘" without a clear scope yet.

Assesses market-entry and GTM-execution risk — competitive response, regulatory exposure, channel dependency, timing risk — and rates each by likelihood x impact. Triggers on "리스크는?", "이거 위험한가", "진입 장벽".

Estimates ROI ranges for each channel in the GTM plan's channel mix — CAC, expected conversion, payback period — and flags which channels are speculative vs. evidence-backed. Triggers on "채널별 ROI", "이 전략 수익성은?", "CAC 계산".

Safety gate. Weighs whether a conclusion (positioning gap, GTM recommendation, market-entry verdict) is actually balanced against evidence on both sides, not just the evidence that supports it. Run after `decide` or `execute` produces a verdict, before `critic`.

obsidian-forge (alias `of`) CLI skill for creating and managing Obsidian vaults, processing inbox notes with AI classification, strengthening knowledge graphs, syncing plugins across vaults, and running the background daemon. Use when initializing vaults, automating PARA routing, managing MOCs, or debugging vault sync and graph operations.

Knowledge graph strengthening pipeline — runs `of graph health` then `of strengthen-graph` to add backlinks, bridge notes, and auto-tags. Trigger: strengthen graph, graph health, fix orphans, improve connections.

AI inbox classification pipeline — runs `of process-all` to classify inbox notes with AI, inject frontmatter, and route to PARA folders. Trigger: process inbox, classify notes, empty inbox, PARA routing.

Batch vault repair — runs `of check-tags --fix`, `of check-links --fix`, and `of normalize-frontmatter --fix` to repair tag health, broken wikilinks, and YAML malformations. Trigger: fix vault, repair tags, fix links, fix frontmatter, vault maintenance.

Obsidian vault diagnostics — runs `of doctor` and reports vault status, AI connectivity, inbox state, and git health. Trigger: vault health check, diagnose vault, vault status.

Full vault sync cycle — runs `of sync` to rebuild MOCs, strengthen graph, and commit to git. Trigger: sync vault, update mocs and graph, commit vault changes.

Software engineering knowledge graph — patterns, laws, refactorings, smells. Activates on any code quality concern, design decision, architecture review, or engineering question — even when the user describes problems informally. Uses Episteme HTTP API (resolved via `epis api env`) via curl.

Use this agent when you need to evaluate a system architecture or technology decision — identifying scalability risks, structural smells, law violations (Conway, Amdahl, Gall), and design pattern misuse using the Episteme knowledge graph.

Code smell detector and refactoring advisor. Triggers on: file path mentioned, 'review this code', 'find smells', 'analyze', 'refactor', or code shared. Calls Episteme HTTP API (analyze + refactor endpoints) immediately. Outputs smell table, ranked refactorings, principle violations.

Use this agent when the user faces an engineering decision, trade-off, or architecture question — choosing between patterns, resolving design conflicts, or applying refactorings. Grounds every recommendation in the Episteme knowledge graph.

Use this agent when you need to explore or research software engineering knowledge — finding design patterns, refactorings, laws, and code smells, or mapping relationships between concepts in the Episteme knowledge graph.

Recursive web crawling with depth control, pagination detection, and sitemap discovery. Use for comprehensive site coverage when you need to follow links across multiple levels.

Fetch a single web page using Obscura headless browser. Best for public, read-only pages. Returns HTML, plain text, links, or JS-evaluated results.

Multi-step web data collection pipeline using Obscura. Combines fetch + scrape for discover-then-collect workflows. Use when you need to crawl a site systematically.

Scrape multiple URLs in parallel using Obscura. Use when you have a list of URLs to process in batch. Returns JSON or text per URL.

Skill creator and optimizer following Agent Skills Open Standard. Triggers on skill create, make skill, skill forge, new skill, skill audit, skill optimize, SKILL.md.

vault-doctor CLI wrapper for Obsidian vault diagnostics and auto-fix. Triggers on vault health, scan vault, fix vault, check vault, frontmatter issues, tag conflicts, vault-fix, vault-scan.

HTTP API-based documentation server (always running). Questions about project architecture, conventions, decisions, code structure, tech debt, env config, progress, or doc health. Also: init project, audit docs, lint, validate, promote note, rebuild index, search vaults.

Analyze Claude Code usage patterns and provide actionable insights. Triggers on "usage insights", "analyze usage", "analytics insights", "usage patterns", "cost analysis". Reports on costs, model distribution, tool usage, cache efficiency, and session patterns.

Trigger when the user asks to build, create, generate, or customize their own AI agent harness ("make my harness", "build my harness"). Routes the conversation to the BYOH MCP tools and drives the full profile → build → install flow. The conversation IS the interview/wizard.