lythoskill

Weekly synthesis writer. Distills the past 7 days' core thread + quest DAG into a frontmatter-rich short doc. Never replays git log or cortex INDEX (those are direct queries already). Forms the weekly counterpart to project-scribe (daily) and project-onboarding (read-side).

Weekly synthesis writer. Distills the past 7 days' core thread + quest DAG into a frontmatter-rich short doc. Never replays git log or cortex INDEX (those are direct queries already). Forms the weekly counterpart to project-scribe (daily) and project-onboarding (read-side).

Session context loader. Reads the latest daily handoff to restore project context without redundant file exploration. Three-layer loading: CLAUDE.md (stable) → daily handoff (session state) → git verification (ground truth). Degrades to file exploration when no handoff exists. CQRS read-side pair with project-scribe.

Analyzes SKILL.md files against Agent Skills best practices. Reviews body size, description quality, progressive disclosure, frontmatter usage, and context efficiency. Provides actionable optimization advice.

夜有所梦 — project memory consolidation. Daily (scribe) captures raw experience, weekly extracts patterns and anomalies, dreaming consolidates the accumulated memory into SSOT. External review via ZK subagent (the same de facto standard pattern used by Hermes Curator's forked-agent review), with cross-model validation via arena for critical docs.

Session context loader. Reads the latest daily handoff to restore project context without redundant file exploration. Three-layer loading: CLAUDE.md (stable) → daily handoff (session state) → git verification (ground truth). Degrades to file exploration when no handoff exists. CQRS read-side pair with project-scribe.

Session context dump. Self-assess what the conversation contains that has NO other carrier (no task, no ADR, no epic) — pitfalls, working-tree anomalies, why-we-chose-this, specific next steps — and write to daily/YYYY-MM-DD.md. Things WITH structured carriers go to their carriers. Things WITHOUT carriers but needed by the next agent go here. Forms CQRS write-side pair with project-onboarding (read-side).

User-acceptance-driven release workflow using heredoc patch files. Each iteration produces a timestamped pr-<timestamp>-<desc>.sh that self-archives after execution. No tag without explicit user LGTM. Supports rollback via archived backups.

Stay clear-headed. Before you commit to a conclusion, check your basis. Is this evidence or assumption? Do the sources agree? What's the confidence per claim — not in general, but specifically? When you catch yourself listing risks without checking, or hesitating between directions without data, that's the signal to stop and verify.

Human-first documentation writer and reviewer. Reviews README, wiki, ADR, daily handoff, showcase, and reference docs for information density, structural rhythm, and anti-template patterns. Ensures human readers get clear prose, not AI-flavored filler.

Test play for skills and deck configurations. DEFAULT: agent reads config, spawns parallel subagents via native Agent tool, judges outputs. Single-deck test AND multi-deck A/B comparison both run agent-orchestrated (no CLI). Cross-player comparison (kimi vs codex) is the ONLY case that needs the CLI runner. Always restores parent deck. No install, no working-set pollution, no deck overwrite. Subagent-friendly: resumes interrupted runs from saved state. CRITICAL: experiments run in `/tmp`, never in committed directories. Subagent inherits parent CWD — prompt must explicitly set workDir.

Scaffold and build projects for the lythoskill ecosystem only. Creates thin-skill monorepos: heavy logic in npm packages (Starter), agent-facing instructions in lightweight SKILL.md (Skill), build output committed to skills/ (Output).

Skill 策展者/买家秀 (curator's perspective). Scans your local cold pool, indexes SKILL.md frontmatter into REGISTRY.json + catalog.db. CLI is mechanical glue (scan/query/tag/audit/find) — YOU are the agent who combines curator's local cache with WebSearch, deep research, and arena testing to discover, annotate, fact-check, and recommend. Curator = 查卡器 + 备注 + 组卡审美. Reconciler-style: any filesystem state → scan → converges to clean index. Auto-backup; rollback via `restore`.

Declarative skill deck governance. `deck link` reconciles the working set to match skill-deck.toml — deny-by-default removes undeclared skills. DEFAULT patterns: PHASE SWITCH via separate deck files (deck link --deck phase<N>.toml), SEED bootstrap (minimal deck → agent self-expands via deck add + curator discovery). Always restores parent deck. Zero state pollution between phases. INNATE FIRST: after compaction, session reload, or skill refresh — read every innate skill's full SKILL.md before any tool skill. Innate = eagerly loaded, always full context. Tool = lazy, read only on trigger. CRITICAL when writing a deck for others: verify every locator path against the REAL repo structure before publishing.

A minimal hello-world skill. No scripts, no dependencies. Demonstrates the simplest possible thin-skill: just a SKILL.md.

Register tasks BEFORE fixing bugs. Close epics when work is done. Always use CLI for state transitions — never mv files by hand. CLI moves update Status History; manual mv causes probe mismatches. Never leave state drift for the next agent to discover via git log. GTD-style governance: ADR, Epic, Task, Wiki. Numeric-prefixed directories enforce workflow order. Timestamp IDs prevent collision. ZK task review (WHAT/WHY/HOW) ensures subagent-readable requirements before assignment. probe detects stale backlog, epic drift, and history mismatches. **Task-git discipline**: One task = 2-5 commits (core, tests, docs, trailer). Every commit includes TASK-xxx for traceability. Task card includes Commit Map and Verification Matrix.

Register tasks BEFORE fixing bugs. Close epics when work is done. Always use CLI for state transitions — never mv files by hand. CLI moves update Status History; manual mv causes probe mismatches. Never leave state drift for the next agent to discover via git log. GTD-style governance: ADR, Epic, Task, Wiki. Numeric-prefixed directories enforce workflow order. Timestamp IDs prevent collision. ZK task review (WHAT/WHY/HOW) ensures subagent-readable requirements before assignment. probe detects stale backlog, epic drift, and history mismatches. **Task-git discipline**: One task = 2-5 commits (core, tests, docs, trailer). Every commit includes TASK-xxx for traceability. Task card includes Commit Map and Verification Matrix.

Session context dump. Self-assess what the conversation contains that has NO other carrier (no task, no ADR, no epic) — pitfalls, working-tree anomalies, why-we-chose-this, specific next steps — and write to daily/YYYY-MM-DD.md. Things WITH structured carriers go to their carriers. Things WITHOUT carriers but needed by the next agent go here. Forms CQRS write-side pair with project-onboarding (read-side).

Test play for skills and deck configurations. DEFAULT: agent reads config, spawns parallel subagents via native Agent tool, judges outputs. Single-deck test AND multi-deck A/B comparison both run agent-orchestrated (no CLI). Cross-player comparison (kimi vs codex) is the ONLY case that needs the CLI runner. Always restores parent deck. No install, no working-set pollution, no deck overwrite. Subagent-friendly: resumes interrupted runs from saved state. CRITICAL: experiments run in `/tmp`, never in committed directories. Subagent inherits parent CWD — prompt must explicitly set workDir.

Scaffold and build projects for the lythoskill ecosystem only. Creates thin-skill monorepos: heavy logic in npm packages (Starter), agent-facing instructions in lightweight SKILL.md (Skill), build output committed to skills/ (Output).

Declarative skill deck governance. `deck link` reconciles the working set to match skill-deck.toml — deny-by-default removes undeclared skills. DEFAULT patterns: PHASE SWITCH via separate deck files (deck link --deck phase<N>.toml), SEED bootstrap (minimal deck → agent self-expands via deck add + curator discovery). Always restores parent deck. Zero state pollution between phases. INNATE FIRST: after compaction, session reload, or skill refresh — read every innate skill's full SKILL.md before any tool skill. Innate = eagerly loaded, always full context. Tool = lazy, read only on trigger. CRITICAL when writing a deck for others: verify every locator path against the REAL repo structure before publishing.

Analyzes SKILL.md files against Agent Skills best practices. Reviews body size, description quality, progressive disclosure, frontmatter usage, and context efficiency. Provides actionable optimization advice.

Skill 策展者/买家秀 (curator's perspective). Scans your local cold pool, indexes SKILL.md frontmatter into REGISTRY.json + catalog.db. CLI is mechanical glue (scan/query/tag/audit/find) — YOU are the agent who combines curator's local cache with WebSearch, deep research, and arena testing to discover, annotate, fact-check, and recommend. Curator = 查卡器 + 备注 + 组卡审美. Reconciler-style: any filesystem state → scan → converges to clean index. Auto-backup; rollback via `restore`.

夜有所梦 — project memory consolidation. Daily (scribe) captures raw experience, weekly extracts patterns and anomalies, dreaming consolidates the accumulated memory into SSOT. External review via ZK subagent (the same de facto standard pattern used by Hermes Curator's forked-agent review), with cross-model validation via arena for critical docs.

A minimal hello-world skill. No scripts, no dependencies. Demonstrates the simplest possible thin-skill: just a SKILL.md.

User-acceptance-driven release workflow using heredoc patch files. Each iteration produces a timestamped pr-<timestamp>-<desc>.sh that self-archives after execution. No tag without explicit user LGTM. Supports rollback via archived backups.

Stay clear-headed. Before you commit to a conclusion, check your basis. Is this evidence or assumption? Do the sources agree? What's the confidence per claim — not in general, but specifically? When you catch yourself listing risks without checking, or hesitating between directions without data, that's the signal to stop and verify.

Human-first documentation writer and reviewer. Reviews README, wiki, ADR, daily handoff, showcase, and reference docs for information density, structural rhythm, and anti-template patterns. Ensures human readers get clear prose, not AI-flavored filler.

Test-driven development with red-green-refactor loop. Use when user wants to build features or fix bugs using TDD, mentions "red-green-refactor", wants integration tests, or asks for test-first development.