## Phase 97: Skills System — Unified Skill Framework, Autonomous Generation & Cross-Domain Composition

Build an industry-leading, SOTA skill system that unifies the 30+ independent
skill-like systems across the Oshun ecosystem into a single framework with
autonomous skill generation, self-evolution, cross-domain composition, quality
tracking, MCP integration, and skill-specific model training. Integrates
patterns from VOYAGER, Memento-Skills, OpenSpace, SkillX, MCP-Zero, xLAM, and
the Agent Skills Open Standard.

---

### 97.1 Unified Skill Framework

#### 97.1.1 Canonical Skill Schema & Registry

- [x] **97.1.1.1** Design canonical UniversalSkill schema: {skillId, name,
      domain, category, description, version, parameters (typed with Zod),
      returns (typed), handler (async function), timeoutMs, retryPolicy,
      qualityMetrics[], dependencies[], conflicts[], requiredPermissions[],
      sandboxLevel, executionHistory[], qualityScore, learningRate,
      breakingChanges[], minModelVersion, tags[]}
- [x] **97.1.1.2** Implement skill registry service backed by PostgreSQL with
      full CRUD, semantic search (pgvector on description embeddings), filter by
      domain/category/capability, and discovery API
- [x] **97.1.1.3** Build skill discovery engine with progressive disclosure
      (metadata always loaded ~100 tokens, full instructions loaded on
      activation, resources loaded on demand — following Agent Skills Open
      Standard pattern)
- [x] **97.1.1.4** Implement skill versioning with semver (major.minor.patch),
      immutable version snapshots, changelog tracking, breaking change
      detection, automatic compatibility checking between skill versions
- [x] **97.1.1.5** Build skill dependency resolver that automatically satisfies
      prerequisites (if skill B requires skill A's output, ensure A runs first),
      detect circular dependencies, and compose dependency DAGs
- [x] **97.1.1.6** Implement skill access control with RBAC (who can
      create/edit/execute/delete skills per domain), audit logging of all skill
      operations, and enterprise skill governance workflows (draft → review →
      approved → published)
- [x] **97.1.1.7** Build SKILL.md file format support conforming to Agent Skills
      Open Standard (YAML frontmatter + Markdown body + optional scripts/ +
      references/ + assets/ directories) for interoperability with Claude Code,
      Codex CLI, Cursor, GitHub Copilot, Gemini CLI, and 10+ other tools
- [x] **97.1.1.8** Implement skill tagging and categorization taxonomy: domain
      (iris, psyche, isis, maya, etc.), modality (text, image, audio, video, 3d,
      code), complexity (atomic, composed, workflow), interaction (sync, async,
      streaming), safety (safe, caution, dangerous, blocked)

#### 97.1.2 Skill Execution Runtime

- [x] **97.1.2.1** Build skill execution engine with sandboxed runtime
      (process-level isolation per skill execution), timeout enforcement,
      resource limits (memory, CPU, network), and graceful cancellation
- [x] **97.1.2.2** Implement skill execution telemetry: per-execution capture of
      {skillId, version, inputHash, outputHash, durationMs, tokensUsed, cost,
      success/failure, errorType, retryCount, callerContext} — feeding into data
      flywheel (Phase 85)
- [x] **97.1.2.3** Build skill retry and fallback system (configurable retry
      policies per skill: exponential backoff, circuit breaker, fallback to
      alternative skill, graceful degradation)
- [x] **97.1.2.4** Implement skill caching layer (cache skill outputs by input
      hash for idempotent skills, configurable TTL, cache invalidation on skill
      version update)
- [x] **97.1.2.5** Build skill rate limiting and quota management (per-user,
      per-domain, per-skill rate limits, burst allowances, quota tracking with
      alerts)

### 97.2 Cross-Domain Skill Composition

#### 97.2.1 Composition Engine

- [x] **97.2.1.1** Build skill composition engine supporting: sequential chains
      (A → B → C), parallel execution (A | B | C → merge), conditional branching
      (if X then A else B), loops (repeat A until condition), and map/reduce
      (apply A to each item in collection)
- [x] **97.2.1.2** Implement cross-domain skill pipelines: enable composing
      skills from different domains (e.g., "research topic (Sophia) → generate
      illustration (Isis) → verify safety (Nous) → add to lesson (Metis) →
      publish (Yemaya)") with typed input/output contracts between steps
- [x] **97.2.1.3** Build skill pipeline definition language (YAML/JSON) with:
      step definitions, input/output mappings, error handlers, timeout per step,
      retry per step, conditional logic, parallel groups, and variable
      interpolation
- [x] **97.2.1.4** Implement pipeline visualization as interactive DAG (nodes =
      skills, edges = data flow, colors = domain, status indicators for
      running/complete/failed)
- [x] **97.2.1.5** Build pipeline template library with pre-composed
      cross-domain workflows for common tasks: content creation pipeline,
      research-to-publication pipeline, game asset pipeline, educational content
      pipeline, news production pipeline
- [x] **97.2.1.6** Implement SkillX-style three-tiered skill hierarchy: planning
      skills (high-level task organization), functional skills (reusable
      tool-based subroutines), atomic skills (execution-oriented patterns) —
      with automatic tier classification based on skill complexity and
      composition depth
- [x] **97.2.1.7** Build ReAcTree-style hierarchical agent tree execution:
      complex tasks decomposed into subgoals within a dynamically constructed
      agent tree with control flow nodes (sequential, parallel, conditional) and
      agent nodes (reason, act, expand into further subgoals)

#### 97.2.2 Inter-Domain Skill Bridge

- [x] **97.2.2.1** Build domain skill adapters that expose each domain's
      internal skill system through the unified interface: Iris tools →
      UniversalSkill, Maya abilities → UniversalSkill, Nous training methods →
      UniversalSkill, etc. for all 30+ existing systems
- [x] **97.2.2.2** Implement skill translation layer that converts between
      domain-specific skill formats and the canonical schema (preserving
      domain-specific metadata while enabling cross-domain discovery)
- [x] **97.2.2.3** Build cross-domain skill catalog that indexes all skills from
      all domains with unified search, showing which skills can interoperate and
      suggesting composition opportunities
- [x] **97.2.2.4** Implement skill transfer learning detector: identify when a
      skill pattern learned in one domain could benefit another (e.g., Maya
      combat balance tuning patterns could inform Euterpe music mixing balance)

### 97.3 Autonomous Skill Generation & Growth

#### 97.3.1 Skill Auto-Generation from Documentation

- [x] **97.3.1.1** Build OpenAPI-to-skill auto-generator: parse all internal
      OpenAPI/Swagger specs across Oshun services → generate typed skill
      definitions with parameter schemas, return types, error handling, and
      usage examples (following Gorilla/ToolFactory pattern, with aggressive
      pruning per Stainless recommendations)
- [x] **97.3.1.2** Build documentation-to-skill pipeline: parse README.md, API
      docs, JSDoc/TSDoc comments → extract capability descriptions → generate
      SKILL.md files with instructions, parameter definitions, and example
      invocations
- [x] **97.3.1.3** Build code-to-skill extractor: analyze exported
      functions/classes across all domains → identify skill-worthy operations
      (public, documented, side-effect-free or safely sandboxed) → generate
      skill wrappers automatically
- [x] **97.3.1.4** Build database-schema-to-skill generator: parse
      Prisma/Drizzle schemas → generate CRUD skills for each entity with
      validation, access control, and audit logging
- [ ] **97.3.1.5** Docstring-driven skill invocation (NexusRaven-style). Split
      2026-09-18.
  - [ ] **97.3.1.5.a** Build the runtime in
        `libs/oshun/skill-system/src/discovery.ts`: extract a tool schema (name,
        params, types) from each skill's TypeScript/Python docstring and
        signature, and let the assistant model (OpenRouter
        `deepseek/deepseek-v4-flash-0731`) invoke skills from those schemas with
        no per-skill glue. Verify: a spec registers three skills by docstring,
        asks the model to invoke one, and asserts the right skill runs with the
        parsed arguments.
  - [ ] **97.3.1.5.b** Train a specialized function-calling model on
        docstring→call data. Needs an owner decision on a compute and data
        budget (stated 2026-09-18). `blocked:governance`
- [x] **97.3.1.6** Build CREATOR-style LLM tool creation: when no existing skill
      matches a task, LLM generates both the skill code and its SKILL.md
      documentation, verifies via execution, then stores in the skill library

#### 97.3.2 VOYAGER-Pattern Autonomous Skill Library Growth

- [x] **97.3.2.1** Build VOYAGER-style automatic curriculum that proposes
      increasingly complex tasks based on current skill library coverage and
      known capability gaps
- [x] **97.3.2.2** Implement skill code generation + verification loop: LLM
      generates skill implementation → execute in sandbox → capture results and
      errors → feed back to LLM for refinement → repeat until self-verification
      confirms success → commit to library
- [x] **97.3.2.3** Build skill library indexing via description embeddings
      (store each skill's natural language description as vector → retrieve
      top-k similar skills when facing new task → compose retrieved skills into
      solution)
- [x] **97.3.2.4** Implement skill deduplication via semantic similarity (when
      new skill's description embedding is within threshold of existing skill,
      compare execution behavior → merge, override, or keep both with
      differentiating tags)
- [x] **97.3.2.5** Build compositional skill creation: when existing atomic
      skills can be combined to solve a new task, automatically generate a
      composed skill that imports and chains them (skills as composable
      subroutines)
- [x] **97.3.2.6** Implement Memento-Skills pattern: skills stored as structured
      markdown files serving as persistent, evolving memory — start with minimal
      seed skills and grow organically from execution experience, with domain
      clusters forming autonomously

#### 97.3.3 Self-Evolving Skill System

- [x] **97.3.3.1** Build OpenSpace-style self-evolving skill engine with three
      evolution modes: FIX (repair broken skills from error logs), DERIVED
      (create enhanced/specialized versions from successful executions),
      CAPTURED (extract novel patterns from successful action sequences and
      codify as new skills)
- [x] **97.3.3.2** Implement evolution cycle: intercept task executions via
      hooks → log inputs/outputs to skill execution store → run diff analysis on
      failures and success metrics → trigger git-like commits to skill files →
      safety: confirmation gates, anti-loop guards, validation before
      replacement
- [ ] **97.3.3.3** Active tool discovery (MCP-Zero-style). Split 2026-09-18.
  - [ ] **97.3.3.3.a** Build the discovery runtime in
        `libs/oshun/skill-system/src/mcp.ts`: from a structured tool request
        (capability, inputs, constraints), route hierarchically over the skill
        registry by semantic match and return ranked candidates. Verify: a spec
        issues three requests against a fixture registry and asserts the top
        match and that an impossible request returns none.
  - [ ] **97.3.3.3.b** Train the request-generation policy via RL. Needs an
        owner decision on a compute and data budget (stated 2026-09-18).
        `blocked:governance`
- [x] **97.3.3.4** Implement skill quality feedback loop: execution telemetry
      (97.1.2.2) feeds into skill quality scores → low-quality skills flagged
      for evolution → FIX mode attempts repair → if repair fails, skill
      deprecated and alternatives promoted
- [x] **97.3.3.5** Build ALAS-style skill mastery tracking: per-skill accuracy
      tracking across execution history → when accuracy drops below threshold,
      trigger skill refinement cycle → converge when all skills exceed mastery
      threshold
- [x] **97.3.3.6** Implement network-effect skill sharing: when one domain
      evolves a skill improvement, propagate the improvement to all domains
      using equivalent skills (collective intelligence pattern from OpenSpace)

### 97.4 MCP Skill Ecosystem

#### 97.4.1 MCP Server Infrastructure

- [x] **97.4.1.1** Build MCP server generator that auto-creates an MCP server
      for each Oshun domain, exposing domain capabilities as MCP tools with
      typed schemas, resource endpoints, and prompt templates
- [x] **97.4.1.2** Implement MCP server for Iris: expose conversation tools
      (start, message, branch, summarize), agent tools (search, analyze,
      generate), memory tools (store, retrieve, forget), knowledge tools (query,
      cite, verify)
- [x] **97.4.1.3** Implement MCP server for Isis: expose generation tools
      (text-to-image, image-to-image, text-to-video, text-to-3d, text-to-audio),
      workflow tools (create, execute, monitor), quality tools (evaluate,
      compare, benchmark)
- [x] **97.4.1.4** Implement MCP server for Sophia: expose search tools
      (semantic, keyword, hybrid), knowledge tools (entity extract, relation
      extract, graph query), evaluation tools (RAG eval, citation verify)
- [x] **97.4.1.5** Implement MCP server for Maya: expose world tools (create
      zone, spawn NPC, modify terrain), game tools (grant item, trigger event,
      query state), analytics tools (player metrics, economy metrics, combat
      metrics)
- [x] **97.4.1.6** Implement MCP server for Nous: expose training tools (launch
      training, monitor progress, evaluate checkpoint), safety tools (check
      content, classify risk, detect injection), model tools (serve, quantize,
      merge)
- [x] **97.4.1.7** Implement MCP server for Veritas: expose news tools
      (research, write, fact-check, translate), editorial tools (prioritize,
      schedule, publish), analytics tools (engagement metrics, bias analysis)
- [x] **97.4.1.8** Implement MCP server for remaining domains: Psyche (avatar,
      voice, emotion), Hathor (world, NPC, narrative), Metis (lecture,
      assessment, tutor), Yemaya (project, asset, pipeline), Kuan Yin (moderate,
      protect, heal), Euterpe (compose, arrange, produce), and all other domains
- [x] **97.4.1.9** Build MCP server registry (internal): catalog all Oshun MCP
      servers with capabilities, versioning, health status, and usage metrics —
      discoverable via semantic search

#### 97.4.2 MCP Skill Wrapping & Marketplace

- [x] **97.4.2.1** Build MCP tool-to-skill wrapper: automatically generate
      SKILL.md files that wrap MCP tools with domain-specific instructions,
      examples, error handling, and prompt templates — transforming raw MCP
      tools into composable skills
- [x] **97.4.2.2** Build internal skill marketplace (Oshun Skills Hub): browse,
      search, install, rate, and review skills created across all domains — with
      quality scores, usage statistics, and domain categorization
- [x] **97.4.2.3** Integrate Skills.sh (Vercel) for importing
      community-contributed skills compatible with the Agent Skills Standard
- [x] **97.4.2.4** Integrate ClawHub for publishing Oshun skills to the broader
      AI agent ecosystem (semver, security scanning, vector search)
- [x] **97.4.2.5** Build skill security scanner: validate all skills (internal
      and imported) for dangerous operations, data exfiltration risks, resource
      abuse, and supply-chain attacks (following Snyk integration pattern)
- [x] **97.4.2.6** Implement skill sandboxing tiers: Tier 1 (read-only, no
      network, no filesystem), Tier 2 (limited filesystem, no network), Tier 3
      (filesystem + controlled network), Tier 4 (full access with audit logging)

### 97.5 Skill Quality, Evaluation & Benchmarking

#### 97.5.1 Quality Tracking & Evaluation

- [x] **97.5.1.1** Build unified skill quality dashboard tracking per-skill:
      success rate, average latency, cost per execution, user satisfaction (from
      data flywheel), error rate by type, retry rate, and quality trend over
      time
- [x] **97.5.1.2** Implement skill regression testing framework: maintain test
      cases per skill (input + expected output), run on every skill version
      update, alert on regressions, prevent publishing if tests fail (following
      Anthropic Skill Creator 2.0 eval pattern)
- [x] **97.5.1.3** Build BFCL-style skill evaluation benchmark for Oshun: test
      function calling accuracy, parallel calling, multi-turn tool use, agentic
      reasoning, error recovery, and format sensitivity across all domain skills
- [x] **97.5.1.4** Implement tau-2-Bench-style agent-user interaction
      evaluation: simulated users interact with Oshun agents using domain
      skills, measuring task completion, consistency across repetitions, and
      user satisfaction
- [x] **97.5.1.5** Build MCPAgentBench-style MCP tool use evaluation: test
      agents using Oshun MCP servers on real domain tasks, measuring tool
      selection accuracy, composition quality, and end-to-end success
- [x] **97.5.1.6** Implement skill degradation detection: statistical process
      control on skill quality metrics → alert when quality drops below 2
      standard deviations from baseline → trigger auto-improvement cycle
      (97.3.3)

#### 97.5.2 Process Reward Models for Skill Execution

- [ ] **97.5.2.1** Build an AgentPRM-style process reward model that scores each
      step of a skill chain using Monte-Carlo rollouts, and train it. Needs an
      owner decision on a compute and data budget — this is model
      training/fine-tuning, not a coding task (stated 2026-09-18).
      `blocked:governance`
- [x] **97.5.2.2** Implement ToolPRMBench-style evaluation for Oshun skill
      chains: convert multi-step skill execution trajectories into step-level
      test cases, evaluate both single-step and multi-step failure detection
- [x] **97.5.2.3** Build hierarchical reward model that evaluates individual
      skill steps AND consecutive step sequences at fine-grained and
      coarse-grained levels
- [x] **97.5.2.4** Implement step-level skill improvement: when PRM identifies a
      weak step in a skill chain, trigger targeted improvement (regenerate that
      step's skill code, fine-tune the model on that step's data, or replace
      with an alternative skill)

### 97.6 Skill-Specific Model Training

#### 97.6.1 Tool-Use Fine-Tuning Pipeline

- [x] **97.6.1.1** Build skill execution data collector that transforms skill
      telemetry (97.1.2.2) into tool-use fine-tuning format: {system prompt with
      tool schemas, user query, assistant response with tool calls, tool
      results, final response} — feeding Phase 87 dataset registry
- [x] **97.6.1.2** Build APIGen-MT-style multi-turn skill-use training data
      generator: Phase 1 (task blueprint generation with ground-truth skill
      actions, validated by LLM committee), Phase 2 (simulated dialogue
      generating realistic multi-turn conversations, retaining only trajectories
      matching expected ground-truth)
- [x] **97.6.1.3** Build ToolACE-R-style model-aware iterative training for
      skill use: progressively adjust training data difficulty based on model's
      evolving capabilities, with adaptive self-refinement at inference time
- [ ] **97.6.1.4** Implement the xLAM-style tool-use fine-tuning pipeline and
      fine-tune Oshun's self-hosted LLMs (Phase 89) into domain-expert tool-use
      models. Needs an owner decision on a compute and data budget — this is
      model training/fine-tuning, not a coding task (stated 2026-09-18).
      `blocked:governance`
- [x] **97.6.1.5** Build ToolMind-style function graph construction: build graph
      of Oshun skills based on parameter correlations → sample skill chains via
      random walks → generate diverse multi-step training scenarios → turn-level
      quality filtering
- [ ] **97.6.1.6** Gorilla RAFT for Oshun skills. Split 2026-09-18.
  - [ ] **97.6.1.6.a** Build the retrieval half at inference in
        `libs/oshun/skill-system/src/discovery.ts`: index skill documentation
        and, at invocation time, retrieve the up-to-date doc for the chosen
        skill and put it in the model's context. Verify: a spec asserts the
        retriever returns the correct skill doc for a query and that a changed
        doc is reflected without re-embedding the whole corpus.
  - [ ] **97.6.1.6.b** Retrieval-augmented fine-tuning (RAFT) of the invocation
        model. Needs an owner decision on a compute and data budget (stated
        2026-09-18). `blocked:governance`

#### 97.6.2 Skill Learning from Demonstrations

- [x] **97.6.2.1** Build skill learning from demonstration (LfD) pipeline:
      capture expert user sessions → extract skill invocation patterns →
      identify successful action sequences → codify as new skills or improve
      existing ones
- [ ] **97.6.2.2** Implement behavioral cloning for skill acquisition: train
      models to replicate successful skill-usage patterns from large-scale
      cross-domain user interaction data. The data does not exist (large-scale
      user interaction data); no production users or history exist to learn from
      (stated 2026-09-18). `blocked:corpus`
- [ ] **97.6.2.3** Build AutoTIR-style autonomous tool-integrated reasoning:
      train via RL to decide whether and which skill to invoke, with a hybrid
      reward. Needs an owner decision on a compute and data budget — this is
      model training/fine-tuning, not a coding task (stated 2026-09-18).
      `blocked:governance`
- [ ] **97.6.2.4** Implement o3/o4-pattern RL-trained skill reasoning so the
      model uses skills natively within its chain of thought. (Restated
      2026-09-18: the original had no measurable target; the owner decision must
      also set the eval and the success bar.) Needs an owner decision on a
      compute and data budget — this is model training/fine-tuning, not a coding
      task (stated 2026-09-18). `blocked:governance`

### 97.7 Per-Domain Skill Enhancement

#### 97.7.1 Iris Agent Skill Enhancement

- [x] **97.7.1.1** Migrate Iris tool framework to unified skill registry (adapt
      ToolId/ToolVersion/ToolDefinition to UniversalSkill, preserve existing
      tool discovery and composition capabilities)
- [x] **97.7.1.2** Build VOYAGER-style autonomous tool library growth for Iris:
      automatic curriculum proposes new tool-use challenges → generate tool
      compositions → verify via execution → add to library
- [x] **97.7.1.3** Implement MCP skill wrapping for all Iris tools: each tool
      gets a SKILL.md wrapper with instructions, examples, and context
- [ ] **97.7.1.4** Build the tool-use fine-tuning pipeline for the Iris
      conversation model from production tool-call data. The data does not exist
      (production tool-call traffic from Iris); no production users or history
      exist to learn from (stated 2026-09-18). `blocked:corpus`
- [x] **97.7.1.5** Implement confidence-based tool selection: when Iris model is
      uncertain which tool to use, expose multiple options to user or run
      parallel tools and compare results

#### 97.7.2 Maya Game Skill Enhancement

- [ ] **97.7.2.1** Build dynamic ability generation: from player combat data,
      find coverage gaps, procedurally generate balanced abilities and playtest
      them. The data does not exist (player combat telemetry and playtest
      cohorts); no production users or history exist to learn from (stated
      2026-09-18). `blocked:corpus`
- [ ] **97.7.2.2** Implement ML-based combo discovery from player combat
      sequences, codifying effective combinations as named combos. The data does
      not exist (player combat sequence data); no production users or history
      exist to learn from (stated 2026-09-18). `blocked:corpus`
- [ ] **97.7.2.3** Build NPC skill learning from player behavior for emergent
      difficulty scaling. The data does not exist (observed player combat
      strategies); no production users or history exist to learn from (stated
      2026-09-18). `blocked:corpus`
- [ ] **97.7.2.4** Implement Eureka-style automatic reward-function design for
      NPC AI: the assistant model (OpenRouter `deepseek/deepseek-v4-flash-0731`)
      generates candidate reward functions for a stated NPC behaviour goal, each
      is scored in the existing simulator, and the loop iterates on the best.
      This needs no training and no production data. Verify: a spec runs the
      loop for one goal (e.g. "guard patrols and investigates noises") over a
      fixed simulator seed and asserts the selected reward's simulated score
      beats the initial candidate and a hand-written baseline.

#### 97.7.3 Nous Training Skill Enhancement

- [x] **97.7.3.1** Build training recipe auto-selector: given a fine-tuning task
      description + dataset characteristics + compute budget → automatically
      recommend optimal training method (LoRA vs QLoRA vs full FT), optimizer
      (AdamW vs Muon), and hyperparameters
- [x] **97.7.3.2** Implement training recipe composition: chain multiple
      training stages (SFT → DPO → GRPO) as a composed skill with automatic data
      handoff, checkpoint management, and progressive evaluation
- [x] **97.7.3.3** Build safety ensemble optimizer: analyze safety check
      accuracy across 25+ detectors → optimize ensemble weights and routing →
      maximize detection rate while minimizing false positives

#### 97.7.4 Isis Generation Skill Enhancement

- [x] **97.7.4.1** Build auto-workflow generator: given a generation task
      description → automatically compose a ComfyUI/pipeline workflow from
      available generation skills → estimate quality and cost → execute
- [ ] **97.7.4.2** Implement learned parameter optimization from production
      generation-parameter → quality correlations. The data does not exist
      (production generation-quality data); no production users or history exist
      to learn from (stated 2026-09-18). `blocked:corpus`
- [x] **97.7.4.3** Build workflow quality predictor: given a workflow definition
      and parameters → predict expected output quality before execution → route
      to cheaper/faster workflow if predicted quality exceeds threshold

#### 97.7.5 Veritas Editorial Skill Enhancement

- [ ] **97.7.5.1** Build editorial skill improvement: fine-tune agent prompts on
      the highest-engagement/accuracy/trust newsroom outputs. The data does not
      exist (production article-performance data); no production users or
      history exist to learn from (stated 2026-09-18). `blocked:corpus`
- [ ] **97.7.5.2** Implement beat-specific skill specialization through
      continuous execution feedback. The data does not exist (a history of agent
      execution feedback per beat); no production users or history exist to
      learn from (stated 2026-09-18). `blocked:corpus`
- [ ] **97.7.5.3** Build editorial skill transfer: when one agent records an
      effective reporting technique, propagate it to agents on other beats.
      Implement the mechanism in `libs/oshun/skill-system/src/composition.ts` —
      a technique registry an agent publishes to and others subscribe from, with
      beat-tagging and conflict handling. Verify: a spec has agent A publish a
      technique, asserts agents B and C (different beats) receive it, that a
      beat-scoped technique is not force-applied where it does not fit, and that
      a duplicate publish is idempotent. The upstream discovery of which
      techniques work is 97.7.5.1 (corpus-blocked); this item is the propagation
      plumbing and stands alone.

#### 97.7.6 All Remaining Domains

- [x] **97.7.6.1** Migrate Psyche expression/accessibility/safety skills to
      unified framework with quality tracking and MCP exposure
- [x] **97.7.6.2** Migrate Hathor NPC behavior/narrative skills to unified
      framework with quality tracking, personality consistency scoring, and MCP
      exposure
- [x] **97.7.6.3** Migrate Metis educational agent/assessment skills to unified
      framework with learning outcome tracking and MCP exposure
- [x] **97.7.6.4** Migrate Yemaya production pipeline/QA skills to unified
      framework with pipeline efficiency tracking and MCP exposure
- [x] **97.7.6.5** Migrate Bellona code generation skills to unified framework
      with compilation success tracking and MCP exposure
- [x] **97.7.6.6** Migrate Euterpe music generation/production skills to unified
      framework with audio quality tracking and MCP exposure
- [x] **97.7.6.7** Migrate Shakti fitness/training skills to unified framework
      with form quality tracking and MCP exposure
- [x] **97.7.6.8** Migrate Aglaea beauty/styling skills to unified framework
      with user satisfaction tracking and MCP exposure
- [x] **97.7.6.9** Migrate Maat strategic analysis skills to unified framework
      with prediction accuracy tracking and MCP exposure
- [x] **97.7.6.10** Migrate Aphrodite game strategy/streaming skills to unified
      framework with engagement tracking and MCP exposure
- [x] **97.7.6.11** Migrate Uzume live production skills to unified framework
      with broadcast quality tracking and MCP exposure
- [x] **97.7.6.12** Migrate Kuan Yin safety/moderation skills to unified
      framework with intervention effectiveness tracking and MCP exposure
- [x] **97.7.6.13** Migrate Demeter/Asase agriculture skills to unified
      framework with yield prediction tracking and MCP exposure
- [x] **97.7.6.14** Migrate Cybele real estate/construction skills to unified
      framework with valuation accuracy tracking and MCP exposure
- [x] **97.7.6.15** Migrate Themis governance/compliance skills to unified
      framework with governance effectiveness tracking and MCP exposure
- [x] **97.7.6.16** Migrate Nisaba/Mnemosyne language/learning skills to unified
      framework with proficiency improvement tracking and MCP exposure
- [x] **97.7.6.17** Migrate Calliope creative/artist skills to unified framework
      with creative consistency tracking and MCP exposure
- [x] **97.7.6.18** Migrate Galatea robotics skills to unified framework with
      motor control accuracy tracking and MCP exposure
- [x] **97.7.6.19** Migrate Oya drone/swarm skills to unified framework with
      flight safety tracking and MCP exposure
- [x] **97.7.6.20** Migrate all remaining domain skills (Nyx, Tara, Meditation,
      Aje, Freya, Brigid, Saraswati, Lakshmi, Annapurna, Athena, Seshat, Airmid)
      to unified framework with domain-appropriate quality tracking and MCP
      exposure
