How ruc-datalab/skilladam works
SkillAdam is an agent-facing tool for improving a skill file: a user requests an optimization, the product prepares tasks and scoring, runs the current and revised skill, and validates proposed edits before applying them. The central runtime connects product services to rollout/model execution and the SkillAdam optimization loop, then supports persisted runs and optional per-edit review. The README documents Codex, Claude Code, Cursor Agent, and Copilot support. Edges are limited to relationships shown in the README or source index/excerpts; several adapter and persistence details are unsampled.
Main entry points: service.py Run Service · skilladam.py SkillAdam Runner · patch_review.py Patch Selection · evaluate.py Evaluation Runner
Components 28
Agent Product
- Product Adapters
skilladam/product_mcp.py - Run Service
skilladam/product/service.py - Task Setup
skilladam/product/tasking.py - Product Workflow
skilladam/product/workflow.py - Session State
skilladam/product/session_store.py - Hunk Review
skilladam/product_review_mcp.py - Patch Selection
skilladam/product/patch_review.py - Product Model
skilladam/product/model.py
Optimization Core
- SkillAdam Runner
skilladam/methods/skilladam.py - Feedback Loop
skilladam/core/feedback_loop.py - Case Sampling
skilladam/core/case_sampler.py - Optimization Memory
skilladam/core/momentum.py - Acceptance Gate
skilladam/core/acceptance_gate.py - Trajectory Condenser
skilladam/core/trajectory_condenser.py - Diff Application
skilladam/core/diff_apply.py - SkillOpt Baseline
skilladam/baselines/skillopt/runner.py
Evaluation Runtime
- Evaluation Runner
skilladam/execution/evaluate.py - Skill Resolver
skilladam/execution/skills.py - Model Backend
skilladam/execution/backend.py - OpenAI-Compatible
skilladam/backends/openai_compatible.py - Usage Ledger
skilladam/cost/ledger.py
Experiments
- Experiment Planning
skilladam/experiments/planning.py - API Preflight
skilladam/experiments/api_preflight.py - Main Results
skilladam/experiments/main_results.py
Other components
- User
- Agent Host
- Skill File
- Model Provider
Connections 29
- User → Agent Host: requests optimization
README.md - Agent Host → Product Adapters: invokes workflow
- Product Adapters → Run Service: delegates
- Run Service → Task Setup: prepares tasks
skilladam/product/service.py - Run Service → Product Workflow: runs workflow
skilladam/product/service.py - Run Service → Session State: persists session
skilladam/product/service.py - Run Service → Evaluation Runner: creates evaluator
skilladam/product/service.py - Run Service → SkillAdam Runner: coordinates optimization
- Run Service → Product Model: records rollouts
skilladam/product/service.py - Product Model → Model Backend: uses backend
- SkillAdam Runner → Feedback Loop: runs loop
skilladam/methods/skilladam.py - SkillAdam Runner → Case Sampling: samples cases
skilladam/methods/skilladam.py - SkillAdam Runner → Optimization Memory: updates memory
skilladam/methods/skilladam.py - SkillAdam Runner → Trajectory Condenser: condenses rollouts
skilladam/methods/skilladam.py - SkillAdam Runner → Evaluation Runner: evaluates skill
skilladam/methods/skilladam.py - Feedback Loop → Acceptance Gate: checks candidate
skilladam/core/feedback_loop.py - Evaluation Runner → Skill Resolver: resolves skill
skilladam/execution/evaluate.py - Evaluation Runner → Model Backend: executes evaluation
skilladam/execution/evaluate.py - Model Backend → OpenAI-Compatible: selects implementation
skilladam/execution/backend.py - OpenAI-Compatible → Model Provider: requests completion
skilladam/backends/openai_compatible.py - Agent Host → Hunk Review: may review edits
README.md - Hunk Review → Patch Selection: submits selection
skilladam/product_review_mcp.py - Patch Selection → Diff Application: applies selected hunks
skilladam/product/patch_review.py - Run Service → Patch Selection: offers proposal review
skilladam/product/service.py - SkillAdam Runner → Skill File: proposes skill changes
README.md - Patch Selection → Skill File: updates after validation
README.md - SkillOpt Baseline → Usage Ledger: records usage
skilladam/baselines/skillopt/runner.py - Experiment Planning → Main Results: uses result config
skilladam/experiments/planning.py - API Preflight → Main Results: validates matrix
skilladam/experiments/api_preflight.py