How xinyuez/a-skill-trainer works
A-Skill-Trainer is an interactive skill-incubation framework. A user submits a task in the Streamlit app; an inference engine executes it and records traces, the Wiki Maintainer distills patterns, and the Skill Proposer uses patterns, traces, and prior audit feedback to produce a proposal. The Harness applies the proposal for human review, then commits or rolls back the change and records the decision for later iterations. README descriptions establish the broad lifecycle; sampled source confirms several implementation links, while unsampled wiring is left out or shown as documented.
Main entry points: harness.py HITL Harness · app.py Training UI · inference_agent.py Inference agent · skill_proposer.py Skill Proposer
Components 22
Training workflow
- Training UI
app.py - HITL Harness
trainer/harness.py - T module
T.py
Agent stages
- Inference agent
trainer/inference_agent.py - Antigravity engine
trainer/inference_antigravity.py - Wiki Maintainer
trainer/wiki_maintainer.py - Skill Proposer
trainer/skill_proposer.py
Workspace records
- Execution traces
trainer/utils/traces_writer.py - Trace access
trainer/layers/raw_layer.py - Trace reader
trainer/utils/traces_reader.py - Wiki patterns
trainer/wiki_maintainer.py - Skill proposal
trainer/skill_proposer.py - Skill audit trail
trainer/harness.py - Skill library
trainer/harness.py
Agent runtime
- Python sandbox tool
trainer/utils/run_python.py - Agent dispatch
trainer/utils/run_agent.py - Stream formatting
trainer/utils/stream_format.py - Tool-call guard
trainer/utils/awrap_tool_call.py - Local datetime
trainer/utils/get_current_datetime.py - Session IDs
trainer/utils/session_creator.py
Other components
- Task author
- Human reviewer
Connections 32
- Task author → Training UI: submits task
README.md - Training UI → Inference agent: starts inference
README.md - Inference agent → Execution traces: writes traces
README.md - Inference agent → Python sandbox tool: uses tool
trainer/inference_agent.py - Inference agent → Tool-call guard: installs guard
trainer/inference_agent.py - Inference agent → Session IDs: creates session
trainer/inference_agent.py - Inference agent → Agent dispatch: dispatches agent
trainer/inference_agent.py - Antigravity engine → Python sandbox tool: uses tool
trainer/inference_antigravity.py - Antigravity engine → Agent dispatch: dispatches agent
trainer/inference_antigravity.py - Antigravity engine → Session IDs: creates session
trainer/inference_antigravity.py - Training UI → Antigravity engine: selects engine
- Training UI → Wiki Maintainer: starts pattern extraction
README.md - Execution traces → Trace access: read via
trainer/layers/raw_layer.py - Trace access → Trace reader: delegates read
trainer/layers/raw_layer.py - Wiki Maintainer → Trace access: reads traces
trainer/wiki_maintainer.py - Wiki Maintainer → Wiki patterns: updates patterns
README.md - Training UI → Skill Proposer: starts proposal
README.md - Skill Proposer → Trace access: reads traces
trainer/skill_proposer.py - Skill Proposer → Wiki patterns: uses patterns
README.md - Skill Proposer → Skill audit trail: reads prior decisions
README.md - Skill Proposer → Skill proposal: generates proposal
README.md - Training UI → HITL Harness: submits proposal
README.md - HITL Harness → Skill proposal: applies proposal
trainer/harness.py - HITL Harness → Skill library: writes skill files
trainer/harness.py - HITL Harness → Human reviewer: requests decision
README.md - Human reviewer → HITL Harness: accepts or rejects
README.md - HITL Harness → Skill library: commits or rolls back
README.md - HITL Harness → Skill audit trail: records feedback and diff
README.md - Agent dispatch → Stream formatting: formats stream
trainer/utils/run_agent.py - Skill Proposer → Agent dispatch: runs agent
trainer/skill_proposer.py - Wiki Maintainer → Agent dispatch: runs agent
trainer/wiki_maintainer.py - Antigravity engine → Local datetime: adds current time
trainer/inference_antigravity.py