Workflow Automation
A registry-driven worker harness that turns wiki-guided context, commands, scheduled jobs, and queue tasks into visible local AI workflows.
The early version was useful because it proved that local capture and processing could work. The current version is broader: tasks enter through commands or schedules, persist in a queue, and route to workers discovered from a registry.
`run.py` creates shared resources, starts the `WorkerManager`, and then gets out of the way. The manager reads `.harness/registry/sub-agents/*.json`, starts essential workers, and wakes on-demand workers by task type.
Architecture Walkthrough
Pick a flow to see which components are active. The board shows the system as a set of explicit paths instead of one script trying to own every workflow.
Conductor
Shared Resources
Worker Manager
Essential Workers
On-Demand Workers
External Systems
`BotWorker` handles command intake for URL, YouTube, GitHub, wiki, benchmark, and coding work. `SchedulerWorker` adds recurring jobs. Both paths persist work into `TaskQueue`.
`BotWorker`, `QueueWorker`, and `SchedulerWorker` start on boot. Coding, digest, lint, benchmark, dashboard, PCR, improvement, and code-eval workers start when matching work appears, then stop after the idle grace window.
Heavy pipeline and improvement work routes to `gemma4:31b`. Medium coding and digest work routes to `qwen3-coder:30b`. Intake uses `claude-haiku-4-5` through `IntakeAgent`.
The architecture is broad enough to run multiple workflow families, but still constrained by registry entries, task types, shared resources, and worker lifecycle rules. Wiki-derived topic bibles guide the harness work, then logs, failures, and operational learnings feed back into the wiki so the next version has better context.
Worker Registry
Essential
Start on boot and keep the interface, pipeline executor, and recurring job scheduler alive.
Wiki
Generate daily vault digests and clean wiki pages through scheduled or queued tasks.
Coding
Run autonomous coding tasks and evaluate models against coding benchmarks.
Ops
Refresh dashboards, collect operational data, and analyze failures for candidate improvements.
Benchmark
Runs ingestion and pipeline quality evaluations across model choices.