Skip to main content

Loader and allowlists

The loader is how a large pack stays cheap. run-workflow runs:

python3 .pipeline/loader/load_workflow.py --workflow {name} --step {step} --slug {slug}

It writes allowed_reads, seed_reads, and seed_bundle (features/{slug}/step-context.md) under {pack}/state/active-context.json. The specialist must not open other .pipeline/skills/ files.

What is allowlisted​

Each .pipeline/workflows/{name}.json lists context.parent.files and context.steps.{agent}.files — agent briefs, SKILL.md, config, wiki. Do not list skill assets/ there.

The loader expands allowed_reads from those skills (the skill's assets/ folder plus markdown links to other skill assets) so specialists may Read a template when the skill names it. seed_reads is the preload set: briefs, skills, and config. Pipeline-state and other skill assets stay on allowed_reads only. The loader concatenates seed_reads into seed_bundle so the specialist Reads once, not one tool call per file.

BA does not ingest the deploy runbook. The developer does not ingest Jira intake. Wiki pages are loaded only when INDEX triggers match — not as a folder.

Receptionist​

AGENTS.md is a routing table. The agent does not load the pack until run-workflow runs the loader. Weather, locations, and other unrelated asks skip the pipeline entirely.

Do not edit​

.pipeline/loader/ is kit machinery. Add files by listing them on a workflow allowlist, not by teaching the parent to glob .pipeline/.