Agent author guide
This guide is for people building agents or workflow runners on top of AetherFS.
Your job is to make good use of sessions and the public control surface without leaking internal assumptions into your design.
What matters most in this role
As an agent author, the highest-signal concepts are:
- session lifecycle
- manifest-driven discovery
- mutation routes
- cache versus bus versus annotations
- checkpoints
- approvals
- commits and promotions
Your core workflow
The default agent workflow is:
- create or receive a session
- inspect session detail and manifest
- mutate files or paths
- publish progress or derived state
- checkpoint before risky changes
- request approval when required
- commit or promote after the workflow allows it
Read these pages first
Start here:
Then use these references:
Discovery model
Do not start by guessing path shapes or assuming a generic remote disk API.
Use:
- session detail for current workspace context
- manifest for tree discovery
That is the correct public model.
Mutation model
Choose the simplest mutation surface that matches the job.
Use:
- directory and rename routes for path changes
- batch when one logical action contains several path changes
patchSegmentswhen you truly need partial or sparse mutation- expected hashes when stale clients must fail safely
State model
Choose the right metadata surface.
Use:
- cache for keyed structured workflow state
- bus for live progress and event-style coordination
- annotations for durable review notes
- events for durable history
Do not use one of these surfaces as a substitute for all the others.
Approval model
Use approvals when your agent is about to cross a meaningful workflow boundary.
Good triggers:
- commit to a durable target
- large-scale delete
- source promotion
- risky transformation
When requesting approval:
- include a clear machine-readable reason
- include useful human-facing details
- expect the session to pause in
SESSION_STATUS_PENDING_APPROVAL
Persistence model
Agents should not treat commit as a save button.
Use:
- checkpoints for recoverability
- commits for durable outcomes
- source promotion for reusable future baselines
Concurrency model
The default order is:
- expected hashes
- collaboration locks when contested edits matter
- POSIX locks only when low-level compatibility is actually required
Common agent-author mistakes
Avoid:
- inventing unsupported REST file routes
- treating the bus as durable state
- treating xattrs as review comments
- skipping checkpoints before risky automated edits
- assuming health, approval, and persistence are interchangeable concepts
Fastest useful agent-author toolkit
If you only remember one reading path:
- sessions
- manifest
- mutation routes
- cache/bus/annotations
- approval
- commit/promotion
That is the core of a strong public agent integration.