SmolFS looks useful for the part of agent work that is usually hand-waved away: a workspace survives the agent process.
One small feature that would make the boundary more auditable: a generated workspace handoff manifest that an agent can attach when it stops, and the next agent/process can read before touching files.
Shape I mean:
{
"volume": "demo",
"mount_path": "./workspace",
"generated_at": "2026-06-25T01:00:00Z",
"last_flush": {
"at": "...",
"status": "ok"
},
"paths": {
"changed_since_mount": ["src/foo.ts", "notes/plan.md"],
"large_files_omitted": [],
"ignored_or_sensitive_paths": [".env", "node_modules/"]
},
"privacy": {
"file_contents_included": false,
"secrets_scanned": true
},
"next_agent_hint": "read notes/plan.md before editing src/foo.ts"
}Why I think this fits SmolFS:
- Durable workspace is only half the handoff; the next runtime also needs to know what state is safe/relevant to load.
- A manifest avoids making raw file contents, prior transcripts, or ad-hoc summaries the default evidence format.
- It gives runners/SDK users a simple preflight: mount volume -> read manifest -> decide whether to continue, inspect, or ask for a fresh flush.
Could be a command like:
smolfs manifest demo --since-mount --json > .smolfs-handoff.json
or a field on smolfs status demo --json if you want to keep the surface smaller.
The important invariant is not the exact schema; it is that durable agent files come with a tiny, privacy-safe load boundary so the next process does not have to infer state from the whole folder.
SmolFS looks useful for the part of agent work that is usually hand-waved away: a workspace survives the agent process.
One small feature that would make the boundary more auditable: a generated workspace handoff manifest that an agent can attach when it stops, and the next agent/process can read before touching files.
Shape I mean:
{ "volume": "demo", "mount_path": "./workspace", "generated_at": "2026-06-25T01:00:00Z", "last_flush": { "at": "...", "status": "ok" }, "paths": { "changed_since_mount": ["src/foo.ts", "notes/plan.md"], "large_files_omitted": [], "ignored_or_sensitive_paths": [".env", "node_modules/"] }, "privacy": { "file_contents_included": false, "secrets_scanned": true }, "next_agent_hint": "read notes/plan.md before editing src/foo.ts" }Why I think this fits SmolFS:
Could be a command like:
smolfs manifest demo --since-mount --json > .smolfs-handoff.jsonor a field on
smolfs status demo --jsonif you want to keep the surface smaller.The important invariant is not the exact schema; it is that durable agent files come with a tiny, privacy-safe load boundary so the next process does not have to infer state from the whole folder.