Skip to content

roadmap.md

Latest commit

 

History

History
55 lines (42 loc) · 3.77 KB

roadmap.md

File metadata and controls

55 lines (42 loc) · 3.77 KB

sandfs Roadmap

This document captures the practical roadmap for sandfs so we can measure progress toward a "VFS-first" agent environment where the virtual tree feels like the whole world.

Where we are today

  • Core VFS – in-memory directories, provider-backed nodes, optimistic versions, snapshots, host export/import helpers.
  • Policy layer – read/write/append policies plus visibility views and path hooks so hosts can mirror activity elsewhere.
  • Command surface – built-in shell commands: pwd, cd, ls, tree, cat, touch, mkdir, rm, cp, mv, write, append, echo, printf, wc, grep, rg, search, stat, head, tail, find, python, python3, host, bash, sh, and help.
  • Host bridge – host materializes the tree to a temp dir and runs native binaries; host fallback is opt-in for untrusted users.

Principles for upcoming work

  1. Pure VFS illusion by default – agents should accomplish >80% of tasks without touching host fallback.
  2. Small, opinionated command set – implement the utilities agents actually need (read/search/light edits) instead of replicating full POSIX semantics.
  3. Explicit escape hatches – keep host fallback as an opt-in convenience and clearly document its risks.
  4. Pluggable execution backends – over time, allow swapping in WASI/containers/namespaces without changing the shell API.

Near-term backlog (v0.x)

Focus: expand the pure-Python toolchain so normal agent workflows never require host commands.

  • Navigation & metadata
    • Add stat/info for per-path metadata (size, policy, version, timestamps).
    • Support ls -a/ls -R flags plus a du-lite summary command.
  • Viewing & paging
    • Implement head, tail, and a pager-style view for long files.
    • Extend cat with line numbers and highlighting options useful to LLMs.
  • Search & filtering
    • Ship a find/fd-style command for filtered traversals (name/glob/extension filters).
    • Enhance rg with include/exclude globs, max-results, and JSON output for agents.
  • Editing & diffs
    • Introduce cp, mv, apply_patch, and diff so agents can restructure files safely.
    • Provide sed-lite replacements and multi-file write helpers.
  • File creation ergonomics
    • Extend write/append with heredoc (<<EOF) support, timestamp helpers, and batch command execution so agents can create structured notes without /bin/sh.
  • Snapshots & history
    • Surface shell commands for snapshot, restore, and diff-snap <id> (wrapping the existing VFS APIs).
  • Ergonomics
    • Structured command responses (rich metadata that the CLI/UI can pretty-print) and better error messaging.
    • Diagnostics helpers like status/info to expose visible roots, policies, and current mounts without falling back to host shells.

Mid-term explorations

  • Structured data helpers – select/jq-style view of JSON/YAML/TOML, CSV previewers, formatting utilities.
  • Archive / packaging – zip/tar helpers entirely within the VFS so agents can bundle outputs without host tooling.
  • Execution backends – research a WASI-based backend (reuse BusyBox/coreutils) and, for Linux deployments, a namespace/container-backed host runner for "hard" isolation.

Long-term ideas

  • Policy-aware agents – surface node policies directly in command responses so planners can avoid forbidden paths proactively.
  • Observability – pluggable telemetry for command usage, timing, and write hooks that integrate with host logs.
  • Multi-tenant sandboxes – carve multiple views over the same VFS (per agent/principal) with isolated shells.

This roadmap will evolve as we learn what commands agents rely on most. Contributions or feedback can be filed as issues referencing the relevant section above.