feat: pi extension - #1

add --sync to sync cwd with sandbox

fb21567 — full CreateOS surface + tool prompt rewrite

What prompted this

The extension registered 33 tools whose prompts had each been written in isolation. Every tool described itself; nothing described how they relate. The model learned 33 verbs and no grammar — it could not route a request to the right tool, and several product capabilities were invisible to it because no tool existed to expose them.

Audited the prompt surface against createos sandbox --help (CLI v0.0.20) and website-04/content/docs/Sandbox/. Findings below.

1. sandbox_firewall never worked

It shelled out to sandbox edit --egress. sandbox edit has no --egress flag — it accepts only --ingress on|off, --add-ssh-key and --auto-pause. Every call died on unknown-flag. Egress control lives under a different command group entirely:

--egress exists only on create and fork. Additionally, an empty rules array produced a no-op rather than clearing the allowlist, which is the opposite of what the prompt promised.

sandbox_list had the same class of bug: it called sandbox list with no flags, which hides everything not active, while its own prompt claimed it listed paused sandboxes. It now takes all and status.

2. The prompt surface had no layering

Restructured into three layers, each with a distinct job:

Two facts were absent from every prompt and cost the model turns on every session:

• The user's files are not in the sandbox. It starts empty and cannot see their disk, so reading a /Users/… path always fails. The model would discover this by failing, then reach for sandbox_sync.
• Exec output is buffered. createos sandbox exec returns only when the command exits, so a foreground server (npm run dev, flask run) hangs the tool call forever. Nothing said to background it.

The system prompt now also carries the cross-tool routing rules, which are the part the model genuinely cannot derive:

• Port access: sandbox_preview_url → sandbox_tunnel → device VPN
• Many identical machines: sandbox_pause + sandbox_fork, not N × sandbox_create
• Anything that must outlive the sandbox: sandbox_sync, sandbox_disk_attach, or git
• bash is always the session sandbox; every other sandbox needs sandbox_exec

Per-tool guidelines now name the failure they prevent rather than restating the description — e.g. a firewall allowlist without a DNS resolver such as 1.1.1.1:53 breaks name resolution and every subsequent request.

3. Features with no tool behind them

A prompt cannot make the model aware of a capability that has no tool, so these were added. 33 tools → 47.

sandbox_create gained env, disks, egress, auto_pause, ssh_keys and disk_mib — all documented CLI flags the tool previously dropped.

4. Docs drift

packages/pi-extension/README.md listed five agent tools that do not exist (sandbox_manage, tunnel, preview_url, network, device). Both READMEs carried stale tool counts, and CLAUDE.md's inventory was off by seven. All corrected.

Verification

Drove the real src/cli.ts wrappers against a live sandbox, not mocks:

PASS  createSandbox(env, egress, autoPause) — sb-01kz6510wh2bbkb3z9kczekw30
PASS  ingress enabled at create            — https://01kz…-<port>.app.sb.createos.sh
PASS  create --env reaches exec            — "yes"
PASS  showFirewall reflects create --egress — {"rules":["pypi.org","1.1.1.1:53"]}
PASS  setFirewall replaces allowlist       — {"rules":["github.com","1.1.1.1:53"]}
PASS  clearFirewall empties allowlist      — {"rules":[]}
PASS  editSandbox --auto-pause off
PASS  listSandboxes(status)                — 12 running
PASS  listTemplates
cleaned up sb-01kz6510wh2bbkb3z9kczekw30

9/9, sandbox destroyed afterwards. tsc --noEmit and oxlint both clean.

Structure

src/tools-resources.ts (networks, disks, templates, devices) split out of tools.ts to stay under the 1100-line file limit. tools.ts calls registerResourceTools(); the import back from tools-resources.ts is type-only and erased at compile time, so the cycle is not real.

Deliberately left out

• sandbox exec --stream — pi.exec() returns only on process exit, so streaming changes nothing observable.
• sandbox shell — an interactive PTY, which a tool call cannot host. sandbox_editor covers the "let the user in" case.
• REST-only features (self-signal, metrics, bandwidth recharge, resize, webhooks) — no createos sandbox subcommand exists for them, and this extension is CLI-only by design. They stay out until the CLI carries them.

The scope question, the three options weighed, and these exclusions are recorded in packages/pi-extension/docs/decisions.md.

Note for reviewers

The oxfmt pre-commit hook reformats any .ts file it touches, and this repo had never been oxfmt'd (it was single-quote / no-semicolon). Roughly 80% of the diff in index.ts, cli.ts and tools.ts is therefore style, not substance. src/ops.ts, find-tool.ts and grep-tool.ts were left untouched, so the package is currently half-formatted — worth a separate mechanical commit to make it uniform.