Files
PaperClipAI/doc/plugins/SANDBOX_FILE_SYNC_HOOKS.md
b247bf7150 feat(runtime): opt-in sandbox file-sync lifecycle hooks (API + provider docs) (#10013)
## Thinking Path

> - Paperclip is an open-source AI-agent management platform; agents run
tasks inside sandboxed environments (Daytona, Kubernetes, E2B, etc.)
> - The control-plane ↔ sandbox file-transfer path flows through the
`environmentExecute` seam in `protocol.ts` — the only verb available to
plugins — which forces a base64-over-exec chunked loop for every file
move: workspace files, assets, Codex home sync
> - This transport is correct and safe, but it bypasses provider-native
bulk/streaming APIs (Daytona `uploadFiles`, K8s `FastUploadInterceptor`
/ volume mounts), leaving significant throughput on the table for large
workspaces
> - The right fix is an opt-in seam extension: providers with faster
native transfer declare two optional verbs; providers that do not opt in
stay on the existing fallback with zero code or behavior change required
> - This PR adds the first layer of that extension — two optional verbs
(`environmentSyncIn` / `environmentSyncOut`) in the plugin SDK, the
runtime plumbing to prefer the native path for the two clean
destroy-then-replace cases, and a doc for the contract
> - The core correctness invariant is byte-identical fallback: if no
provider opts in, execution is exactly what ships today;
`assertSyncOperationsConfined` enforces host-side path confinement for
providers that do opt in
> - No provider advertises the verbs yet → zero production behavior
change; future PRs wire up Daytona and K8s providers against this
contract

## Linked Issues or Issue Description

No public GitHub issue exists for this feature. Description follows the
`feature_request` issue template:

**Subsystem affected:**
packages/plugins — plugin system; packages/adapter-utils — adapter
runtime; server/ — EnvironmentRuntimeService

**Problem or motivation:**
Sandbox file transfers currently always use a base64-over-exec chunked
loop regardless of what the underlying provider supports. For workspaces
larger than a few MB this becomes the dominant wall-clock cost of every
sandbox run, and it bypasses bulk/stream APIs that providers like
Daytona already expose natively.

**Proposed solution:**
Add two optional, opt-in plugin hooks — `onEnvironmentSyncIn` /
`onEnvironmentSyncOut` — to the plugin SDK. When a provider defines both
hooks and both are advertised via the existing `supportedMethods`
negotiation, the runtime prefers the native path for the two clean
destroy-then-replace transfer cases; all other cases fall back to the
existing byte-identical base64 transport.

**Alternatives considered:**
An unconditional verb would require every provider to implement or stub
the verb. The opt-in / `METHOD_NOT_IMPLEMENTED` pattern (already used by
`environmentExecute`) preserves backward compatibility with zero
provider changes required.

**Roadmap alignment:**
Consistent with the ✅ "Cloud / Sandbox agents" and ✅ "Plugin system"
milestones; extends the plugin seam rather than adding
control-plane-level logic.

**Additional context:**
Searched open pull requests and issues for duplicate sandbox file-sync /
native-transfer work; none found.

## What Changed

- **`packages/plugins/sdk`**
- `protocol.ts`: two new optional `HostToWorkerMethods` —
`environmentSyncIn` / `environmentSyncOut` — plus generic
`SyncOperation`, `SyncFileMapping`, and `SyncOutcome` types
- `define-plugin.ts`: optional `onEnvironmentSyncIn` /
`onEnvironmentSyncOut` fields on `PluginDefinition`; worker advertises
each verb only when its hook is defined (else `METHOD_NOT_IMPLEMENTED`,
mirroring `environmentExecute`)
  - `worker-rpc-host.ts`: route new verbs to plugin hooks
  - `index.ts`: re-export new public types
- **`packages/adapter-utils`**
- `command-managed-runtime.ts`: expose optional `syncIn` / `syncOut` on
`CommandManagedRuntimeRunner` (available only when both verbs are
advertised); add `assertSyncOperationsConfined` host-side
path-confinement guard
- `sandbox-managed-runtime.ts`: `SandboxManagedRuntimeClient` gains
optional `syncIn` / `syncOut`; orchestrator prefers native path for
default-provision asset inbound and workspace-download-into-fresh-dir
outbound; all other paths keep the existing base64 fallback
- `sandbox-file-sync.test.ts` (new): 234-line characterization suite —
native-opt-in branch, fallback branch, `assertSyncOperationsConfined`
escape-path rejection, `followSymlinks` → tar `-h`
- `command-managed-runtime.test.ts`: negotiation + native-sync +
confinement tests
- **`server/src/services/environment-runtime.ts`**:
`EnvironmentRuntimeService` delegates to `syncIn` / `syncOut`, gated on
advertised support
- **`server/src/services/environment-execution-target.ts`**: minor
typing fix alongside the new verbs
- **`doc/plugins/SANDBOX_FILE_SYNC_HOOKS.md`** (new): documents the full
contract — opt-in / no-op guarantee, operation ordering,
provider-may-tar, atomicity, `followSymlinks`, secret modes (0600, no
window), path confinement, `operationId` opacity, resource bounds,
shell-quoting

## Verification

```bash
# SDK suite
pnpm --filter packages/plugins/sdk test

# Adapter-utils suite (includes new sandbox-file-sync characterization tests)
pnpm --filter packages/adapter-utils test
# Expected: 255 pass / 4 skip

# Type-check across affected packages
pnpm --filter packages/plugins/sdk typecheck
pnpm --filter packages/adapter-utils typecheck
# Server changed-file spot check:
cd server && npx tsc --noEmit --skipLibCheck 2>&1 | grep -E "environment-(runtime|execution-target)" | head -20
```

Key behavioral invariant to spot-check: with no provider opting in (the
current state), run any sandbox task and confirm file-transfer behavior
is byte-for-byte identical to what the pre-PR code produces. The
characterization tests assert this at the unit level.

## Risks

- **Zero production risk today**: no provider advertises
`environmentSyncIn` / `environmentSyncOut`, so the new code paths are
unreachable in production; all real traffic stays on the existing base64
fallback
- **Path confinement**: `assertSyncOperationsConfined` rejects any
`targetPath` that escapes the declared root — this is the primary
security boundary for future providers. The test suite covers
escape-path rejection
- **Atomicity**: the contract delegates atomicity to providers; the doc
explicitly calls out that directory-level ops are not guaranteed atomic
- **Secret transport**: credential assets (e.g., Codex `auth.json`,
directory mappings) continue to use the existing tar path — they do not
go through the new verbs in any current provider

> For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and
discuss it in `#dev` before opening the PR. Feature PRs that overlap
with planned core work may need to be redirected — check the roadmap
first. See `CONTRIBUTING.md`.

## Model Used

Provider: Anthropic  
Model: `claude-sonnet-4-6` (Claude Sonnet 4.6)  
Context window: 200 K tokens  
Capabilities: extended tool use, multi-file code generation, agentic
reasoning via the Paperclip agent framework

## Checklist

- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` /
`Closes: #` / `Refs: #` OR (b) described the issue in-PR following the
relevant issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] My branch name describes the change (e.g. `docs/...`, `fix/...`)
and contains no internal Paperclip ticket id or instance-derived details
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] All Paperclip CI gates are green
- [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge

---------

Co-authored-by: Harold Kim <harold@paperclip.ing>
Co-authored-by: Paperclip <noreply@paperclip.ing>
2026-07-22 10:08:03 -07:00

9.3 KiB

Sandbox file-sync lifecycle hooks

A sandbox environment provider moves workspace and asset files between the host and the sandbox around every run. By default the runtime synthesizes that transfer over the single environmentExecute verb: it base64-encodes bytes and pipes them through base64 -d shell commands, one bounded round-trip per chunk. That works everywhere but is slow for large workspaces because it cannot use a provider's native bulk file transport.

The two optional, opt-in hooks documented here let a provider replace that base64-over-exec transfer with its own native mechanism:

  • onEnvironmentSyncIn — before execution: place a set of host files/directories at target sandbox paths.
  • onEnvironmentSyncOut — after execution: copy a set of sandbox files/directories back to target host paths.

They are entirely opt-in. A provider that does not define them keeps the exact base64 fallback, byte-for-byte — there is zero behavior change for existing providers.

Opt-in / no-op semantics

A hook is opted into exactly like onEnvironmentExecute: defining it on your PluginDefinition makes the worker advertise the matching verb in InitializeResult.supportedMethods; leaving it undefined omits the verb and the guarded handler throws METHOD_NOT_IMPLEMENTED if it is ever called.

Both hooks are advertised and consumed as a pair. The host runtime uses the native path only when the worker advertises both environmentSyncIn and environmentSyncOut; if a provider advertises only one, the orchestrator keeps the base64 fallback for both directions. Define both or neither.

export default definePlugin({
  async setup() { /* ... */ },
  async onEnvironmentSyncIn(params) {
    return { operations: await transferInbound(params) };
  },
  async onEnvironmentSyncOut(params) {
    return { operations: await transferOutbound(params) };
  },
});

The operation / file-mapping contract

Each hook receives an ordered list of operations. Each operation carries an opaque id and a list of source→target file mappings:

interface PluginSyncOperation {
  operationId: string;               // opaque, non-sensitive; do NOT interpret it
  files: PluginSyncFileMapping[];
}

interface PluginSyncFileMapping {
  sourcePath: string;                // absolute
  targetPath: string;                // absolute
  kind: "file" | "directory";
  mode?: number;                     // POSIX mode to apply at the target
  exclude?: string[];                // glob excludes for a directory mapping
  followSymlinks?: boolean;          // directory symlink handling; see below
}

interface PluginEnvironmentSyncResult {
  operations: { operationId: string; filesTransferred: number; bytesTransferred: number }[];
}

For onEnvironmentSyncIn, sourcePath is a host path and targetPath a sandbox path. For onEnvironmentSyncOut the direction is reversed. All sandbox paths are POSIX. Return per-operation filesTransferred / bytesTransferred for observability.

Ordering

Operations are applied strictly in array order, and the orchestrator invokes the hooks in a fixed lifecycle order (inbound before execution, outbound after). The orchestrator owns what and when; a provider only executes the opaque transfers it is handed and must not reorder them.

A provider may tar internally

The contract only describes the observable source→target result. How you move the bytes is yours: bulk upload API, an internal tar stream, per-file enumeration — all are fine. Whatever you do, the materialized target must be observationally identical to the mapping (same files, same contents, same modes, same symlink treatment) so the native and fallback paths are interchangeable.

operationId is opaque

operationId is an opaque, non-sensitive token authored by the orchestrator. It is not derived from any secret or user data, it is safe to log and safe to expose to the sandbox, and a provider must not parse or depend on its value. Do not echo it into a path or a place where it could collide with real data.

followSymlinks applies to kind: "directory" mappings and has exactly the meaning of tar's -h flag:

  • falsy (default) — archive and recreate symlinks as links (preserve).
  • true — dereference each symlink to its target bytes.

A provider honoring a directory mapping MUST reproduce this: preserve links when falsy, dereference to bytes when true. The orchestrator passes the same value it passes to its own tar create step, so native and fallback are observationally identical. There is no separate extract-side symlink flag and no execution-time special case — symlink handling lives entirely in this one flag.

Atomicity contract

The required guarantee level is deliberately equal to the base64 fallback's floor, so opting in never weakens integrity and never over-promises.

  • Single-file mappings (kind: "file") MUST be atomic-replace (REQUIRED). Stage the bytes to a provider-chosen temporary path, then atomically rename onto targetPath, so an interrupted transfer never leaves a truncated file at targetPath. This mirrors the fallback, which stages to <path>.paperclip-upload and then mv -f.

    • The temp file MUST live in the same directory (same filesystem) as targetPath. A cross-device rename degrades to copy-then-unlink and reintroduces the truncation window it is meant to close.
    • Reserve the .paperclip-upload* scratch names: a provider-chosen temp must not collide with the fallback scratch name or with a real target.
  • Directory mappings and the sync as a whole are NOT atomic / NOT transactional. A directory transfer is destroy-then-replace: a crash mid-transfer can leave a partial tree, and the runtime does not roll back already-moved bytes across operations. This matches today's behavior; do not assume a directory operation is atomic. Where an individual file must be integrity-protected, deliver it as its own kind: "file" mapping so it inherits the single-file atomic-replace guarantee.

  • Every operation is fail-loud. An operation either completes or raises to the orchestrator; never report partial success silently. The orchestrator may then retry or fall back.

Secret material and file modes

This seam can carry credential-bearing files (for example an auth directory). Treat mode as mandatory for such mappings:

  • Apply the requested mode (e.g. 0o600) with no world-readable window — create the target with the mode, or chmod before writing any bytes, never after.
  • mode MUST be honored for files inside a directory mapping too, not only for kind: "file" mappings. If you tar internally, preserve permissions; if you enumerate, set the mode as each file lands. A credential that rides a directory mapping otherwise silently loses its 0o600 guarantee.
  • A directory mapping is not atomic (see above). If a directory carries an individually-sensitive secret whose integrity matters, prefer delivering that secret as a kind: "file" mapping so it gets atomic-replace, or protect its integrity out of band.

Host-side path confinement (required of the orchestrator)

The sandbox is untrusted relative to the host, so the orchestrator — not the provider — owns and confines every path. Before an operation is handed to a provider, the runtime canonicalizes each mapping's sourcePath/targetPath and confines it to an orchestrator-owned root (the workspace directory or a specific asset directory), rejecting absolute escapes and .. traversal fail-closed. A provider receives only already-confined, orchestrator-authored paths and MUST NOT widen them (for example by following a sandbox-planted symlink out of the intended root on an outbound write). Confinement is a host-side complete-mediation guard and is never delegated below the trust boundary.

Resource bounds

The base64 fallback enforces transfer caps so a runaway payload cannot exhaust memory. A native provider MUST keep an equivalent bound — stream or chunk large transfers rather than buffering unboundedly, and fail closed on an oversized inline payload rather than silently removing the cap.

Shell safety (native providers that shell out)

If your native transfer builds shell command strings (for example a pod-exec tar/base64/mv pipeline), single-quote every interpolated path so a path containing shell metacharacters is transferred literally, never interpreted. Providers whose transport is a non-shell API (a typed bulk-upload call) do not need this, but any shell interpolation must quote.

Reference: minimal shape

async onEnvironmentSyncIn({ operations }) {
  const results = [];
  for (const op of operations) {          // apply in order
    let filesTransferred = 0;
    let bytesTransferred = 0;
    for (const f of op.files) {
      if (f.kind === "file") {
        // stage to a same-dir temp, then atomic rename onto f.targetPath,
        // applying f.mode with no world-readable window
      } else {
        // materialize f.sourcePath at f.targetPath (destroy-then-replace),
        // honoring f.exclude and f.followSymlinks, applying per-file modes
      }
    }
    results.push({ operationId: op.operationId, filesTransferred, bytesTransferred });
  }
  return { operations: results };
}