Files
paperclip/packages/plugins/sandbox-providers/kubernetes/src/upload-interceptor.ts
T
Jannes Stubbemann 05ab45225a feat(plugin-kubernetes): self-hostable Kubernetes sandbox provider (stage 1/3: plugin package) (#5790)
## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work
> - Sandbox providers are the seam that lets agent runs execute in
isolated environments; today the only first-party remote provider is
Daytona, a hosted third-party service
> - Self-hosters running Paperclip on their own infrastructure (often
Kubernetes already) have no first-party way to run agent sandboxes on a
cluster they control
> - That gap matters for teams with data-residency, sovereignty, or cost
constraints who cannot or will not send workloads to a hosted sandbox
service
> - This pull request adds a Kubernetes sandbox-provider plugin as a
standalone, workspace-excluded package: it implements every
SandboxProvider hook the Daytona provider does, on infrastructure the
operator owns
> - The benefit is that any Paperclip deployment with a Kubernetes
cluster gets multi-tenant, network-isolated, quota-bounded agent
sandboxes with zero new external dependencies

## Linked Issues or Issue Description

No existing issue. Following the feature template:

- **Problem:** Paperclip's remote sandbox execution requires a hosted
third-party provider. Self-hosters cannot run agent sandboxes on their
own Kubernetes clusters with a first-party provider.
- **Proposed solution:** A `@paperclipai/plugin-kubernetes`
sandbox-provider plugin with two backends: long-lived sandboxes via the
[kubernetes-sigs/agent-sandbox](https://github.com/kubernetes-sigs/agent-sandbox)
CRD (multi-command exec, adapter-install pattern) and one-shot
`batch/v1` Jobs (stable APIs only, no extra controllers).
- **Alternatives considered:** Driving kubectl from a generic shell
provider (no lifecycle/lease semantics), or requiring a hosted provider
(exactly the constraint this removes).

## What Changed

This is **stage 1 of 3** of a staged contribution (direction agreed with
maintainers): the plugin package alone. Stage 2 (server integration:
lease params, provider registration) and stage 3 (agent runtime images +
CI) are companion PRs that will be cross-linked from a comment here.

- New package `packages/plugins/sandbox-providers/kubernetes`
(workspace-excluded, like the path already carved out in
`pnpm-workspace.yaml`): src, unit + kind integration tests, operator
prerequisite manifests, README, smoke-test guide
- Implements the full SandboxProvider hook surface the Daytona provider
implements: `validateConfig`, `probe`, `acquireLease`, `resumeLease`,
`releaseLease`, `destroyLease`, `realizeWorkspace`, `execute`
- Two backends: `sandbox-cr` (default; long-lived pod via the
agent-sandbox `Sandbox` CR, supports multi-command exec) and `job`
(one-shot `batch/v1` Job; nothing beyond k8s 1.27+ required)
- Per-run adapter resolution: one environment serves mixed harnesses;
the per-run `adapterType` hint is read through a local optional type
extension, so the plugin typechecks and builds against the current
plugin SDK and simply falls back to the environment's configured default
adapter until stage 2 lands
- Exec-env wrapping: the Kubernetes exec API carries no environment, so
commands are wrapped to receive the run's env
- Fast-upload interception for workspace realization, scoped per lease
- Per-tenant isolation: derived namespace per company, RBAC,
ResourceQuota, restricted-PSS pod security (runAsNonRoot, drop ALL,
seccomp RuntimeDefault, no SA token automount)
- Network egress policy in two flavors: native `NetworkPolicy` and
`CiliumNetworkPolicy` (FQDN allowlists)
- Image allowlist with glob matching, registry override, and per-run
image override validation
- Per-run Kubernetes Secrets carrying agent credentials, ownerRef'd to
the Job or Sandbox CR for cascade GC

## Verification

- Standalone build, exactly as the README documents:
  ```bash
  cd packages/plugins/sandbox-providers/kubernetes
  pnpm install --ignore-workspace
  pnpm test        # 147 unit tests, 17 files, all green
  pnpm typecheck   # clean against the in-repo plugin SDK on master
pnpm build # dist/ emitted, manifest + worker entrypoints present
  ```
- A kind-cluster end-to-end integration test is included
(`RUN_K8S_INTEGRATION_TESTS=1 pnpm test
test/integration/end-to-end-run.test.ts`)
- Beyond CI: this provider has been verified in a production
multi-tenant deployment against five harnesses (opencode, pi, codex,
gemini, claude code) with real billed runs

## Risks

- **Zero behavior change for any existing deployment.** The package is
workspace-excluded; nothing in the server imports or loads it until
stage 2's integration lands. No existing code paths are touched.
- The default `sandbox-cr` backend depends on an alpha CRD
(`agents.x-k8s.io/v1alpha1`); the README flags this and the `job`
backend uses only stable APIs as a fallback.
- Risk surface is confined to deployments that explicitly install and
configure the plugin.
- The default runtime images (`ghcr.io/paperclipai/agent-runtime-*`) are
published by the stage 3 companion PR (#7934); until that lands,
deployments must point `runtimeImage` at their own images.

## Model Used

Claude Opus 4.8 (1M context), extended thinking, with tool use (Claude
Code).

## 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 run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots (no UI changes)
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green (pending this push)
- [ ] 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: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-10 21:07:00 -07:00

147 lines
5.6 KiB
TypeScript

/**
* Fast-upload interceptor for the chunked-shell file transfer protocol used by
* `@paperclipai/adapter-utils`'s `command-managed-runtime.writeFile()`.
*
* The default protocol uploads a binary file by:
* 1. INIT: mkdir -p '<DIR>' && rm -f '<B64>' && : > '<B64>'
* 2. CHUNKS: printf '%s' '<base64-chunk>' >> '<B64>' (repeated N times)
* 3. FINALIZE: base64 -d < '<B64>' > '<TARGET>' && rm -f '<B64>'
*
* Where `<B64>` is `<TARGET>.paperclip-upload.b64`. Over k8s exec each call costs
* ~50-100ms (new WebSocket + container exec), so N chunks = N round trips. For
* a 1MB payload split into 64KB base64 chunks that's ~250 round trips.
*
* This interceptor short-circuits the protocol entirely:
* - INIT: capture `<B64>`, start a buffer, RESPOND success without exec.
* - CHUNK: append the literal base64 chunk to the buffer, RESPOND success.
* - FINALIZE: decode the buffered base64 → bytes → one exec `cat > '<TARGET>'`
* with stdin = bytes. Drop intermediate `.b64` entirely.
*
* If any step doesn't match the protocol exactly, we abandon the buffer and let
* the caller go through the slow path (passthrough). This keeps the optimization
* **transparent and safe**: pattern drift → graceful fallback.
*
* Reference: see packages/adapter-utils/src/command-managed-runtime.ts.
*/
import { posix as pathPosix } from "node:path";
const INIT_RE =
/^mkdir -p '([^']+)' && rm -f '([^']+)\.paperclip-upload\.b64' && : > '\2\.paperclip-upload\.b64'$/;
const CHUNK_RE =
/^printf '%s' '([A-Za-z0-9+/=]+)' >> '([^']+)\.paperclip-upload\.b64'$/;
const FINALIZE_RE =
/^base64 -d < '([^']+)\.paperclip-upload\.b64' > '\1' && rm -f '\1\.paperclip-upload\.b64'$/;
const MAX_BUFFER_BYTES = 100 * 1024 * 1024; // 100MB safety cap
export interface InterceptStateFlush {
/** Final target path on the remote (decoded payload goes here). */
targetPath: string;
/** Decoded payload bytes. */
payload: Buffer;
}
export type InterceptDecision =
/** Pattern recognized; respond with success without running the command. */
| { action: "ack"; reason: string }
/** Pattern recognized; flush via single exec. Caller does the exec. */
| { action: "flush"; flush: InterceptStateFlush }
/** Pattern not recognized; pass through to the normal exec path. */
| { action: "passthrough"; reason: string };
interface BufferState {
targetPath: string;
chunks: string[];
totalChars: number;
}
/**
* Stateful interceptor. Keyed by the base64 temp path (`<TARGET>.paperclip-upload.b64`).
* One instance per plugin worker is fine — concurrent uploads to different paths
* don't interfere.
*/
export class FastUploadInterceptor {
private buffers = new Map<string, BufferState>();
/**
* Inspect a single shell command (the literal argument to `sh -c <cmd>`).
* Returns the action the plugin should take.
*/
decide(command: string): InterceptDecision {
// 1. INIT
const initMatch = INIT_RE.exec(command);
if (initMatch) {
const dir = initMatch[1];
const targetPath = initMatch[2];
const b64Path = `${targetPath}.paperclip-upload.b64`;
// Sanity: dir should be the parent of target. If not, fall through.
if (pathPosix.dirname(targetPath) !== dir) {
return { action: "passthrough", reason: "init dir/target mismatch" };
}
this.buffers.set(b64Path, { targetPath, chunks: [], totalChars: 0 });
return { action: "ack", reason: `init upload to ${targetPath}` };
}
// 2. CHUNK
const chunkMatch = CHUNK_RE.exec(command);
if (chunkMatch) {
const base64 = chunkMatch[1];
const targetPath = chunkMatch[2];
const b64Path = `${targetPath}.paperclip-upload.b64`;
const state = this.buffers.get(b64Path);
if (!state) {
// Chunk arrived without init — must passthrough (the upload was started
// some other way, perhaps before we started tracking).
return { action: "passthrough", reason: "chunk without prior init" };
}
if (state.totalChars + base64.length > MAX_BUFFER_BYTES * 4 / 3) {
// base64 is ~4/3 of binary size; cap memory.
this.buffers.delete(b64Path);
return { action: "passthrough", reason: "buffer cap exceeded" };
}
state.chunks.push(base64);
state.totalChars += base64.length;
return { action: "ack", reason: `buffered ${base64.length} b64 chars` };
}
// 3. FINALIZE
const finalizeMatch = FINALIZE_RE.exec(command);
if (finalizeMatch) {
const targetPath = finalizeMatch[1];
const b64Path = `${targetPath}.paperclip-upload.b64`;
const state = this.buffers.get(b64Path);
if (!state) {
return { action: "passthrough", reason: "finalize without buffered state" };
}
this.buffers.delete(b64Path);
// Decode in one go. base64.decode handles concatenated base64 strings
// because each chunk was emitted from the encoder in fixed-size pieces.
const joined = state.chunks.join("");
const payload = Buffer.from(joined, "base64");
return {
action: "flush",
flush: {
targetPath: state.targetPath,
payload,
},
};
}
return { action: "passthrough", reason: "no upload pattern" };
}
/**
* Drop all buffered state. Called on releaseLease so a new lease doesn't
* inherit stale buffers (lease IDs are unique so target paths shouldn't
* collide, but cleanup is hygiene).
*/
reset(): void {
this.buffers.clear();
}
/** Number of in-flight uploads being tracked. Mostly for tests / diagnostics. */
get pendingCount(): number {
return this.buffers.size;
}
}