dsh-Go review Encrypted sharing and link lifecycle
Source-level tests of dsh-Go conversation sharing cover decryption, deletion, expiry and quota failures, with clear limits on Android and remote-agent claims.
AgentIndex · Published 2026-11-05 · Updated 2026-11-05
dsh-Go is a third-party Android client for DeepSeek Harness. It can prepare a Linux environment inside the phone or connect to DSH running on a computer. That makes it relevant to people who want to inspect and continue agent work away from a desktop. Its conversation-sharing system addresses a separate question: how can someone read a replay without receiving access to the working agent?
We examined that sharing boundary. The actual viewer decrypted three upstream vectors, rejected a wrong key and rejected modified ciphertext. The actual Worker API stored, retrieved and deleted ciphertext against local storage stand-ins. These results support a concrete sharing workflow; they provide no measurement of coding ability, phone performance or remote-session reliability.
Version and reproducible scope
The review pins commit 16cb44b7fd652d3545cfe6596a127a7968c18eb3 on November 5, 2026. The Flutter manifest reports 1.0.11+1; the build suffix is a source placeholder, not an installed APK identity. The project documents Android 8 or later and an AGPL-3.0 license. Project and installation requirements
Thirteen named checks ran on macOS arm64 with Node.js 26.6.0. They import the unmodified viewer and selected TypeScript API modules. R2, KV and rate-limit bindings are in-memory stand-ins; requests never reach Cloudflare. The fixture, source fingerprints and result make those substitutions explicit. We did not build Flutter, install an APK or run the Dart encryption tests.
| Input or operation | Observed result |
|---|---|
| Empty, short Unicode and conversation vectors | Viewer recovered exact text |
| Wrong 32-byte key or modified ciphertext | Viewer rejected decryption |
| Valid upload and read | 201, then byte-identical 200 response |
| Missing or wrong deletion credential | 401 or 403; object retained |
| Correct deletion credential | 204; following read returned 404 |
| Seven-day object dated eight days ago | Read returned 404 while fake storage still held it |
| Structurally framed random ciphertext | API accepted it; viewer could not decrypt it |
| Failed KV reads and writes | Both attempted uploads were accepted |
A replay link carries its decryption key
The documented design compresses conversation text, encrypts it with AES-256-GCM and uploads a framed ciphertext blob. A browser viewer obtains the blob and decrypts it using a key in the link fragment. An ordinary HTTP request does not include that fragment, which separates the stored object from the decryption material in the intended flow. Sharing design
Our test executed the viewer's decryption function, rather than recreating the algorithm in the test. It recovered the empty, Unicode and conversation fixtures exactly. Flipping a ciphertext byte or supplying another valid-length key caused rejection. This is useful evidence for the selected format and client implementation. It is not an audit of the delivery website, browser extensions, clipboard handling or every file embedded in a conversation.
For a recipient, the complete URL is the access material. Anyone receiving it may obtain readable content, and a later copy can outlive the original link. Review a conversation before sharing it: removing a live agent connection does not remove secrets already recorded in messages or attachments. This follows from the sharing model, rather than from a demonstrated leak.
Accepted storage is different from readable content
The Worker checks the content type, size, minimum length and DSG1 marker. We supplied a 32-byte object with the correct marker but no valid encrypted payload. Upload returned 201, while the actual viewer rejected it. That is consistent with a service that does not possess the content key. API implementation
Consequently, a successful upload is not a full replay acceptance test. Open a newly shared link in the intended recipient browser and confirm that the relevant conversation appears. Our fixture did not mount the browser UI or render attachments, and the supplied Node-generated vectors do not independently establish Android-to-browser interoperability.
Oversized bodies were refused both through a declared length and through actual bytes without that header. Invalid TTL, content type and framing were also rejected. These checks are useful input boundaries, but do not measure behavior under concurrent uploads or large decompression workloads.
Deletion and expiration have different responsibilities
The upload response supplies a deletion token. The stored metadata contained its SHA-256 digest rather than the token itself. Missing and incorrect credentials failed; the correct bearer token deleted the fake object and subsequent reads returned 404. Losing that token is therefore operationally different from losing the replay key: one manages the stored object, while the other opens it.
For expiry, we changed only the fake object's upload time. A seven-day object became unavailable after eight days even though it remained in the bucket stand-in. This validates the selected read-time check. It does not validate real R2 lifecycle cleanup, which is a separate storage operation. The forever tier returned no expiry. None of these actions retracts a recipient's downloaded copy.
Quotas favor availability during storage failure
With an hourly allowance of one, the second ordinary upload returned 429 and a positive Retry-After. An explicit zero allowance also rejected uploads. A denying burst limiter returned 429. When both KV reads and writes threw errors, uploads continued, matching the documented fail-open policy. Quota implementation
This tradeoff matters if you operate the sharing service: the hourly counter is not a guaranteed spending ceiling, and our in-memory sequential test says nothing about distributed consistency. Treat provider-level resource controls and monitoring as a separate operational decision.
Who should try it
dsh-Go is a plausible choice for existing DSH users who want an Android interface and replay links. The README distinguishes local engine, SSH and DSH Remote capabilities; they should not be assumed interchangeable. Phone control requires additional permissions, and configured model providers still receive model requests. Local execution does not establish offline operation.
Start by testing your intended connection mode and a nonsensitive conversation on a spare project. This review supports selected sharing mechanics; Android lifecycle, battery use, SSH reconnection, screen control and model output remain adoption checks for the actual device. For a desktop-centered alternative, see the PI-Desktop storage review. For browser-controlled agent work, see the Browser Harness review.