SIP encryption library as a dsh plugin — XChaCha20-Poly1305 AEAD encrypt/decrypt agent tools plus streaming encrypted file pack/unpack (.sipft artifacts), backed by sip_protocol (Python). Zero-deploy: no service process, tools spawn one-shot Python per call.
Server has 8 tools with significant definition gaps. Tools 1-2 and 3-4 are duplicates (sip_encrypt/sip_decrypt appear twice with different schemas). Descriptions exist but are inconsistent in quality and language (mix of English and Chinese). Input schemas are present but lack proper type declarations for several parameters. No output schemas documented. Parameters lack descriptions in some cases (e.g., tool 5's 'recipient_id' has minimal guidance). Error handling is absent from tool definitions. Naming is reasonable (verb_noun pattern) but composition is poor due to duplication. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels in metadata.
Pack (encrypt) a local file into a self-contained .sipft artifact (SIPFT1.0: streaming XChaCha20-Poly1305 AEAD, per-chunk HKDF keys, authenticated tag chain — constant memory for any file size). Pass a 32-byte hex key and the input file path; get the artifact path plus chunk metadata. The artifact travels over ANY channel (dsh, git, cloud drive) and unpacks only with the same key.
Unpack (decrypt + verify) a .sipft artifact produced by encrypted_file_pack. Streaming per-chunk verification: any tampering, reordering, truncation or cross-artifact splicing is rejected at the offending chunk; a wrong key is rejected at the authenticated header. Output file name comes from the authenticated header (path traversal sanitized; name conflicts auto-renamed).
Decrypt SIP XChaCha20-Poly1305 AEAD ciphertext (local, one-shot). Pass key/ciphertext/nonce/auth_tag hex fields from sip_encrypt; returns plaintext, or an auth error if the tag does not match (tampered or wrong key).
使用SIP协议解密一条消息。要求通道已通过sip_handshake建立。输入加密消息的JSON字符串,返回解密后的明文。
Encrypt UTF-8 text with SIP XChaCha20-Poly1305 AEAD (local, one-shot). Pass a 32-byte hex key and plaintext; get nonce/ciphertext/auth_tag hex fields. Ship all three fields together with the key designation to the decrypt side. For full handshake/rekey channels use the SIP MCP server instead.
Duplicate tool definitions: sip_encrypt and sip_decrypt appear twice (tools 1-2 vs 5-6) with different input schemas and descriptions. Tool 1-2 expect 32-byte hex keys; tools 5-6 expect plaintext/encrypted_message. This creates ambiguity for LLMs and violates single-responsibility principle.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., what does sip_handshake return? Is it a JSON object with 'message' and 'session_id' fields?). This violates pattern:tool and forces LLMs to guess.
Mixed language descriptions (English and Chinese) without consistent structure. Tool 1 description is 200+ chars; tool 5 is 80 chars. Inconsistency makes it harder for LLMs to parse intent. Descriptions should be 50-200 chars, consistent in language, and follow WHAT/WHEN/RETURN structure.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
使用SIP协议加密一条消息。要求通道已通过sip_handshake建立。返回加密后的payload(base64编码的JSON)。
执行SIP握手协议的三重DH密钥交换。支持三种角色:initiator(发起方)、responder(响应方)、complete(发起方完成)。initiator: 返回握手Hello消息,等待对方的Auth响应。responder: 输入收到的Hello消息,返回Auth响应并完成握手。complete: 输入收到的Auth消息,完成发起方握手。
执行SIP密钥轮换(Rekey),实现前向保密。要求通道已通过sip_handshake建立。支持initiator(发起轮换)和responder(响应轮换)两种角色。
Parameter descriptions are minimal or missing context. Tool 5's 'recipient_id' says 'optional, defaults to handshake peer' but does not explain what happens if the peer is unknown or if the parameter is required for multi-peer scenarios. Tool 7's 'message' parameter lacks format guidance (is it JSON string? Base64? Raw bytes?).
No error handling guidance in tool definitions. Tools declare risk levels (READ_ONLY, WRITE) but do not document what errors can occur or how to recover. E.g., sip_decrypt can fail with 'auth error if tag does not match' but no guidance on retry strategy or user-fixable vs fatal classification.