Loopback-only MCP control endpoint for DeepSeek Harness sessions. Exposes seven tools that start, steer, cancel, list, and read already-known DSH Sessions through the native Session Controller and subagent services.
Strong schema definitions with comprehensive parameter descriptions and clear naming conventions. All 5 tools have explicit input schemas with type constraints and detailed descriptions (100-300+ chars each). Tool names follow verb_noun pattern (session_start, session_send, session_cancel, sessions_list, events_read). However, output schemas are not documented in the visible code, only input schemas are defined. Error handling descriptions exist but lack actionable recovery guidance. No tool annotations (readOnlyHint/destructiveHint) despite clear WRITE vs READ_ONLY risk classifications.
Forward-only, lossless paging over the durable raw Session event log, plus a chunk mode that retrieves one oversized event in pieces through the same tool. The page never advances past an event it did not deliver whole, never activates a cold root, and keeps no listener, cursor, snapshot, or cache between requests — the caller's after_seq is the only cursor that exists.
Cancel one queued prompt or the nearest step of an existing root session. Returns the native cancellation receipt. Cancelling a step that has already started running does not interrupt it; the cancellation takes effect when the step finishes.
Submit one text prompt to an existing root DSH session. delivery "queue" (the default) adds a later turn; "steer" targets the nearest step. Returns the native acceptance receipt only, never a completion or a scheduling order.
Create a root DSH session in the MCP client's actual project directory, or adopt the existing session with the supplied id, then submit one text prompt. cwd selects project context and Workspace grouping; using a temporary directory creates an ungrouped temporary context and does not enforce read-only access. If the directory belongs to a registered DSH workspace, attach the session there; otherwise leave it ungrouped. Omit agent_preset to use the deployment's configured default; set it only as an intentional, known override. Set permission_preset to a native DSH preset such as "read-only"; it is applied before the first prompt, while cwd still names the real project. Returns once DSH accepts the prompt and never waits for the turn to finish. Reusing the same request_id links a retry to the message the first attempt persisted. A failure that reports stage "create" still carries the session_id it tried to create, so a retry can adopt that id instead of creating a second session. A stage "permission" failure identifies a created Session whose prompt was not submitted.
Output schemas not documented. Tool descriptions explain what is returned (e.g., 'Returns the native acceptance receipt only'), but the actual response structure is not formally defined in visible code. LLMs cannot plan downstream tool calls without knowing response fields.
Tool annotations missing. Despite clear risk classifications (WRITE vs READ_ONLY), tools lack MCP destructiveHint and readOnlyHint annotations. This prevents clients from applying appropriate safety guards or UI warnings.
Error recovery guidance incomplete. Descriptions mention failure stages (e.g., 'stage "create"') but do not explain what the LLM should do next. E.g., 'A failure that reports stage "create" still carries the session_id it tried to create, so a retry can adopt that id instead', this is in the description but not structured as actionable error guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 69 | 2026-07-28+ | v2 |
List all root sessions known to this endpoint, in creation order. Returns an empty list when no sessions exist. The list never includes subagent children, only root sessions.
events_read parameter complexity. The 'address' parameter is an object with undocumented structure ('Durable address of one root Session or one direct subagent child'). The schema does not show what fields 'address' contains, forcing LLMs to guess or fail.
sessions_list description is minimal (27 chars). 'List all root sessions known to this endpoint, in creation order.' lacks context on when to call it, what the response structure is, or how to use results with other tools.