23
Published Tools
5
Total Stars
0
Weekly Downloads
39
GitHub Followers
215
Public Repos
100/100
Avg Security
Published Tools
4 MCP Servers19 Skillsacross 5 categoriesTurns three sandbox environment failures that name neither their cause nor a remedy into a diagnosis with a path forward. Family 1, the Windows workspace ACL: nine reports (#7538, #7622, #7646, #7720, #7750, #7735, #7771, #7804, #7816) of one signature —
...moreTwo public tools.restrict() use cases for the dsh harness, applied per agent scope: a per-preset tool ALLOWLIST (#5786: a preset agent inherits the global layer and every ancestor scope layer, so plugin-registered commands leak into a minimal preset) and
...moreProactive session-log completeness audit for the dsh harness: detects seq gaps, index-reuse, format-version drift, unknown-required types (hard-fail on reload), empty text blocks (#5773), orphan tool calls, and open-turn tails — before replay crashes.
...moreDeployment-level tool-call deadlines for dsh: a tool that never declares ToolDefinition.timeoutMs is dispatched with no deadline at all, so one wedged call (a delegation whose child never settles, a third-party MCP tool that never returns) leaves the turn
...moreMakes an unattributed model attempt retryable, for the one seam that can tell. Three reports, one missing evidence. Discussion #8321: a `dsh-llm-pi-ai` route that relays an upstream empty response ends the turn, because `classifyPiAiError` has no rule mat
...moreCron-driven autonomous task runs for the dsh harness: fire a calendar cron rule and spawn a fresh agent run via ctx.agents.create(), unlike the official @deepseek-ai/dsh-schedule which only messages the live session.
...moreSession-scoped model and effort selection for dsh. Every entry point that names a model for one session also rewrites the deployment default, so a lane cannot launch on a model without re-pointing every later session in the same DSH home: the session mode
...moreargszero
Writes the activation report the harness already composes but never files. When a third-party plugin fails to activate — most often because a tool or prompt-section name it registers is already taken — the fiber that threw is unloaded whole, so every tool
...moreKeeps a compaction call from dying on the very route it was meant to shrink away from. When the harness summarizes an oversized history and the summary request is refused by the session's current model - because that model's context window is smaller than
...moreA tool parameter declared with a type list (["boolean","string"]), an `anyOf` or a `oneOf` is legal JSON Schema, and third-party MCP servers publish them — a code-search server declaring `anyOf [array|string, null]` on its main parameters is the reported
...moreargszero
`dsh-mcp-client` registers every tool a connected MCP server lists, unconditionally, so a server advertising 161 tools puts 161 tools and their full JSON schemas into every request that assembles the tool surface (reported as #9385). `ctx.tools.restrict()
...moreargszero
MCP tool calls rendered as the calls they are: the dsh Web Tool row is a keyed seat dispatched by wire Tool name, so a family of names that only exists after an MCP server answers tools/list falls back to the generic card — which flattens `mcp__<server>__
...moreargszero
A session whose request exceeds the model's context window and whose summarizer cannot fit that same window has no way out: every later message replays the same oversized history into the same 400, so the session is unusable until someone edits the persis
...moreAn empty-identity tool call can be advertised but never answered: tool/call is admitted with callId "", the harness's own assembler only synthesises a fallback id for a missing one (nullish, so an empty string defeats it), and the paired tool/result is th
...morePins a tool call to the identity it was approved with. The registry hands tools/pre-execute listeners the live execution object and documents name/agent/parent/callId/signal/token as readonly, while the object is a plain writable literal with only the arg
...moreRunaway tool-call argument guard for dsh: wrap the llm/stream waterfall and cut a tool call that streams more than a configurable number of arguments bytes, terminating with a routed TOOL_CALL_ARGUMENTS_TOO_LARGE error finish so the oversized call is neve
...moreThinking-loop guard for dsh: observes the llm/stream waterfall and breaks a model that degrades into a loop. Detects three per-call shapes (reasoning-only calls, restated-material calls, self-repeating low-entropy reasoning) and re-fires instead of latchi
...moreActionable read_image refusal for dsh: when the harness refuses to read an image because the resolved model route does not declare image input, turn that opaque 'switch to an image-capable model' message into concrete config guidance (declare "input: [tex
...moreContext-overflow classifier guard for dsh: observes the llm/stream waterfall and reclassifies a provider request-size rejection that names an *input* budget (max_prompt_tokens / max_input_tokens) rather than the model's context window. Unrecognized, such
...moreSteer and new-output preemption for dsh: a new user input (next-step steer) or new output from the awaited job interrupts a blocking job_output(wait: true) read immediately instead of letting it block up to the job-wait cap, so the agent loop reaches its
...moreThinking-level fallback for dsh: wraps the llm/stream waterfall and re-issues an auxiliary model call the provider rejected for naming an unsupported thinking level (e.g. Google: "Thinking level MINIMAL is not supported for this model"). On that provider
...moreCost budget for dsh web_search: bounds the number of billable provider search requests a `web_search` call may initiate (per call / per turn / per session / whole subagent tree) and denies the call with an actionable error once the budget is exhausted. Ex
...morePer-delegation reasoning effort for dsh subagents: makes AgentOptions.reasoningEffort reach child requests (the agent loop silently ignored it) and adds a subagent_effort tool that accepts reasoning_effort and model per call
...more