9 results
DSH tool-call compatibility: strips sandbox_permissions/justification from tool-call arguments for third-party models (GPT etc.), and lets the user skip possibly-stuck tool calls with an LLM-facing notice.
Fail-closed DSH compatibility guard for redundant GPT/Codex sandbox escalation arguments
Strip sandbox_permissions / justification from model tool-call arguments so sandbox escalation is never triggered when the session already has sufficient permission.
Normalize invalid non-widening sandbox escalation arguments before DSH tool dispatch.
DeepSeek Harness plugin: redacts known secret shapes in tool arguments and results behind stable placeholders, and appends preimage-free JSONL incidents
Neutralize tool calls with invalid JSON arguments on the wire, so one malformed model generation cannot brick a session against strict OpenAI-compatible servers (vLLM et al).
DeepSeek Harness plugin: a circuit breaker that denies the Nth identical tool call in a row, by exact hash of tool name and canonical arguments
Fixes malformed tool-call arguments (nested {"arguments":{…}} wrappers, wrapper-key aliases, param-key aliases) at the llm/stream seam — repairs execution, persistence and history replay in one place. 修复模型产出的畸形工具参数。
Hard break on repeated identical tool calls: a synchronous ctx.tools.guard monotonic gate that denies the 2nd identical (tool name + canonical arguments) call per agent, with an optional same-path read guard and advisory first-repeat warning.