dsh-plugin-sandbox-escalation-fix
Normalize redundant sandbox requests and malformed justifications in DeepSeek Harness tools
21 results
Normalize redundant sandbox requests and malformed justifications in DeepSeek Harness tools
Rule-gated automatic approval for DeepSeek Harness sandbox escalations with an optional LLM reviewer and native human fallback.
为 DeepSeek Harness 增加介于 Workspace Write 与 Full access 之间的自动批准权限档:例行沙箱升级由分类模型一次性放行,危险或不确定的操作仍转人工审批。An auto-approval permission preset for DeepSeek Harness between workspace-write and full access: routine sandbox escalations are granted once by a classifier, while dangerous or uncertain requests still go to a human.
RigorQuant for DeepSeek Harness: session-scoped unattended, context-isolated multi-agent research for empirical/computational mathematics with a four-part pre-implementation check battery and a jacobian/Lean escalation lane.
LLM auto-review approval answerer for DeepSeek Harness — decides sandbox escalations without a human prompt via a deterministic filter and a clean-context LLM safety review. REQUIRES a patched harness core (see core-patches/)
Fail-closed DSH compatibility guard for redundant GPT/Codex sandbox escalation arguments
Agent-decided approvals for DeepSeek Harness: a workspace-write base permission mode where an independent approval subagent judges every sandbox escalation (risky operations are rejected), with a configurable approval model and a per-session audit trail in the conversation window's 审批 tab.
DeepSeek Harness bundle that routes real sandbox escalation prompts through a dedicated approval reviewer
DeepSeek Harness plugin (temporary, pre-upstream-fix): automatically injects a CPython tempfile shim (PYTHONPATH -> sitecustomize) into every confined shell command on the Windows sandbox, so python/pytest tempfile use works with zero extra tools, zero model-context overhead, and zero escalation.
LLM pre-review for sandbox-escalation approvals: an independent-context LLM gate answers sandbox escalation requests before they reach the user, falling back to the user on any failure.
DeepSeek Harness plugin: Windows toast notification with sound when a session requests elevation approval — shows which session, why, and what is being elevated
Strip sandbox_permissions / justification from model tool-call arguments so sandbox escalation is never triggered when the session already has sufficient permission.
Approval timeout for DeepSeek Harness: when a sandbox/escalation approval is not answered within a configurable period, the request is automatically rejected instead of waiting forever.
DSH web plugin: play a notification chime when the agent asks the user a question (ask_user_question / plan review) or requests approval (sandbox escalation, tool approval).
DSH Web GUI plugin: session-scoped "Allow for this session" option in the approval dialog (per-mode standing grants, localStorage)
Normalize invalid non-widening sandbox escalation arguments before DSH tool dispatch.
Static security audit for dynamic Cordis plugins in DeepSeek Harness (DSH): rule-based source scanning, risk reports injected into tool results, and user-approval escalation before activating high-risk plugin packages.
Tolerate redundant sandbox_permissions requests: a tool call asking to escalate to a mode that is not strictly wider than the session's standing mode runs as-is at the standing mode instead of failing with "not strictly wider". Genuine escalations still route through the approval flow.
Press Enter to approve the sandbox permission-escalation prompt in the DeepSeek Harness WebUI instead of clicking the Allow once button.
Ring a real system sound when a DeepSeek Harness agent finishes a task OR needs your input: custom sounds, 4 intensity levels, auto escalation, and a settings card in the GUI.
Approval-free writes into whitelisted directories: wraps ctx.fs.checkedTarget for write/edit, auto-approves bash escalations that touch whitelisted paths, with a Settings GUI. 白名单目录写入免审批(write/edit + bash),设置页可视化调整。