fix: prevent runtime permission stalls and Agent reconnect failures

Execute could stall on workspace-external access or native child-session
permissions, while structured questions could be lost or answered incorrectly.
Route owned OpenCode descendant interactions, correct Continue question and
custom-answer handling, and resolve DSH Execute ACP permissions without a
second approval. Ask remains read-only and unrelated sessions stay isolated.

Restore the required DSH session projection service. Reuse offline OpenCode
configuration by plugin version, remove unused npm shims from remote payloads,
normalize dependency modes, and accept real compound descriptors up to 4 MiB.
Fix intermittent Agent attach rejection and closed-handshake waits.

Add focused and real-Runtime regressions, document interaction boundaries and
validation evidence, and include the existing application/tool navigation
design drafts. Those navigation changes are design-only, not implemented UI.

Validation: full suite 3597 passed, 0 failed, 61 skipped across 338 files;
typecheck, lint, release-note verification, whitespace and relative-link checks
passed. Current-source Linux x64 installation, Ask/Execute, native subagent
external writes, reconnect and stop/bootstrap passed with 10 completed real
model calls, 13 native endpoint tests and 20 consecutive handshakes.
System sleep/wake and release CI packaging remain unverified.

Release note: 修复 Execute 访问工作区外目录或运行原生子代理时卡在权限等待的问题，并恢复 OpenCode 与 Continue 结构化问答的正确转交。修复远程 Agent 安装与重连失败；Ask 仍保持只读，任意终端密码或业务确认不会自动代答。

FEATURES.md

@@ -119,8 +119,8 @@ otherwise.
   directories, and create Ask or Execute projects. Both modes start the
   signed Runtime directly without bubblewrap. Ask applies OpenCode's Ask
   permission configuration and permits only native read approvals at the
-  Agent tool-dispatch boundary, while Execute receives no such override and
-  retains all permissions of the selected SSH account. The Agent owns accepted Prompts,
+  Agent tool-dispatch boundary, while Execute explicitly allows native permissions
+  and retains all permissions of the selected SSH account. The Agent owns accepted Prompts,
   provider/tool rounds, Runtime processes, a stable model ledger, and a bounded
   semantic transcript over a private Unix socket and ACP v5. Work continues on
   the Host after Desktop exit, network loss, or local-process termination.
@@ -217,7 +217,13 @@ otherwise.
   code is not subject to Ask initialization isolation.
 - [x] **Ask and Execute work modes**: Ask remains read-only. Execute is the
   user's authorization for all tools, processes, network access, and writable
-  paths available to the current local or SSH account.
+  paths available to the current local or SSH account, including paths outside
+  the workspace and native subagent work.
+- [x] **Native Runtime interaction routing**: OpenCode and Continue questions
+  support choices, yes/no, free-text answers, and skipping through the existing
+  question card. OpenCode also routes questions from owned child sessions.
+  Execute permission confirmations are handled automatically rather than
+  waiting for another approval. See the [interaction boundaries](./docs/features/assistant-workbar/runtime-interactions.md).
 - [x] **Experts and Subagents**: Supports explicit experts, team analysis, and
   up to three read-only experts running in parallel. Chat shows each
   expandable full expert response first and the main Agent's synthesis below,