dsh-model-compat-guard

Plugging thinking-heavy models like GPT-5.x or Qwen-thinking into DSH almost guarantees three recurring failures: compaction summaries hitting a hard 8,192-token cap and failing closed, speculative sandbox_permissions payloads triggering impossible escalation errors, and tool calls rejected wholesale for a missing description field. dsh-model-compat-guard exists to fix exactly those three with zero configuration out of the box.

It sets three interception points along the request waterfall. For compaction, it raises maxTokens to 32,768 and downgrades the reasoning effort to the cheapest tier the model actually supports, so thinking tokens no longer blow through the output budget. Before tool execution, it resolves the current sandbox mode; if an escalation request is destined to fail (same or narrower tier, no sandbox service, unknown mode) it strips the redundant sandbox_permissions and justification so the call proceeds under permissions already granted, while legitimate wider upgrades still flow through the normal approval gate. For bash, run_code, and subagent invocations, it auto-synthesizes a description from the first meaningful line before strict schema validation, leaving already-valid descriptions untouched.

Ideal for DSH users running long conversations with thinking models, engineering teams that rely on sandboxed executors for code generation, and developers juggling multi-model workflows daily. No config changes are required; a log line containing compat-guard: tuned compaction request confirms the guard is active.