[Bug] Subagents always fail with 400 "Reasoning is mandatory" when parent uses a thinking-required model (Ox Alpha via OpenRouter)

[Sgcan]

"## Environment\n\n- dsh: 0.1.0-rc.6 (core), DSH Desktop 2.0.0 (Electron shell)\n- macOS 26 (Apple Silicon)\n- Provider: OpenRouter custom provider (llm-pi-ai), model stealth/ox-alpha (thinking-required model)\n- Parent session configured with reasoningEffort: high (works fine for the parent)\n\n## Symptom\n\nEvery subagent spawn from a parent session using stealth/ox-alpha fails immediately. The parent session log receives:\n\n\nBackground subagent <id> failed before it finished. It left no closing message.\n\n\nThe child session's log ends with:\n\n\nturn/end: {\"turn\": 1, \"reason\": {\"kind\": \"error\", \"error\": {\"message\": \"400: {\\\"message\\\":\\\"Reasoning is mandatory for this endpoint and cannot be disabled.\\\",\\\"code\\\":400,\\\"metadata\\\":{\\\"provider_name\\\":null}}\", \"code\": \"INVALID_REQUEST\"}}}\n\n\n## Root cause (traced in source)\n\nThree independent findings in the unpacked app:\n\n1. dsh-subagent resolveChildAgentOptions() only forwards provider, model, and maxTokens from the parent to the child \u2014 reasoningEffort is silently dropped:\n\njs\nfunction resolveChildAgentOptions(parent, requested, childDepth) {\n const parentProvider = parent.options.provider;\n const parentModel = parent.options.model;\n const parentMaxTokens = parent.options.maxTokens;\n return {\n ...parentProvider !== void 0 ? { provider: parentProvider } : {},\n ...parentModel !== void 0 ? { model: parentModel } : {},\n ...parentMaxTokens !== void 0 ? { maxTokens: parentMaxTokens } : {},\n ...requested,\n subagentDepth: childDepth\n };\n}\n\n\n2. dsh-tool-subagent agentOptions schema also has no reasoningEffort field, so callers cannot pass one explicitly:\n\njs\nagentOptions: z.object({\n provider: z.string(),\n model: z.string(),\n maxTokens: z.number().step(1).min(1).max(Number.MAX_SAFE_INTEGER)\n}).default(void 0),\n\n\n3. dsh-llm-pi-ai resolveModel() tries to derive a default effort from the child's own profile (describableReasoningLevel(resolvedModel, profile.reasoning)), but a spawned child's profile carries no reasoning level \u2014 so defaultEffort is undefined and the outgoing request has no reasoning_effort at all.\n\nResult: the child request goes to OpenRouter without reasoning_effort; for a thinking-required model like Ox Alpha that is a hard 400, so every child dies on its very first turn. Any parent model that mandates thinking (reasoning-required models on OpenRouter, etc.) makes the subagent tool 100% unusable.\n\n## Expected behavior\n\n- Children should inherit the parent's reasoningEffort (or a sensible per-model default) so thinking-required models work in delegation, OR\n- agentOptions should accept a reasoningEffort so callers can pin it, OR\n- The model declaration should support a defaultEffort for third-party llm-pi-ai providers (currently reasoningEfforts list exists but there is no default), so children get a valid effort automatically.\n\n## Workarounds used\n\n- Switching the parent session to a model that tolerates a missing effort (e.g. SiliconFlow DeepSeek V4 Flash) makes subagents work.\n- Community plugin foxfasf/dsh-delegate-effort appears to target exactly this gap (not yet tested).\n\n## Related\n\n- dsh-reasoning-settings plugin (independent reasoning-effort page for llm-pi-ai providers) \u2014 shows effort config for third-party providers is currently limited."

[weijiafu14]

Your source trace matches what we see: resolveChildAgentOptions() forwards only provider / model / maxTokens, and reasoningEffort never reaches the child — so a thinking-required endpoint rejects the child's first request. That's DSH's own dsh-subagent, and we can't fix it.

What we can offer is a different subagent runtime that carries reasoning effort per child. pi2dsh runs unmodified Pi-ecosystem plugins as native DSH plugins:

dsh plugin --profile <your-profile> add pi2dsh
dsh plugin --profile <your-profile> add @tintinweb/pi-subagents
# restart dsh

Every child it spawns is a real DSH session created through ctx.agents, and:

• a per-child thinking level becomes that child's own reasoningEffort, applied through the official agent/request waterfall — it does not leak to siblings or to the parent;
• the child's model resolves in three tiers: an explicitly requested model → otherwise the caller's live route (read from the caller's durable session log's last request/header, so a mid-session model switch in the UI is inherited) → otherwise the creation-time snapshot.

So for stealth/ox-alpha, the child request carries reasoning instead of dropping it.

The honest boundary: this does not repair DSH's native subagent / subagent_fork tools — it is a parallel runtime you delegate through instead.

Evidence, run on a real DSH loop: examples/subagents · acceptance report.