Subagent thinking effort stuck at "high" for third-party models: support_efforts ignored and global/inherited effort always wins

What version of Kimi Code is running?

Reported on 0.38.x; mechanism verified against main @ 9a71582.

Which open platform/subscription were you using?

Third-party model (custom OpenAI-compatible provider configured via config file).

What issue are you seeing?

Original user report (Chinese):

kimi code调用子代理的思考强度有严重bug。我配置了第三方模型，无论我修改config文件里的描述还是web端的设置，或者我直接在提示词里直接说明使用max思考强度都没用。调用子代理都只会使用high思考强度（可能和主代理思考强度有关）最离谱的是，我把config文件中子代理模型中的思考强度删掉了low和high，只保留了max，调用子代理时依然会使用high思考强度

Summary: with a third-party model configured, subagents always send thinking effort high, and none of the user's countermeasures work:

• editing the model's thinking fields in the config file;
• changing the thinking setting in the web UI;
• asking for "max" thinking explicitly in the prompt;
• even removing low/high from the subagent model's support_efforts so only max remains — subagents still send high.

What steps can reproduce the bug?

1. Configure a third-party (non-Kimi protocol) model and a [secondary_model] pool for subagents.
2. Set that model's support_efforts to ["max"] only.
3. Spawn a subagent (Agent tool) bound to that model.
4. Inspect the outgoing request: reasoning_effort is "high".

What did you expect to happen?

• A subagent model that declares support_efforts: ["max"] should resolve to max (or at least to a declared effort), regardless of protocol.
• Global/inherited effort should not silently override a subagent model's declared constraints; subagent binding should clamp to the target model's declared efforts.
• The max→high startup migration should not rewrite a value the user deliberately set via the UI.

Root cause analysis (main @ 9a71582)

How a subagent's thinking effort is resolved (v2 engine):

1. Subagent inheriting the caller's model (primary, or no [secondary_model.models] pool) inherits the caller agent's current thinking level — resolveSubagentBinding returns own.thinkingLevel, and fork binds thinking ?? sourceData.thinkingLevel (agent-core-v2/src/session/subagent/configSection.ts, agent-core-v2/src/session/agentLifecycle/agentLifecycleService.ts).
2. Subagent bound to a pool/force model gets thinking = undefined; at bind time it resolves via resolveThinkingEffortForModel(undefined, global [thinking] config, model, strict) — the global thinking.effort wins over the model's own default_effort / support_efforts (agent-core-v2/src/agent/profile/profileService.ts, agent-core-v2/src/kosong/model/thinking.ts).
3. Prompt-level requests cannot help: the subagent spawn path has no thinking parameter at all — only a model choice.

Why high survives every countermeasure:

• Non-Kimi protocols never validate support_efforts. strictThinkingValidation is only enabled for the kimi provider (kosong/provider/providers/kimi/kimi.contrib.ts:96). For OpenAI/Anthropic-compatible third-party providers the resolved effort is passed through verbatim as reasoning_effort (kosong/provider/bases/openai/openai-legacy.ts:616), so editing support_efforts changes nothing. (This half is addressed by PR #3144.)
• The web UI writes thinking effort to the global config (POST /config), so both main-agent and subagent resolution pick it up.
• Startup migration migrateThinkingEffortMaxToHigh rewrites a persisted thinking.effort = "max" to "high" once (agent-core-v2/src/app/config/migrations.ts), because max is session-only now — a user who set max via the web UI is silently moved to high on the next start.
• When a model declares support_efforts containing high but no default_effort, registration defaults default_effort to high (kosong/model/modelAuth.ts; v1: agent-core/src/config/model.ts).

Additional context

PR #3144 implements "declared support_efforts ⇒ fall back to a declared effort" independent of protocol, which fixes the verbatim-passthrough half once merged. The precedence question (global/inherited high winning over a subagent model's declaration), the web-UI-writes-global behavior, and the migration behavior are not covered by it.

Start with resolveSubagentBinding in agent-core-v2/src/session/subagent/configSection.ts and the fork logic in agent-core-v2/src/session/agentLifecycle/agentLifecycleService.ts. Then trace resolveThinkingEffortForModel in agent-core-v2/src/agent/profile/profileService.ts and kosong/model/thinking.ts, along with the migration in agent-core-v2/src/app/config/migrations.ts. Done means subagent effort respects the target model's declared support_efforts, while inherited and global values no longer override valid model constraints or silently rewrite the user's setting.