[Bug] 子代理默认委派失败：子代理继承的是过期的 base 默认模型，而非父代理实际运行的模型

[epiloguess]

[Bug] 子代理（subagent）默认委派失败：子代理继承的模型与主代理实际模型不一致——根因包含 options 创建时快照定格、全局 agent-default-model 被会话级切换污染，且错误被吞没

环境

• 产品：DeepSeek Harness（dsh），web profile
• 涉及包（源码仓库路径 / 发布包）：
  • packages/host/apiproxy（发布包 @deepseek-ai/dsh-host-apiproxy）
  • packages/subagent/subagent（发布包 @deepseek-ai/dsh-subagent）
  • packages/subagent/subagent-in-process-driver（发布包 @deepseek-ai/dsh-subagent-in-process-driver）
  • packages/core/agent（发布包 @deepseek-ai/dsh-agent）
  • packages/subagent/agent-default-model（发布包 @deepseek-ai/dsh-agent-default-model）
  • packages/bundle/base（发布包 @deepseek-ai/dsh-base）

摘要

通过 subagent / subagent_fork 或 workflow 的 agent() 委派子代理时，默认情况下失败，表现为 Error: subagent run failed（或 agent() 返回 null、子代理产出 0 token / contextWindow 不一致）。子代理在首轮 LLM 调用即挂掉——它拿到的模型不是父代理实际运行的模型。

复现步骤

1. dsh web 使用默认 web profile（base 默认 agent-default-model 指向未配置凭证的 deepseek-official/deepseek-v4-flash）。
2. 主会话正常（所选模型按请求注入）。
3. 调用 subagent 工具或 workflow 的 agent("...")，不指定 provider/model override。
4. 观察：Error: subagent run failed。
5. 对照：显式 agent("...", { provider, model }) override → 成功。

证明执行后端正常，坏的是到达子代理的默认模型选择。

期望行为

父代理可指定子代理模型；未指定时，子代理应使用父代理实际运行的同一模型，而非来自其他来源的硬编码/未知默认。

实际行为

• 子代理经 resolveChildAgentOptions 继承 parent.options 快照（packages/subagent/subagent/src/child-agent.ts:68-83）。
• web 入口在创建主代理时，把 parent.options 定格为 base 默认 deepseek-official：selection.current getter 的兜底返回 defaults.defaultModelSelection() 而非 agentDefaultModel.currentSelection()（packages/host/apiproxy/src/api-proxy.ts 的 selectionFor，对应发布包 dsh-host-apiproxy/lib/index.js:1697-1706）。
• 主代理自身正常，因其模型经 installModelSelection 在 agent/request 时按请求注入（packages/core/agent/src/model-selection.ts，对应 dsh-agent/lib/index.js:255-303）——这是与 options 快照完全分离的路径。
• 子代理真实失败（如 deepseek-official 无凭证）被折叠为 stopReason: "error" 且无 diagnostic（dsh-subagent/lib/index.js:191-212；dsh-subagent-in-process-driver/lib/index.js:124-132），表面只剩 subagent run failed。

根因分析

1. resolveChildAgentOptions 只继承 parent.options 创建时快照
parent.options 是 ReactLoopAgent 构造时保存的不可变快照（只设一次）。主代理实际模型通过 agent/request waterfall 在每次 LLM 调用时覆盖，但不回写 parent.options。UI 切换后父代理用新模型，子代理却继承旧快照。

2. web 入口把 options 定格为 base 默认
selectionFor 的 selection.current getter 兜底到 defaults.defaultModelSelection()（base 硬编码 deepseek-official），而非 agentDefaultModel.currentSelection()（用户实际的默认/选择）。对照 packages/host/headless（发布包 dsh-headless/lib/index.js:66-75）正确地用 agentDefaultModel.currentSelection() 填充 options——无头模式子代理大概率正常，web 模式不正常。

3. agent-default-model 全局配置被会话级切换污染（更深的根因）
selectModel RPC（packages/host/apiproxy/src/api-proxy.ts，对应 dsh-host-apiproxy/lib/index.js:2596-2630，第 2614 行）在每次切换模型时既设置会话级 selection.current，又调用 saveDefaultModelSelection 把值持久化进全局 settings.yaml 的 agent-default-model 命名空间（packages/subagent/agent-default-model，saveSelection → settings.replace）。
→ 名为"默认模型"的全局单值配置，实际是"用户最后一次切换"。在多会话产品里语义错误：会话 A 切模型静默改写全局默认，被会话 B/子代理读取。且 web 入口创建新会话/子代理时并不消费该全局值（用 base 默认），使其既被污染又没被当默认使用。

4. 配置层：tool-subagent 缺 agentOptions
packages/bundle/base 的 cordis.patch.yml（第 313-329 行）中 tool-subagent / tool-subagent-fork 均未设 agentOptions；其 Config 类型虽已支持该字段并会透传，但配置缺失导致子代理无模型兜底。

5. 错误被吞（症状放大器）
toStopReason/readResult 在出错时不附加 diagnostic，使失败几乎无法从表面报错诊断。

涉及组件

包（源码 / 发布）	角色	问题
host/apiproxy / dsh-host-apiproxy	web 入口 / 模型选择网关	主因：selection.current 兜底错（1697-1706）；selectModel 每次写全局默认（2596-2630, 2614）
subagent/subagent / dsh-subagent	子代理继承 + 错误折叠	继承 options 快照；吞 diagnostic
subagent/subagent-in-process-driver / dsh-subagent-in-process-driver	子代理运行结果	出错无 diagnostic
core/agent / dsh-agent	installModelSelection 按请求注入	本身正常，但是父代理唯一路径，凸显与 options 分裂
subagent/agent-default-model / dsh-agent-default-model	默认模型服务 + 持久化	被动：API 被误用于把会话级状态持久化为全局默认
bundle/base / dsh-base	base 默认 agent-default-model	deepseek-official/deepseek-v4-flash 默认值来源；tool-subagent 缺 agentOptions

修复建议

1. 子代理模型兜底：resolveChildAgentOptions 在 parent.options.provider/model 缺失时，从 agentDefaultModel.currentSelection() 读取当前默认作为 fallback（注意：需配合下方 F2/F3，否则在 web 模式下读到的可能是被污染的"最后一次切换"值）。
2. web 入口正确填充 options：创建主代理时，用 agentDefaultModel.currentSelection()（感知 settings/覆盖）填充 options，与 dsh-headless 对齐，而非 defaults.defaultModelSelection()。
3. 分离会话级选择与全局默认：selectModel 只设会话级 selection.current；仅当用户**显式"设为默认"**时才持久化到全局 agent-default-model。恢复该 key 的真实语义。
4. tool-subagent 配置加 agentOptions，但不可硬编码，应从 agent-default-model 读取或做成模板变量（不同部署默认模型不同）。
5. 暴露真实错误：toStopReason 返回 "error" 时附加 diagnostic，使 subagent run failed 可操作。

补充（架构）

两个症状同源：模型选择没有单一事实来源。四套机制重叠共存——options 快照（创建时）、selection.current（会话级/请求时）、请求 config（installModelSelection 注入）、全局 agent-default-model settings。仅修 resolveChildAgentOptions 兜底会暴露"全局默认被污染"；仅修 selectModel 持久化而不修继承链，子代理仍坏。需跨 core/agent / subagent / host/apiproxy / agent-default-model 协调对齐语义。

---

English Version

[Bug] Subagent delegation fails by default: child inherits a model inconsistent with the parent's actual model — root cause spans a frozen options snapshot, a globally-polluted agent-default-model, and a swallowed error

Environment

• Product: DeepSeek Harness (dsh), web profile
• Packages (source path / published package):
  • packages/host/apiproxy (@deepseek-ai/dsh-host-apiproxy)
  • packages/subagent/subagent (@deepseek-ai/dsh-subagent)
  • packages/subagent/subagent-in-process-driver (@deepseek-ai/dsh-subagent-in-process-driver)
  • packages/core/agent (@deepseek-ai/dsh-agent)
  • packages/subagent/agent-default-model (@deepseek-ai/dsh-agent-default-model)
  • packages/bundle/base (@deepseek-ai/dsh-base)

Summary

Delegating via subagent / subagent_fork or workflow's agent() fails by default, surfacing as Error: subagent run failed (or agent() returning null, or the child producing 0 tokens / a contextWindow mismatch). The child dies on its first LLM call — it receives a model that is not what the parent actually runs.

Steps to reproduce

1. dsh web with the default web profile (base agent-default-model points at unconfigured deepseek-official/deepseek-v4-flash).
2. The main session works (selected model injected per request).
3. Invoke subagent / workflow's agent("...") without a provider/model override.
4. Observe: Error: subagent run failed.
5. Control: agent("...", { provider, model }) with explicit override → succeeds.

Proves the execution backend is fine; what's broken is the default model selection reaching the child.

Expected behavior

The parent may specify the child's model; when unspecified, the child should use the same model the parent actually runs on, not a hardcoded/unknown default from another source.

Actual behavior

• The child inherits parent.options via resolveChildAgentOptions (packages/subagent/subagent/src/child-agent.ts:68-83).
• The web entry freezes parent.options to the base default deepseek-official: the selection.current getter falls back to defaults.defaultModelSelection() instead of agentDefaultModel.currentSelection() (packages/host/apiproxy/src/api-proxy.ts, selectionFor; published dsh-host-apiproxy/lib/index.js:1697-1706).
• The parent itself works because its model is injected per request via installModelSelection (packages/core/agent/src/model-selection.ts; dsh-agent/lib/index.js:255-303) — a path completely separate from the options snapshot.
• The child's real failure (e.g. no credentials for deepseek-official) is collapsed to stopReason: "error" with no diagnostic (dsh-subagent/lib/index.js:191-212; dsh-subagent-in-process-driver/lib/index.js:124-132), leaving only subagent run failed.

Root cause analysis

1. resolveChildAgentOptions inherits only the creation-time parent.options snapshot
parent.options is an immutable snapshot saved in the ReactLoopAgent constructor (set once). The parent's actual model is overridden per LLM call via the agent/request waterfall but is never written back to parent.options. After a UI model switch the parent uses the new model while the child inherits the stale snapshot.

2. The web entry freezes options to the base default
selectionFor's selection.current getter falls back to defaults.defaultModelSelection() (base hardcoded deepseek-official), not agentDefaultModel.currentSelection() (the user's actual default/selection). Contrast packages/host/headless (dsh-headless/lib/index.js:66-75), which correctly populates options from agentDefaultModel.currentSelection() — headless children likely work, web children do not.

3. The global agent-default-model config is polluted by per-session switches (deeper root cause)
The selectModel RPC (packages/host/apiproxy/src/api-proxy.ts; dsh-host-apiproxy/lib/index.js:2596-2630, line 2614) on every switch both sets the session-scoped selection.current and calls saveDefaultModelSelection, persisting the value into the global settings.yaml agent-default-model namespace (packages/subagent/agent-default-model, saveSelection → settings.replace).
→ The globally-named "default model" actually holds "the user's last switch". In a multi-session product this is semantically wrong: switching in session A silently rewrites the global default read by session B / children. Moreover, the web entry does not consume this global value when creating a new session/child (it uses the base default), so the value is both polluted and unused as a real default.

4. Config layer: tool-subagent lacks agentOptions
In packages/bundle/base cordis.patch.yml (lines 313-329), neither tool-subagent nor tool-subagent-fork sets agentOptions. The Config type already supports and forwards the field, but its absence leaves the child with no model fallback.

5. Errors are swallowed (symptom amplifier)
toStopReason/readResult attach no diagnostic on error, making the failure nearly impossible to diagnose from the surface message.

Affected components

Package (source / published)	Role	Issue
host/apiproxy / dsh-host-apiproxy	web entry / model-selection gateway	primary: selection.current fallback wrong (1697-1706); selectModel writes global default every switch (2596-2630, 2614)
subagent/subagent / dsh-subagent	child inheritance + error folding	inherits options snapshot; swallows diagnostic
subagent/subagent-in-process-driver / dsh-subagent-in-process-driver	child run result	no diagnostic on error
core/agent / dsh-agent	installModelSelection per-request injection	fine, but the parent's only path — highlights divergence from options
subagent/agent-default-model / dsh-agent-default-model	default-model service + persistence	passive: API misused to persist per-session state as global default
bundle/base / dsh-base	base default agent-default-model; tool-subagent config	source of deepseek-official/deepseek-v4-flash; missing agentOptions

Suggested fixes

1. Child model fallback: in resolveChildAgentOptions, when parent.options.provider/model is absent, read the current default from agentDefaultModel.currentSelection() (note: must be paired with F2/F3 below, otherwise the value read in web mode may be the polluted "last switch").
2. Web entry populates options correctly: when creating the main agent, fill options from agentDefaultModel.currentSelection() (settings/override-aware), aligning with dsh-headless, instead of defaults.defaultModelSelection().
3. Separate per-session selection from global default: selectModel should only set the session-scoped selection.current; persist to the global agent-default-model only on an explicit "set as default" action. Restore the key's true meaning.
4. Add agentOptions to tool-subagent config, but never hardcode — read from agent-default-model or use a template variable (default models differ per deployment).
5. Surface the real error: attach diagnostic when toStopReason returns "error", so subagent run failed is actionable.

Additional note (architecture)

The two symptoms share one root: model selection has no single source of truth. Four overlapping mechanisms coexist — options snapshot (creation-time), selection.current (session/request-time), request config (installModelSelection), and the global agent-default-model settings. Fixing only the resolveChildAgentOptions fallback exposes the "polluted global default"; fixing only selectModel persistence without the inheritance chain leaves children broken. Coordinated semantic alignment across core/agent / subagent / host/apiproxy / agent-default-model is required.

[wangyue9011]

我今天被这问题deepseek余额变成负数了~

[xryanv]

We independently reproduced this subagent model inheritance bug and implemented a focused source-level fix with full test coverage.

Root Cause

resolveChildAgentOptions (packages/subagent/subagent/src/child-agent.ts) and continuable descriptor snapshotting read parent.options.model (the immutable session-creation snapshot) instead of the effective active model from parent.session.requestHeader()?.config.model. When the parent session switches models in the Web GUI, the header is updated, but fresh default subagents inherit the stale creation snapshot.

Fix & Verification

We published a clean standalone fix branch against master:

• Branch: fix/subagent-live-model-inheritance
• Commit: 8467ff75218d6e35feeeea4db24ee0eafe575775
• Tests: 667/667 passed in packages/subagent/ (including regression tests for one-shot and continuable parent model switches, cold resumes, and descriptor fidelity).