[建议 / Feature] Add Advisor Tool / Dual-Model Agent Mode for ZCode Agent

提交前确认 · Pre-submission checklist

• 我已搜索过现有 issue，确认这不是重复提议 / I searched existing issues and confirmed this isn't a duplicate.
• 我已阅读 CONTRIBUTING.md / I've read CONTRIBUTING.md.

问题类别 · Category

UI / 界面体验 · UI / UX

涉及的 Agent 框架 · Agent framework

ZCode Agent（自研）

使用场景 · Use case

For long-running coding tasks, using the strongest available model for every tool call and implementation step can be unnecessarily expensive and slower than using a faster model for routine execution.

I would like ZCode to support an Advisor architecture similar to Anthropic's Advisor tool , where the primary executor model performs most of the coding and tool work but can periodically consult a stronger model for higher-level reasoning.

A typical configuration could be:
Executor: GLM-5.3-Flash
Advisor: GLM-5.3

The executor would remain responsible for editing files, running commands, searching the codebase, implementing changes, and performing most routine reasoning. When it reaches a difficult architectural decision, debugging problem, ambiguous requirement, or needs to reconsider its approach, it could consult the advisor model and then continue execution using the advisor's recommendations.

Doing something like this today requires lots of manual switching models then asking a stronger model to review the work. This interrupts the autonomous agent workflow and makes it harder to take advantage of different model strengths.

建议方案 · Proposal

Add an optional Advisor/Executor mode to ZCode Agent.

The executor remains the primary agent and retains control of the coding session, while an independently configurable advisor model can be called when additional reasoning is useful.

Suggested functionality:

Independently select the Executor and Advisor models.
Allow both models to come from Z.ai / GLM Coding Plan/API or Custom providers, including mixing providers where supported.

Recommended default configuration:
Executor: GLM-5.3-Flash
Advisor: GLM-5.3

Allow the executor to automatically decide when consulting the advisor would be useful.
Support explicit advisor consultation for cases such as architecture decisions, implementation planning, difficult debugging, code-review checkpoints, failed approaches, and course correction.
Allow users to configure an advisor call limit/budget per task or session.
Optionally expose modes such as Automatic, Manual only, or Disabled.
Return the advisor's recommendation to the executor rather than transferring ownership of the agent session.
Preserve enough relevant task/code context for the advisor to provide useful recommendations without requiring the advisor to perform all of the executor's token-heavy work.

Conceptually:

User → Executor (GLM-5.3-Flash) → tools / code changes

When additional reasoning is needed:

Executor → Advisor (GLM-5.3) → recommendation → Executor → continue execution

This would be similar in concept to Anthropic's Advisor architecture.

预期价值 · Expected value

This unlocks the speed and efficiency of GLM-5.3-Flash for routine implementation work.

For large or long-running coding sessions, the majority of tokens are often spent reading files, making edits, running tools, analyzing command output, and iterating on straightforward implementation details. Those operations do not necessarily require the strongest model on every turn.

An Advisor architecture would allow ZCode to reserve the stronger model for the points where its additional reasoning capability provides the most value.

The expected benefits are:

Lower cost — most token-heavy execution can use the faster/cheaper executor model.

Higher speed — Flash can handle routine implementation and tool interaction without requiring the stronger model for every step.

Better reasoning quality — GLM-5.3 can still be consulted for architecture, planning, difficult debugging, and strategic decisions.

Better long-running agent reliability — periodic advisor reviews could identify incorrect assumptions or unproductive approaches before the executor spends significant time following them.

Greater model flexibility — advanced users could create combinations optimized for speed, cost, reasoning quality, or their own custom providers.

Overall, this could provide a useful middle ground between running the entire coding agent on GLM-5.3 and running the entire task on GLM-5.3-Flash.

你认为的优先级 · Your perceived priority

中 · Medium

你使用的 ZCode 版本 / 环境 · ZCode version / environment

v3.10.2/ Windows

补充材料 · Additional context

One especially useful extension would be configurable advisor policies, for example:

Advisor calls: Automatic | Manual | Disabled

Maximum advisor calls per task: 0 / 1 / 3 / 5 / Unlimited

Consult on: Planning | Architecture | Repeated failures | Debugging | Code review | Before completion

A further improvement could allow the executor to provide the advisor with a compact summary of the current task, relevant files, attempted approaches, errors, and proposed next action. The advisor could then return recommendations without needing to take over tool execution itself.

Start by reading CONTRIBUTING.md and then locate the ZCode Agent implementation and its model-selection or session-control entry points; none are named in this issue. Map where executor configuration, provider selection, tool calls, and task context are handled. Done should include a scoped design and implementation path for optional advisor consultation, budgets, modes, and returning recommendations to the executor.