[FEATURE] `keep-coding-instructions` field to agent configuration for unified non-coding workflows

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

With the introduction of --agent flag and agent setting in v2.0.59, users can now configure the main thread with an agent's system prompt, tool restrictions, and model. However, there's no way to disable coding-specific instructions within the agent configuration itself.

Currently, to create a non-coding agent (e.g., business analyst, researcher, writing assistant), users must:

1. Create an agent file in .claude/agents/my-agent.md (for prompt, tools, model)
2. Create a separate output-style file in .claude/output-styles/non-coding.md with keep-coding-instructions: false
3. Configure outputStyle in settings.json
4. Run claude --agent my-agent

This creates unnecessary complexity and file sprawl for a common use case.

Current State

Agent frontmatter supports:

---
name: my-agent
description: ...
tools: Read, Write, Bash
model: sonnet
permissionMode: default
skills: skill1, skill2
---

Output-style frontmatter supports:

---
name: my-style
description: ...
keep-coding-instructions: false  # ← Only available here!
---

These are two parallel systems that don't integrate.

Proposed Solution

keep-coding-instructions field to agent configuration:

---
name: business-analyst
description: Strategic analysis and planning
tools: Read, Write, Bash, WebSearch
model: opus
permissionMode: default
keep-coding-instructions: false  # ← ADD THIS
---

You are a business analyst specializing in...

This would allow a single --agent business-analyst command to fully configure both the agent persona AND disable software engineering instructions.

Alternative Solutions

1. Current workaround: Maintain parallel agent + output-style files (complex, error-prone)
2. Shell aliases: alias myagent='claude --agent x && /output-style y' (hacky, doesn't persist)
3. SessionStart hooks: Add instructions via hooks (less effective than system prompt modification, loses power over conversation length)

Priority

High - Significant impact on productivity

Feature Category

CLI commands and flags

Use Case Example

• High - This affects anyone using Claude Code for non-software-engineering tasks
• Growing use case - With Output Styles, Anthropic acknowledged Claude Code can be "any type of agent"
• Simplification - Reduces configuration from 3 files to 1 file

Additional Context

Related discussions:

• Issue #10694: Ability to "turn off" default system prompt after output-styles deprecation
• Issue #10721: Concerns about output-styles deprecation and plugin alternatives
• Issue #5361: Main thread direct agent invocation (partially addressed by --agent flag)

The --agent flag in v2.0.59 is a great step forward. Adding keep-coding-instructions would complete the unification of agent and output-style systems.