Plugin-defined agent types with tools:all silently block Write/Edit tool calls

Bug Description

When using a plugin-defined agent type with tools: all in the agent definition frontmatter, the sub-agent's Write/Edit tool calls are silently blocked. The agent reports creating files successfully, but nothing is actually written to disk. No error is surfaced.

Reproduction

1. Create a plugin with an agent definition file (agents/executor.md):

``yaml
---
name: executor
tools: all
---
# Executor Agent
...
``

1. In a skill's SKILL.md, instruct the Orchestrator to call the Agent tool with:

``
subagent_type: review-loop:executor
``

1. The sub-agent attempts to create/edit files via Write/Edit tools — it reports success, but no files are written.

1. Switching to subagent_type: general-purpose (built-in type) with the same prompt and manually embedding the agent's system prompt — Write/Edit works correctly.

Expected Behavior

tools: all in plugin agent definition frontmatter should grant the same tool access as the built-in general-purpose agent type. If it intentionally doesn't, Write/Edit calls should fail with a clear error, not silently succeed.

Actual Behavior

Write/Edit tool calls are silently blocked — the sub-agent's output claims files were created/modified, but the filesystem is unchanged. No error is returned to the agent or the caller.

Impact

This is particularly problematic because:

• The silent failure makes debugging very difficult (agent thinks it succeeded)
• There's no documentation about this restriction
• The tools: all declaration in frontmatter implies full tool access

Environment

• Claude Code CLI
• Plugin: review-loop v1.3.0
• Workaround: use subagent_type: general-purpose and manually embed agent system prompt in the prompt parameter

Workaround

Use subagent_type: general-purpose for any agent that needs file write access, and include the agent definition body manually in the prompt. Plugin-defined agent types work fine for read-only agents (e.g. tools: read-only).