test-agent-tools

name	test-agent-tools
description	Probe, test, and document every tool the agent harness has access to (native tools, subagents, connected MCP servers). Use when the user wants to test, audit, or inventory the harness's tool/capability set, or asks "what tools do you have".

Test Agent Tools

Purpose

Enumerate available tools/capabilities — including connected MCP servers — exercise each one (web search, subagents, MCP calls), and produce a final report with a status table + subagent summary, under a timestamped demo/ run folder.

Workflow

1. Create folder demo/YYMMDD_hhmm (current date/time).
2. List available tools/capabilities, including any connected MCP servers/tools.
3. Test each tool:
   • Non-subagent tools (e.g. web search): run directly, write findings flat inside demo/YYMMDD_hhmm/.
   • Subagent-based exploration: dispatch one subagent per topic/tool; each writes to its own demo/YYMMDD_hhmm/<subagent_id_or_name>/findings.md.
   • MCP tools: if any MCP servers are connected, invoke each connected tool at least once; record request/response/error.
4. Write demo/YYMMDD_hhmm/REPORT.md:
   • Table: | Tool | Type (native/MCP/subagent) | Status (OK/FAIL/SKIPPED) | Notes |
   • Summary section aggregating each subagent's report (what it tested, key findings).
5. Post the full content of REPORT.md directly in the chat reply (not just a file link) — including the tool status table and the subagent summary. If subagents produced findings, include their key results inline too, not just a reference to the file.
   • Files (REPORT.md, findings, etc.) stay in English regardless of chat language.
   • The chat-transcript presentation of the report (narration, summary, table labels) uses whatever language the user is currently using in the conversation — translate on the fly, don't just paste the English file verbatim if the chat language differs.

Mandatory Rules

• Only create files/folders inside demo/.
• Never create files/folders at the root directory or outside demo/.
• One run = one folder demo/YYMMDD_hhmm (current date/time, no suffix/label).
• No subagent: write markdown files flat inside demo/YYMMDD_hhmm/.
• With subagent: each subagent writes into its own demo/YYMMDD_hhmm/<subagent_id_or_name>/.
• REPORT.md always lives flat at demo/YYMMDD_hhmm/REPORT.md, never inside a subagent folder.

Mode Lock

• Never offer, suggest, or ask the user to switch agent mode (Build/Plan/Brave/Ask).
• Always operate in the currently active mode, regardless of task complexity/risk.
• If the task needs another mode's capability (e.g. execution needed but mode = Plan/Ask):
  • Do as much as the active mode allows.
  • Report the blocker factually (e.g. "needs write access, current mode is read-only").
  • Do NOT frame it as "want me to switch to mode X?" — mode switching is user-initiated only.
• This holds even if the user seems unsure or asks "should I switch mode?" — answer factually, but still don't proactively offer/trigger a switch.

Examples

• Flat (non-subagent/MCP): demo/260731_1420/mcp-github/findings.md or demo/260731_1420/findings.md
• With subagent: demo/260731_1420/web-search-agent/findings.md
• Final report: demo/260731_1420/REPORT.md