Agent Runtime Forensic Prompts

# Agent Runtime Forensic Prompts

Use this reference when user asks for a prompt to send to another agent/runtime to explain a configuration, model route, context window, or tool-access discrepancy.

## Session-derived lesson

If user asks for a prompt for another agent, the correct move is to produce the paste-ready prompt. Do not replace the request with local discovery, remote probing, or a manual-terminal script unless user explicitly asks for that route.

## Prompt pattern

A good cross-agent forensic prompt should:

1. Name the target runtime and mission.
2. Include the known baseline from the current machine.
3. Put red lines before tasks:
   - read-only unless approved;
   - no restarts;
   - no config/model/auth/provider edits;
   - no installs/migrations;
   - no secrets or raw credential files;
   - presence-only credential reporting.
4. Tell the target agent to inventory its available tools/status routes first.
5. Avoid assuming terminal/file tools exist.
6. If terminal/file tools fail, require the exact error plus an agent-side recovery path.
7. Distinguish real backend behavior from local metadata, stale cache, UI estimates, config overrides, or hallucinated assumptions.
8. Require a final report with PASS / INCONCLUSIVE / BLOCKED.

## Output sections to request