Subagent Evidence

Subagent Evidence

Prerequisites & Dependencies

• Access to explore agent dispatch workflow
• Power to post-process agent summaries before acceptance
• No external runtime dependencies

Dispatch-Time Contract

When dispatching an explore agent, enforce the following contract:

1. Observation-only return: The agent returns raw observations, never inferred conclusions.

2. Each observation must include:

   • file:line citation — exact file path and line number where the observation originates
   • excerpt — the raw text snippet (1-3 sentences) from that location
   • tag — one of: [OBSERVED], [INFERRED], [HEURISTIC]

3. Inference tagging requirement: Any conclusion drawn by the agent must be explicitly tagged [INFERRED] and accompanied by at least 2 supporting observations with file:line citations.

4. No bare claims: Statements without file:line citations are rejected at dispatch time.

Post-Dispatch Verification

After receiving an agent summary:

1. Sample-cite: For each substantive claim in the summary, locate and verify at least 2 file:line citations from the original observations.

2. Flag unsupported conclusions: Claims with fewer than 2 verified citations are marked [UNVERIFIED] in the chat.

3. Reject inference without evidence: Any [INFERRED] claim without 2 supporting observations is discarded and re-reported to the user.

Verification Workflow Example

Agent summary: "The codebase uses Factory pattern extensively"
→ I verify: 2+ citations showing Factory pattern usage
→ Result: Claim accepted with [VERIFIED] tag

Agent summary: "Component X is unused"
→ I check: only 1 citation found for "unused"
→ Result: Claim marked [UNVERIFIED], re-reported to user

Windows PowerShell Notes

• Citation format: C:\path\to\file.py:42 or relative paths as appropriate
• Excerpt extraction: use sed -n '42p' "C:\path\to\file.py" to pull exact line
• Tag markup: wrap inferences in [INFERRED] brackets explicitly in the summary text