Add Phase 5.5: verify mismatches via agent discussion

[hua7450]

Summary

• Adds Phase 5.5 to the audit-state-tax command that verifies reported mismatches via agent-to-agent discussion before including them in the final report

Motivation

During a full audit of Iowa's 2025 income tax PR (#7389 in policyengine-us), an audit agent reported a false positive: it flagged the QBI fraction parameter as wrong for 2023+ (0.75 instead of 1.0). However, tracing the code revealed that ia_qbi_deduction (which uses the fraction) is only called in the pre-2023 indiv/joint path — the 2023+ consolidated path starts from federal taxable income directly, where the QBI deduction is already included. The parameter was correct; the agent just didn't trace the code path.

What Phase 5.5 Does

For each MISMATCH reported by an audit agent:

1. Creates a verification team (TeamCreate)
2. Spawns a verifier agent that greps for parameter usages and traces the call chain from the top-level tax variable down to the flagged parameter
3. Resumes the original audit agent as a teammate so it can explain its reasoning
4. Lets them discuss (up to 3-4 round-trips via SendMessage)
5. Collects a verdict: CONFIRMED, REJECTED, or INCONCLUSIVE
   • CONFIRMED → proceeds to Phase 6 (600 DPI visual verification)
   • REJECTED → excluded from final report (noted as "investigated and cleared")
   • INCONCLUSIVE → proceeds to Phase 6 for manual verification

Other Changes

• Updated Key Rules to require both code-path verification AND visual confirmation
• Added Rule 8: "Trace code paths" — a parameter mismatch is only real if the parameter is reachable in the target tax year
• Updated Pre-Flight Checklist

Test plan

• Verified the edit applies cleanly on top of add-audit-state-tax-command
• Run an actual audit with mismatches to test the team discussion flow

🤖 Generated with Claude Code

[MaxGhenis]

The concept here is good (verifying mismatches via code-path tracing before reporting them), but the implementation references TeamCreate and SendMessage which don't exist in Claude Code. The workflow can't run as written. Please rework to use standard Agent tool calls (e.g. spawning a general-purpose subagent to verify the code path) instead of the team API. Happy to review a revised version.

[hua7450]

The concept here is good (verifying mismatches via code-path tracing before reporting them), but the implementation references TeamCreate and SendMessage which don't exist in Claude Code. The workflow can't run as written. Please rework to use standard Agent tool calls (e.g. spawning a general-purpose subagent to verify the code path) instead of the team API. Happy to review a revised version.

I implement this feature in the /review-program command, which design to supersede /review-pr and /audit-state-tax