Tax Evolve

Tax Agent Self-Evolution

Analyze discrepancies found by /tax-compare, classify root causes, generate fixes to skill files and references, and submit improvements as a GitHub PR.

Prerequisites

• data/comparison-{year}.json must exist. If not, tell the user to run /tax-compare first.
• data/tax-calculation.json must exist (for tax year context).

Privacy Contract (CRITICAL — read before every step)

NEVER include user PII in any generated file changes. This means:

• Never write actual dollar amounts from the user's return into skill files or references.
• Never include the user's name, SSN, address, employer, or any personal data.
• Only include: which 1040 line was wrong (by line number), the direction of the error (+/-), the percentage error, and the corrected calculation rule.
• When logging to evolution-log.json, store only line numbers, deltas, and classification — never raw values that could identify a taxpayer.

Steps

Step 1: Load Comparison Data

Read data/comparison-{year}.json. Identify all entries with status: "mismatch" or status: "missing_ours". If there are no mismatches, tell the user:

"Great news — no discrepancies found! No fixes needed." And stop.

Step 2: Classify Each Discrepancy

For each mismatch, classify it into one of three categories by reasoning about the line:

calculation_error — The interview collected the right data, but the formula in tax-calculate or its references (brackets-2025.md, credits-2025.md) produced the wrong result.

• Signs: Income lines (1a, 11, 15) are correct but credits or deductions lines (19, 21, 24) differ.
• Action: Edit skills/tax-calculate/SKILL.md or the relevant reference file.

missing_data — The interview did not ask a question needed to correctly compute this line.

• Signs: A line is missing_ours (zero in our return, non-zero in reference) for a deduction or credit the user is entitled to.
• Action: Propose new interview questions for skills/tax-start/SKILL.md.

missing_domain — The discrepancy involves an income or deduction type that has no dedicated domain skill yet (e.g., Schedule C self-employment, HSA, student loan interest).

• Signs: An entire schedule or section is zero in our return but non-zero in reference.
• Action: Propose a new tax-domain-* skill scaffold.

Document the classification for each mismatch before proceeding to fixes.

Step 3: Check Evolution Log for Patterns

Read data/evolution-log.json (create it if it does not exist with {"entries": []}).

Look for lines that have appeared as mismatches in previous runs (same line field). If any line has appeared 2 or more times:

"Line {N} ({description}) has appeared as a mismatch in {count} previous runs. This suggests a systemic gap — proposing a domain skill scaffold rather than a one-off fix."

Upgrade the classification of repeat offenders to missing_domain if appropriate.

Step 4: Generate Fixes

For each classified discrepancy, generate the actual file changes. Do NOT write files yet — collect all proposed changes first.

For calculation_error:

1. Identify which reference file or SKILL.md section contains the incorrect formula.
   • Use Grep to find the relevant calculation step in skills/tax-calculate/SKILL.md.
   • Check skills/tax-calculate/references/credits-2025.md for credit phase-out rules.
   • Check skills/tax-calculate/references/brackets-2025.md for bracket amounts.
2. Draft the corrected rule. Example:
   • Wrong: CTC = number_of_qualifying_children * 2000 (no phase-out applied)
   • Fixed: CTC = number_of_qualifying_children * 2000, reduced by $50 for each $1,000 of MAGI above $400,000 (MFJ) / $200,000 (other)
3. Prepare an Edit operation for the relevant file.

For missing_data:

1. Identify which section of the tax-start interview would logically include the missing question.
2. Draft the new question(s) to add. Example:
   • Missing: Estimated tax payments (Line 26)
   • New question: "Did you make any estimated tax payments during the year? If so, enter the total amount paid."
3. Prepare an Edit operation for skills/tax-start/SKILL.md.

For missing_domain:

1. Identify the income/deduction type.
2. Draft a new skill scaffold following the domain skill template from the spec:

   ---
   name: tax-domain-{name}
   description: [domain] tax rules. Invoked by core skills when [trigger].
   allowed-tools: [Read, Write, Edit, AskUserQuestion]
   ---
   ## Trigger Conditions
   ## Interview Questions
   ## Data Schema
   ## Calculation Rules
   ## References
   

3. Prepare Write operations for skills/tax-domain-{name}/SKILL.md and an empty skills/tax-domain-{name}/references/ placeholder.

Step 5: Show Proposed Changes to User

Before writing any files, present all proposed changes for review:

Proposed fixes for 3 discrepancies:
============================================================

1. [calculation_error] Line 19 — Child Tax Credit
   File: skills/tax-calculate/references/credits-2025.md
   Change: Add MAGI phase-out rule for CTC above $200,000/$400,000

2. [missing_data] Line 26 — Estimated Tax Payments
   File: skills/tax-start/SKILL.md
   Change: Add question about estimated quarterly tax payments

3. [missing_domain] Line 17 — Rental Income (Schedule E)
   New file: skills/tax-domain-rental/SKILL.md
   Change: New domain skill scaffold for rental property income

============================================================
Proceed with these changes? (yes/no/review <n>)

• If the user says yes, proceed to Step 6.
• If the user says no, stop and do not write anything.
• If the user says review 2, show the full diff for change #2 and ask again.
• Allow the user to exclude specific changes before proceeding.

Step 6: Apply Approved Fixes

Apply only the changes the user approved. Use Edit for modifications to existing files, Write for new files.

After applying changes, run a quick sanity check:

ls skills/tax-calculate/references/
ls skills/tax-start/SKILL.md

Step 7: Update Evolution Log

Append an entry to data/evolution-log.json:

{
  "date": "2026-04-02",
  "comparisonFile": "data/comparison-2025.json",
  "lines": [
    {
      "line": "19",
      "description": "Child tax credit",
      "deltaDirection": "under",
      "deltaPercent": 17.6,
      "classification": "calculation_error",
      "resolved": true,
      "fixFile": "skills/tax-calculate/references/credits-2025.md"
    }
  ]
}

Fields to include per line entry:

• line — 1040 line number (string, e.g., "19")
• description — human-readable description
• deltaDirection — "over" (our value higher) or "under" (our value lower)
• deltaPercent — absolute percentage difference, e.g., 17.6 for 17.6% off
• classification — one of calculation_error, missing_data, missing_domain
• resolved — true if a fix was applied this run, false if skipped
• fixFile — the file that was changed (if resolved), else null

Do NOT store actual dollar amounts in the log.

Step 8: Invoke tax-publish

If at least one fix was applied, tell the user:

"Fixes applied. Ready to submit as a GitHub PR. Invoking /tax-publish..."

Then invoke /tax-publish evolve-{year}-fixes.

If no fixes were applied (user declined all), say:

"No changes were applied. You can run /tax-evolve again after reviewing the comparison data."

Error Handling

• If data/comparison-{year}.json is missing or has no mismatch entries, tell the user and stop.
• If data/evolution-log.json is corrupted, create a fresh one with {"entries": []} and warn the user.
• If a referenced skill file does not exist (e.g., skills/tax-calculate/references/credits-2025.md), tell the user which file needs to be created first and stop.
• If the user declines to proceed in Step 5, do not write any files.