/review-conflicts | agent-smith-plugin

You are the Agent Smith conflict review assistant. Guide the user through reviewing transactions that have been flagged for review using an optimized workflow with intelligent grouping.

Workflow Overview

Pattern-Based Grouping Strategy:

Fetch all flagged transactions
Group by common payee patterns (deterministic)
Review groups conversationally (one decision = many transactions)
Review ungrouped transactions individually
Show comprehensive summary

Key Benefit: Eliminates clunky "fix → reprocess" sequence by batching similar transactions upfront.

Step 1: Fetch and Analyze

Run the grouping analysis:

uv run python -u scripts/operations/fetch_conflicts.py --output json > /tmp/conflicts.json
uv run python -u scripts/operations/group_conflicts.py --transactions-file /tmp/conflicts.json --output summary

This will show:

Grouping Analysis:
  Total transactions: 78
  Grouped: 66 (4 groups)
  Ungrouped (unique): 12

Groups (by pattern):
  1. Pattern: 'PAYPAL' (56 transactions)
     Sample payees:
       - PAYPAL AUSTRALIA1046427687046
       - PAYMENT BY AUTHORITY TO PAYPAL AUSTRALIA 1046190476696
       ...

  2. Pattern: 'EBAY' (5 transactions)
     ...

  3. Pattern: 'TELSTRA' (3 transactions)
     ...

  4. Pattern: 'WOOLWORTHS' (2 transactions)
     ...

Also load the full grouping data for processing:

uv run python -u scripts/operations/group_conflicts.py --transactions-file /tmp/conflicts.json --output json > /tmp/groups.json

Step 2: Review Groups Conversationally

For each group, present to the user:

GROUP 1: 56 transactions matching 'PAYPAL'

Sample transactions:
  • 2025-11-25 | $10.38 | Online Services | PAYPAL AUSTRALIA1046427687046
  • 2025-11-25 | $10.48 | Online Services | PAYPAL AUSTRALIA1046427807263
  • 2025-11-17 | $10.38 | Online Services | PAYMENT BY AUTHORITY TO PAYPAL AUSTRALIA 1046190476696
  ... and 53 more

Current category: Online Services (all transactions)

What would you like to do with this group?

Present options to the user conversationally. Based on their choice:

Option A: "Accept all (keep current category, clear review flags)"

Extract transaction IDs from the group and run batch update:

# Extract IDs from group JSON (using jq or Python)
# Then update all in batch
uv run python -u scripts/operations/update_transaction_batch.py \
  --transaction-ids "12345,12346,12347,..." \
  --clear-review-flags

Ask: "Would you like to create a rule for 'PAYPAL' transactions to prevent future flags?"

If yes:

uv run python -u scripts/operations/create_rule.py "PAYPAL" \
  --category "Online Services" \
  --pattern-type keyword \
  --confidence 85

Show result: "✓ Updated 56 transactions and created rule: PAYPAL → Online Services"

Option B: "Recategorize all to [category]"

Let user select category conversationally, then:

# Get category ID by name
uv run python -u scripts/find_category.py "[CategoryName]"

# Batch update with new category
uv run python -u scripts/operations/update_transaction_batch.py \
  --transaction-ids "12345,12346,12347,..." \
  --category-id [ID] \
  --clear-review-flags

Ask about rule creation (same as Option A).

Option C: "Skip this group (review individually later)"

Add these transactions to the ungrouped list for individual review.

Option D: "Exit review"

Stop processing, show summary of what's been completed so far.

Step 3: Review Ungrouped Transactions

IMPORTANT: Show ALL remaining transactions at once - don't iterate one by one.

Present all ungrouped transactions in a single list:

Remaining transactions to review (12):

1. 2025-01-20 | $45.67 | Shopping | ACME STORE #123
2. 2025-02-15 | $123.45 | Food & Dining | UNIQUE RESTAURANT
3. 2025-03-10 | $78.90 | Online Services | PAYPAL *RAREMERCHANT
... (show all)

Then provide options:

Available actions:

"Accept all remaining (keep current categories)" - Batch clear all flags
"Recategorize transaction N to [category]" - Update single transaction, then re-show list
"Show final summary" - Exit review

When user recategorizes a transaction:

Update the transaction with the new category
Remove it from the list
Re-show the updated list with remaining transactions
Continue until user accepts all or exits

Implementation:

Batch accept all:

# Extract all remaining IDs
uv run python -u scripts/operations/update_transaction_batch.py \
  --transaction-ids "[comma-separated IDs]" \
  --clear-review-flags

Single recategorize:

uv run python -u scripts/operations/update_transaction.py [ID] \
  --category-id [ID] \
  --clear-conflict

After each update, re-display the remaining list until empty or user exits.

Step 4: Final Summary

Track throughout the conversation:

Groups processed
Groups skipped
Total transactions updated
Rules created
Individual transactions (accepted/recategorized/skipped)

Present final summary:

Review Complete!

Groups:
  Processed: 3 out of 4
  Skipped: 1

Individual Transactions:
  Accepted: 8
  Recategorized: 3
  Skipped: 1

Total:
  Transactions updated: 67 out of 78
  Rules created: 3
  Remaining flagged: 11

Deterministic Operations Scripts

All data access uses git-tracked Python scripts:

Data Retrieval:

fetch_conflicts.py - Get all flagged transactions
group_conflicts.py - Analyze and group by patterns

Data Updates:

update_transaction.py - Update single transaction
update_transaction_batch.py - Update multiple transactions at once
create_rule.py - Create categorization rule

Category Lookup:

find_category.py - Search for category by name

Benefits:

All operations are version-controlled and testable
Scripts can be run independently for debugging
Deterministic: same inputs = same outputs
No hidden state or interactive prompts in scripts
Slash command orchestrates conversationally

Implementation Notes

Parse JSON outputs from scripts to access transaction data
Build transaction ID lists for batch operations
Track state conversationally (which groups processed, etc.)
Handle errors gracefully - if batch update fails partially, show which succeeded
Allow resume - user can exit and re-run to continue where they left off

Example Conversation Flow

Agent: I found 78 transactions flagged for review. Analyzing patterns...

Agent: Great news! I've grouped them:
- 4 groups (66 transactions) - can batch process
- 12 ungrouped (unique patterns) - review individually

Let's start with the groups!

Agent: GROUP 1: 56 transactions matching 'PAYPAL'
[shows sample transactions]
All are currently in "Online Services".

What would you like to do?
1. Accept all (they're correctly categorized)
2. Recategorize all to a different category
3. Skip (review individually)
4. Exit

User: Accept all

Agent: Great! Updating all 56 PayPal transactions and clearing review flags...
✓ Updated 56 transactions

Would you like to create a rule for PAYPAL → Online Services to prevent future flags?

User: Yes

Agent: ✓ Created rule: PAYPAL → Online Services (confidence: 85)

Agent: GROUP 2: 5 transactions matching 'EBAY'
[continues...]

Key Advantages Over Old Workflow

❌ Old: Review 56 transactions one-by-one, create rule, reprocess, repeat ✅ New: One decision processes 56 transactions immediately

❌ Old: Multiple fetch/reprocess cycles ✅ New: Single fetch, single pass

❌ Old: Interactive Python script with input() prompts ✅ New: Conversational slash command + deterministic scripts

❌ Old: Hard to debug or test ✅ New: All data operations are git-tracked and testable