Troubleshooting

When Rules Are Triggered

Rules run automatically during family syncs or can be triggered manually. Understanding when rules execute helps troubleshoot categorization issues.

Manual trigger (user-initiated)

Rules can be manually triggered in the following ways:

• Clicking “Re-apply” on an individual rule in /settings/rules
• Clicking “Apply All” in /settings/rules
• Running the rake task: bin/rails rules:apply_all[family_id]

During family sync (only active rules)

Rules with active: true run automatically when a family sync occurs (Family::Syncer.perform_sync):

Scheduled syncs:

• SyncAllJob - runs daily at 2:22 AM for all families
• SyncHourlyJob - runs every hour (for items that opt-in to hourly syncing)

Triggered syncs:

• Provider webhooks (Plaid, etc.) - triggers sync which eventually propagates to family
• Manual sync button on accounts page
• After CSV imports complete (Import model calls family.sync_later)

When rules do NOT run

Rules are not triggered in these scenarios:

• Manual transaction creation - Only triggers an Account sync (not Family sync)
• Manual transaction editing - Same as above, only Account sync runs
• Inactive rules - Rules with active: false never run automatically, only via manual “Re-apply”

How do I see why a rule run failed?

Failures during async rule runs (including auto-categorize provider failures) are written to the debug log. To view them:

1. Go to Settings → Debug
2. Look for entries with category rules

Each failed run includes the error message and the rule that triggered it.

Category Rule Popup Behavior

When a user selects a category for a transaction, a popup may appear to create an automatic categorization rule. This popup is intentionally gated by several conditions:

When does the popup appear?

The popup only shows when all of these conditions are met:

1. User has not disabled rule prompts (rule_prompts_disabled is false)
2. User hasn’t dismissed the popup in the last 24 hours
3. The transaction category actually changed
4. No existing rule already sets this category for similar transactions
5. The transaction has a category assigned

Why isn't the popup showing?

Check these common reasons:

• Recently dismissed: Wait 24 hours after dismissing the popup
• Rule already exists: A matching rule may already be in place
• Prompts disabled: Check if rule_prompts_disabled is enabled for the user

Sure Doesn’t Work Over HTTPS

If Sure behaves incorrectly or fails when accessed over HTTPS, it’s usually due to missing SSL-related configuration between Nginx and Rails.

What causes this issue?

Sure relies on correct protocol headers to determine whether a request is secure.
When HTTPS is terminated at Nginx but not properly forwarded to the app, Rails may treat requests as HTTP.

How do I fix it?

Apply the following configuration changes:

1. Nginx

   Ensure HTTPS is forwarded to the upstream app:

   proxy_set_header X-Forwarded-Proto https;
   

2. docker-compose.yml

   In the x-rails-env: &rails_env section, set:

   RAILS_ASSUME_SSL: "true"
   

   This tells Rails to treat all requests as HTTPS.

Some pages break over HTTPS

Sure uses WebSockets for certain pages. If some pages fail to render correctly or break entirely, this is often caused by missing SSL or upgrade headers in your reverse proxy configuration.

What are the symptoms?

You may see pages partially load, fail completely, or errors similar to:

Failed to upgrade to WebSocket (REQUEST_METHOD: GET, HTTP_CONNECTION: close, HTTP_UPGRADE: )

This indicates that the WebSocket upgrade request is not being forwarded correctly.

How do I fix it?

Ensure your Nginx configuration includes the required WebSocket headers:

proxy_set_header Upgrade $http_upgrade;
proxy_set_header Connection "upgrade";

These headers allow Nginx to properly handle WebSocket upgrade requests over HTTPS.

Budget totals do not update after deleting transactions

If a budget still shows spending from a transaction that was already deleted, the most likely cause is a stale aggregate cache. Sure hard-deletes the transaction entry, so the deleted transaction is gone; the stale value can come from cached budget or income totals that were not invalidated by deleting an older entry.

What causes this issue?

Budget totals are derived from entry aggregation queries. Older versions of Sure used the most recent entry update timestamp as part of the cache key, which meant deleting an older transaction could leave the cache key unchanged.

How was it fixed?

The entry aggregate cache version now includes both:

• The current entry count
• The newest entry updated_at timestamp

Because the count changes when an entry is deleted, budget and income aggregate caches are invalidated even when the deleted transaction was not the most recently updated entry.

How can maintainers validate it?

Confirm the installed version includes the cache invalidation fix for Family#entries_cache_version, then re-check the affected budget after the next request or cache refresh. The regression test is:

bin/rails test test/models/family_test.rb

Chat fails immediately on a fresh family with a strict OpenAI-compatible provider

Some strict OpenAI-compatible providers (not OpenAI itself) reject chat requests with a schema validation error when a family has no accounts, categories, merchants, tags, or tickers yet.

Rule runs fail silently

If auto-categorization or merchant detection rules appear to do nothing, the background job may have failed without surfacing an error in the UI.

How do I see rule run failures?

Async rule run failures are now logged to the debug log. Go to Settings → Debug and look for entries from the rule_run category. Each failed run includes the error class, message, and the rule that triggered it.

You can also check worker logs directly:

docker compose logs worker

What causes rule runs to fail?

Common causes include:

• LLM provider errors — The AI provider returned an error or timed out. Check your OPENAI_ACCESS_TOKEN and provider connectivity.
• Auto-categorize provider failures — The provider returned an unexpected response format. These are now propagated and logged rather than silently swallowed.
• Worker not running — Rules run in the background via Sidekiq. Verify the worker container is healthy.

App crashes for a family with an invalid timezone

If a family’s timezone setting contains an unrecognized IANA zone name (for example, after a tzdata rename like Europe/Kiev → Europe/Kyiv), Sure now falls back to the application default timezone instead of crashing.

Chat fails immediately on a new account with a strict OpenAI-compatible provider

Some strict OpenAI-compatible providers (not OpenAI itself) reject chat requests with a schema validation error when a family has no tags, merchants, or categories yet. This happens because the assistant’s tool schemas include enum fields built from family data — an empty family produces enum: [], which is invalid JSON Schema.

What causes this?

Sure builds tool schemas for the assistant using family data (account names, categories, merchants, tags, tickers). A family with none of this data produces enum: [] for those fields, which is invalid JSON Schema. OpenAI tolerates it, but strict providers reject the entire request before the model runs. Family#timezone is a free-text field. If the stored value becomes stale — due to a tzdata rename, a database restore from an older backup, or a direct database edit — the app previously raised an ArgumentError on every request for that family, including the login page.

How is it handled now?

Sure validates the timezone value against ActiveSupport::TimeZone before using it. If the value is unrecognized, the app falls back to the default timezone and logs a warning to Settings → Debug under the timezone category. The log is debounced to once per day per family so it does not flood the debug log.

To fix the root cause, update the family’s timezone to a valid IANA zone name from Settings → Profile. Sure builds tool schemas dynamically from your family’s data (account names, categories, merchants, tags, tickers). When none of that data exists yet, the schema contains enum: [], which is invalid JSON Schema. OpenAI tolerates it, but strict OpenAI-compatible providers reject the entire request before the model runs.

How was it fixed?

Empty enum arrays are now pruned from tool schemas before the request is sent, falling back to a plain string type. This fix applies to both chat tool definitions and the /mcp endpoint’s tools/list.

If you are on an older version and cannot upgrade, add at least one account, category, or tag to the family — this populates the enum and avoids the empty-array case.

Chat shows “assistant not available” before the model finishes

Self-hosted users running a local model (Ollama, LM Studio, etc.) may see the chat fail with “assistant not available” even though the model is still generating a reply and tokens are being billed.

What causes this?

Sure uses a whole-turn watchdog (AI_RESPONSE_TIMEOUT, default 90 seconds) that starts when the message is queued. Custom OpenAI-compatible providers use a synchronous code path — nothing renders until the full reply is generated — so the watchdog fires before the model finishes.

Tool-using turns make this worse: each tool call adds another full model round. The total time the watchdog must cover is:

(1 + ASSISTANT_MAX_TOOL_CALL_ITERATIONS) × OPENAI_REQUEST_TIMEOUT
  + tool execution time + queue wait

With the defaults (ASSISTANT_MAX_TOOL_CALL_ITERATIONS=5, OPENAI_REQUEST_TIMEOUT=60), a worst-case turn can take up to 360 seconds of model time — far beyond the 90-second default.

How do I fix it?

The most effective approach is to lower ASSISTANT_MAX_TOOL_CALL_ITERATIONS (reducing the worst-case bound) and then size AI_RESPONSE_TIMEOUT using the formula above.

For a local Ollama setup:

OPENAI_REQUEST_TIMEOUT=300
ASSISTANT_MAX_TOOL_CALL_ITERATIONS=2
AI_RESPONSE_TIMEOUT=1200

You can also set Chat Response Timeout from Settings → Self-Hosting → AI Provider in the UI. The environment variable takes precedence.

See Chat response timeout for full sizing guidance.

Why is Sure not running auto-categorization and merchant detection on the same transactions again?

Sure caches AI-generated results to avoid redundant API calls and costs. Once a transaction has been processed by AI rules, it won’t be re-processed unless you explicitly reset the AI cache.

Why does Sure cache AI results?

When AI rules process transactions, Sure stores:

• Enrichment records - Which attributes were set by AI (category, merchant, etc.)
• Attribute locks - Prevents rules from re-processing already-handled transactions

This caching ensures:

• Transactions aren’t sent to the LLM repeatedly
• API costs are minimized
• Processing is faster on subsequent rule runs

How do I force Sure to re-process transactions?

To have AI rules re-process transactions, you need to reset the AI cache:

1. Go to Settings → Rules
2. Click the menu button (three dots)
3. Select Reset AI cache
4. Confirm the action

After resetting, the next time rules run (either manually or during a sync), AI will re-process all transactions.

Note: This will incur API costs if using a cloud provider.

The reset runs in the background. You can track its progress and see how many AI cache entries were removed in Settings → Debug, filtered by the ai_cache_reset category. Each run logs when it starts, how many entries were removed per scope, and any records that could not be cleared.

When should I reset the AI cache?

Common scenarios for resetting the cache:

• Switching LLM models - Different models may produce better categorizations
• After system updates - New versions may have improved prompts
• Fixing miscategorizations - When AI made systematic errors
• Testing - During development or evaluation of AI features

The AI cache is automatically cleared when you change the OpenAI model setting.

How do I verify the reset actually ran?

The reset runs as a background job. You can confirm it completed — and see a breakdown of how many AI cache entries were removed — in the debug log:

1. Go to Settings → Debug
2. Filter by category ai_cache_reset

You will see:

• An info entry when the reset is enqueued from the rules page
• An info entry when the job starts on a worker
• A completion entry with the number of enrichments deleted, broken down by scope
• warn entries for any individual records that could not be cleared
• An error entry if the job could not be enqueued at all

If you see the enqueue entry but no start entry, the job never reached a worker — check that your worker container is running.