OmniFocus MCP Performance Patterns

AppleScript round-trips via osascript are the dominant performance cost. Every optimization in this project reduces the number of subprocess calls.

Known Operation Timings

Timeout defaults: 60s standard, 300s maximum. Configurable per operation.

Pattern 1: Eliminate N+1 Queries

The most impactful optimization in this project. Before the fix, filtering projects by their tasks required one AppleScript call per project:

# BAD: N+1 pattern (33 projects = 33 calls = 7.6s)
for project in projects:
    tasks = get_tasks(project_id=project.id)  # 230ms each!
    if meets_filter(tasks):
        results.append(project)

# GOOD: Batch pattern (33 projects = 1 call = 0.9s)
all_tasks_by_project = _get_tasks_batch_for_filtering()  # Single call
for project in projects:
    tasks = all_tasks_by_project.get(project.id, [])
    if meets_filter(tasks):
        results.append(project)

Implementation: _get_tasks_batch_for_filtering() fetches minimal data (id, projectId, dueDate) for ALL tasks in one AppleScript call, returning a {projectId: [tasks]} dictionary.

Result: 8.5x faster (7.6s -> 0.9s for 33 projects). Scales linearly — 500 projects would go from 115s to 13.5s.

Apply this pattern whenever you need to cross project boundaries for filtering or aggregation.

Pattern 2: Conditional Filter-First Architecture

get_tasks() has a critical optimization: it detects whether "selective filters" are active and changes its execution order accordingly.

Selective filters: flagged_only, overdue, dropped_only, blocked_only, next_only, relative date ranges

IF selective filters active:
    1. Apply filters in AppleScript (cheap — just checking properties)
    2. Extract full properties only for matching tasks (expensive)
    Result: 82-88% faster for selective queries

IF no selective filters:
    1. Extract full properties for all tasks
    2. Apply post-processing filters in Python
    Result: Avoids overhead of empty filter checks

Why this matters: Extracting 27 properties per task is expensive. If you're only looking for 5 overdue tasks out of 188, filtering first means you only extract properties for 5 tasks instead of 188. This is a 19x speedup for selective queries.

When adding new filter parameters: Determine whether the filter is "selective" (likely to match a small subset) or "broad" (likely to match most tasks). Selective filters should go into the filter-first path.

Pattern 3: Minimize Data Per Call

When you don't need all properties, fetch only what you need:

-- GOOD: Fetch only what's needed for filtering
set taskData to {id of t, project of t, due date of t}

-- BAD: Fetch everything when you only need to filter
set taskData to {id, name, note, status, flagged, due date, defer date, ...}

_get_tasks_batch_for_filtering() demonstrates this — it fetches only id, projectId, and dueDate because that's all the project filtering logic needs.

Pattern 4: Python Post-Processing vs AppleScript Filtering

Some filters are better applied in Python after fetching data:

• Tag filtering (OR/NOT modes): Done in Python via _filter_tasks_by_tags() because AppleScript tag matching is limited
• Tag filtering (AND mode): Done in AppleScript for efficiency, then verified in Python for consistency
• Date range filtering: Simpler to compare ISO dates in Python than to build date comparison logic in AppleScript

Rule of thumb: Use AppleScript filtering when it eliminates data transfer (fewer tasks to serialize). Use Python filtering when the logic is complex or when the data is already fetched.

Anti-Patterns

Repeated Safety Checks

Each destructive operation calls _verify_database_safety() which runs an AppleScript to fetch the database name (~100ms). For batch operations that call multiple destructive methods, this check runs multiple times unnecessarily. Consider caching the result for the session.

Large AppleScript Blocks

Very large AppleScript strings (1000+ lines) can be slow to parse. If a script grows unwieldy, consider splitting into two calls: one for data retrieval, one for data processing, with Python bridging them.

Unnecessary Property Extraction

Don't add properties to the "always extract" list unless they're commonly needed. Each additional property adds serialization overhead for every task in every query.

Profiling Infrastructure

The project has performance profiling tests in tests/test_profile_performance.py and test data generation scripts in scripts/local/setup_profiling_data.sh.

To profile a change:

1. Switch to test database: ./scripts/local/switch_to_test_db.sh
2. Generate test data: ./scripts/local/setup_profiling_data.sh
3. Run profiling: pytest tests/test_profile_performance.py -v
4. Switch back: ./scripts/local/switch_to_prod_db.sh

Detailed Analysis

For the complete performance analysis including Phase 1/Phase 2 optimization results and empirical data, see:

• docs/performance/GET_TASKS_PERFORMANCE_ANALYSIS.md
• docs/performance/PERFORMANCE_OPTIMIZATION_SUMMARY.md