Managing Context Size

This skill teaches you how to avoid context exhaustion when working with large Terra datasets, scattered workflows, and verbose Cromwell metadata.

Context Size Risks

These operations can exhaust your context window:

Safe Patterns

1. Always Use Summary Mode for Metadata

# SAFE: Returns ~1-2K tokens
get_job_metadata(workspace_namespace, workspace_name, submission_id, workflow_id)

# DANGEROUS: Can return 100K+ tokens
# (There is no "full" mode - the tool protects you)

Summary mode is the default. It returns structured, actionable information without the verbose raw metadata.

2. Get Log URLs Before Fetching Content

For scattered workflows (50+ shards), fetching all logs can quickly exhaust context.

# Step 1: Get URLs only (fast, small response)
get_workflow_logs(
    workspace_namespace, workspace_name, submission_id, workflow_id,
    fetch_content=False  # Default
)

# Step 2: Identify which specific tasks failed
get_job_metadata(..., mode="summary")

# Step 3: Fetch content only for failed tasks you need
# (Consider fetching logs one at a time for scattered workflows)

3. Check Table Sizes Before Fetching Entities

# Step 1: Check row counts
get_workspace_data_tables(workspace_namespace, workspace_name)
# Response: {"tables": [{"name": "sample", "count": 5000}, ...]}

# Step 2: If count > 100, reconsider whether you need all entities
# Often, specific entity names from workflow inputs are sufficient

4. Limit Workflow Count in Submission Status

# Default: Returns first 10 workflows (manageable)
get_submission_status(workspace_namespace, workspace_name, submission_id)

# For large submissions, the default limit protects you
# Only use max_workflows=0 if you truly need all workflow details

5. Exclude Input Resolutions (Default Behavior)

# Default: Excludes inputResolutions (94% size reduction)
get_submission_status(workspace_namespace, workspace_name, submission_id)

# Only include inputs if you need to debug input values
get_submission_status(..., include_inputs=True)  # Much larger response

Progressive Disclosure Strategy

When investigating workflows, follow this pattern:

1. •

   Start broad, go narrow:

   • •Begin with summaries and overviews
   • •Drill into specific items only when needed
2. •

   Identify targets first:

   • •Get workflow IDs before fetching details
   • •Get failed task names before fetching logs
   • •Get entity names from workflow inputs, not full table dumps
3. •

   Fetch content last:

   • •URLs and summaries first
   • •Actual content only for items you're debugging

Context Budget Estimation

Rough token estimates for planning:

When Large Responses Are Needed

Sometimes you genuinely need large data. In these cases:

1. •Process incrementally: Request data in batches
2. •Extract specific fields: Use get_job_metadata(..., mode="extract", field_path="...")
3. •Work offline: For very large datasets, consider extracting to files and analyzing externally

Anti-Patterns to Avoid

Quick Reference: Safe Defaults

All these tools have context-safe defaults:

• •get_job_metadata: Summary mode by default
• •get_workflow_logs: URLs only by default (fetch_content=False)
• •get_submission_status: 10 workflows max, no inputs by default
• •Log truncation: First 5K + last 20K chars when fetching content