---

name: telemetry-insights description: Analyze AI coding session telemetry for usage patterns, token efficiency, and workflow optimization. Supports current session or historical time periods.

Telemetry Insights

Scope Detection

User Request	Scope
"How efficient am I?"	Current session
"Analyze this conversation"	Current session
"this week" / "last 7 days"	7 days
"today" / "last 24 hours"	1 day
"last month"	30 days
"productivity" / "project" / "traces"	7 days (default historical)

---

Data Access

Database: ~/.blueplane/telemetry.db

Schema Reference

Claude Code (claude_raw_traces table):

• Workspace filter: WHERE cwd LIKE '/path/to/workspace%' (NOT workspace_hash)
• Token fields: input_tokens, cache_creation_input_tokens, cache_read_input_tokens, output_tokens
• Model field: message_model (e.g., "claude-sonnet-4-5-20250929")
• Role field: message_role ('user' | 'assistant')
• Branch field: git_branch

Cursor (cursor_sessions + cursor_raw_traces tables):

• Step 1: Get workspace_hash from cursor_sessions WHERE workspace_path LIKE '/path/to/workspace%'
• Step 2: Get composer_ids from cursor_raw_traces WHERE event_type='composer' AND workspace_hash=?
• Step 3: Filter bubbles by extracting composer_id from item_key: "bubbleId:{composerId}:{bubbleId}"
  • Bubbles have empty workspace_hash in global storage
  • Must join via composer_id from Step 2
• Token field: token_count_up_until_here (cumulative context, NOT per-message tokens)
• Message type: message_type (0=user, 1=assistant, but NULL for ~all events - use heuristics)
• Model: ❌ Not stored in database

Tool Usage Analysis (Claude Code only)

Data source: Decompress event_data blob from claude_raw_traces

import zlib, json
decompressed = zlib.decompress(row['event_data'])
event = json.loads(decompressed)
tools = event['payload']['entry_data']['message']['content']
# Filter: item['type'] == 'tool_use'
# Extract: item['name'], item['input']['file_path'], item['input']['command']

Tool operation churn (detect both productive and unproductive patterns):

Negative patterns (wasted effort):

• Files written then deleted without commit
• Same file edited multiple times in short succession (trial-and-error)
• Bash commands repeated with similar patterns (debugging loops)
• Write operations for files later removed (unwanted artifacts)

Positive patterns (intentional iteration):

• Progressive refinement: Write → Read → Edit (deliberate improvement)
• Test-driven flow: Write test → Run → Edit code → Run (TDD cycle)
• Exploration: Multiple Read/Grep before Write (research-based development)
• Read-before-edit ratio > 0.8 (careful, informed changes)

Workflow efficiency signals:

• Tool distribution (Bash, Write, Edit, Read, Task usage)
• Read-before-edit ratio (higher = more careful)
• File touch count (edits per unique file path)
• Time between tool uses (rapid = reactive, spaced = deliberate)

Cross-reference with git to classify churn:

• Compare Write/Edit file paths against final committed files
• Flag operations on files not in git diff --name-only
• Distinguish exploration (positive) from mistakes (negative)

---

Raw Metrics

A. Flow & Structure

• analysis_scope, session_duration, active_exchanges, total_ai_responses
• context_switch_count, prompt_timing_buckets, session_summaries

B. Prompting Patterns

• prompt_count, average_prompt_length, median_prompt_length
• prompt_complexity_score (low/medium/high), reprompt_loop_count

C. Token Economics

• total_input_tokens, total_output_tokens, input_output_ratio
• estimated_cost_usd: (input/1M × $3) + (output/1M × $15)

D. Patch Behavior

• patch_count, total_lines_added, total_lines_removed, add_remove_ratio

E. Behavioral Signals

• delegation_style: high-level vs step-by-step indicators
• positive_feedback / negative_feedback counts

F. Capability Usage

• capabilities_invoked, agentic_mode_usage

G. Model Strategy (Claude Code only)

• per_model_usage, per_model_tokens, dominant_model

H. Temporal Productivity

• tokens_per_day, tokens_per_hour_utc, prompts_per_minute_by_session

I. Tool Usage & Workflow Patterns (Claude Code only)

• tool_distribution: counts by tool name (Bash, Write, Edit, Read, etc.)
• read_before_edit_ratio: Read operations / Edit operations
• files_created_not_committed: count of Write file_paths not in git
• file_touch_count: edits per unique file_path
• high_churn_files: files edited 3+ times
• pattern_classification: productive_iteration vs trial_and_error counts

---

Derived Insights

1. Effort_vs_Progress_Score (0-1): lines_added / total_tokens
2. Context_Sufficiency_Index (0-1): 1 - (corrections / prompts)
3. AI_Utilization_Quality_Score (0-1): weighted agentic + efficiency + success
4. Predicted_Task_Difficulty: easy/moderate/hard based on query rate, switches
5. AI_vs_Human_Burden_Ratio: ai_output_chars / user_input_chars
6. Persistence_vs_Abandonment: reprompt loops vs topic abandonment
7. Patch_Efficiency_Curve: clean (ratio>3) vs thrashy, lines_per_prompt
8. Intent_Shift_Map: task type transitions count
9. Prompt_Quality_vs_Result: success rate by prompt length
10. Confidence_Trajectory: improving/declining/mixed
11. Stuckness_Prediction: risk_level based on correction rate, loops
12. Prompt_Pacing_Profile: rapid_iterator/balanced/deliberate
13. Model_Strategy_Assessment: single_model/tiered_models, cost_awareness
14. Peak_Performance_Windows: top hours UTC
15. Session_Focus_Profile: short_bursts/long_deep_work/mixed
16. Workflow_Quality_Score (0-1): based on read-before-edit ratio, productive vs wasteful churn
17. Development_Discipline: careful (high read-first) vs reactive (low read-first, high trial-and-error)

---

Output Format

1. RAW_METRICS: { JSON with metrics A-I above }

2. DERIVED_INSIGHTS: { JSON with insights 1-17 above }

3. SESSION_SUMMARY: 6-10 sentences covering:
   - Analysis scope (first sentence)
   - Duration and activity level
   - Workflow patterns
   - Token efficiency
   - Key recommendations

---

Interpretation Guidelines

Input:Output Ratio	Assessment
< 10:1	Excellent
10-25:1	Normal
25-50:1	Context-heavy
> 50:1	Inefficient

Recommendation Triggers

Signal	Recommendation
0% agentic usage	Enable agentic mode for multi-file tasks
>50:1 ratio frequently	Start new conversations for simple queries
High negative feedback	Provide more context upfront
High context switches	Consider task batching

---

Limitations

• Model info unavailable for Cursor (platform limitation)
• Token counts depend on capture completeness
• Task type inference is keyword-based (heuristic)
• Cost estimates based on Claude 3.5 Sonnet pricing