skill-timer

Skill Timer

Timing instrumentation for skill execution measurement with microsecond precision and anomaly detection.

Purpose

Enable comprehensive performance measurement and analysis:

• Wall-clock duration: microsecond precision from start to end.
• Checkpoints: intermediate timing points for bottleneck analysis.
• Token tracking: input/output token counts with cost estimation.
• Anomaly detection: real-time detection of shortcuts and timeouts.
• Baseline comparison: compare against historical averages.
• Cross-analysis: performance across models, agents, and modes.

Use this skill when

• Measuring skill execution duration.
• Comparing performance across models or agents.
• Identifying bottlenecks with checkpoints.
• Tracking token consumption and costs.
• Detecting potential agent shortcuts (suspiciously fast execution).
• Building performance baselines for CI/CD.
• Analyzing historical timing trends.

Do not use for: skills with sub-5-second expected duration, pure syntax tests, or skills that produce no file output (timing metadata needs a file to embed in).

Quick Start

# Measure a skill execution with checkpoints
Use the skill-timer skill.

skill_name: rule-reviewer
target_file: rules/200-python-core.md
model: claude-sonnet-4-6
review_mode: FULL
timing_enabled: true

Output: Timing metadata embedded in output file with duration, checkpoints, token costs, and baseline comparison.

Inputs

Required (timing-start)

• skill_name: string: name of the skill being timed.
• target_file: path: target file path.
• model: string: model slug (e.g., claude-sonnet-4-6).

Required (timing-end)

• run_id: hex string (16 chars): from timing-start output.
• output_file: path: output file for metadata embedding.
• skill_name: string: skill name (for recovery if run_id lost).

Optional (timing-end)

• input_tokens: integer (default: none).
• output_tokens: integer (default: none).
• format: string (default: human): human, json, markdown, quiet.
• dimension_timings: JSON array (default: none): per-dimension timing data.
• auto_dimension_timings: flag (default: off): derive dimension_timings automatically from dim_<name>_start / dim_<name>_end checkpoint pairs. Explicit dimension_timings wins if both are supplied (with WARNING).
• review_mode: string (default: FULL).