dancing-bear: Skill for Claude Code

Improve Decompose Sweep

Post-run calibration for workflows/code/decompose-sweep.yaml. Reads run telemetry, identifies patterns, proposes and applies data-backed changes, then validates and logs the calibration.

When to Use

• After a decompose-sweep workflow run completes
• User says "improve decompose", "calibrate decompose", or "decompose retrospective"
• Recurring calibration pass (monthly, or after 3+ runs)

Step 1: Collect Telemetry

Run telemetry files land in the workflow's workspace_dir. The workflow engine resolves {work_dir} to <cwd>/out/ (see src/workflow/cli.py), so sweep runs write to:

out/decompose-sweep/run-metrics.json
out/decompose-sweep/run-<timestamp>/run-metrics.json

# Find all sweep run telemetry files
find out/decompose-sweep -name "run-metrics.json" | sort

# Read the most recent
find out/decompose-sweep -name "run-metrics.json" | sort | tail -1 | xargs cat

Fallback — if no telemetry files exist (e.g., run was aborted, workspace cleaned), parse agent summary messages from conversation context and note the source in the calibration log entry.

Step 2: Analyze Patterns

From the collected metrics, compute:

Defer rate by reason How often each defer heuristic triggered. High rates indicate a threshold is too aggressive.

Reason	Count	% of candidates
file_not_large_enough	N	N%
has_open_pr	N	N%
recently_modified	N	N%
imports_too_tangled	N	N%

Overshoot rate Files that were split but output modules still exceed the target line count. Indicates the plan-splits stage is over-estimating how cleanly a file divides.

Test failure domains Which domains produced failing tests post-split. Cluster by root cause:

• Import errors (re-export facade incomplete)
• Missing shared fixtures (not extracted to shared submodule)
• Circular dependencies introduced

Strategy accuracy

Strategy	Used	Met target	Avg lines saved	Avg output max
extract-submodules	N	N%	N	N
extract-package	N	N%	N	N
split-tests-by-class	N	N%	N	N
extract-helpers	N	N%	N	N