flywheel-compound-refresh

Consolidate, replace, or archive stale and duplicate entries in a docs/solutions learning store.

3|1|Updated Apr 15, 2026
One-click install
npx skills add https://github.com/burningportra/agent-flywheel-plugin --skill flywheel-compound-refresh-burningportra
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: flywheel-compound-refresh
Source: https://github.com/burningportra/agent-flywheel-plugin/tree/main/skills/flywheel-compound-refresh
Command: npx skills add https://github.com/burningportra/agent-flywheel-plugin --skill flywheel-compound-refresh-burningportra

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? A growing docs/solutions/ knowledge store accumulates contradictory lessons over time — an old "use sync exec" note sits next to a new "use async exec" note with nothing flagging the drift. This Skill runs a scoring-based refresh sweep that detects stale, duplicate, and overlapping learnings and archives the losers without ever deleting anything. ## Core Features & Use Cases - 5-vector overlap scoring: Groups solution docs by (problem_type, component) and classifies each group as Keep, Update, Consolidate, Replace, or Delete using the flywheel_memory MCP tool's refresh_learnings operation. - Rename-aware staleness detection: Runs git log --follow before classifying a component as stale, so renamed files are not falsely flagged for deletion. - Archive-only mutation with confirmation gates: Pruned entries move to docs/solutions/_archive/ via git mv (preserving history), and every Delete classification requires its own explicit AskUserQuestion confirmation. - Use Case: After months of compound-engineering work, your docs/solutions/ folder has 80 entries with contradictory advice for rewritten components. Run the sweep to get a grouped report, confirm the Consolidate/Replace actions, and archive 15 stale entries while keeping the fresh primaries. ## Quick Start Ask the agent to run a compound refresh sweep over the docs/solutions learning store and report which entries should be consolidated, replaced, or archived.

Frequently Asked Questions about flywheel-compound-refresh

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I clean up duplicate or stale solution documents in a docs/solutions folder?

Run a refresh sweep that groups entries by problem_type and component, scores overlap across five dimensions, and classifies each group as Keep, Update, Consolidate, Replace, or Delete. Approved entries are moved to docs/solutions/_archive/ with git mv so history is preserved.

How does the refresh sweep avoid deleting learnings that are still relevant?

It never deletes files — everything pruned is archived via git mv into docs/solutions/_archive/. Delete classifications additionally require an individual AskUserQuestion confirmation per entry, and the first run is read-only, rendering a report before any mutation.

Does the sweep handle renamed components when detecting stale entries?

Yes. Before classifying anything as stale, it runs git log --follow on the component path to detect renames. A renamed file is not treated as staleness, and without this staleProbe callback the algorithm refuses to emit Delete or Replace decisions.

What happens if the flywheel_memory MCP tool is unavailable?

The sweep falls back to invoking the algorithm directly via tsx mcp-server/src/refresh-learnings.ts from a small wrapper. The MCP path is preferred so error envelopes flow through the standard hint channel.

What are the limitations of the learning-store refresh sweep?

It requires an existing docs/solutions/ corpus with parseable frontmatter; malformed files are flagged as Unparseable for manual review. It also will not act on Delete classifications without explicit per-entry confirmation, and thresholds should not be lowered without an audit trail.