From agents-flow
Context management protocol for rot reduction using editing, delegation, and summarization.
How this skill is triggered — by the user, by Claude, or both
Slash command
/agents-flow:compaction-protocolThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Used by `/build` to manage context rot. Read on-demand at compaction steps; do not preload.
Used by /build to manage context rot. Read on-demand at compaction steps; do not preload.
/compact last. Re-summarization compresses, but it also creates a new lower-fidelity anchor the model will over-attend to. Use only when conversation bulk (not tool output) is the source of pressure.Trigger on concept shifts, not percentages:
Use context editing to clear superseded tool outputs (reads after edits, old test runs, stale doc lookups). Reports 84% reduction vs. summarization; no paraphrase artifacts. Default tool — reach for /compact only when context editing can't fix the root pressure.
/compact — last resortWhen conversation bulk (not tool output) is the pressure, write the Keep list to .claude/NOTES.md first:
/compact
Keep: <architectural decisions>; <current task and files in scope>;
<unresolved bugs or failing tests>; <open questions from the issue>.
Drop: <rejected alternatives>; <tool outputs already acted on>;
<API docs already internalized>; <exploration that led nowhere>.
Keep what the model can't reconstruct from the issue or .claude/NOTES.md. Drop anything large and replayable.
After compaction, run: Summarize where we are and what the next step is.
Diff the summary against the Keep list in .claude/NOTES.md. If anything is missing, restate it explicitly before the next tool call. If hollow, abort and re-read the issue + .claude/NOTES.md.
/wrap-up./compact. Never skip steps.npx claudepluginhub misiekhardcore/agents-flow --plugin agents-flowGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.