From consensus-loop
Extracts learnings from conversations, audit history, and git logs; manages memories and cleans stale entries. Use after tracks, retrospectives, or maintenance.
npx claudepluginhub berrzebb/claude-plugins --plugin consensus-loopThis skill uses the workspace's default tool permissions.
Mine the current session for learnings and translate them into durable memories. Sessions generate massive amounts of information — your job is to find the fraction worth remembering, structure it properly, and persist it.
Searches, retrieves, and installs Agent Skills from prompts.chat registry using MCP tools like search_skills and get_skill. Activates for finding skills, browsing catalogs, or extending Claude.
Searches prompts.chat for AI prompt templates by keyword or category, retrieves by ID with variable handling, and improves prompts via AI. Use for discovering or enhancing prompts.
Guides MCP server integration in Claude Code plugins via .mcp.json or plugin.json configs for stdio, SSE, HTTP types, enabling external services as tools.
Mine the current session for learnings and translate them into durable memories. Sessions generate massive amounts of information — your job is to find the fraction worth remembering, structure it properly, and persist it.
| Signal | Mode | Phases | When |
|---|---|---|---|
| Casual question ("뭘 배웠지?", "learnings") | Quick | 1a → 3 → 6 | Simple summary, end-of-session |
| Maintenance request ("메모리 정리", "cleanup") | Full | 1~6 all | Memory audit, bulk operations |
| Track argument ("/retrospect OR") | Full | 1~6 all | Deep track analysis |
| During retrospective (③ step) | Full | 1~6 all | Post-audit protocol |
Mode is selected automatically — do not ask the user.
| Context | Detection | Approval Behavior |
|---|---|---|
| Interactive | Main session, user is responding | Present candidates → wait for approval → execute |
| Headless | Subagent, no human to respond | Auto-approve threshold-met candidates → defer ambiguous → execute → report |
In headless mode, NEVER ask questions or wait for input. Any prompt will block indefinitely. Auto-approve what meets threshold rules, defer the rest as ⏭️ deferred to orchestrator.
Read config: ${CLAUDE_PLUGIN_ROOT}/config.json → plugin.locale
Locate memory: use Glob to find **/memory/MEMORY.md. Read it to understand existing memories (avoid duplicates, check line count < 200).
Read references/gathering.md for detailed source collection methods.
Quick mode: conversation + git log only (skip audit_history). Full mode: conversation + audit_history + git log.
Read existing memory files → build topic index → mark candidates as "new" or "update".
Read references/candidates.md for memory format template, categories, and threshold rules.
4 categories: feedback (corrections + confirmations), project (work context), user (role/preferences), reference (external pointers).
Read references/execution.md for presentation format, approval flow, and integrity checks.
When invoked during post-audit retrospective (③ Memory cleanup step):
Only this skill and agents may read/write memories. Other skills (verify, merge, planner) must not access the memory directory directly. This prevents memory pollution from non-learning contexts.
| Actor | Memory Access |
|---|---|
consensus-loop:retrospect | Read + Write (primary owner) |
| Agents (orchestrator, implementer) | Read only (via auto-memory system) |
| Other skills (verify, merge, planner) | No access — request via orchestrator if needed |