Skip to content

Latest commit

 

History

History
35 lines (27 loc) · 2.29 KB

File metadata and controls

35 lines (27 loc) · 2.29 KB

Knowledge Quality Lifecycle

The platform runs automated maintenance (see worker/config/sidekiq.yml); Claude Code participates in the quality loop. Full MCP tool list: ../../reference/auto/mcp-tools.md.

Automated (background jobs)

Job Schedule Effect
Compound learning decay 3:45 AM daily importance_score decays on stale learnings
Memory consolidation 4:00 AM daily STM→long-term (access≥3), dedup (similarity≥0.92)
Rot detection 4:00 AM daily Auto-archives context entries (staleness≥0.9)
Skill lifecycle daily/weekly/monthly Conflict scan, stale decay, re-embedding, gap detection
Shared knowledge maintenance daily Import from learnings, recalc quality, audit stale
Trajectory analysis 5:00 AM daily Execution-trajectory improvement recommendations
Intervention policy tuning weekly Analyze approval patterns, suggest policy adjustments

Manual (Claude Code responsibilities)

Trigger Tool
Solved a non-trivial bug create_learning (discovery / failure_mode)
Established/confirmed a pattern create_learning (pattern / best_practice)
Documented a procedure create_knowledge (procedure)
Found entity relationships extract_to_knowledge_graph
Implemented a reusable capability create_skill
Used a learning successfully / wrongly reinforce_learning / dispute_learning
Found two conflicting learnings resolve_contradiction
Read useful / outdated shared knowledge rate_knowledge (4-5 / 1-2 + corrected create_knowledge)
Encountering a bug query_learnings first — reinforce if found, fix + create_learning if not

Skip for trivial fixes (typos, renames, formatting), speculative/unverified analysis, or knowledge that already exists.

Tool Evolution

platform.* tools live in server/app/services/ai/tools/platform_api_tool_registry.rb. After adding/modifying, run rails mcp:generate_tool_catalog; rails mcp:sync_docs regenerates the broader fallback docs. For new tools, also add the name to the catalog and create a pattern learning. For deprecations, add a deprecation note in the action definition + a best_practice learning pointing at the replacement.