#What it does
Claudear analyses the outcomes of every fix: PR diffs, review comments, Q&A answers, and execution logs. Over time it builds per-repo knowledge that makes future fixes faster and more accurate.
#Knowledge sources
The learning system draws from multiple signals:
- PR diffs — analyses what changed in merged PRs to learn patterns and conventions (
diff_analysis). - Q&A answers — when the same question is answered consistently, the answer is promoted to standing instructions (
qa_promotion). - Review feedback — classifies reviewer comments to detect recurring patterns like "always add tests" or "use our custom error type" (
review_classification). - Execution logs — extracts learnings from Claude's execution logs to understand what approaches work (
auto_extract_learnings). - Strategy fingerprinting — tracks how Claude approaches different classes of fixes (
strategy_fingerprinting). - Quality scoring — scores fix quality based on merge velocity and review feedback (
quality_scoring).
#Auto AGENT.md
When auto_agent_md = true, Claudear auto-generates an AGENT.md file in each repo from accumulated knowledge. This file is included in Claude's system prompt on future fix attempts, giving it repo-specific context without manual curation.
#Cluster detection
The learning system also detects clusters of correlated issues appearing within a time window. When issues are related (e.g. a deploy broke multiple endpoints), they are grouped so the AI can address the root cause rather than treating each symptom independently.
#Configuration
All settings live under [learning]. See the full reference for every option.
[learning]
# Auto-extract learnings from execution logs (default: true)
auto_extract_learnings = true
# Analyse PR diffs on merge (default: true)
diff_analysis = true
# Promote repeated Q&A answers to instructions (default: true)
qa_promotion = true
qa_promotion_threshold = 2
# Accumulate per-repo knowledge (default: true)
repo_knowledge = true
# Classify review feedback patterns (default: true)
review_classification = true
review_promotion_threshold = 3
# Track fix strategy patterns (default: true)
strategy_fingerprinting = true
# Score fix quality (default: true)
quality_scoring = true
# Cluster correlated issues (default: true)
cluster_detection = true
cluster_window_minutes = 30
min_cluster_size = 3
# Auto-generate AGENT.md from knowledge (default: false)
auto_agent_md = false#CLI commands
# Show what Claudear has learned about a repo
claudear learn show org/my-repoThe Dashboard Learning page shows accumulated knowledge per repo in a browsable format.