Beo Compounding
Overview
Compounding is the per-feature learnings capture phase after review truly finishes. Its job is to turn completed feature work into durable, reusable knowledge without inflating weak observations into institutional rules.
See ../reference/references/shared-hard-gates.md § Shared References Convention.Core principle: preserve what future work should remember, not just what happened.
Hard Gates
Boundary with Reviewing
The reviewing skill's Learnings Synthesis specialist identifies candidate learnings — patterns, surprises, and reusable insights — but does not write the final learnings file or promote anything. Compounding is where formal capture, triage, dedup, and promotion happen. If the reviewing specialist produced staging notes, use them as input to Phase 2 below rather than re-analyzing from scratch.
Default Compounding Loop
Complete these phases in order. The three-agent analysis (Phase 2) runs in parallel.
Phase 1: Gather Context
Load references/compounding-operations.md for the exact artifact collection set, git-history fallback rules, and context-gathering procedure.
Phase 2: Three-Category Analysis (3 Parallel Subagents)
Load references/compounding-operations.md for the exact three-agent analysis split, staging-file targets, and subagent task summaries.
Phase 3: Synthesis & Triage
After all three agents complete, load references/compounding-operations.md for the exact staging-file read order, dedup procedure, triage tags, slugging rule, learnings-file write flow, and QMD refresh step.
Use the format from references/learnings-template.md. Include YAML frontmatter.
Learning Triage
See references/compounding-operations.md § 3. Synthesis and Triage (subsection Triage Tags) for the severity scale and triage procedure.
Phase 4: Promote Critical Learnings
For every severity: critical learning, load references/compounding-operations.md for the exact promotion criteria, approval gate, append format, and post-promotion refresh flow.
Phase 5: Update STATE.json
Load references/compounding-operations.md for the canonical STATE.json completion shape, flush step, and checkpoint rules.
Handoff
Compounding complete.
- Learnings: .beads/learnings/YYYYMMDD-<slug>.md
- Critical promotions: N findings added to critical-patterns.md
- The ecosystem now has [N total] accumulated learnings.
Next feature starts with this knowledge available.After compounding completes, return control to beo-router to pick up the next feature or phase.
Context Budget
Follow ../reference/references/shared-hard-gates.md § Context Budget Protocol. Skill-specific checkpoint: see references/compounding-operations.md for the full checkpoint procedure.
Red Flags & Anti-Patterns
- Never skip compounding — the loop only works if it runs every cycle
- Never promote everything as critical —
critical-patterns.mdcaps at 20-30 entries; only promote learnings that would have saved ≥30 min - Never write generic learnings — "Test more carefully" is worthless; include specific cause-and-effect (e.g., "non-null migration without default fails silently")
- Never fabricate findings — 2 genuine entries beats a long file with invented ones
- Never let subagents author the final learnings file — they produce staging only; orchestrator synthesizes
- Never promote local notes without proving broad reuse value
- Never split one feature's learnings across multiple files
References
references/learnings-template.md: full template for learnings files with YAML frontmatter