Second Brain

Context Engineering Upgrade — 3-Tier Integration from Agent Skills Repo

2 min read

Context Engineering Upgrade — 3-Tier Integration from Agent Skills Repo

Decision

Integrated evidence-backed context engineering findings from muratcankoylan/Agent-Skills-for-Context-Engineering into the system as 3 tiers: (1) PreCompact/PostCompact hook upgrades with 5-section structured summaries and 4-probe compression verification, (2) Tool Intelligence directives in MEMORY.md with namespace discipline, output offload thresholds, and contradiction detection, (3) Evaluation Rigor Protocol in protocols.md with justification-before-score, pairwise double-pass, and rubric anchoring rules.

Rationale

~70% of the source repo’s content was already covered by our system (Vault+Graphiti, OpenClaw hierarchy, 56 skills, 60 MCP servers). The remaining ~30% contained specific evidence-backed thresholds and structural techniques we lacked: artifact trail separation for compression (all methods score 2.2/5.0 without it), observation masking (83.9% of tokens are observations), and evaluation rigor rules (+15-25% reliability from justification-before-score). Adopted only what enhances, zero conflict.

Alternatives Rejected

  1. Install the repo as a Claude Code plugin — rejected: their skills are teaching materials (‘textbooks not toolboxes’ per their own self-analysis), would add 13 conceptual skills that overlap with our 56 operational ones.\n2. Adopt BDI cognitive architecture — rejected: formal ontology overhead without practical benefit; ST + Vault + Graphiti already provides reasoning substrate.\n3. Cherry-pick individual skills — rejected: no single skill was production-ready; the value was in specific numbers and structural techniques embedded across multiple skills.

Outcome

Pending

Graph View

Backlinks