🦛 Hippo Memory Insights — 2026-06-26

[github-actions[bot]]

Summary

• Memories in store: 347
• New lessons learned from commits today: 0
• Highest-confidence memory: Recurring incident: stale workflow .lock.yml files cause churn and CI friction; run make recompile after markdown workflow edits.

Top Memories Surfaced

• Recurring incident: stale workflow .lock.yml files cause churn and CI friction; run make recompile after markdown workflow edits.
• CI validation in .github/workflows/ci.yml scans YAML for ANSI escape sequences before other jobs run.
• Recurring incident: node: command not found on GPU/self-hosted paths when Node runtime/tooling is missing; validate runtimes and PATH early.
• Guard list_code_scanning_alerts calls in workflow prompts with state: open and severity: critical,high to avoid oversized responses.
• The compiler in pkg/workflow/compiler_yaml.go strips ANSI codes from workflow text, but workflow authors should still keep source files clean.
• Use steps.sanitized.outputs.text/title/body directly in workflow markdown; needs.activation.outputs.* is deprecated.
• Claude permission behavior around pkg/workflow/claude_engine.go and pkg/workflow/claude_tools.go remains a high-value architectural constraint.

Suggested Improvements

Error Patterns

• Add or strengthen early environment validation for Node-dependent paths, especially self-hosted/GPU workflows, so failures stop before runtime setup proceeds. The recurring node: command not found incident suggests this is still a brittle edge.
• Keep workflow authoring guidance aggressively centered on make recompile after .github/workflows/*.md edits. This remains the clearest repeated source of avoidable CI churn.
• Continue enforcing bounded list_code_scanning_alerts usage in prompts. A lint or audit rule for required state and severity filters would prevent oversized-tool-response regressions.

Code Quality

• Review workflow markdown and compiler-related docs for remaining references to needs.activation.outputs.*; migrate examples to steps.sanitized.outputs.* to reduce deprecation drift.
• Consider a focused cleanup pass around workflow-engine permission documentation and tests in pkg/workflow/claude_engine.go and pkg/workflow/claude_tools.go so the acceptEdits versus unrestricted modes remain explicit and hard to misuse.
• Keep architectural guidance concentrated in smaller, task-focused files and skills. Memory retrieval still strongly surfaces modularity guidance and typed-domain patterns as durable signals.

CI Health

• Preserve and extend the ANSI-prevention path: the validate-yaml job in .github/workflows/ci.yml is a useful front-line guard, and memories suggest it is worth keeping early in the pipeline.
• Prefer fast local gates (make build && make fmt, then make agent-report-progress) over broad local test sweeps. Memories continue to indicate full make test is better left to CI except for targeted investigations.
• Artifact/cache filenames remain a workflow-health footgun. Files that land in cache-memory or artifacts should keep filesystem-safe timestamp and naming formats.

Quick Wins

• Add a lightweight audit/check that flags workflow markdown still using deprecated needs.activation.outputs.* references.
• Add a prompt-lint rule or reusable snippet for list_code_scanning_alerts filter requirements.
• Surface Node runtime preflight checks earlier in self-hosted/GPU workflow setup paths.
• Refresh docs/examples that mention workflow outputs, lock-file handling, and ANSI-safe YAML editing so the most common operational lessons are visible where authors work.

Longer-term Themes

• The memory store keeps surfacing workflow-authoring reliability as a systemic theme: lock-file churn, oversized tool responses, ANSI contamination, and runtime preconditions all point to a need for stronger author-time guardrails.
• Architectural memories repeatedly highlight tool-permission semantics as a sensitive boundary. This suggests continued investment in explicit tests and docs around engine/tool enforcement behavior.
• A large share of value still comes from imported guidance rather than fresh git-learned incidents. That may indicate the store is stable, but also suggests opportunity to improve commit-to-memory extraction quality when real fixes land.

Memory Health

• After consolidation, the store now holds 347 memories with 0 open conflicts and full embedding coverage (347/347).
• Consolidation removed 191 low-strength memories, merged 30 episodic entries into 12 semantic summaries, and reported 4 low-quality memories worth auditing.
• Pruning appears healthy overall, but the low-quality set should be reviewed with hippo audit in a follow-up run if the team wants tighter memory quality control.

Operational notes

• hippo learn --git found no fix/revert/bug commits in the recent scan window, so there were no new commit-derived lessons today.
• An initial hippo sleep attempt hit a transient database is locked error during concurrent operations; rerunning consolidation serially with sleep --no-share succeeded.
• A pre-consolidation snapshot showed 537 memories; post-consolidation status is 347 memories, reflecting aggressive decay/dedup behavior rather than missing data.

Generated by 🦛 Daily Hippo Learn · 27.3 AIC · ⌖ 21.7 AIC · ⊞ 3.6K · ◷

• expires on Jun 28, 2026, 11:40 PM UTC-08:00