| 1 | --- |
| 2 | name: "agent-history-hygiene" |
| 3 | description: "Keep agent histories as durable context stores by summarizing stable knowledge and removing session chatter." |
| 4 | domain: "team-optimization" |
| 5 | confidence: "high" |
| 6 | source: "manual audit across squad histories during reskill" |
| 7 | --- |
| 8 | |
| 9 | ## Context |
| 10 | |
| 11 | Histories should help the next session start smarter, not longer. Repeated project intros, rollout broadcasts, timestamps, and requester names inflate context without improving recall. |
| 12 | |
| 13 | ## Patterns |
| 14 | |
| 15 | - Use `Core Context` for stable ownership and operating context. |
| 16 | - Keep `Learnings` for durable technical or editorial lessons that will matter again. |
| 17 | - Merge duplicate learnings into one stronger statement. |
| 18 | - Promote team-wide patterns into `.squad/skills/` once they recur across several agents. |
| 19 | - Prefer summaries of outcomes over transcript-style chronology. |
| 20 | |
| 21 | ## Examples |
| 22 | |
| 23 | - Good: summarize a whole migration as "automation now writes through the branch-protection workflow skill". |
| 24 | - Good: condense repeated onboarding updates into one `Core Context` bullet. |
| 25 | - Good: keep one canonical lesson about a pipeline contract instead of storing the same insight in multiple dated entries. |
| 26 | |
| 27 | ## Anti-Patterns |
| 28 | |
| 29 | - Repeating "PRD now available" or milestone broadcasts in multiple histories after the team has absorbed them. |
| 30 | - Keeping long timestamped session logs inside an agent history. |
| 31 | - Storing branch names, requester names, or one-off PR trivia that does not change future decisions. |