compact-hygiene

name	compact-hygiene
description	Guides proactive /compact usage with preservation instructions to prevent context rot in long Cascade sessions. Triggers when context reaches ~50% of window, before large tool outputs, or when the user mentions "context", "compact", "drift", or "forgetting".
version	1.0.0
trigger	Context window is approaching 50% utilization, OR user asks about: context management, drift, "the agent keeps forgetting", long sessions, "how do I stop the AI from losing track", or explicitly mentions /compact.

Compact Hygiene

Long sessions rot. Most users only run /compact when the warning light is already red — and by then the earliest (most foundational) context is already being summarized away. You run it proactively, with explicit preservation instructions.

When to Run

• Proactive (best): at 50–60% context utilization, before the model starts approximating
• Pre-output: right before a large tool output (big file read, long test run) you don't need to keep verbatim
• Checkpoint: after finishing a major step, before starting the next
• Never: at 95%. By then you're compressing context that already degraded the last 20 responses.

How to Compact With Preservation

Run /compact with a clear list of what to keep verbatim and what to drop. Example:

/compact
Preserve verbatim:
- The current plan file path: plans/2026-04-17-stripe-webhooks.md
- The 3 AGENTS.md invariants I cited this session (testing ≥80%, no PII logging, API shape)
- The 2 design decisions from @architect: (1) use Stripe PaymentIntents not Charges, (2) webhook idempotency via Redis INCR
- The current plan's checkbox state (which are done, which are next)
- All error messages from the last failed test run

Discard:
- Tool outputs from exploration that didn't lead anywhere
- File contents I've already edited
- Previous grep results we superseded
- Verbose npm install logs

Preservation Checklist — What You Almost Always Want to Keep

• Current plan file path and checkbox state
• Cited AGENTS.md invariants
• Design decisions made this session
• Current blocker (if any)
• Recent error messages
• Vault pages being actively referenced

Things Safe to Drop

• Tool outputs from rejected paths
• File contents you've already modified
• Verbose command logs (test runs, installs)
• Exploration that didn't pan out
• Superseded grep/search results

Agent Protocol

1. Monitor context size silently. At ~50%, suggest a compact and list what you'd preserve.
2. Ask first if the session has made novel design decisions not yet written to the plan. Compact can summarize these away.
3. Write decisions to vault/ FIRST. Anything worth preserving long-term goes to vault/decisions/ before you compact — the vault survives forever, the session doesn't.
4. After compact, re-read the preserved plan file and any cited vault pages to ground yourself before continuing.

Never

• Never run /compact without explicit preservation instructions — you lose the thread
• Never wait until context is at 95% to compact — by then the earliest (and most foundational) context is already degraded
• Never compact mid-tool-call — finish the in-flight step first
• Never rely on compact as long-term memory — write durable facts to vault/ before compacting; the vault survives forever, the session doesn't
• Never compact away an active blocker or the current plan file path — always keep those in the preserve list

Credits

Pattern from MindStudio's /compact writeup (Apr 2026) and Anthropic's automatic context compaction cookbook.