Session Guard

Overview

Long sessions corrupt silently. Context compaction drops instructions unannounced. Hooks don’t fix it (confirmed by multiple developers on GitHub issues #19471, #9796, #64171). This skill prevents both through behavioral self-enforcement: monitor health, anchor critical rules through compaction, split before damage occurs. No packages, no databases - pure behavioral enforcement that works in every harness.

When to Use

• Session exceeds 40 tool calls
• Agent contradicts earlier decisions
• Style/naming conventions start drifting
• After any context compaction event
• Task scope growing unbounded

Health Signals

Protocol

Green Zone (0-40 tool calls)

Normal operation. No intervention needed.

Yellow Zone (40-60 tool calls)

1. CHECKPOINT - summarize progress in one paragraph
2. RECITE - state the 3-5 most critical active rules aloud: “Active rules: [naming convention], [file structure], [error handling pattern], [testing requirement]”
3. ASSESS - almost done? Push through. Not done? Prepare split.
4. REDUCE - no exploratory reads. Only targeted operations.

Red Zone (60+ tool calls OR drift signal)

1. STOP - do not make more tool calls
2. VERIFY - re-read project rules (don’t trust memory)
3. CHECKPOINT - write state to handoff document
4. SPLIT - create handoff, suggest fresh session

Context Anchoring (anti-compaction)

When compaction has occurred (sudden loss of earlier context, or after /compact):

1. RE-READ the project’s rules file immediately
2. RECITE the 3-5 critical rules aloud in your response
3. VERIFY your planned next action matches those rules before executing
4. If uncertain about ANY prior decision, RE-READ the source file - don’t guess

What survives compaction:

• Most recent user messages (high priority)
• Currently-invoked skill body (capped at 5K tokens, oldest dropped first)
• Git status and project structure
• File contents read AFTER compaction

What gets LOST in compaction:

• Decisions made early in conversation
• Architectural rules stated only verbally (not in files)
• Context from tool outputs (file reads, command outputs)

Compaction-Safe Pattern

Keep critical instructions in FILES (CLAUDE.md, CONTEXT.md), NOT in conversation. If a rule matters, it must live in a file the agent can re-read - not in something agreed on earlier.

Common Mistakes

• Trusting that you remember the rules after 50+ tool calls (you don’t - re-read)
• Re-reading EVERYTHING to be safe (wastes tool calls - be targeted)
• Feeling fine therefore assuming context is fine (compaction is SILENT)
• Splitting AFTER noticing problems (split BEFORE - prevention, not recovery)

Why This Matters

Context compaction is the #1 unsolved platform problem in 2026. Hooks don’t fix it (confirmed: the agent ignores post-compaction injections because the compaction summary creates narrative momentum). This skill is the lightweight behavioral countermeasure: no infrastructure, no packages - disciplined self-monitoring that works in every harness.