Session Health Monitor

Session Health Monitor

Monitor your OpenClaw agent context window health, get warnings via Telegram when usage is high, save critical facts before compaction, and keep memory directories clean.

Overview

Four capabilities for OpenClaw agent sessions: Context Threshold Warnings — Agents append usage footer to Telegram messages and warn at configurable thresholds Compaction Detection — Track usage drops to infer when context was compacted Pre-Compaction Snapshots — Save key facts and decisions to daily memory files before they're lost Memory Rotation — Archive old daily memory files to prevent clutter

1. Add shared skill reference

Add to your shared/INDEX.md: | Context window health, compaction detection, pre-compaction snapshots | `skill-session-health.md` |

2. Create shared skill doc

Create shared/skill-session-health.md: # Session Health Monitor ## Context Health Thresholds | Level | Condition | Action | |--------|------------------------------------|-------------------------------| | GREEN | <50% used AND 0 compactions | Normal operation | | YELLOW | >=50% used OR >=1 compaction | Save key facts via snapshot | | RED | >=75% used OR >=2 compactions | Save facts NOW, session ending| ## Behavioral Rules 1. When context reaches YELLOW+, extract 3-5 key facts (decisions, files changed, blockers) 2. Run: `bash scripts/snapshot.sh "fact1" "fact2"` 3. Append footer to Telegram messages at YELLOW+: `X% Context Window | Nx compacted` 4. Do this BEFORE session ends or context gets compacted 5. After any detected compaction, immediately snapshot what you remember

3. Add heartbeat step

Add to your agent heartbeat/loop: **Context health check**: Run `session_status` → always append context % to Telegram messages as footer: `📊 X% Context Window`. If Context >50% OR Compactions >=1, add: "⚠️ consider /restart after current task." If Context >75% OR Compactions >=2, flag as urgent.

Context Health Thresholds

LevelConditionActionGREEN<50% used AND 0 compactionsNormal operationYELLOW>=50% used OR >=1 compactionConsider saving key factsRED>=75% used OR >=2 compactionsSave facts NOW, session ending

Telegram Message Footer

Agents append a footer to every outgoing Telegram message: 📊 42% Context Window # GREEN — no extra warning 📊 63% Context Window | 1x compacted # YELLOW — consider restart ⚠️ 📊 81% Context Window | 2x compacted # RED — urgent, save facts This keeps the user informed about session health without requiring manual checks.

Pre-Compaction Snapshot Protocol

• When context reaches YELLOW or above, the agent SHOULD:
• Extract 3-5 key facts from the current session (decisions made, files changed, blockers found)
• Write them to memory/YYYY-MM-DD.md using scripts/snapshot.sh
• Include any unfinished work or next steps
• Do this BEFORE the session ends or context is compacted
• Example snapshot content:
• ## Pre-Compaction Snapshot (14:32)
• Refactored auth module to use JWT instead of sessions (files: src/auth.ts, src/middleware.ts)
• Bug found in rate limiter: counter resets on deploy, not on TTL expiry
• Next: write tests for new auth flow, fix rate limiter reset logic
• Decision: using RS256 for JWT signing (user preference)
• When to trigger:
• Context hits 50%+ for the first time in a session
• After any detected compaction
• Before ending a long session
• When the agent detects it has accumulated significant context

context-check.sh

Standalone health check, useful in heartbeat loops. bash scripts/context-check.sh # Human-readable output bash scripts/context-check.sh --json # Machine-readable JSON echo '{"context_window":{"used_percentage":72}}' | bash scripts/context-check.sh # Exit codes: 0=GREEN, 1=YELLOW, 2=RED

snapshot.sh

Save facts to daily memory file. bash scripts/snapshot.sh "Fact one" "Fact two" "Fact three" echo -e "Fact one\nFact two" | bash scripts/snapshot.sh -

rotate.sh

Archive old daily memory files. bash scripts/rotate.sh # Archives files older than 3 days (default) KEEP_DAYS=7 bash scripts/rotate.sh # Keep 7 days instead

Configuration

All configuration is via environment variables with sensible defaults: VariableDefaultDescriptionMEMORY_DIRAuto-detect (see below)Where to write daily memory filesKEEP_DAYS3Days to keep before archivingHEALTH_GREEN_MAX50Max % for GREEN statusHEALTH_RED_MIN75Min % for RED statusCOMPACTION_DROP30% drop that indicates compaction Memory directory auto-detection order: $MEMORY_DIR environment variable ~/.openclaw/workspace/memory (if exists) ~/.claude/memory (fallback)

jq not installed

# macOS brew install jq # Linux sudo apt-get install jq

Reset compaction state

rm /tmp/session-health-*.json

Agent not appending footer

Check shared/INDEX.md references skill-session-health.md Check heartbeat includes the context health step Verify session_status tool is available to the agent