context-monitor

Context Monitor Skill

Generate a context backpressure system that monitors tool call volume and warns the agent before context rot degrades session quality.

Step 1: Generate scripts/harness/context_backpressure.sh

Create the file scripts/harness/context_backpressure.sh with the following content:

#!/usr/bin/env bash
# Harn context monitor — prevents context rot in long sessions
# Why: https://harn.app/kb/context.html — "Context as finite resource with diminishing returns"
# Pattern: https://harn.app/kb/context.html — "Context-Efficient Backpressure"

set -euo pipefail

TRACE_LOG=".claude/agent-trace.jsonl"

# Dependency check
if ! command -v jq &>/dev/null; then
  echo "Harn: context_backpressure.sh requires jq (not installed, skipping)" >&2
  exit 0
fi

# Graceful fallback if trace log doesn't exist
if [[ ! -f "$TRACE_LOG" ]]; then
  exit 0
fi

# Count tool calls in current session
TOOL_CALLS=$(wc -l < "$TRACE_LOG" | tr -d ' ')

if [[ "$TOOL_CALLS" -ge 75 ]]; then
  echo "Harn: 75 tool calls — high context usage. Summarize progress before continuing." >&2
elif [[ "$TOOL_CALLS" -ge 50 ]]; then
  echo "Harn: 50 tool calls — recommend compacting: update CHECKPOINT.json, then compact" >&2
elif [[ "$TOOL_CALLS" -ge 30 ]]; then
  echo "Harn: 30 tool calls — consider delegating research to sub-agents" >&2
fi

exit 0

Make sure scripts/harness/ directory exists before writing the file.

Step 2: Add context management section to AGENTS.md

Append the following section to the project's AGENTS.md file. If AGENTS.md does not exist, create it with this content. If it already exists, add this section at the end:

## Context Management
- Delegate research and grep tasks to sub-agents — only summaries return
- After 30+ tool calls, consider compacting: summarize progress, update CHECKPOINT.json
- Use LEARNED.md for discoveries that must survive compaction
- Prefer just-in-time retrieval (read files when needed) over pre-loading everything