Skip to main content
ClaudeWave
Skill70 repo starsupdated 2mo ago

context-budget-guard

Context Budget Guard proactively monitors token consumption during extended conversations and initiates context cleanup when usage reaches 70 percent capacity, preventing model degradation before it occurs. Deploy this skill at the beginning of sessions expected to exceed 30 minutes or before each major task phase in multi-hour workflows to estimate current context load and trigger compression before the model's reasoning coherence declines.

Install in Claude Code
Copy
git clone --depth 1 https://github.com/ArchieIndian/openclaw-superpowers /tmp/context-budget-guard && cp -r /tmp/context-budget-guard/skills/openclaw-native/context-budget-guard ~/.claude/skills/context-budget-guard
Then start a new Claude Code session; the skill loads automatically.

SKILL.md

# Context Budget Guard

State file: `~/.openclaw/skill-state/context-budget-guard/state.yaml`

Don't wait for the model to go incoherent. Act at 70%, not 95%.

## When to Use

- At the start of any long-running session (>30 min expected)
- Before each major task step in a multi-hour workflow
- When `long-running-task-management` advances to a new stage

## The Process

### Step 1: Initialize (Session Start)
Write to state: `session_start` timestamp, `compaction_count: 0`, `status: monitoring`, `threshold_pct: 70`.

### Step 2: Check Budget (Before Each Major Step)
Estimate current context usage as a rough percentage:
- **Low (<50%)** — continue normally
- **Medium (50–70%)** — note in state, proceed with caution (avoid loading large files)
- **High (>70%)** — trigger `context-window-management` NOW, before continuing

To estimate: count approximate tokens from recent messages, loaded file contents, and active task context.

### Step 3: After Compaction
- Increment `compaction_count` in state
- Write `last_compacted_at`, `strategy_used` (passed from `context-window-management`)
- Reset mental estimate to ~20% (post-compaction baseline)

### Step 4: Session End
Update state: `status: idle`, `session_end` timestamp.

## Key Principles

- The 70% threshold is the trigger — earlier is always better than later
- Check budget BEFORE loading large files or running subagents, not after
- If in doubt: compact. A clean context costs less than a confused one.
- Works best paired with `long-running-task-management` — check budget at every checkpoint