From hatch3r
Monitors and maintains conversation context health during long sessions. Assesses degradation with checklists, identifies issues like stale files or repeated errors, applies fixes like refreshing context or delegating tasks, and verifies improvements.
npx claudepluginhub hatch3r/hatch3rThis skill uses the workspace's default tool permissions.
```
Manages context window health in long Claude sessions: monitors triggers, assesses states, compresses stale content for subagents, recommends fresh starts to prevent degradation.
Monitors context window health during multi-step agent tasks to ride peak quality for high-fidelity output and prevent drift in long-running sessions or complex refactors.
Performs structured self-healing assessment of AI subsystems (memory, reasoning, tools, communication) to detect drift, rebalance, and integrate learnings. Use mid-session for formulaic responses, error chains, or context overload.
Share bugs, ideas, or general feedback.
Task Progress:
- [ ] Step 1: Assess current context health
- [ ] Step 2: Identify degradation signals
- [ ] Step 3: Apply corrective action
- [ ] Step 4: Verify health improvement
Run through the self-assessment checklist:
| Check | Healthy | Degraded |
|---|---|---|
| Task recall | Can state requirements from memory | Need to re-read issue |
| Progress | Clear forward momentum | Cycling or stuck |
| Errors | Occasional, different causes | Repeated, same cause |
| Files | Recently read and current | Stale, may have drifted |
| Scope | Aligned with acceptance criteria | Drifted to tangential work |
platform in .agents/hatch.json)After corrective action:
During context health checks, also scan for signs of context poisoning -- stale or incorrect information that has accumulated in the conversation:
| Signal | Detection | Action |
|---|---|---|
| Outdated file content | You reference a file's content but the file has been modified since you last read it | Re-read the file before continuing |
| Stale assumptions | A decision was made based on information that has since changed (e.g., a function was refactored) | Re-verify assumptions against current state |
| Contradictory context | Two pieces of context in the conversation disagree (e.g., "the API uses REST" vs. code showing GraphQL) | Resolve by reading the actual source of truth |
| Accumulated errors | Multiple tool calls have failed, suggesting the mental model of the codebase is wrong | Reset context by re-reading key files from scratch |
Context poisoning is more dangerous than missing context because it leads to confident-but-wrong decisions.
hatch3r-context-health -- full monitoring protocol with integration pointshatch3r-board-pickup -- auto-advance mode uses context health for session management