One research request grew to 261,834 tokens and failed against a 128,000-token model limit. The context did not become large in one prompt. Replayed tool results accumulated across checkpoints.
INC-002 exposed three missing controls: no hard token budget, no message cap, and no pruning policy for old tool observations. Large search and file results were also inserted without compression.
A practical prevention stack
- Set a maximum context-token budget on every graph.
- Cap messages and replayable tool results.
- Compress observations before model re-entry.
- Check the budget after every node, not only before invocation.
- Test a long multi-tool request with realistic result sizes.
The important metric is context size at each boundary. A final failure log is too late to explain where growth started.
Run a heuristic review with:
prevention-scanner scan . --format json Then confirm the result with a test that simulates 50 or more tool calls and asserts the context remains bounded.
Read the full INC-002 incident.
