mcp-tool-output-offload-threshold-2000-tokens
MCP responses exceeding 2,000 tokens should be offloaded to a temp file with a summary reference returned to context rather than held in-context verbatim. This prevents individual verbose API responses from consuming disproportionate context budget. The pattern is: write full output to /tmp/toolname-TIMESTAMP.json, return ‘Full output at /tmp/…, summary: [3-5 lines]’ to context.
Related
- sub-agent-return-budget-hard-cap-1000-2000-tokens
- mcp-tool-namespace-discipline-fully-qualified-server-prefix
- clawteam-openclaw-multi-agent-swarm-evaluation
- memory-md-size-limit-causes-truncation-on-load
- precompact-artifact-trail-must-be-verbatim-index
- evaluation-justification-before-score-prevents-anchoring-bia
- anchored-iterative-merge-prevents-compaction-drift
- observation-masking-zero-overhead-vs-summarization
- permanent-mcp-registration-requires-claudejson-not-mcpjson