[icesixgod]

Summary

MCP startup previously prewarmed task history, and compressed metadata probes decompressed entire logs into temporary disk files. Multiple plugin processes could therefore write tens of GB before the viewer was opened.

Serve MCP without history prewarming and stream bounded compressed headers for task discovery. Preserve complete frame/tail validation, pagination, filesystem checks, and task-local read budgets when a trajectory is requested. Add regression tests and update the English and Chinese documentation.

Privacy impact

Metadata and format probes read only bounded headers and create no decompression spool. Full history reads retain their existing validation and limits. Returned content, redaction, and full-detail consent remain unchanged; native read-limit retries still require a verified matching task identity.

Validation

• uv run ruff format --check .
• uv run ruff check .
• uv run mypy and Windows platform type checks
• uv run pytest --cov --cov-report=term-missing — 629 passed on the isolated fix snapshot; 91.99% coverage
• Python 3.14 tests with resource warnings treated as errors — 629 passed
• Release validation and packaged MCP stdio smoke test
• English and Chinese documentation are synchronized
• No real logs, credentials, base instructions, encrypted reasoning, or unredacted screenshots are included

A local probe of 803 synthetic compressed archives representing 13.47 GB of uncompressed history made no decompression-spool calls and recorded 0 B of disk writes during header discovery.