MCP lifecycle and limits

Heavy operations (open, query, decode, Manifest/resources) share one bounded queue. Discovery and metadata stay responsive. Timeouts include queue time. Every heavy tool accepts timeout_ms in 1..= the server timeout. For progress notifications, set the MCP request's _meta.progressToken; progress reports phase and elapsed seconds, not a completion percentage.

Running timeout/cancellation closes only the affected dataset. Reopen it to continue. A timeout result includes dataset_closed: true; the SDK may suppress the response to an explicit cancellation, so check list_datasets afterward. Cancelling or timing out while queued preserves existing datasets. A cancelled open never registers a dataset. Closing rejects new work; accepted work may finish within its deadline and continues to count against quotas. EOF closes the session and its datasets.

Server defaults can be configured after mcp on the command line:

Option	Default
`--timeout-ms`	60000
`--max-datasets`	4 (open/opening/retained by accepted work)
`--max-pending`	16 (running plus waiting heavy requests)
`--max-input-bytes`	1073741824 source bytes per dataset
`--max-dex-bytes`	1073741824 uncompressed DEX bytes per dataset
`--max-total-input-bytes`	2147483648 source bytes across datasets
`--max-output-bytes`	131072 serialized data-tool result bytes, including text and structured copies
`--max-resource-bytes`	67108864 uncompressed bytes per resource entry

open_dataset.inputs accepts 1–64 paths. Per-call max_bytes must be at least 4096 and at most the configured --max-output-bytes; it limits the whole serialized tool result, including both JSON copies.

Input requests are limited to 256 KiB per JSON-RPC frame; malformed/oversized frames close the session. Query calls allow at most 16 statements, 16 temporary queries and 64 KiB of SQL. Archives allow at most 100000 entries. Resource limits apply before decompression, and files are never extracted. Source/uncompressed byte quotas are not an RSS limit: parsers and prepared data use additional memory, and the engine's existing per-operator materialization limits still apply. server_busy and session_quota_exceeded are recoverable; retry later or close unused datasets.