Second Brain

Patch the 4 MCP gaps surfaced in the EA Sales+CRM scoping handoff with minimal-surgical-scope changes (snowflake-mcp v22.0.5→v22.0.6: drop spurious CORTEX from DESCRIBE AGENT + default describe_semant

2 min read

Patch the 4 MCP gaps surfaced in the EA Sales+CRM scoping handoff with minimal-surgical-scope changes (snowflake-mcp v22.0.5→v22.0.6: drop spurious CORTEX from DESCRIBE AGENT + default describe_semant

Decision

Patch the 4 MCP gaps surfaced in the EA Sales+CRM scoping handoff with minimal-surgical-scope changes (snowflake-mcp v22.0.5→v22.0.6: drop spurious CORTEX from DESCRIBE AGENT + default describe_semantic_view_v2 to YAML form via SYSTEM$READ_YAML_FROM_SEMANTIC_VIEW; tableau-mcp v6.0.5→v6.0.6: add new get_workbook_runtime_binding tool + cumulative-warning on legacy get_workbook_upstream_tables). Reclassify Gap 3 (result-size cap) as harness-side, mitigated by Gap 2 YAML form. Do NOT git-commit because both repos have pre-existing in-flight modifications from a separate upgrade cycle; instead extract per-file patch artefacts to a dedicated mcp-patches/ folder + apply-procedure README so the maintainer can fold cleanly later. Containers rebuilt and restarted on new images. All 4 fixes live-verified against the exact failure cases in the original gap report.

Rationale

Each gap fix was verified live against the original failure case. Gap 1: DESCRIBE AGENT returns RUNWAL_BOT full agent_spec with planning_instructions + tools + resources, 50 KB. Gap 2: RNWL_SALES_TARGET YAML inline at 25.7 KB with no Pydantic error and 26 module rules visible at YAML tail. Gap 4: Welcome Call Dashboard cumulative returns RNWL_PROD plus RNWL_DEV which is the documented false-positive while runtime returns RNWL_PROD only which is the truth. Surgical scope was chosen over deeper architectural changes such as switching to a REST API for cortex agents, because the live data and Snowflake docs confirmed DESCRIBE AGENT is valid SQL and the bug was only the spurious CORTEX qualifier. Not-committing was chosen over force-committing because the alternative would mix Gap-fix changes into a separate upgrade-cycle commit and harm reviewability. Isolated patch artefacts preserve clarity. Three handoff-quantum corrections surfaced as side-effects: 27 not 26 Sales module rules with 21/27 inter-view divergent, 16 not 22 CRM rules, 96 not 64 verified queries — flagged for CTO. Risk: maintainer may inadvertently revert via git restore — mitigated by patches README warning.

Alternatives Rejected

Outcome

Pending

Graph View

Backlinks