MCPg v0.6.6 — release notes

Released: 2026-07-01 Tool surface: 252 tools across 19 capability buckets Tests: 2623 pass (PG 14 / 15 / 16 / 17 / 18 / 19 / WarehousePG) CI: PG 14-19 + a now-fully-passing WarehousePG lane

This is a patch-level bump (0.6.5 → 0.6.6). It completes the entire feature roadmap tracked in docs/feature-shortlist.md — six new read-capable tools spanning live-ops, vector retrieval, and graph projection — plus a set of real reliability fixes surfaced by finally getting the WarehousePG CI lane to run against a genuine server. Every addition is backward-compatible and additive.

New tools

Live database operations

• analyze_table_bloat — per-table and per-index estimated bloat % (catalog estimate by default; precise pgstattuple/pgstatindex mode when installed), sorted worst-first, so VACUUM/REPACK can be targeted precisely instead of guessed at.
• dry_run_ddl — runs a proposed ALTER TABLE / non-concurrent CREATE INDEX under a tight lock_timeout, captures the lock mode held, wall-clock duration, and WAL bytes generated, then always rolls back. CREATE INDEX CONCURRENTLY and other non-transactional DDL are rejected up front as ineligible, since they can’t run inside the wrapping transaction.
• run_select_tuned — runs a read-only SELECT with an elevated, bounded work_mem (hard-capped at 2 GiB) scoped to that single statement, for analytical queries that would otherwise spill to disk.

Retrieval & vector

• retrieve_with_context — “one-shot RAG”: a pgvector k-NN search that also expands each hit one hop along foreign keys (parent + child rows), packing the match and its relational context into a single response.
• recommend_ivfflat_probes — the IVFFlat counterpart to the existing recommend_hnsw_ef_search: sweeps probes, measures recall@k and latency, and recommends the smallest value clearing a target recall.

Graph

• generate_graph_projection — projects a relational schema into an Apache AGE property graph by generating the openCypher CREATE/MERGE statements (rows→vertices, FKs→edges) for review, never executed — the same pattern as generate_test_data and recommend_redistribute.

Enhancements

• audit_database now folds in sequence-exhaustion and postgresql.conf-sanity checks as two additional scorecard categories, so the comprehensive scan surfaces at-risk sequences and dangerous GUCs directly — no separate audit_sequences/audit_settings call needed.
• dump_database gains an optional schemas parameter to scope a pg_dump to specific schemas instead of the whole database. Purely additive — omitting it is byte-for-byte unchanged.

Reliability fixes

• The warehousepg-latest CI lane is now genuinely green — for the first time. It had been failing since the day it was added: the pinned image never existed on Docker Hub, so the build failed before a single test ever ran, silently, for every PR. Repointed at a real published image, corrected its connection wiring, and fixed three real Greenplum/WarehousePG dialect incompatibilities the working lane then surfaced (distribution-key-aware uniqueness constraints, a PG13+-only DROP DATABASE flag, a default-schema collision in dump/restore).
• cancel_query / terminate_backend could target the wrong server. Both were marked “safe to route to a read replica,” but a PID names a process on one specific physical server — with MCPG_REPLICA_URLS configured, the signal could silently miss or hit the wrong backend. Fixed: these are primary-only actions now, always.
• The MCP Registry publish step had been silently failing since 0.6.1 — the version it submitted was never bumped past the original registry-launch value, so every later release was rejected as a duplicate while PyPI/GHCR/GitHub Releases kept succeeding unnoticed. It now derives the version from the release tag automatically.

Upgrade

pip install --upgrade mcpg   # once published to PyPI
docker pull ghcr.io/devopam/mcpg:0.6.6   # or :latest

No configuration changes required.

Full changelog

See ../CHANGELOG.md [0.6.6] for the complete itemised list.