MCPg v0.6.4 — release notes
Released: 2026-06-27 Tool surface: 243 tools across 19 capability buckets Tests: 2499 pass (PG 14 / 15 / 16 / 17 / 18 / 19) CI: PG 14-19 + an experimental WarehousePG lane
This is a patch-level bump (0.6.2 → 0.6.4). The intervening 0.6.3 cut was staged but never published; everything it would have carried plus the work since rolls into 0.6.4. The no-deprecation rule held across every PR — PG ≤ 18 paths stay advertised by per-tool status probes, and no consumer-visible return shape was removed or renamed.
Why 0.6.4 (skipping 0.6.3): the 0.6.3 release PR sat open while ~30 more PRs landed on
main; rather than re-base a stale cut, the next published version is 0.6.4 with the complete rolled changelog. The minor bump stays reserved for a stability commitment closer to 1.0.
| Theme | Highlights |
|---|---|
| PG 19 readiness | SQL/PGQ, in-server REPACK, async-I/O advisor, lock+recovery analytics, runtime toggles, partition MERGE/SPLIT, skip-scan advisor, WAIT FOR LSN + RYW, DDL helpers, characterisation tests (3.4) |
| WarehousePG (§15) | MPP detection probe + read introspection (distribution policies, segment health, AO tables, resource groups) + MPP advisors (analyze_mpp_query_plan, recommend_redistribute) + experimental CI lane |
| Logical replication writes (2.1) | create_publication / drop_publication / create_subscription / drop_subscription with DSN-redacting result repr |
| Agent ergonomics | generate_test_row_for (8.1), analyze_session_cost (8.7), session-intent surface filter (8.8), recommend_headline_tools (14.4) |
| Config & sizing advisors (§16) | audit_sequences (overflow), audit_settings (postgresql.conf sweep), recommend_postgres_conf (pgtune calculator) |
| Vector | recommend_hnsw_ef_search — multi-query recall@k / latency sweep with index verification (9.1) |
| LangChain / LangGraph integration | outputSchema sweep (8.6) — 192 tools now emit a typed JSON output schema on the wire |
| Process | Roadmap-row linkage tooling (14.6) — PR template + tools/roadmap_linkage.py validator |
Headline: structured outputs (outputSchema) across the surface
The 8.6 sweep converted every tool that returns a single helper
dataclass verbatim from a dict[str, Any] return to its typed
frozen-dataclass return, so FastMCP auto-derives a JSON outputSchema
that LangChain / LangGraph clients validate against. The
structured-output manifest floor climbed from 5 → 192 across seven
batches. List-returning tools wrap into a {"result": [...]}
envelope; the ~17 remaining dict-typed tools are code-emitting /
result-restructuring handlers (ORM generators, describe_self,
prepare_migration) that legitimately can’t auto-derive a clean
schema and are documented opaque exceptions.
Backward-compatible: FastMCP keeps emitting the legacy text content
array alongside the new structuredContent, so clients reading only
content see no change.
Upgrade
pip install --upgrade mcpg # once published to PyPI
No configuration changes required. Every new capability is additive
and gated behind the same MCPG_ACCESS_MODE / MCPG_ALLOW_DDL
controls as before. The new MCPG_SESSION_INTENT env var (8.8) is
opt-in — unset means the full tool surface, exactly as today.
Full changelog
See ../CHANGELOG.md [0.6.4] for the complete
itemised list (Added / Changed / Fixed), including the Phase 3 PG 19
readiness block previously staged for 0.6.3.