fix(storage): persist MeshNode.SyncBehavior in Postgres (decouple survives restart) by rbuergi · Pull Request #119 · Systemorph/MeshWeaver

rbuergi · 2026-06-29T21:59:58Z

Problem

The static-repo "Not synced" decouple claim (MeshNode.SyncBehavior) had no Postgres column. PostgreSqlStorageAdapter persists only the typed content JSONB plus a fixed set of metadata columns (state, version, …) — SyncBehavior was dropped on write and the row→node mapper never read it, so every reload from PG defaulted to Include.

Consequence: a partition an admin decoupled via Admin → Partition Sync (or per-node Stop Sync) silently re-coupled on the next restart/redeploy, and the next static-repo import re-clobbered admin-managed content. This is the root cause behind the client-i Provider/Anthropic API key reverting to the shared Azure Foundry default (2026-06-29): the key was entered + the partition decoupled, but the decouple evaporated on the next deploy and the importer re-seeded our key.

The decouple feature appeared to work because SyncBehavior survives in-memory (the live MeshNodeStreamCache keeps the full node) — it just never made it to storage. Sqlite was unaffected: it serializes the whole node as JSON, so the claim round-trips there. The bug is Postgres-only.

Fix

Persist SyncBehavior as a mesh_nodes column (it mirrors state — a node-level metadata enum), the only table whose nodes participate in static-repo sync:

Schema — add sync_behavior SMALLINT NOT NULL DEFAULT 0 to mesh_nodes in every provisioning path of PostgreSqlSchemaInitializer.
Migration V39 — ALTER TABLE … ADD COLUMN IF NOT EXISTS across every existing schema (public + auth + all partitions), idempotent, one dynamic-SQL pass.
Adapter — write it in the UPSERT (mesh_nodes only — satellite tables don't carry it), surface it on the single-node / prefix reads (satellite-safe 0 AS sync_behavior so UNION branches stay aligned), and read it via a defensive mapper (PgMeshNodeReader.ReadSyncBehavior — column absent ⇒ Include) shared by the storage adapter and the cross-schema query provider. The defensive read means no existing SELECT that omits the column can break.

The authoritative decouple read (StaticRepoImporter.ReadClaimedRoots → GetMeshNodeStream → the adapter's exact read) now returns the persisted claim, so a decoupled partition root durably skips import across restarts.

Repro test

SyncBehaviorPersistenceTests (Testcontainers PG): round-trips ExcludeThisAndChildren through PostgreSqlStorageAdapter.Write/Read. Fails before this change (mapper returns Include), passes after. Plus a default-Include case.

Verification

MeshWeaver.Hosting.PostgreSql.Test: 509 passed, 1 skipped, 0 failed (the change touches core read/write SELECTs, so the whole PG suite is the regression gate).
Migration + PG projects build clean (0 warnings).

Scope notes

Whole-partition decouple (the documented Admin → Partition Sync flow) is fully covered: the partition-root claim is read authoritatively and skips the entire subtree.
The mesh_node_history trigger and auth-mirror trigger are intentionally left untouched (they insert explicit column lists; history/audit doesn't need the claim).
Also includes the /provider-keys skill (src/MeshWeaver.AI/Data/Skill/) documenting AI provider key administration — the decouple step it prescribes is now durable.

🤖 Generated with Claude Code

…vives restart) The static-repo "Not synced" decouple claim (MeshNode.SyncBehavior) had no Postgres column. PostgreSqlStorageAdapter persisted only the typed content plus a fixed metadata column set, so SyncBehavior was dropped on write and the row->node mapper defaulted every reload to Include. A partition an admin decoupled therefore silently re-coupled on the next restart, and the next static-repo import re-clobbered admin-managed content — e.g. a customer's Provider/Anthropic API key reverting to the shipped Azure Foundry default (atioz, 2026-06-29). Sqlite was unaffected: it serializes the whole node as JSON. - Add sync_behavior SMALLINT column to mesh_nodes (every provisioning path in the schema initializer) + V39 migration (ADD COLUMN IF NOT EXISTS across every existing schema: public, auth, and all partitions). - Write it in the UPSERT (mesh_nodes only — satellite tables don't carry it), surface it on the single-node/prefix reads (satellite-safe `0 AS sync_behavior` in the UNION branches), and read it via a defensive mapper (column absent -> Include) shared by the storage adapter and the cross-schema query provider. - Repro: SyncBehaviorPersistenceTests round-trips ExcludeThisAndChildren through the PG adapter (fails before this change, which mapped it back to Include). - Docs: add the /provider-keys skill for administering AI provider keys. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

github-actions · 2026-06-29T22:19:10Z