What happened?

Upgrading from a pre-v0.26.3 PGLite brain wedges on column "agent_name" does not exist and the user can never reach schema v36 unaided.

applyForwardReferenceBootstrap in src/core/pglite-engine.ts:227 is missing a probe for mcp_request_log.agent_name. Migration v33 (admin_dashboard_columns_v0_26_3, shipped in v0.26.3) introduces the column, but PGLITE_SCHEMA_SQL already references it in:

// src/core/pglite-schema.ts:445
CREATE INDEX IF NOT EXISTS idx_mcp_log_agent_time
  ON mcp_request_log(agent_name, created_at DESC);

On a brain whose mcp_request_log table predates v33, the upgrade path collapses:

1. connectEngine → initSchema() runs.
2. applyForwardReferenceBootstrap() runs but doesn't bootstrap agent_name (not in REQUIRED_BOOTSTRAP_COVERAGE).
3. PGLITE_SCHEMA_SQL replays. CREATE TABLE mcp_request_log is IF NOT EXISTS → no-op on the pre-existing table; the column is not added.
4. CREATE INDEX idx_mcp_log_agent_time ...(agent_name, ...) → column "agent_name" does not exist → schema replay aborts.
5. Migration v33 (the one that would ALTER TABLE ... ADD COLUMN IF NOT EXISTS agent_name TEXT) never runs because initSchema() failed before the migration runner.

This is the same forward-reference bug class CLAUDE.md describes (#239/#243/#266/#357/#366/#374/#375/#378/#395/#396) — the maintenance contract on applyForwardReferenceBootstrap was missed when v33 landed.

What did you expect?

gbrain upgrade (or gbrain init --migrate-only) should walk a pre-v0.26.3 PGLite brain forward to v36 cleanly, the same way the existing four probes do for pages.source_id, links.link_source/origin_page_id, content_chunks.symbol_name/language, and pages.deleted_at.

Steps to reproduce

1. Have a PGLite brain that predates v0.26.3 (i.e. mcp_request_log table exists without agent_name column). In my case the brain was at schema_version = 4.
2. gbrain upgrade (or any path that triggers initSchema()).
3. Watch the upgrade chain fail:

   === Applying migration v0.11.0: GBrain Minions — durable background agents ===
     Pre-v0.21 brain detected, applying forward-reference bootstrap
   column "agent_name" does not exist
   Phase A (schema) failed: Command failed: gbrain init --migrate-only. Aborting; re-run after fixing.
   Migration v0.11.0 reported status=failed.
   

4. gbrain doctor confirms the wedge:

   [FAIL] minions_migration: MINIONS HALF-INSTALLED (partial migration: 0.11.0). Run: gbrain apply-migrations --yes
   [WARN] schema_version: Version 4, latest is 36. Fix: gbrain apply-migrations --yes
   

   …but the suggested fix loops back to the same error.
5. gbrain apply-migrations --force-retry 0.11.0 does not help — the failure is in initSchema() upstream of the migration runner, not in v0.11.0's orchestrator.

Environment

• gbrain version: 0.28.1
• OS: macOS 26.3 (Darwin 25.3.0)
• Bun version: 1.3.12
• Database: PGLite (embedded)

Suggested fix

Two-line change matching the existing pattern in applyForwardReferenceBootstrap:

1. Probe. Extend the SELECT EXISTS(...) AS ... round-trip in src/core/pglite-engine.ts:229-249 with:

   EXISTS (SELECT 1 FROM information_schema.tables
           WHERE table_schema='public' AND table_name='mcp_request_log') AS mcp_log_exists,
   EXISTS (SELECT 1 FROM information_schema.columns
           WHERE table_schema='public' AND table_name='mcp_request_log' AND column_name='agent_name') AS agent_name_exists

2. Bootstrap. Add the parallel branch to the if (...) chain:

   const needsAgentName = probe.mcp_log_exists && !probe.agent_name_exists;
   // ...
   if (needsAgentName) {
     // v33 (admin_dashboard_columns_v0_26_3) adds full column set + backfill.
     // Bootstrap only adds enough for PGLITE_SCHEMA_SQL's
     // `CREATE INDEX idx_mcp_log_agent_time ON mcp_request_log(agent_name, ...)`
     // not to crash. v33 runs later via runMigrations and is idempotent.
     await this.db.exec(`
       ALTER TABLE mcp_request_log ADD COLUMN IF NOT EXISTS agent_name TEXT;
     `);
   }

3. Pin it in CI. Add to REQUIRED_BOOTSTRAP_COVERAGE in test/schema-bootstrap-coverage.test.ts:47:

   // v0.26.3 — forward-referenced by `CREATE INDEX idx_mcp_log_agent_time
   // ON mcp_request_log(agent_name, created_at DESC)`.
   { kind: 'column', table: 'mcp_request_log', column: 'agent_name' },

   And extend the strip-block at test/schema-bootstrap-coverage.test.ts:78-98 to drop the index + column so the test exercises the new branch:

   DROP INDEX IF EXISTS idx_mcp_log_agent_time;
   ALTER TABLE mcp_request_log DROP COLUMN IF EXISTS agent_name;

The mirror change to PostgresEngine.applyForwardReferenceBootstrap is likely warranted too — the same failure mode applies to any old Supabase/self-hosted brain that pre-dates v0.26.3 and goes through connectEngine on a v0.28.x binary. (I didn't hit it; only the PGLite path bit me.)

Workaround for affected users

For anyone hitting this in the wild: gbrain export still works against the wedged brain (the read path doesn't trigger initSchema()'s schema replay). Export markdown, move the broken brain dir aside, run gbrain init --pglite on a fresh dir (initializes cleanly at v36), then gbrain import the rescued markdown. Embeddings + links + timeline get rebuilt from scratch via gbrain embed --stale && gbrain extract all.

Why CI didn't catch it

test/schema-bootstrap-coverage.test.ts is the right shape (declarative REQUIRED_BOOTSTRAP_COVERAGE + drift guard), but it only covers what's listed in that array. v33's agent_name column was added to pglite-schema.ts and migrate.ts but never propagated to the bootstrap probe or the coverage list. The maintenance contract in the docstring at pglite-engine.ts:222-225 is the right place — it just got missed when v33 landed. Worth considering: a stricter CI check that greps pglite-schema.ts for CREATE INDEX ... ON <table>(<col>) patterns and asserts each <table>.<col> either appears in the original CREATE TABLE before the migration adds it, or has a matching entry in REQUIRED_BOOTSTRAP_COVERAGE. Out of scope for this issue, but flagging.