Schema Unification (gbrain-base → gbrain-base-v2)

v0.41.22 ships gbrain-base-v2 — a 15-type DRY/MECE taxonomy (14 canonical + note catch-all) — as the install default for new brains. Existing brains on gbrain-base can opt in via the pack_upgrade_available onboard finding + the unify-types PROTECTED Minion handler.

This skill is the playbook for that migration.

brain_first: exempt

This skill is ABOUT the brain's shape — it can't depend on the brain it's reshaping. No gbrain search lookup first; jump straight to onboard.

When this skill fires

• Agent runs gbrain onboard --check and sees pack_upgrade_available or type_proliferation warnings
• User asks "what is the canonical taxonomy / how do I clean up my page types / migrate to v2"
• A dangling_aliases finding surfaces (post-unify GC)
• An agent ingesting from a custom pack wants to consult the v2 taxonomy as a reference

Mental model (one paragraph)

A production gbrain brain accreted 94 distinct pages.type values over years of ingestion: tweet / tweet-thread / tweet-bundle / tweet-single / media/x-tweet/bundle / tweet-stub all coexisting; 5.5K concept-redirect pages; atom-partner-link pages that should be links; civic / framework / insight / memo / anecdote one-offs. The cure: collapse to 15 canonical types (person, company, media, tweet, social-digest, analysis, atom, concept, source, deal, email, slack, writing, project, note) with subtypes/format/origin pushed to frontmatter, alias-rows for redirects, real link-rows for edge-shaped pages, and a catch-all that bins long-tail unknowns to note with frontmatter.legacy_type = <original> for rollback.

Workflow

Phase 1: Discovery

Confirm the brain is actually on gbrain-base (not already on v2).

gbrain schema active --json | jq -r '.identity'

Expected: gbrain-base@1.0.0+<sha>. If you see gbrain-base-v2@..., the brain is already on v2 — skip the migration.

Then run onboard to see what would change:

gbrain onboard --check

Look for the pack_upgrade_available finding. If it's ok, there's no successor declared for the active pack — done.

Phase 2: Preview

Run the per-cluster narrative:

gbrain onboard --check --explain

This invokes the unify-types handler in dry-run mode and prints:

• How many pages would retype per cluster (tweets, articles, companies, etc.)
• How many concept-redirect pages would become alias rows
• How many edge-shaped pages would convert to real links
• The synthesized catch-all rules for unknown types

Review the output. If the proposed changes look wrong, don't proceed — file an issue or write a custom pack with adjusted mapping_rules.

Phase 3: Apply

The handler is PROTECTED (manual_only) — autopilot will never auto-fire it. Submit explicitly:

gbrain jobs submit unify-types \
  --params '{"target_pack":"gbrain-base-v2","apply":true}'

On PGLite (the install default), or on any setup without a running gbrain jobs work worker or supervisor daemon, add --follow so the job executes inline:

gbrain jobs submit unify-types \
  --follow \
  --params '{"target_pack":"gbrain-base-v2","apply":true}'

The persistent worker daemon is Postgres-only. Without --follow on PGLite, the job sits queued forever and the migration never runs.

apply defaults to false (dry-run) per the handler contract, so "apply":true is required here or the job reports success having retyped nothing and left the active pack unflipped. Omit it to preview.

Watch progress per phase (worker-daemon runs; with --follow the same progress streams inline):

gbrain jobs get <job_id>      # one job: status, progress, result
gbrain jobs watch --follow    # live dashboard of the whole queue

A job that stays queued here means no worker is running; resubmit with --follow to execute it inline.

On a 186K-page brain expect ~10 minutes. The handler runs:

1. Preflight (validate target pack has mapping_rules:)
2. Stats snapshot (pre-state for celebration summary)
3. Acquire gbrain-unify db-lock (60min TTL)
4. Apply phases:
   • Explicit retype rules (tweets, articles, companies, etc.)
   • Catch-all retype (unknown types → note with legacy_type)
   • Page-to-link rules (atom-partner-link, symlink)
   • Page-to-alias rules (concept-redirect)
5. Final sync (untyped rows by path-prefix)
6. Flip active pack to gbrain-base-v2
7. Verify + celebration summary

Phase 4: Verify

gbrain onboard --check
gbrain schema stats

Expected:

• pack_upgrade_available → ok (active pack is now v2)
• type_proliferation → ok (≤16 distinct typed values)
• dangling_aliases → ok (slug_aliases all point at active canonicals)
• gbrain schema stats shows ≤16 distinct types

Phase 5: Post-migration

Anything that used --type article keeps working post-unify if your CLI calls go through the expandTypeFilter helper (it expands article to media+subtype=article automatically). Direct SQL against pages.type needs updating to the canonical types.

Search queries get a small ranking signal: pages reached via slug_aliases (canonicals of one or more aliases) get a 1.05x boost. Visible via gbrain search --explain.

Rollback

Every retyped page preserves frontmatter.legacy_type = <original>.

Restore types in bulk (Postgres/Supabase deployments only; requires direct DB access):

UPDATE pages SET type = frontmatter->>'legacy_type'
WHERE source_id = 'default' AND frontmatter->>'legacy_type' IS NOT NULL;

On PGLite there is no SQL shell, so use the CLI surface instead: frontmatter.legacy_type persists per page, so individual retypes can be reverted through the normal put_page/CLI surface, and the soft-delete restore and pack-flip revert below work on every engine.

Page-to-alias and page-to-link source pages soft-delete with 72h TTL. Restore within that window:

gbrain restore <slug>

Revert the active pack flip:

gbrain schema use gbrain-base

Anti-patterns

• Don't run unify-types under autopilot. It's manual_only by design. Autopilot remediation should never silently change your taxonomy.
• Don't expect mapping_rules to cover every legacy type explicitly. Use the catch-all (*unknown*) for the long tail. Pages get retyped to note with legacy_type preserved.
• Don't rewrite body-text wikilinks. The slug_aliases table IS the resolver. [[old-redirect-slug]] keeps working via engine.resolveSlugWithAlias short-circuit.
• Don't bypass the dry-run. Always run --explain before applying. The trust delta is real.
• Don't run two unify jobs concurrently. The gbrain-unify db-lock serializes them; the second submission rejects with "already in progress."

Decision tree

Active pack already gbrain-base-v2?
  → Skip migration.

Custom pack with own mapping_rules?
  → Run --check --explain to see if your pack declares migration_from
    for the active pack. If yes, target_pack = your pack name.

Brain has many custom types not covered by gbrain-base-v2 mapping_rules?
  → The catch-all retype binds them to `note` with legacy_type preserved.
    Review by inspecting frontmatter.legacy_type after the migration.

Federated brain (multiple sources)?
  → Add --params source_id to scope the migration per-source. Each
    source can be migrated independently.

Worried about a specific cluster's mapping?
  → Fork gbrain-base-v2 (`gbrain schema fork gbrain-base-v2 my-pack`),
    edit mapping_rules in your fork, then target the fork.

Contract

Inputs:

• A brain on gbrain-base (or any pack with migration_from: gbrain-base-v2).
• Trusted local CLI access on the brain host: gbrain jobs submit grants the PROTECTED-handler opt-in itself for protected names; the remote MCP submit_job op cannot.
• ~10 min wallclock on a 186K-page brain.

Outputs:

• Pages retyped to canonical types with frontmatter.legacy_type preserved (per-page rollback signal).
• slug_aliases rows for concept-redirect pages (alias table IS the resolver — no link rewrite).
• Real links rows for edge-shaped pages (atom-partner-link, symlink, etc.).
• Active pack flipped to gbrain-base-v2 atomically at end of successful run.

Side effects:

• Source pages soft-deleted with 72h restore TTL (gbrain restore <slug>).
• One-time cache invalidation on KNOBS_HASH_VERSION bump (5→6); self-healing in cache.ttl_seconds.
• Query-time --type X alias-expands via expandTypeFilter (back-compat).

Failure modes:

• Concurrent submission rejected by the gbrain-unify db-lock; second call exits gracefully.
• Catch-all retype excludes page_to_link + page_to_alias source types (caught in E2E pre-merge).
• Phase failures abort the run before active_pack_flipped; partial state restorable via op_checkpoint resume.

Anti-Patterns

DON'T:

• Submit unify-types via the remote MCP submit_job op. PROTECTED handlers require trusted local callers (gbrain jobs submit on the brain host); remote MCP rejection is the intentional trust boundary.
• Edit mapping_rules in gbrain-base-v2.yaml to skip clusters you don't trust. Fork the pack instead (gbrain schema fork) so the source-of-truth migration stays consistent across brains.
• Run unify-types from inside an autopilot tick. The check is manual_only — autopilot deliberately never auto-fires it because pack upgrades are one-time consenting taxonomy decisions.
• Hard-delete soft-deleted source pages before the 72h restore window. Use gbrain restore <slug> first if rollback is needed.
• Assume frontmatter.legacy_type survives every roundtrip. The marker is canonical for the immediate post-migration window; downstream re-imports may overwrite it.

Output Format

Per phase, the handler emits to stderr:

[unify-types] phase=retype-explicit applied=N skipped=M  cost=USD  ttl=Ns
[unify-types] phase=retype-catch-all applied=N
[unify-types] phase=page-to-link converted=N pages soft-deleted
[unify-types] phase=page-to-alias aliased=N pages soft-deleted
[unify-types] phase=sync residual=N
[unify-types] active_pack flipped from gbrain-base to gbrain-base-v2

Final celebration summary to stderr:

═══════════════════════════════════════════════════════════
  gbrain-base-v2 migration complete
═══════════════════════════════════════════════════════════
  Before: 94 distinct page types
  After:  15 canonical types
  Retyped:      25,632 pages
  Aliased:       5,521 redirects → slug_aliases table
  Linkified:        65 ghost pages → real link rows
  Soft-deleted:  5,586 pages (restorable for 72h)
═══════════════════════════════════════════════════════════

For structured JSON, gbrain call get_job '{"id": <id>}' returns the job row; its result field carries the UnifyTypesResult shape with per_phase, pack_identity_after, active_pack_flipped (gbrain jobs get <id> prints the same result inline).

Reference

• Architecture: https://github.com/garrytan/gbrain/blob/master/docs/architecture/type-taxonomy.md
• Pack-upgrade mechanism: https://github.com/garrytan/gbrain/blob/master/docs/architecture/pack-upgrade-mechanism.md
• Issue: https://github.com/garrytan/gbrain/issues/1479