version: "0.28.0" title: "v0.28 — Takes + Think + Unified Model Config" status: published

v0.28 — Takes, Think, Unified Model Config

You upgraded from v0.27 (or earlier) to v0.28. New surface area:

Takes — typed/weighted/attributed claims stored as fenced markdown tables and indexed in Postgres. Four kinds: fact | take | bet | hunch. Holders: world | garry | brain | <slug>.
gbrain takes CLI — list / search / add / update / supersede / resolve.
Unified model config — models.default replaces per-phase dream.synthesize.model etc. Aliases: opus, sonnet, haiku, gemini, gpt. CLI flag --model <name> overrides per-call.
Per-token MCP allow-list — access_tokens.permissions.takes_holders controls which holders an MCP-bound token sees. Default: ["world"].
Three new MCP ops — takes_list, takes_search, think.
gbrain think op surface — registered now; pipeline lands in v0.28.x.

What the migration did automatically

gbrain apply-migrations --yes runs three orchestrator phases:

Schema verify — schema migrations v37 (takes + synthesis_evidence) and v38 (access_tokens.permissions JSONB) ran; tables are present. (v31 was claimed by v0.25's eval_capture_tables, v32+v33 by v0.26's OAuth + admin dashboard, v34 by v0.26.5's destructive-guard columns, v35 by v0.26.8's auto-RLS event trigger, and v36 by v0.27's subagent provider-neutral persistence before v0.28 landed.)
Backfill takes — walked every page in the brain, parsed any fenced  table found, populated the takes index in Postgres. Idempotent; safe to re-run.
Re-chunk TODO queued — ~/.gbrain/migrations/pending-host-work.jsonl gained an entry asking you to re-chunk pages that had takes BEFORE v0.28. The chunker now strips fenced takes content (so per-token MCP visibility actually works) but the strip rule only applies to NEW imports. Legacy pages still have takes content baked into their content_chunks rows — re-chunk them to apply the strip retroactively.

What you need to do as the host agent

1. Verify the upgrade

gbrain doctor

Expected: takes_backfill_complete and takes_fence_chunk_leak checks both green. If either is red, follow the doctor's fix hint.

2. Re-chunk pages with pre-v0.28 takes (one-time, cosmetic for fresh installs)

For brains with existing fenced takes content in pre-v0.28 markdown:

# Identify pages with takes content that's been chunked into content_chunks
gbrain doctor --check takes_fence_chunk_leak

# If the check is RED, re-import those pages so the new chunker rule applies.
# Run only on the affected slugs (the doctor output enumerates them):
gbrain extract takes --rebuild
gbrain sync   # picks up the markdown delta if any

Fresh installs (no pre-v0.28 takes content): nothing to do.

3. Migrate to the unified model config (optional, mechanical)

Old per-phase keys still work in v0.28 with a deprecation warning. v0.30 will remove them. Migrate when convenient:

# Old (deprecated):
gbrain config set dream.synthesize.model claude-sonnet-4-6
gbrain config set dream.patterns.model claude-sonnet-4-6

# New (one key controls everything):
gbrain config set models.default sonnet

# Per-op override (optional):
gbrain config set models.dream.synthesize opus

# Cleanup deprecated keys:
gbrain config unset dream.synthesize.model
gbrain config unset dream.patterns.model

4. Configure MCP token visibility (optional, security-relevant)

Existing tokens default to permissions.takes_holders=["world"] — they see public claims only, never private hunches. Tokens for trusted agents (e.g., your own OpenClaw deployment) need explicit broader visibility:

# List existing tokens
gbrain auth list

# Grant a token visibility into garry's takes (and brain-derived takes)
gbrain auth permissions <token-name> set-takes-holders world,garry,brain

# Or create a new token with the wider set up-front
gbrain auth create my-claude-desktop --takes-holders world,garry,brain

Tokens for third-party agents (claude.ai, public integrations) should keep the default ["world"] — Garry's hunches stay private.

5. Try the takes layer

# Add a take by hand
gbrain takes add people/alice-example \
  --claim "Strong technical founder I have ever met" \
  --kind take --who garry --weight 0.85 \
  --source "OH 2026-05-01"

# List
gbrain takes people/alice-example

# Search
gbrain takes search "technical founder"

# Resolve a bet
gbrain takes resolve people/alice-example --row 3 --outcome true \
  --value 50000000 --unit usd --source crustdata

Verify

gbrain doctor         # all checks green; schema_version >= 32
gbrain stats          # numbers stable

# If any step failed, file an issue:
# https://github.com/garrytan/gbrain/issues
# Include: gbrain doctor output, ~/.gbrain/upgrade-errors.jsonl if it exists,
# and which step broke. The migration is designed to be idempotent — safe
# to re-run after fixing the underlying issue.

What ships in v0.28.x as follow-ups

gbrain think synthesis pipeline (gather → RRF → cite → synthesize)
gbrain takes seed <slug> — LLM extracts claims from page prose
Dream auto_think + drift phases (opt-in)
Cross-page reference resolution in source columns

The op surfaces are registered in v0.28.0 so MCP/SDK callers can detect them; the pipelines fill in incrementally without breaking the contract.