Summary

Fixes the v0.16.0+ jsonb double-encode bug in src/core/minions/handlers/subagent.ts that breaks dream synthesize's orchestrator slug-collection. Symptom: transcripts_processed=1, pages_written=0 even when child subagents successfully wrote pages to the DB.

Root cause

The four subagent_* jsonb writers (persistMessage, persistToolExec{Pending,Complete,Failed}) used engine.executeRaw (which routes to postgres.js's unsafe()) with JSON.stringify(value) plus $N::jsonb cast. Verified empirically:

P1 JSON.stringify(obj) + $::jsonb  →  jsonb_typeof='string'   (broken — current code)
P2 raw object         + (no cast)  →  jsonb_typeof='object'   (correct)
P3 raw object         + $::jsonb   →  jsonb_typeof='object'   (correct — this PR)
P4 sql.json(obj) via unsafe        →  jsonb_typeof='object'   (correct)

postgres.js v3 auto-encodes objects/arrays for jsonb columns. queue.ts:233 already passes raw objects for minion_jobs.data and was never affected. PGLite handles raw values identically (verified — all three patterns produce object jsonb on PGLite).

Symptom this resolves

collectChildPutPageSlugs in src/core/cycle/synthesize.ts:459 queries:

SELECT input->>'slug' FROM subagent_tool_executions WHERE ...

On string-typed jsonb the ->> operator returns NULL, so pages_written reports 0 even when child subagents successfully put_page'd. The reverse-write step that mirrors DB → markdown then has no targets, leaving real synthesized pages stranded in the DB. After this fix, the orchestrator picks up slugs correctly and the dream cycle closes the loop end-to-end.

Companion changes in this commit

• repair-jsonb extended from 5 → 8 columns (subagent_messages.content_blocks, subagent_tool_executions.{input, output}); each target guarded with to_regclass() so pre-v0.15 brains that lack the subagent_* tables don't throw at upgrade time.
• doctor's jsonb_integrity check extended to the same 8 columns with the same guard. Any user hitting this bug will see "X rows double-encoded ... Fix: gbrain repair-jsonb" instead of silent dream failures.
• Header comment of repair-jsonb.ts updated: the prior claim that the parameterized $N::jsonb form was always safe was wrong — safety came from queue.ts/sources.ts/etc. passing raw objects, not from the binding form itself.

Field evidence

On the developer's brain at the time of fix (production gbrain deployment, postgres engine):

• subagent_messages.content_blocks: 40/40 corrupt
• subagent_tool_executions.input: 39/39 corrupt
• subagent_tool_executions.output: 39/39 corrupt

gbrain repair-jsonb after this patch repaired all 118 rows. Subsequent gbrain dream --input <transcript> --json produced pages_written=2, reverse_write_count=2, child job completed, and the rendered markdown landed on disk.

Test plan

• bun run typecheck clean
• 245 tests pass across minions, subagent-handler, cycle-synthesize, cycle-patterns, repair-jsonb, migrations-v0_12_2, doctor
• End-to-end: fresh transcript → gbrain dream --input → DB has object-typed jsonb → orchestrator returns correct slugs → markdown lands on disk
• Reviewer should consider whether to ship a one-shot migration that auto-runs repair-jsonb for affected tables, or rely on gbrain doctor to nudge users — current PR ships the detection + repair tool but does not auto-trigger it

Upstream context

PR #525 (chmod fix, by another author) was unrelated to this jsonb work despite name overlap from local commits that mis-attributed it. PR #539 (open) does post-hoc UPDATE-based data repair in synthesize.ts/patterns.ts but does NOT patch the write path — this PR is the missing complement. They can land in either order: this PR stops the bleeding, #539 cleans existing brains.

🤖 Generated with Claude Code

^{Need help on this PR? Tag @codesmith with what you need.}

• Let Codesmith autofix CI failures and bot reviews