fix: gstack-memory-ingest calls non-existent gbrain put_page CLI verb

[smithjoshua]

Symptom

/setup-gbrain Step 7.5 (transcript & memory ingest gate) writes 0 pages on a clean install. With --quiet removed, every page errors with:

Unknown command: put_page
Run gbrain --help for available commands.

Root cause

bin/gstack-memory-ingest.ts:767-779 shells out via:

const args = ["put_page", "--slug", page.slug, "--title", page.title,
              "--type", page.type, "--tags", page.tags.join(",")];
execFileSync("gbrain", args, { input: page.body, ... });

The gbrain CLI (v0.18+ through current v0.26.7) does not expose a put_page subcommand. The verb is gbrain put <slug> and it accepts only <slug> positionally plus content via stdin or --content. There is no --title, --type, or --tags flag on the CLI. put_page is the MCP tool name; this looks like confusion between the MCP and CLI surfaces.

$ gbrain put --help
Usage: gbrain put <slug> [options]
Options:
  <slug>     Page slug (required)
  --content  Full markdown content with YAML frontmatter (required)

Repro (current main)

bun bin/gstack-memory-ingest.ts --bulk --include-unattributed --limit 1

Result: written: 0, failed: 1.

Confirmed on a clean install with gbrain v0.26.7 + gstack v1.26.3.0.

Fix

gbrain put reads tags/title/type from YAML frontmatter, so:

1. Switch to the correct CLI verb (put instead of put_page).
2. Inject title, type, tags into the frontmatter that buildTranscriptPage / buildArtifactPage already produce, instead of passing them as CLI flags.

Bundled ergonomic improvements in the same function:

• timeout: 30000 -> 60000. Auto-link reconciliation on dense transcripts hits 30s once the brain has a few hundred existing pages; the original limit caused timeouts in real backfills.
• maxBuffer: 16 MB. Without this, Node truncates gbrain's stderr at the default 1 MB and callers see only Command failed: with no detail. Cost real debugging time on this very bug.
• Surface stderr/stdout in the returned error message instead of the bare exception, for the same reason.

Tests

• bun test test/gstack-memory-ingest.test.ts -> 15/15 pass
• bun test on the three test files touching this code path (gstack-memory-ingest, gstack-memory-helpers, skill-validation) -> 362/362 pass

Validated end-to-end on a real machine: 122 Claude Code transcripts + 108 Codex transcripts written cleanly with this patch (via a temporary one-off wrapper that exercises the same gbrain put codepath this PR introduces).

Versions

• gstack: v1.26.3.0 (current main, db9447c)
• gbrain: v0.26.7
• Bun: 1.3.12
• macOS arm64
• Affects every release since v1.26.0.0 (where transcript ingest landed)

[metaWuming]

+1, confirming this on macOS / gstack v1.26.3.0 / gbrain v0.18.2 / PGLite.

Independent diagnosis matched yours exactly: gstack-memory-ingest.ts:762-784 shells out with put_page --slug ... --title ... --type ... --tags ... while gbrain ≥0.18.x only registers put <slug> --content (slug positional, metadata in YAML frontmatter).

Concrete impact data point on my install — mcp__gbrain__get_health reports page_count: 0, embed_coverage: 0 despite git-sync to my brain repo working fine. So the brain manifests load empty across all v1.26 skills, exactly as #1305 describes.

Your fix (swap to put + inject title/type/tags into frontmatter) matches Brett's Option A in #1305 and is the minimum-diff path. Hoping this lands soon — the headline v1.26 transcript ingest feature is non-functional on every fresh setup until then.

[chriskedz]

+1, independently hit and patched the identical bug on a clean install (gstack v1.26.3.0 / gbrain v0.27.0 / Supabase tier). Without this fix, bulk ingest writes 0/104 pages. With it, 104/104, 0 failures.

One edge case worth folding in before merge: the body.startsWith("---\n") branch only catches transcript pages (where buildTranscriptPage prepends the agent/session_id frontmatter). buildArtifactPage returns raw.slice(0, 200000) of a markdown file — and plenty of ~/.gstack/projects/*/learnings.jsonl, *-design-*.md, and ceo-plans/*.md don't start with ---. Under your current patch, those would still ingest, but title/type/tags get silently dropped, which breaks gbrain list --type learning and tag filters downstream.

Two-line addition handles it:

if (body.startsWith("---\n")) {
  // ...existing inject path...
} else {
  body = `---\ntitle: ${JSON.stringify(page.title)}\ntype: ${page.type}\ntags:\n${page.tags.map(t => `  - ${t}`).join("\n")}\n---\n\n` + body;
}

Verified shape end-to-end: artifact pages roundtrip through gbrain get <slug> with all three fields visible in the rendered frontmatter.

Otherwise this is exactly the patch we want — the headline v1.26 transcript ingest feature is non-functional on every fresh setup until it lands. Hoping for a merge soon.

[smithjoshua]

Closing — superseded by #1344, which Garry merged in v1.26.5.0 (2026-05-07). Same put_page → put fix plus the source-id validator bug we noticed in passing. Thanks for shipping the fix wave.

Tested end-to-end against my real transcript pile (228 files) using a temporary workaround script while #1328 was in review; the upstream v1.26.5.0 path replaces that workaround cleanly.