Summary

brain-aware planning (v1.52) advertises support for "any thin-client MCP" brain, but the
classifier lib/gbrain-local-status.ts freshClassify() probes via gbrain sources list --json — a local-DB-only command. Under thin-client mode this is intentionally not
routable, so the probe throws, the catch falls through to the defensive broken-config
default, and gen-skill-docs --respect-detection suppresses every brain block. A healthy
thin-client brain therefore never gets brain-aware planning.

Environment

• gstack v1.52.1.0 (current main, 070722a)
• gbrain 0.41.28.0, thin-client mode (remote MCP over HTTP; no local DB)
• Cause is OS-independent (first hit on Windows, but not platform-specific)

Reproduction

1. Configure gbrain as a thin-client (remote mcp_url, no local engine). Doctor is healthy:
   gbrain doctor --json --fast → {"mode":"thin-client","status":"ok",...}
2. The probe command is refused by design:

   $ gbrain sources list --json
   `gbrain sources` is not routable. sources commands manage local DB + config rows.
   Per-subcommand thin-client routing lands in v0.31.x. For now: use `sources_list` /
   `sources_status` MCP tools, or run on the host.   (exit 1)
   

3. gstack-config gbrain-refresh → ~/.gstack/gbrain-detection.json shows
   "gbrain_doctor_ok": true, "gbrain_mcp_mode": "remote-http", "gbrain_local_status": "broken-config"
   (healthy thin-client, yet broken-config).
4. bun run gen:skill-docs:user → planning SKILL.md files contain no brain blocks
   (gen-skill-docs.ts:50 un-suppresses only when gbrain_local_status === 'ok').

Root cause

In lib/gbrain-local-status.ts freshClassify(), once the on-PATH check (L224) and
config-exists check (L227) pass — both true for a thin-client — it probes with
execFileSync("gbrain", ["sources","list","--json"]) (L239). gbrain sources is local-DB-only
and explicitly not routable in thin-client mode, so the call throws; the stderr matches neither
"Cannot connect to database" nor "config.json", hitting the defensive broken-config
default (L260). The 5-state enum has no healthy-thin-client path and freshClassify never
checks the mode.

Suggested fix

Detect thin-client mode before the sources list probe and treat a healthy thin-client as
ok. The signal already exists: gbrain doctor --json --fast returns mode: "thin-client" +
status: ok|warnings, and detect already calls it in detectDoctor() (bin/gstack-gbrain-detect
:121) — the mode field is one parse away. detectMcpMode() independently returns remote-http
too. e.g. in freshClassify: if doctor reports thin-client mode and ok/warnings status, return
"ok" without running the host/local-DB-only sources list probe.

Workaround (until fixed)

Override detection externally, no gstack edits: confirm health via gbrain doctor, then write
"gbrain_local_status": "ok" into ~/.gstack/gbrain-detection.json (keeping the other real
detect fields) and run bun run gen:skill-docs:user. Must be re-run after every ./setup or
gstack-config gbrain-refresh, since both rewrite detection.json from the broken probe.

Secondary (minor)

scripts/resolvers/gbrain.ts generateGBrainContextLoad tells the agent to run
gbrain get_page "<slug>", but the thin-client CLI has no get_page subcommand
(Unknown command: get_page); the CLI verb is gbrain get. (get_page is the MCP op name.)