Summary

/sync-gbrain --full against a GitHub-hosted repo fails at the code stage because deriveCodeSourceId generates source IDs that violate gbrain's [a-z0-9-]{1,32} validator.

Command / context

• gstack: v1.26.3.0
• gbrain: v0.27.0 (Postgres engine, Supabase pooler)
• OS: macOS Darwin 25.4.0
• Repo: https://github.com/kengwei/unerase
• Invocation: /sync-gbrain --full (Claude Code skill)

Observed

$ bun run ~/.claude/skills/gstack/bin/gstack-gbrain-sync.ts --full

[gbrain-sync] mode=full engine=unknown

gstack-gbrain-sync (full):
  ERR   code         source registration failed: gbrain sources add gstack-code-github.com-kengwei-unerase failed: Invalid source id "gstack-code-github.com-kengwei-unerase". Must be 1-32 lowercase alnum chars with optional interior hyphens (e.g. "wiki", "yc-media").
 (0.7s)
  OK    memory       ingest pass complete (29.7s)
  OK    brain-sync   curated artifacts pushed (0.2s)

  2 ok, 1 error, 0 skipped

Memory + brain-sync stages run fine. Only the code stage fails, so the source never gets registered and code indexing for the repo never happens.

Root cause

deriveCodeSourceId() in ~/.claude/skills/gstack/bin/gstack-gbrain-sync.ts (around line 170 in v1.26.3.0):

function deriveCodeSourceId(repoPath: string): string {
  const remote = canonicalizeRemote(originUrl());
  if (remote) {
    return `gstack-code-${remote.replace(/[\/\s]+/g, "-").replace(/-+/g, "-")}`;
  }
  // fallback...
}

For https://github.com/kengwei/unerase.git:

• canonicalizeRemote() → github.com/kengwei/unerase
• Regex /[\/\s]+/g replaces slashes and whitespace but leaves dots
• Result: gstack-code-github.com-kengwei-unerase — 38 chars AND contains .

gbrain's sources add validator requires [a-z0-9-]{1,32} so this fails. Affects essentially every github.com/<org>/<repo> remote.

Expected

deriveCodeSourceId should always produce IDs that satisfy gbrain's [a-z0-9-]{1,32} validator, regardless of the remote URL shape.

Suggested fix

Two viable approaches:

// Option A: include dots in the replace, then truncate
return `gstack-code-${remote.replace(/[\/.\s]+/g, "-").replace(/-+/g, "-")}`.slice(0, 32);

// Option B: strip the host, use org-repo only
const orgRepo = remote.split("/").slice(-2).join("-");  // "kengwei-unerase"
return `gstack-code-${orgRepo}`.slice(0, 32);

Option B is more readable and produces stable IDs across hosts (gitlab.com/kengwei/unerase and github.com/kengwei/unerase would collide, but that's vanishingly rare in practice). Option A preserves the host distinction.

Either option should still wrap with a final length-bound or hash-truncate to handle pathological long org/repo names.

The fallback branch for repos without an origin (line ~174) already does .replace(/[^a-z0-9-]+/g, "-") correctly — that pattern would work as a model for the with-remote branch too.

Workaround

Register the source manually with a short ID:

gbrain sources add gstack-code-<repo-basename> --path "$(pwd)" --federated
gbrain sync --source gstack-code-<repo-basename>

Related

Filed alongside garrytan/gbrain#457 (gbrain's first-sync path doesn't honor --strategy code even when the source ID is valid). The two bugs stack: gstack-side ID generator fails first; if it didn't, the gbrain-side first-sync would still skip code files. End-user impact is the same in either case (no code indexing).