v1.40.0.0 fix wave: gbrain sync hardening (8 community PRs + migration)#1547
Open
garrytan wants to merge 9 commits into
Open
v1.40.0.0 fix wave: gbrain sync hardening (8 community PRs + migration)#1547garrytan wants to merge 9 commits into
garrytan wants to merge 9 commits into
Conversation
…1414) Cherry-picked from #1468 by 0xDevNinja and extended with the hostname-fold migration that codex review surfaced. Pre-fix `deriveCodeSourceId` hashed the absolute repo path alone, so two machines with identical home-dir layouts (chezmoi-managed dotfiles, ansible-provisioned VMs) derived the same id and clobbered each other's `local_path` in a federated brain. Last-writer-wins, with cryptic "Not a git repository" errors on the loser. Hash key is now `\${hostname}::\${path}`. Conductor worktrees on a single host stay distinct (path entropy unchanged within a host); cross-machine federations stop colliding. Migration (D1=B + codex refinements): every existing user has a pre-#1468 path-only-hash source id in their brain that no longer matches what `deriveCodeSourceId` produces. Without migration, the next sync registers a fresh source and orphans the old one. This commit adds: - \`derivePathOnlyHashLegacyId\` — separate helper for the pre-#1468 form. Distinct from \`deriveLegacyCodeSourceId\` (pre-pathhash v1.x form); both probes run. - \`planHostnameFoldMigration\` — feature-checks \`gbrain sources rename <old> <new>\` (exact argument shape, not just \`--help\`), gates on path-drift (skip migration if old source's \`local_path\` differs from current repo root), and falls back to register-new + sync-OK + remove-old when rename is unsupported. As of gbrain 0.35.0.0 the rename subcommand does not exist, so users go through the cleanup path; the rename path stays dormant until gbrain ships it. - \`removeOrphanedSource\` — called only AFTER new-source sync verifies page_count > 0. Closes the data-loss window codex flagged where "register new, remove old before sync" can wipe pages if sync fails. - \`sourceLocalPath\` — looks up a source's \`local_path\` from \`gbrain sources list --json\` for the drift gate. - Helpers accept an optional \`env\` parameter so tests can inject a gbrain shim via PATH without process-wide PATH mutation (Bun's spawnSync doesn't pick up runtime PATH changes). Pre-positions for commit 4's centralized gbrain-exec helper. - \`if (import.meta.main)\` guard around \`main()\` so the helpers can be imported for in-process unit tests. Tests cover: pure derivation, ids-match degenerate case, no-legacy short-circuit, path-drift skip path, rename path with shim, cleanup fallback when rename unsupported, cleanup fallback when rename call itself fails, source-lookup happy/missing/error paths. \`GSTACK_HOSTNAME\` env var is a test-only knob; production uses \`os.hostname()\`. Fixes #1414 Co-Authored-By: Claude <noreply@anthropic.com>
Cherry-picked from #1481 by drummerms and extended with the explicit HTTPS-remote regression case for #1357 (decision D2=A). `constrainSourceId` truncated the slug with `slug.slice(-tailBudget)`, which cut mid-word when the boundary fell inside a token. For a repo where the combined `prefix-org-repo-pathhash` exceeded 32 chars, this produced embarrassing artifacts like `gstack-code-kill-270c0001-c32152` (from `drummerms-av-sow-wiz-skill-270c0001`). Two changes carried from #1481, adapted for the #1468 hostpathhash: 1. `constrainSourceId` now walks hyphen-separated tokens from the right, accumulating whole tokens until adding the next would exceed `tailBudget`. When no token fits, falls through to the existing `${prefix}-${hash}` form. 2. `deriveCodeSourceId` now retries with `repo-only-hostpathhash` (dropping the org segment) when the full `org-repo-hostpathhash` triggers truncation. Keeps the repo name readable when it fits at all. Plus a new test asserting the source id is period-free for the exact HTTPS-with-.git remote shape from #1357 (`https://github.com/foo/bar.git`). canonicalizeRemote strips `.git`; the sanitizer strips any residual non-alnum. The test closes #1357 by pinning the property. Closes #1357 Co-Authored-By: Claude <noreply@anthropic.com>
Cherry-picked from #1508 by jasshultz, restructured per codex review #4 and #7 to widen scope and centralize the spawn surface. The bug: gbrain auto-loads .env.local from cwd via dotenv. When /sync-gbrain runs inside a Next.js / Prisma / Rails project whose .env.local defines its own DATABASE_URL (pointing at the app's local DB), gbrain reads that value instead of its own ~/.gbrain/config.json — auth fails, code + memory stages crash. This commit: - Adds lib/gbrain-exec.ts: buildGbrainEnv, spawnGbrain, execGbrainJson, execGbrainText, spawnGbrainAsync (the last one for memory-ingest's streaming gbrain import call). buildGbrainEnv seeds DATABASE_URL from ${GBRAIN_HOME:-$HOME/.gbrain}/config.json, returns a fresh env object (never the caller's by identity — codex review #11), and honors the GSTACK_RESPECT_ENV_DATABASE_URL=1 escape hatch. - Routes every gbrain spawn in bin/gstack-gbrain-sync.ts and bin/gstack-memory-ingest.ts through the helpers. Both files now own zero direct spawnSync("gbrain"|spawn("gbrain"|execFileSync("gbrain" call sites. - Threads buildGbrainEnv into the spawnSync("bun", [memory-ingest], ...) grandchild in runMemoryIngest (codex review #7). Without this, the parent fix is half-baked — the bun child inherits a clean env but needs DATABASE_URL pre-seeded too. spawnGbrainAsync inside memory-ingest provides defense in depth for standalone invocations. - Adds GBRAIN_HOME support — aligns with detectEngineTier (already honors GBRAIN_HOME) so all gstack-side gbrain calls agree on which config file matters. Resolves baseEnv.HOME first, then homedir(), so test injection works without process-wide HOME mutation. - Adds test/build-gbrain-env.test.ts: 10 unit tests covering all five env-seeding branches (seed from config / override caller / GSTACK_RESPECT escape hatch / missing config / unparseable config / no database_url field / GBRAIN_HOME path / object-identity guard / unrelated-vars preservation / idempotent-when-matches). - Adds test/gbrain-exec-invariant.test.ts: static-source check that greps both bin/gstack-gbrain-sync.ts and bin/gstack-memory-ingest.ts for direct spawnSync("gbrain"|spawn("gbrain"|execFileSync("gbrain"| execSync(...gbrain matches and fails the build if any are found. Refactor-proof against future contributors adding a new gbrain spawn without env threading. The invariant is intentionally narrow — only the two files where the DATABASE_URL bug actually hurts users are guarded. Migrating the spawn sites in lib/gbrain-local-status.ts, lib/gstack-memory-helpers.ts, and bin/gstack-brain-context-load.ts is a follow-up. Co-Authored-By: Jason Shultz <jasshultz@gmail.com> Co-Authored-By: Claude <noreply@anthropic.com>
The v1.29.0.0 changelog promised .gbrain-source would be added to the consuming repo's .gitignore so the per-worktree pin stays local, but the change actually only added it to gstack's own .gitignore. Without the consumer-side entry, the pin gets committed and Conductor sibling worktrees of the same repo + branch step on each other's pin every time anyone commits. Add ensureGbrainSourceGitignored after a successful gbrain sources attach in runCodeImport. Idempotent on repeat runs (line-trim match), creates .gitignore if missing, logs a warning and continues on permission errors so a read-only checkout doesn't fail the sync. Gate the top-level main() call behind import.meta.main so tests can import the helper without triggering a full sync run on module load. Tests in test/gbrain-source-gitignore.test.ts cover: create-when-missing, append-without-trailing-newline, append-with-trailing-newline, idempotent on repeat, recognize whitespace-surrounded entry, no-throw on read-only file. 6 pass.
Supabase free-tier cold-starts can push `gbrain sources list --json` past 10s (observed 14.5s in the wild), causing probeSource() to throw ETIMEDOUT during /sync-gbrain code stage even though the underlying CLI was healthy. Matches the 30s ceiling already used by `sources add` / `sources remove` in the same file. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
#1452) Cherry-picked from #1465 by genisis0x and extended with the v1.40.0.0 upgrade migration that codex review #5 surfaced. #1465 alone only patches bin/gstack-artifacts-init, which means fresh installs and re-inits pick up the new pattern. But existing users who already ran v1.38.1.0 have a `.migrations/v1.38.1.0.done` marker — that migration won't re-run no matter what we change. So their installed `.brain-allowlist`, `.brain-privacy-map.json`, and `.gitattributes` stay without the new pattern, and `/plan-eng-review` artifacts continue to silently drop out of their federation queue. This commit: - bin/gstack-artifacts-init: adds projects/*/*-eng-review-test-plan-*.md to the three managed blocks. v1.38.1.0 covered design + test-plan; this completes the set for /plan-eng-review. - gstack-upgrade/migrations/v1.40.0.0.sh: targeted in-place repair for existing installs. Same idempotent jq-based shape as v1.38.1.0. Adds the new pattern to .brain-allowlist (before the USER ADDITIONS marker), .brain-privacy-map.json (as class=artifact), and .gitattributes (as merge=union). NEVER commits + pushes — the user controls when the patches ship to their federated artifacts repo. - test/artifacts-init-migration.test.ts: 5 new tests covering the v1.40.0.0 migration applied on top of a post-v1.38.1.0 state, jq patching, gitattributes append, idempotent re-run, and done-marker write when files are missing entirely. Co-Authored-By: Claude <noreply@anthropic.com>
…stall probe Cherry-picked from #1487 by genisis0x and extended with the post-install subcommand probe per T6 / codex review #19. `bun install` in $INSTALL_DIR fails on Windows MSYS/MINGW/Cygwin shells because gbrain's native postinstall script mis-parses path arguments and aborts with a non-zero exit, breaking gstack-gbrain-install for Windows users running git-bash/MSYS2. The package installs cleanly without scripts. This commit: - Adds Windows shell detection via `uname -s` matching MINGW*/MSYS*/CYGWIN*/Windows_NT (#1487's case statement already covers all four — codex review #18 confirmed MINGW* is included). Windows paths get `bun install --ignore-scripts`; macOS and Linux unchanged. - Adds a post-install probe of `gbrain sources --help`. `gbrain --version` already runs (D19 PATH-shadowing validation), but version success doesn't prove the subcommand surface is reachable — and `--ignore-scripts` may have skipped artifacts that subcommands need. Probe failure logs a clear warning (with Windows-specific remediation pointing at re-running `bun install` outside MSYS) but does NOT exit non-zero; users may still get value from gbrain even if the probe fails transiently. Refs #1271 Co-Authored-By: Claude <noreply@anthropic.com>
Bumps VERSION 1.39.2.0 → 1.40.0.0 (MINOR — substantial gbrain capability hardening across sync pipeline, install path, federation allowlist; ~600 net LOC added across 8 community PRs + plan-review refinements). CHANGELOG entry follows the release-summary format: two-line headline, lead paragraph, "numbers that matter" with before/after table across 8 user-visible surfaces, "what this means for builders" closer, itemized Added/Changed/Fixed/NOT fixed/For contributors sections. Per-commit contributor credits: 0xDevNinja, drummerms, Jayesh Betala, Jason Shultz, genisis0x. Also names NikhileshNanduri and realcarsonterry in the wave's "Fixed" section for independent submissions of the .gbrain-source gitignore bug. Co-Authored-By: Claude <noreply@anthropic.com>
This was referenced May 16, 2026
E2E Evals: ✅ PASS2/2 tests passed | $.28 total cost | 12 parallel runners
12x ubicloud-standard-8 (Docker: pre-baked toolchain + deps) | wall clock ≈ slowest suite |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
6 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Consolidates the 8 highest-volume gbrain-sync bugs from the open backlog into one release. Each contributor's PR is credited in commit metadata; codex outside-voice review surfaced 7 additional P1 refinements (migration sequencing, env scope widening, allowlist upgrade migration) that landed alongside the cherry-picks.
Closes #1384, #1414, #1357, #1452 (follow-on), #1386 (partial — Windows ingest at scale remains separate work), #1271 (follow-on).
Supersedes #1521, #1501, #1464, #1481, #1485, #1508, #1507, #1465, #1487, #1463 — closing those PRs after merge with credits to each contributor.
What lands
sources renameand register-new + sync-OK + remove-old fallback. Path-drift detection skips migration when the old source'slocal_pathdiffers from the current repo root.skill→kill. Repo-only fallback when org prefix forces overflow. Folds in deriveCodeSourceId produces invalid slug for github HTTPS remotes (contains ".", exceeds 32 chars) #1357 (HTTPS-with-.gitslug period regression test).commandbuiltin (fix(gbrain): probe CLI without command builtin #1485, jbetala7) — works on Windows shells.lib/gbrain-exec.tscentralized helper. Threads env through every gbrain spawn AND into thebun gstack-memory-ingest.tsgrandchild. Static-source invariant test fails the build if a future contributor adds a direct gbrain spawn outside the helper. HonorsGBRAIN_HOME;GSTACK_RESPECT_ENV_DATABASE_URL=1escape hatch..gbrain-sourcein consumer.gitignore(fix(gbrain-sync): add .gbrain-source to consumer repo .gitignore #1521, genisis0x) — Conductor sibling worktrees stop stomping each other's pin. Three other contributors independently filed this fix (NikhileshNanduri fix(gbrain-sync): append .gbrain-source to consumer repo's .gitignore after attach #1501, realcarsonterry v1.33.3.0 fix: add .gbrain-source to .gitignore and create utility to ensure it in consumer repos #1464); credited in CHANGELOG.gbrain sources --helpprobe that flags missing native artifacts with Windows-specific remediation.Test plan
bun testgreen (full free suite)bun run test:gate— runs as part of full suite, all greenbuildGbrainEnvunit tests — 10 cases covering seed/override/escape-hatch/missing/unparseable/GBRAIN_HOME/object-identity/preservation.gbrain-sourcegitignore tests — create / append / idempotent / whitespace / read-only checkouthttps://github.com/foo/bar.gitcaseskill→killregression pinnedspawnSync("gbrain"/spawn("gbrain"/execFileSync("gbrain"outsidelib/gbrain-exec.tsPlan + reviews
Plan:
/Users/garrytan/.claude/plans/ok-now-lets-look-stateless-plum.md(eng-reviewed + codex-reviewed; both CLEAR).🤖 Generated with Claude Code