Grok Build 1.0.42

• [new] grok worktree create [NAME] builds the same worktree grok -w would, without opening a session. It prints only that directory on stdout (cd "$(grok worktree create my-fix)" works). Default is HEAD plus your uncommitted changes; --ref starts from a clean branch, tag, or commit. If the checkout does not contain the directory you ran from, it prints the worktree root.
• [new] grok clone works on Windows via ProjFS. Enable the Client-ProjFS optional feature (grove doctor reports it). Build-output dirs such as target and node_modules redirect out of the projected tree as NTFS junctions ([redirects] windows_junction, on by default).
• [new] Memory-v2 can run Dream in batches (memory_v2.batch_dream_enabled, off by default and not remotely enableable) with caps on run time, calls per batch, and note bytes. memory_v2.compact_index_enabled injects a titles-only memory index so more topics fit the 8 KiB budget; a local value wins over remote settings.
• [changed] With a custom models host, the startup model-list fetch authenticates with XAI_API_KEY (Authorization: Bearer) and never with your grok login session. If auth_provider_command is set and XAI_API_KEY is not, that provider token is sent instead; otherwise a missing key fails the fetch. Inference still uses each model's api_key / env_key. With an external auth provider and models_base_url / models_list_url, that list is what the picker offers, and models.allowed_models in user config is ignored (it still applies from requirements.toml).
• [changed] /learn writes run artifacts under ~/.grok/learn/runs instead of the OS temp directory, and deletes runs older than 30 days, so Windows temp cleaners no longer wipe the report before you curate it.
• [fix] /learn will not launch on fewer than five kept sessions, stops spending when every agent in a map wave fails (quota exhaustion or API outage) so you can resume later, treats a silent hour with no files changing as a stalled phase, and moves a leftover ~/.grok/workflows/learn-traces.rhai aside so it cannot shadow the bundled workflow after an upgrade. The report step writes report.md with edit tools instead of shell heredocs, and nothing is retired or trashed until the report is on screen.
• [changed] On a WinGet install, grok update prints winget upgrade --id xAI.GrokBuild -e and does not change the binary. Quit Grok first, then upgrade with WinGet. The package tracks stable and can lag a few days.
• [changed] Permission mode /auto is labeled Auto-review in the TUI, Shift+Tab cycle, slash-command help, and dashboard.
• [changed] Memory records observations after each turn, including a turn you stop before it finishes.
• [changed] Restoring a stashed draft (Ctrl+S / Alt+S) places the cursor at the end of the restored text.
• [changed] The fullscreen subagent transcript opens from a scrollback block or the tasks pane, not from a dashboard row.

Under the hood

/learn's map-reduce workflow fans mappers in waves and pauses with an infra reason when a whole wave fails, so finished work replays from the journal on resume. state.py wait can exit 4 (phase_stalled) after --phase-stall-min (default 60) with no file changes under the run directory. Skills /learn creates include origin: learn in frontmatter. Shell completions cover grok worktree create.

Upgrade notes

• WinGet installs: quit Grok, then winget upgrade --id xAI.GrokBuild -e. Do not expect grok update to replace the binary.
• Custom models: keep XAI_API_KEY set for the model-list fetch; per-model keys are only for inference. If you use auth_provider_command with a custom models endpoint, the picker follows that list.
• /learn reports now live under ~/.grok/learn/runs. A stale ~/.grok/workflows/learn-traces.rhai may be moved to ~/.grok/learn/trash/ so the bundled workflow is used.