[danielss-dev]

https://linear.app/daniels-dev/issue/DAN-89

Default interval

Daniels chose 10 minutes. That is the shipped default, not a placeholder. The single constant is DEFAULT_AUTO_FETCH_INTERVAL_MIN in ui/src/lib/autoFetch.ts. Settings → Git still offers Off, 5 min, 10 min, and 30 min, and each open repository can opt out.

What Daniels should check live

Open two repositories. From another machine, push a commit to one remote. Wait one 10-minute interval. The behind-count for that branch should update, and a notice should appear (N new commits on origin/<branch>). The other repository fetches a few seconds later. Confirm a background fetch never flashes a console window on Windows. A remote with no saved credential should fail once, quietly, and then wait instead of prompting every tick.

Proof

• Scheduler interval, no second concurrent fetch for the same repo, and skip while a manual fetch/pull is running: ui/src/lib/autoFetch.test.ts (fires at the interval…, does not start a second fetch…, never runs two fetches…).
• One tick after a push to a bare remote updates the behind-count and the new-commit count the notice uses; HEAD, porcelain status, git ls-files -s, git write-tree, and file bytes stay the same: background_fetch_one_tick_reports_new_commits_without_touching_worktree. Notice text: does not spawn a fetch until the interval, then surfaces a new-commit notice once.
• Auth-required remote, no cached credential, no prompt, one quiet failure, then backoff: background_fetch_auth_failure_does_not_prompt (HTTP 401 plus a credential helper that would block) and records one quiet auth failure and does not spawn again until backoff ends.
• Setting Off spawns no background fetch: spawns no background fetch when the setting is off and spawns nothing when the interval is off… (wakeAt is null, the fetch spy is never called).
• Windows console: background fetch goes through prepare_git_command → crate::git_command() → CREATE_NO_WINDOW (background_fetch_command_uses_hidden_git_and_cannot_prompt). On Windows, console_probe::hidden_console_child_has_no_window re-executes the test binary and checks GetConsoleWindow() is null.
• Local: cargo clippy -p strand-core -p strand-tauri -p strand-ops -p strand-headless -p strand-azdo-protocol -p strand-azdo -- -D warnings passed; pnpm --filter ./ui exec tsc --noEmit passed; pnpm --filter ./ui test (auto-fetch and settings) passed; cargo test -p strand-core --lib background_fetch passed (4 tests).
• CI on https://github.com/danielss-dev/strand/actions/runs/39228389030 is green: Frontend (tsc + vitest), Rust (test + clippy) on ubuntu-22.04 and macos-latest, Azure DevOps Server helper (Windows), and Native agent review (Windows).

Fetch only. No auto-pull, no fetch of closed recent repos, no OS notification.

Review fixes

• Background fetch keeps the user's core.sshCommand / GIT_SSH_COMMAND / GIT_SSH. It only forces ssh -o BatchMode=yes when no ssh command is set.
• Background fetch passes -c core.askPass=, so a configured askpass helper can't prompt.
• The manual-op skip follows the repository the op started on, not the tab that is active when the tick runs.
• After a fetch, the refresh goes through the store's coalesced refreshUpstream. That refreshes meta per tab, and refs through refreshRefs when the repo is active.
• A fetch that printed nothing and brought no commits writes no activity entry.
• The 10-minute default wording (autoFetch.ts:6) is unchanged, pending confirmation.
• Tests:
  • Rust: background_fetch_keeps_the_users_core_ssh_command, plus askPass assertions.
  • UI: autoFetchIsNoop, and a store test for refreshUpstream.
• Round 2:
  • Tests now cover the busy-repo skip: a busy non-active repo is skipped while the active one fetches, a tab switch mid-op keeps the skip, and the op ending clears it. They run through the extracted manualOpPath / isManualBusyRepo (no behaviour change).
  • The background-fetch command tests run in a fresh temp repo, with the ssh probe blind to GIT_SSH_COMMAND, GIT_SSH and global/system config.

Merge order 157 → 153 → 155 → 154 → 156 → 158: no later PR in the stack touches these files, so none needs adapting for this one.