v0.39.0

Features

• Git-branch artifact backend (t1231_1): ait artifact can now keep artifact blobs on a dedicated branch of your repository (the gitbranch backend), shared through origin. ait artifact gitbranch-migrate --from <old> --to <new> moves a store to a new branch and leaves the old one readable but closed to writes. If it fails partway, it prints a command to resume.
• Artifacts tab in Settings (t1231_2): ait settings has a new Artifacts tab (k) where you choose the default artifact backend and the gitbranch branch name. Branch names are checked before saving, and changing one shows the exact migrate command to run.
• Test-map variant axes (t1853_4): Runners can now declare a variant axis, so one test unit runs once per value (such as a platform or flavor). The new ait testmap axes command lists, explains and checks the declared axes. A change to a value that is not declared never selects fewer tests.
• Test-map scopes, areas and classify (t1853_5): Test annotations can now give a test a kind and scope it to named areas. ait testmap areas lists areas and imports them from a codemap, and ait testmap classify --suggest proposes kinds. scan records the scoped rows, and check reports dead, conflicting or invalid scopes.
• Test-map adoption tables (t1853_6): Covers lines written by annotate and verify now get provenance rows, written before the test file. ait testmap check prints SEEDED: / ADOPTED: summaries and reports shadowed, rejected or orphaned rows.
• Test-map cost ledger (t1854_4): ait testmap run now records how long each test unit takes. The new ait testmap costs command reports per-unit estimates, the full-suite p95 and a suggested gate timeout. Only runs on a tree that can be trusted count toward these numbers.
• Gate verifier errors and environment (t1859_1): A build or test command run by a gate can now report an error, which is distinct from a failure: exit 75 means refused, exit 3 means errored. Commands also receive AIT_GATE_TASK_ID, AIT_GATE_RUN_ID and AIT_GATE_RESULT_FILE. A one-line result written to that file appears in the gate ledger, cleaned and capped at 200 characters.
• Scratchpad notes store (t2073_1): Each task can now have a private notes store, kept out of git and deleted automatically when the task is archived.
• ait scratchpad notes editor (t2073_2): Added ait scratchpad, a small notes editor for a task. It autosaves shortly after you stop typing, saves when you quit (ctrl+q) or close the pane, and opens read-only if another pad already holds the notes.
• Notes pane from minimonitor (t2073_4): Press o in ait minimonitor to open a notes pane bound to the agent you're following. It closes with its agent and never appears in agent lists. Its position and height are set with tmux.scratchpad_position and tmux.scratchpad_height.

Bug Fixes

• ./ait git pull cleans up after itself (t1739): ./ait git pull now always runs without autostash. It carries your -c options through every step and warns when it cannot confirm that the pull finished cleanly. It refuses options it cannot handle safely: interactive rebase, --edit, --autostash, and -C / --git-dir / --work-tree.
• ./ait git -c … push works (t2063): Pushes with options in front, such as ./ait git -c http.proxy=… push, now go through the normal best-effort push with your config applied. A repository config that git cannot read is reported as a failure, not as "no remote".
• Merge-mode pulls abort their own stopped merge (t2064): With pull.rebase=false, a ./ait git pull whose merge stops (on a conflict or a signing failure, for example) is now aborted, so the data worktree is not left mid-merge. A deliberate --no-commit pause is left as it is.
• Sync reports a failed remote check (t2070): When git remote itself fails, for example on a config git cannot read, ait sync and task-data convergence now report a failure with git's error. Before, they silently treated the repository as having no remote.
• Archive temp dirs always cleaned up (t2032): Commands that read a task from an archive bundle no longer leave temporary directories behind when they set up their own exit handlers. ait note and ait gate now refuse a task that exists only inside an archive bundle.
• Exit handlers see the real exit status (t2071): Cleanup steps added in front of a transaction's exit handler now receive the command's real exit status instead of a reset one.
• Remote drift check reports its own failures (t2061): When the remote drift check before implementation cannot run (EXTRACT_FAILED or an incomplete run), the workflow now says so and asks how to proceed. It no longer reads that as drift or as no drift, and the deferred-plan marker is cleared.
• Drift check status survives errexit (t2085): The remote drift check step now records its exit status even under set -e, so a failing check reports its result instead of stopping early.
• ait ide restores every project's frozen agents (t2039): ait ide now offers frozen agents from every registered project, grouped by project with a prompt for each, and every restored agent gets its minimonitor. Two minimonitors are no longer spawned for the same window.
• Task title on frozen agent rows (t2042): Frozen agents in ait monitor and ait minimonitor now show their task title, just like live ones.
• Restored agents can be frozen again (t2045): An agent whose restore was confirmed during reconcile no longer refuses a new freeze with "already frozen".
• Restore checks that frozen marks were cleared (t2062): A restore now confirms that the pane's frozen stamp and attempt mark were actually removed, and reports any that remain. Reconcile repairs a stale stamp on a live agent, so the agent can be frozen again.
• Attempt mark cleared after a retried reconcile (t2066): A restore confirmed after an earlier failed reconcile attempt now also clears the pane's restore-attempt mark.
• Frozen-agent records read exactly (t2067): Restore, reopen and drop now read stored agent records exactly as written. Values with spaces, newlines or | (such as a window name padded with spaces) are no longer changed.
• Codex freeze for daemon-routed agents (t2046): Freezing a hand-started Codex agent that runs through Codex's shared app-server daemon no longer keeps a session id it cannot verify. Such agents come back by re-picking the task, as the freeze docs now explain.
• Frozen restore keys only in the viewer (t2054): You now restore or re-pick a frozen agent only from ait frozenagent. The minimonitor's R and the monitor's p are gone, and the monitors point you to the viewer instead. Picking a task while following a frozen agent no longer offers to kill it.
• Monitors stop treating frozen agents as running (t2057): The pick dialog no longer warns that a frozen task is "already running", and picking the next task never kills a frozen stand-in. If the frozen state cannot be checked, nothing is killed.
• Frozen agents protected at kill and restart (t2065): An agent that gets frozen while a kill or restart confirmation is open is no longer killed, which would have lost its capture. Applink refuses remote kills of frozen panes.
• Restart never starts a duplicate (t2075): In ait monitor, if the old agent cannot be killed during a restart, you get a warning and nothing new is launched.
• Failed sibling-pick kill reported (t2078): When picking a sibling task cannot kill the old agent, ait monitor and ait minimonitor now say so and keep focus on that agent.
• Helper panes close with their agent (t2088): Killing an agent directly from the monitors, or dropping a frozen agent, now also closes its shadow and scratchpad panes. Each kill re-checks the pane first, so a pane that was re-bound or reused in the meantime is left alone.
• Shadow phase stays on its own tmux server (t2043): A shadow's phase lookup no longer picks up a value from a pane on another tmux server that happens to have the same pane id.
• Minimonitor returns to its width (t2050): When its window grows, the minimonitor pane now goes back to its configured width instead of staying wide.
• Trail card double-click crash fixed (t2089): Double-clicking a parent card in ait trails or the board's By-Trail view no longer crashes with an expanded_tasks error.
• Kotlin shebang scripts scanned correctly (t2030): ait testmap now skips a leading #! line in .kt and .kts files instead of letting it hide the file's dependencies. A file with lone carriage-return line endings is flagged as one the scanner cannot analyse.
• Test-map runs from a subdirectory registry (t2080): When aitestmap/ lives in a subdirectory, ait testmap run now records modified files relative to that directory, so such repositories get cost estimates. Changes outside the registry root still keep a run out of the estimates.
• Count checks pass on macOS (t1751): Several tests no longer fail on macOS because BSD wc -l pads its counts with spaces.
• Test suites no longer leak temp dirs (t2031, t2055): The docs-vocabulary test and 20 more test suites no longer leave fixture directories in the temp folder after each run. A repo-wide check stops new leaks.
• Pipefail test checks no longer flaky (t2053, t2056): Test checks in 19 files no longer fail at random because grep -q exits early and the writer feeding it dies of SIGPIPE under pipefail.
• Stale test guards fixed (t2060, t2079): The merge-driver guard now finds the driver wherever it is called. Three tests now pass the cd-guard lint.

Improvements

• Crew git can't hang or prompt (t2024): Every crew fetch and push now has a timeout (AIT_CREW_GIT_TIMEOUT, 10s by default) and can no longer stop on a password or SSH prompt. On a timeout, brainstorm's add-work reports the state it left behind. Release blog posts are now dated by the exact release time, so a late-night release no longer gets a post dated in the future.
• Crew branch deletion can't hang (t2036): Deleting crew branches on origin (crew cleanup, brainstorm delete, and the brainstorm TUI) now times out instead of hanging on a remote it cannot reach.
• Sync and drift check use the shared git timeout (t2035): ait sync, the remote drift check and release lookups now run git through the shared time-limited runner. It kills hung transports and blocks credential prompts. Set AIT_SYNC_ALLOW_PROMPT=1 to allow prompts for ait sync.
• Codex restores confirmed from the open session (t2040): A restored frozen Codex agent is now confirmed by the session it opens, so the restore finishes quickly without sending a prompt. A late hook acknowledgement is never refused.
• Session hook records only its own pane (t2047): The SessionStart hook now records a session only when it fires from inside its pane's own process tree. Codex agents routed through the shared app-server daemon are no longer recorded against the wrong pane.
• Frozen-agent viewer re-pick key is now P (t2049): In ait frozenagent, re-pick moved from p to P.
• Shadow >q helps you decide (t2094): The shadow's >q is now "Help me decide what to do now". It explains the situation in simple words, then gives options and a suggestion when a question is on screen, or next moves and a recommendation otherwise. >e now only explains.
• Shadow concern blocks labelled (t2072): Each shadow concern block now has a line above it naming the kind of review: plan challenge, plan assumptions, error diagnosis or implementation.
• One rule for helper panes (t2073_3): Shadow and scratchpad panes are now recognised by one shared rule, so they never show up as agents. The minimonitor's sibling fallback skips them too.
• Test-map eligibility covers implicit inputs (t2081): Cost tracking now accounts for files a test reads without naming them, such as Python __init__ / conftest files, Go module and package files and Gradle scripts, and for dynamic source / exec. A run is never counted toward estimates when such an input has changed.

Documentation

• Artifact docs (t1231_3): New concept and command pages cover ait artifact and its local, dir and gitbranch backends. The settings docs now describe the Artifacts tab and list all nine tabs.
• Settings key reference completed (t2068): The settings reference now lists the Profiles and Shortcuts tab keys.
• Scratchpad docs (t2073_5): New pages document the scratchpad notes pane, how to open it from the minimonitor, its keys and how autosave works.

Performance

• Faster frozen-agent listing (t2041): Listing frozen agents across projects now reads the session store once instead of once per agent. With 100 agents, this takes about 0.2s instead of 6s.

Tests

• Exhaustive merge-driver guard (t2069): The merge-driver guard test now catches every call to the driver in the framework's shell scripts. Any call it cannot classify fails the test.