v0.38.0

Features

• Test-map registry with scan and check (t1853_1): Added the test-map registry (the aitestmap/ files) and the ait testmap scan and ait testmap check commands. scan reconciles the registry with the test units found in the repository; check validates it (ownership, duplicate rules, expired waivers, drift) and reports problems by severity, with --strict failing on more of them.
• Test-map annotations with annotate and verify (t1853_2): Tests can declare what they cover in testmap: comment annotations in bash, Python, Go and Kotlin files. ait testmap annotate adds a stamped covers line and ait testmap verify re-stamps it after the source changes, touching only lines proven to be comments and keeping the file's mode and line endings.
• Test-map dependency scanners (t1853_3): The test-map engine now works out which files each test depends on, scanning bash, Python, Go, Kotlin, Java and Gradle sources plus external scanner plugins, with results cached by file content. ait testmap annotate --suggest proposes missing covers lines, and --from-body writes them for one test member (--dry-run previews).
• Test-map run command and runners (t1854_1): Added ait testmap run, which runs test units through project runners declared in aitestmap/runners.yaml, with per-unit timeouts and cleanup of every process a runner starts. Each run records its start and end commit and the files changed; scan and check now take their unit lists from these runners.
• Cross-repo blockers on trail cards (t1957_1): Trail cards in the board's By-Trail view and in ait trails now show when a member is blocked by a task in another registered project. Members from another project show their live status, or UNREACHABLE, from a background check.
• Open a trail by handle (t1957_2): ait trails --trail <handle> opens that trail at startup, and a running ait trails in tmux can be told to switch to a given trail.
• Find linked trails in other projects (t1957_3): Added the search behind the linked-trail jump: it finds trails in other registered projects that share members or cross-repo dependencies with the current one, and says so when it could not fully search a project instead of reporting no links.
• Jump to a linked trail (t1957_4): Press L in the board's By-Trail view or in ait trails to list linked trails in other projects. Picking one starts that project's tmux session if needed, opens the trail in its ait trails, and switches you there.
• Prune dimensions in brainstorm (t1983): On the Browse tab of ait brainstorm, press A → Prune Dimensions to create a child node that keeps only the dimensions you choose — the top N of a ranked list or a hand-picked set. No agent runs; the new node becomes the subgraph's HEAD and the source node is left unchanged.

Bug Fixes

• Worktrees no longer left dirty by .aitask-data (t1711): Linked worktrees no longer show .aitask-data as untracked, which blocked worktree removal and task archiving. New installs get the correct ignore rule, and ait setup repairs existing ones.
• Setup .gitignore commits stay scoped (t1959): The .gitignore steps in ait setup now commit only .gitignore, so work another session has staged no longer rides along. If you had already edited .gitignore, the new rule is left on disk uncommitted.
• Data-branch setup commits only its own changes (t1963): When ait setup moves task data to a separate branch, its commit on main contains only its own .gitignore rules and the removal of aitasks/ and aiplans/. A failed commit now warns instead of aborting setup.
• Setup protects pre-existing changes in more cases (t1972): ait setup now records your uncommitted changes whenever it might commit, so task-metadata edits on the data branch and a staged .gitignore in a repo with no commits are no longer swept into its commits. A staged rename in framework files is no longer turned into an added file.
• Upgrades leave your uncommitted edits alone (t1964): ait upgrade no longer puts your uncommitted edits to framework files or task metadata into its commits; it lists them as left alone. Upgrades now also apply the symlink-safe .gitignore repair, so linked task worktrees stop showing .aitask-data as untracked.
• Sync never auto-stashes or overwrites ignored files (t1792): ait sync now ignores git's auto-stash settings, so it no longer stashes and re-applies files held by live sessions. A behind-only fast-forward refuses with a clear error rather than overwrite local ignored files such as userconfig.yaml.
• Task-data pulls never autostash (t1987): With rebase.autoStash set, the task-data pull when you pick a task, and the retry after a rejected push, no longer stash and re-apply other sessions' task files. The fast-forward refuses to overwrite an ignored local file and tells you to move it aside and run ./ait sync.
• Crew pulls leave others' files alone (t2008): ait crew addwork, ait crew setmode and the crew runner no longer stash and re-apply other agents' uncommitted heartbeat, status and command files when rebase.autoStash or merge.autoStash is on.
• Crew branches published to origin (t2016): ait crew init, addwork, setmode, the runner and ait crew status set now push the crew branch to origin with an upstream. A failed push shows git's own error, and a repository with no origin skips the push quietly.
• Crew runner sees remote stop requests (t2018): The crew runner now pulls before writing its heartbeat, so a stop request pushed from another machine is merged and honoured. Pulls never leave a half-done merge in the crew worktree, and merging diverged crew histories needs git 2.38 or newer.
• Dir artifact store refuses a bad root (t1283): The dir artifact backend now stops with a clear error when its store directory is unset, missing or not marked as a store, instead of writing blobs into an empty, unmounted share. Run ait artifact init-store dir once to mark a store — existing dir stores need this step too.
• Safer shared-directory artifact store (t1960): The dir store no longer writes into the empty mount point when the share is unmounted mid-operation, and no longer reports a file as deleted when the store itself has vanished.
• Artifact rm cleans up moved artifacts' blobs (t1286): ait artifact rm now deletes leftover local blobs even after the artifact was moved to another backend. If a blob it would delete has local edits or unusual git index flags, it refuses with a repair hint and changes nothing.
• Artifacts with newlines in file names (t1967): ait artifact create and ait attach no longer fail on a file whose name contains a newline.
• Artifacts attach only to active tasks (t2004): ait artifact create now refuses an archived owner task and names the archived file or bundle in its error. ls, get and rm still accept archived owners.
• Attach refuses archived tasks (t2011): ait attach add now refuses an archived task with a clear error, instead of committing into the archived file or failing with a confusing "outside repository" error. ls and get still accept archived tasks.
• Archived-task lookups clean up temp files (t2023): Commands that read a task packed into an archive bundle no longer leave a temporary directory behind. ait attach rm and ait artifact rm on such a task now say plainly that it is packed and cannot be edited.
• Truncated archive bundles are skipped (t1978): The board, stats, codebrowser and trails no longer crash on a truncated .tar.gz archive bundle; they skip it and keep reading the others.
• Engine check handles unusual file names (t1970): Verifying the test-map engine binary during install no longer fails when its path contains a newline or backslash.
• Folding keeps folded tasks' history separate (t1962): Folding no longer copies the folded tasks' notes and gate runs into the primary task as if they were its own; they go into a separate ## Folded History section, and file paths in them no longer count as files the task touches.
• Older folded tasks keep their requirements (t1966): In tasks folded with the older layout, requirements after the notes or gate-run sections are counted again by trails and by overlap and scope checks, and folding such a task again keeps them as requirements.
• Board edits and syncs no longer overlap (t1947): The board refuses to open an editor while a sync is running, and a sync waits until the editor closes. A timed-out sync is stopped along with all its git processes.
• Board and syncer wait out running syncs (t1955): When the board suspends itself to run an agent or an interactive sync, it now refuses with a notice if a sync is running or a timed-out sync may still be changing task data. ait syncer blocks new syncs and recovery agents for that repo until such a sync is gone.
• Syncer interactive sync refreshes after finishing (t1961): In ait syncer, an interactive sync run from the conflict dialog without a separate terminal no longer errors when it finishes.
• Board task operations no longer dropped (t1994): A sync or agent launch started at the same moment can no longer silently cancel an archive, delete, rename or commit you confirmed on the board, which also left a loading overlay stuck on screen.
• Trail jumps and launches no longer stall (t1988): In the board's trail view and ait trails, a background sync or dialog command can no longer cancel a linked-trail jump, an agent launch or a refresh, which used to leave later jumps refused or R disabled until restart.
• Trail owners checked consistently (t1989): /aitask-trail can now refresh a trail whose owner task was archived, and creating a trail refuses an owner that is archived, in another project (proj#42) or nonexistent. After a refusal, the skill asks you to choose a live owner from the trail's own tasks.
• Brainstorm focus and switcher stay in your session (t1951): Opening a brainstorm from the board no longer jumps to another project's brainstorm window with the same task number, and the TUI switcher (j) no longer mistakes a session on another tmux server for yours.
• Codebrowser focus stays in this project (t1952): Opening a file reference in the codebrowser from a board task targets only this project's tmux session and warns instead of guessing when there is none. The launch runs in the background, so the board stays responsive.
• Switcher agent dialogs use the selected session (t1953): In the TUI switcher (j), the explore (X) and agent (e) dialogs now preselect the session chosen with Left/Right, and launching into another session moves you there.
• Narrow launch dialog shows Run/Cancel (t2013): In ait minimonitor, the narrow agent launch dialog now fits its Run in tmux and Cancel buttons on screen, and scrolls when the content is still too tall.
• Wide launch dialog scrolls on short terminals (t2020): The agent launch dialog in the board, monitor, codebrowser, trails, brainstorm and the TUI switcher now scrolls on short terminals, so Tab or the mouse wheel reaches Run and Cancel.
• Window name shown for a new tmux session (t2025): In the agent launch dialog, choosing "+ Create new session" keeps the window-name field visible, so you can see and edit the new window's name.
• Window name shown for a blank window choice (t2027): Choosing the blank window entry on an existing session now shows the window-name field, and the launch uses the name typed there instead of a hidden default.
• Bracketed window names no longer crash monitor (t2006): ait monitor now shows window names and task titles exactly as written in agent cards, the preview header and the kill dialog. A name like agent-[/]x no longer crashes the TUI, including on k in minimonitor.
• Monitor shows all user text literally (t2017): ait monitor and ait minimonitor now show session names, task titles and pane commands exactly as written in dialogs, bars, rows and toasts, without square brackets being swallowed as styling.
• Closed shadow pane no longer reads as fresh (t1971): The monitors no longer treat a shadow pane that has just closed as a fresh, never-analysed shadow, which could clear a standing staleness verdict.
• Review loop waits out shadow copy-mode (t2001): In ait minimonitor, the auto-recheck loop (L) no longer types into a shadow pane in tmux copy-mode, where the keys ran as copy-mode commands and disarmed the loop. It now holds, stays armed, and sends the recheck once you leave the mode.
• Auto-recheck hold survives a stale delivery (t2009): Disarming and re-arming the auto-recheck loop while a recheck is being delivered no longer resets the new loop's hold banner to ARMED, or records its next hold twice.
• Brainstorm refreshes after section decompose (t1995): In ait brainstorm, Module Decompose "From section markers" now refreshes the node list, the DAG and the module status as soon as it finishes, including after a partial failure.
• Failed section decompose no longer stuck waiting (t2005): A from-sections Module Decompose now checks every module's section before writing anything, and one message names every missing module. A partial failure marks the operation Error with the step it reached, instead of Waiting forever.
• Brainstorm prune names an incomplete node (t2019): If pruning dimensions fails while the new node is being created, the error and the operation record now call that node incomplete instead of created.
• Settings survives mistyped config files (t1945): Settings no longer crashes when a config file holds a value of the wrong type (such as "defaults": null); it names the bad file and keeps its current state. A broken file at startup gives a one-line error instead of a traceback.
• Frozen-agent viewer selection and copy fixes (t2003): In ait frozenagent, selected text is readable again, the / search box shows what you type, and ctrl+c and y copy the same way, showing in the footer only while something is selected.
• Crew log viewer search box visible (t2014): Pressing / in ait crew logview now shows the search box and what you type. Remapping a shortcut onto a key a TUI already uses no longer changes that key for other copies of the same screen.
• Test-map annotations respect byte-order marks (t1969): The test-map engine reads testmap: annotations in files that start with a UTF-8 byte-order mark, and never inserts an annotation above the mark, which used to break such Python files.
• Gradle scanning reads strings correctly (t1975): ait testmap no longer drops Gradle project dependencies, or marks a build file as dynamic, when a string contains /* or // such as a glob or URL. A file it cannot read safely is flagged so it widens test selection rather than hiding dependencies.
• Kotlin constructor parameters not indexed (t1976): ait testmap no longer records constructor parameters, as in data class X(val email: String), as top-level declarations, which made annotate --from-body attach tests to the wrong files.
• Delegated Kotlin properties named correctly (t1977): A delegated property such as val retrofit by lazy { … } is now indexed as retrofit, not lazy, so imports of it resolve and no phantom lazy symbol appears.
• Kotlin multi-dollar strings fail safely (t1980): ait testmap no longer silently skips Kotlin code after a multi-dollar string such as $$"…"; such a file is flagged as one it cannot analyse, so it widens test selection.
• Kotlin multi-dollar strings in test annotations (t1990): ait testmap now understands Kotlin 2.2 multi-dollar strings when it reads and writes test annotations, so an annotation edit can no longer land inside a raw string.
• Kotlin fun interface recorded once (t1997): The Kotlin dependency scanner records fun interface X as one declaration, and no longer reads a backticked name such as `fun` as a keyword.
• Backticked Kotlin string templates read correctly (t1999): The scanners for .kt and .gradle.kts files now read a $`name` template inside a string, so a quote in the name no longer hides later declarations or project(...) dependencies.
• Backticked Kotlin keywords no longer misread (t2002): A backticked `import`, `package` or modifier name is no longer treated as the keyword, which could record a phantom import, set the wrong package or hide the next declaration.
• Kotlin inline checks see past annotations (t2012): ait testmap now flags Kotlin inline functions and const values even when an annotation or another modifier sits between keyword and declaration (e.g. inline @JvmSynthetic fun).
• Kotlin scanner ignores brackets in strings (t2022): The Kotlin scanner no longer reads brackets, dots or ? inside string literals as code, which had hidden DI annotations, declarations and references. Inline properties and accessors are now flagged like inline functions.
• Sync recovery hint pinned by a test (t1696): Added a regression test confirming that the ./ait sync hint shown after a partial metadata update really recovers, even while other agents hold edited task files.
• Resume-session test survives model refreshes (t1949): A test that broke whenever the OpenCode model list was refreshed now reads its model from a fixed test registry.
• Markup scanner test passes again (t2021): A test fixture's deliberately broken title was flagged by the repo-wide markup check, failing the Python suite; it is now split so the check passes.

Improvements

• Gate lock exit handling unified (t1692): The gate commands now release their lock through the shared, guarded exit handler, so a later edit cannot make a failed step report success. No behaviour change.
• Compact elapsed time on monitor cards (t1996): The PROMPT/DONE/IDLE badge in ait monitor and ait minimonitor now shows elapsed time in at most four characters (IDLE 5h instead of IDLE 18734s), so it no longer wraps a minimonitor card.
• Continue prompt names the new task (t2015): After /aitask-explore, /aitask-pr-import, /aitask-revert or /aitask-review creates a task, the "continue or save for later" prompt now names it; when review creates several, it names their parent.

Documentation

• Trail skill explains uncommitted-file refusal (t1709): The /aitask-trail skill now says what to do when creating a trail is refused because of uncommitted changes: commit or revert the file the error names.
• Cross-repo trail docs (t1957_5): Documented cross-repo blockers on trail cards, ait trails --trail and the L linked-trail jump on the cross-project dependencies page, linked from the trails and board pages.
• Auto-recheck holds documented (t2010): The minimonitor how-to now lists every auto-recheck hold banner and what to do about it, the causes that disarm the loop, and how to review past holds with ait minimonitor --loop-log.

Performance

• Faster agent-window focus (t1950): Pressing p or g on the board, or p in ait trails, now searches only your own tmux session and the default one for an existing agent window, instead of every session on the server.
• Repeated L presses don't pile up searches (t1979): Pressing L repeatedly in By-Trail or ait trails runs one linked-trail search at a time, and only the newest result is shown.

Tests

• Directory-guard lint passes again (t1965): Fixed the nine test-script sites the cd-guard lint flagged, so the lint passes again.
• Trails live test ignores your tmux config (t1974): The ait trails live test now starts its private tmux server without the developer's tmux configuration.
• Hung runner tests fail fast (t1985): A watchdog makes a hung test-map runner test fail within its own time limit with a clear HANG: message, instead of stalling the run for 10 minutes, and cleans up the processes it left behind.
• Annotation conformance test suite (t1991): Added 189 bash, Python, Go and Kotlin cases that run through the real path ait testmap uses to read and write annotations, kept identical to the research corpus by a CI check.
• Watchdog reads test output in chunks (t1993): The hang watchdog reads test output in fixed-size chunks instead of loading whole logs into memory twice.
• Hang guard enforced for runner tests (t1998): ait engine test and CI now fail when a runner test that starts processes is not protected by the hang watchdog, before the full suite can hang.

Maintenance

• Minimonitor rows fit 40 columns (t1351): Minimonitor agent cards no longer wrap at 40 columns: names are clipped by on-screen width (CJK included), the shared agent prefix is dropped (agent-pick-1351_12 shows as pick-1351_12), and room is kept for the scrollbar.
• Parser research harness and baseline (t1954_1): Added a research harness and a 213-file, 8-language corpus to measure how safely the test-map engine reads and writes testmap: annotations, with baselines for the current readers and three native tools.
• Tree-sitter parsers evaluated (t1954_2): Measured three tree-sitter runtimes as annotation readers; all were 1.3–3× over the 300 ms scan budget, and the pure-Go one refused about a quarter of real Kotlin files.
• Highlighters, compilers and LSPs evaluated (t1954_3): Measured syntax highlighters, native parsers, ANTLR lexers, language servers and sidecar files as annotation readers. Highlighters cannot safely authorise writes; compiler front ends are accurate but need a runtime the release does not ship.
• Annotation-parser research report (t1954_4): The test-map engine keeps its own readers for bash, Python, Go and Kotlin, since none of the twelve alternatives tested was safe, runtime-free and within budget. More languages wait for a reader that passes the onboarding contract.
• Kotlin lexers pinned to the real lexer (t1992): Measured ait testmap's three Kotlin readers against the real Kotlin lexer and kept them separate, with a shared test table that fails if any of them drifts from it.