v0.27.0

Features

• Machine-verified quality gates (t635_12): Build, test, and lint gates now run automatically before a task can complete; each gate is skipped when its command isn't configured for the project.
• Async human gates (t635_15): Sign off on review/merge approval gates out-of-band with ait gate pass, using code-bound signatures so an approval can't be silently reused against different code.
• Documentation gate (t635_19): New docs_updated gate walks you through updating project docs as part of finishing a task.
• Multi-stage completion stats (t635_20): Stats now track in-flight (done-but-awaiting-gates) tasks and per-phase pipeline timings.
• Task attachments (t1030_1, t1030_2, t1030_3): New ait attach command attaches files to tasks — content-addressed storage, reference counting, garbage collection, and correct handling across archive and fold.
• Learn a skill from anywhere (t1071_2): New /aitask-learn-skill command generates a complete skill from a local file, URL, repo path, or a live terminal pane.
• Shadow error diagnosis & learner spawning (t1071_1, t1071_5): The shadow companion can diagnose a followed agent's errors on request and spawn a learner agent to capture a skill from what it's doing.
• Applink scrollback history (t1057): The mobile companion can request older terminal scrollback beyond the live viewport.
• Platform-agnostic chat abstraction (t1074_1): Added a ChatAdapter core (with a full mock) as the foundation for chat-platform integrations.
• Unified artifact storage design (t1065): Landed the design for native artifact storage/sharing built on a stable-handle / mutable-manifest model.

Bug Fixes

• Reliable installs & upgrades (t1075): Downloads release tarballs from the CDN instead of the rate-limited API, and honors an explicitly requested version.
• Clearer ID-counter errors (t1077, t1087): A counter fetch failure is now distinguished from a genuinely uninitialized setup on both the claim and peek paths.
• ID counter drift hardening (t1079): Task ID assignment self-heals against counter drift.
• Attachment reference cleanup (t1093, t1096): Hard-deleting a task now releases (or rebinds, for folded tasks) attachment references so blobs aren't orphaned.
• TUI session identity collisions (t1099): Session identity is keyed on project path, fixing collisions between repos that share a default session name.
• Applink viewport-only rows (t1054): Live terminal frames stream only the visible viewport, fixing row misalignment.
• Applink pause verb (t1055): The mobile client can pause the content stream.
• Project registry race (t1073): Concurrent registry writes no longer lose group assignments.
• ait stats tui (t1083): Now routes to the stats TUI instead of erroring.
• Stats TUI registered repos (t1098): Shows all registered repos.
• Model picker ranking (t1082): Recently-verified models rank ahead of legacy fallbacks.
• Board in-flight card polish (t635_22): Fixed op-hint brackets being swallowed and clarified empty-gate copy.
• Fail-closed folded-task delete (t1102): The board fails closed if unfolding children during a delete errors.
• Agent launch window default (t1115): Agent/create launches default to a new tmux window instead of reusing a remembered one.
• Minor fixes (t1002, t1095): Silenced a benign shellcheck info; fixed a stale shadow-spawn test assertion.

Improvements

• Unified gate declaration (t635_14): Profiles declare their gates via a single default_gates list.
• Risk evaluation gate (t635_13): Verifies a task's plan documents code-health and goal-achievement risk.
• Concurrent gate-ledger merges (t635_21): Gate records made on different machines auto-merge on sync instead of conflicting.
• Board by-topic sort modes (t1035): By-Topic view offers recency, topic-id, size, and alphabetical sorting (press o).
• Shadow feedback staleness indicator (t1104): Shows when the shadow's analysis is stale relative to the followed agent's current screen.
• Deeper shadow plan review (t1071_3): A --deep flag captures more scrollback during plan review.
• Configurable skill-authoring guide (t1071_6): The learn-skill flow uses a configurable authoring-standards guide.
• Optional learn-skill wrappers/commit (t1100): Cross-agent wrappers and committing are now optional when generating a skill.
• Applink history depth limit (t1092): History capture depth is configurable and bounded independently of live monitoring.
• Applink pane titles (t1114): Mobile pane status includes task titles.
• Attachment metadata layout (t1030_5): Evaluated and settled on per-blob attachment metadata.

Documentation

• Topic anchoring concepts (t1034): Documented anchor/topic grouping.
• Learn & shadow docs (t1071_7): Added website docs for the learn-skill command and shadow agent.

Performance

• Focused-pane content streaming (t1045): Applink streams binary terminal content only for the focused/subscribed pane.
• Monitor gate-summary cache (t1111_1): Caches gate summaries by file mtime instead of re-reading every tick.
• Faster focus switching (t1111_2): Reduced redundant re-rendering when switching focus in the monitor.
• Non-blocking monitor refresh (t1111_3): Removed synchronous tmux calls from the monitor refresh path.

Tests

• Git-tag resolver drift guard (t1084): Guards parity between the installer and library git-tag resolvers.
• Remove geminicli from fixtures (t1086): Fixtures updated to only expect supported agents.

Maintenance

• Applink data-plane limits (t1007): Hardened applink with resource limits (frame size, pane counts, keyframe cadence).
• Port learn-skill cross-agent (t1071_4): Ported the learn-skill command to Codex and OpenCode.