What we shipped.
Every release, newest first. Morning updates itself — this is what arrived.
v0.1.532026-10-24
New
- A failure goes to the stage that owns it. A test-side finding from the code review is carried forward to Testing instead of sending the task back to Development; a problem with the test environment is fixed in Testing; and a problem the task cannot fix itself (a wedged Docker container, a missing tool on the machine) becomes a card in the chat with the exact command, so you run it once and approve. The retry budget now counts repeats of the same failure, so a new problem found late still gets its round, and a finding the Brain already patched is re-checked instead of blocking the task again.
- Skills can file and update issues in your connected tracker (Redmine, Jira, GitHub, GitLab, Linear), with attachments where the tracker accepts them. The chat assistant can propose an issue from a conversation; nothing is written until you approve the card.
- Task proposals from the chat say what runs after what and whether the task gets a design stage, and a bar above the composer points you to proposals still waiting for your decision.
- Redraft a project's whole design from the Design Studio. The design constructor from the new-project wizard now opens inside the Studio: pick a voice and a palette separately, see example screens for each voice, then apply the result as a draft or to your registered screens. Every generate step in the new-project wizard has a Stop button.
- DESIGN.md is one document and every design run reads all of it: identity, tokens, imagery, composition rules and voice. The composition rules (how screens are arranged) live in the document and are edited from the Design tab.
- Messenger chats: tables arrive as readable blocks on a phone, and the assistant can send files such as screenshots into Telegram, Slack and Discord.
- Launch setup detects one rebuildable Docker stack instead of listing every backend a second time as a separate service, and a recovery fix from the Brain changes only the services it diagnosed.
- A new task branches from your local base branch when it is ahead of the remote (for projects that do not push the base branch), with a note in the task log.
- The license agreement was updated (the October 2026 edition). The desktop app asks you to accept it once on the next launch.
Fixed
- The test gate shows a long browser suite's progress and no longer kills it at 15 minutes while it is merely slow; a fresh task worktree installs the test package's dependencies before running; a file a report only mentions (a deleted probe, a quoted diff) is no longer read as a claimed test; and a stage retry is granted once per round instead of again and again.
- Build output in your main checkout (`target/`, `build/`, an Angular cache) no longer blocks the next task as "uncommitted changes", and the files Morning snapshots before a task are actually committed, including ones in nested folders. When the root really is dirty, the Brain lists the files and the git command instead of investigating for a quarter of an hour.
- The implementer sees the approved design again on every workflow, so it builds the mock you approved instead of drawing one from the plan's prose.
- Tasks paused by a provider usage limit resume when the limit lifts (they stayed paused for good), a network outage during a run is retried after a cooldown, and resolving a parked task's last blocker by hand puts it back in the queue.
- The Brain retries on its own after installing a missing package instead of asking you to click.
- Delivery no longer blocks on a merge conflict where one side deleted the file.
- A pasted screenshot attaches once in every chat composer, not twice.
- Screenshot capture recovers when the headless browser has died, instead of failing until Morning restarts.
- The Launch button's "Detecting…" can no longer spin forever.
- The design chat at the approval gate works on an older version of the design instead of answering "Design artifact not found".
- Planning a roadmap no longer leaves `specs/` files in your project.
- A plan that lists `**Files:** none` on a verification-only task is accepted.
v0.1.522026-10-16
New
- The Documentation workflow writes user documentation. A documentation task now explores the product's screens and flows first, then one technical writer decides the pages and writes them under `docs/<language>/`, with every screenshot marked for capture. After the task merges, Morning photographs the marked screens from your app and lands them on a `morning/docs-refresh` branch. The per-task documentation update at the end of every other workflow is unchanged; the old developer-facing "docs" skill is gone, and an existing board's documentation workflow follows the new shape on the next start.
- The MCP server row in Settings shows a chevron, so it is clear the row opens for editing.
Fixed
- Desktop: the Studio's screen chat and the MCP tool servers start again in the packaged app. The v0.1.51 build shipped MCP children that could not find their own dependencies, so every design chat answered "the design-chat server failed to connect".
- A slow tracker is no longer reported as a bad token. Importing an issue retries a timed-out read once, fetches the one issue you picked instead of listing the whole tracker, and the error names what actually happened: a timeout to retry, a connection problem, a token to check, an issue outside the provider scope, or a board with no Backlog column.
- A diagram in a plan that does not parse shows its source with the parser's own line and reason, instead of a "Syntax error in text" box filling the plan.
- "Draft with AI" on the Design tab's Voice section reads a reply that wrapped its JSON in a code fence or carried a line break inside a quote, instead of failing with "the draft could not be read".
- Inserting an MCP tool name from the skill editor's picker lands where your cursor was (or on its own line at the end when you never placed one), never glued onto the first word of the file.
- Design and domain-research stages no longer get added to a documentation workflow; a documentation task plans pages, not screens.
v0.1.512026-10-15
New
- Connect external tools (MCP servers) by pasting their config. Settings → External MCP servers takes the standard `mcpServers` snippet a server's README ships, and **Connect** shows the tools the server answers with. The chat assistant can set a server up for you from its README; nothing is registered or run until an admin approves. Remote servers now work with Copilot and Codex as well as Claude Code.
- A skill says which external tools it uses. In the skill editor, tick the servers a skill relies on and insert their tool names; the editor shows, per workflow stage, which of them would be missing and why. A stage receives a server only through a skill that declares or names it, so no step is handed a tool it has no instructions for. External tools now reach workflow stages, not only the chat.
- Importing an issue from your tracker brings its comments and attachments along, and the import dialog filters by number, title, label or the tracker's own status.
- Screenshots from test runs: the task shows up to sixty of the newest, they refresh on every run (a failing run contributes the screens from its passing tests), links in the Implementation Summary keep working across runs, and the docs updater can copy them into your documentation.
- Brand voice. The product's personality, audience and tone live in DESIGN.md, are edited from the Design tab, and reach every design surface — the design stage, the Studio and the design chat.
Fixed
- The Claude provider's Test no longer reports "not signed in" after a re-login. It asks the CLI directly instead of reading the Keychain, which could wait on a permission prompt.
v0.1.502026-10-14
New
- Reviews decide the gate. When Morning reviews work at an approval step, its verdict is the decision: a pass approves, a fail sends the task back with the review's findings as the asks. The review itself appears at that step as a named Approval Review.
- A task's base branch can be changed from the task's edit dialog right up until work has started on its branch. After that it is shown as pinned, with the reason.
- The commits Morning makes on a task branch are named after the task, so a branch log says whose work it holds.
Fixed
- A task waiting for screens another task holds resumes on its own under auto mode once they are free, instead of sitting paused until someone presses Play.
- A long task chat turn is stopped only when it is genuinely idle: a tool that is still working keeps it alive, and every chat has the same three-hour ceiling.
- When you choose a base branch for a task, its submodules follow it too (a branch of the same name on their origin, otherwise their own default) instead of always integrating their main line.
- A failed UI check such as "Element Menu not found" is no longer mistaken for a missing tool on the machine.
v0.1.492026-10-12
New
- Install missing tools from one button. When a task stops because a tool such as Docker, Node or pnpm is missing on this machine, the blocker, the Brain's message and Settings → Instance capabilities all offer "Install in Terminal". Morning writes the install script and opens it in your own terminal, where you run it and enter your password; Morning never installs anything itself. A tool with no known recipe gets a command drafted for you to read and edit first, and a command that worked is remembered on this machine.
Fixed
- A brand-new Mac no longer loops on a missing tool. A task stops once, names what is missing, and resumes with Retry after you install it, with no restart of Morning. Before, the same missing tool failed Testing again and again and sent the task back to Development for a problem no code could fix.
- The web test suite runs from the package that holds its Playwright config (a monorepo's `apps/web`) instead of the repository root, where it found nothing. A missing `npx` or a missing Playwright config is named instead of being reported as "no report".
- Auto mode's automatic reviewer works again: the gate pauses the task before the reviewer runs, so its approval is accepted instead of refused and shown to you as a dialog full of blockers. Turning Auto on for the first time selects the automatic reviewer on every approval step that does not insist on a person.
- The Execution tab fills the panel, and a stage that re-runs (a plan revised after review) shows as running again instead of appearing only once it finishes.
- The Brain is visible while it works on a blocked task: a pulsing brain in the task header and in the simple mode's "Needs you" list, and a line in the Blockers section saying it is triaging.
- A long closing report from a testing stage no longer lands in the task chat as raw JSON split across two messages; it reads as one verdict line with any follow-ups.
- The Brain's message under a blocker no longer says it changed nothing right under a proposed fix.
- Project setup: one folder picker everywhere, adopting a shared project checks access first and clones into a named folder, wizard errors are visible, required provider fields are enforced, and a launch with no provider says so.
- Cloud sign-in: teams you join after signing in are discovered, re-linking an account rotates the held session instead of killing it, and a "no team" answer is no longer reported as "unavailable".
v0.1.482026-10-09
New
- Every workflow stage now runs on Morning's own engine. Planning, development, review and testing go through one runtime with the same turn, time and budget limits as before, so a stage no longer stops mid-work saying it "did not write its output".
- The Execution tab is a timeline. One row per attempt of each stage, one bar per model call with its duration and estimated cost, quality gates as points, idle waits compressed, and the card running right now named from live evidence. A failed attempt shows the reviewer's or gate's verdict under the call that wrote it, and where the task was sent back. The Progress and Graph views and the separate Cost tab are folded into it.
- The task Chat narrates the run as it happens: model messages and tool calls per stage. A stage that needs a decision from you asks it as a quiz that pauses the task until you answer.
- Choose the model per skill. Each skill can run on its own provider and model, picked as a tier (showing the model it lands on) or an exact model, with a fallback or a wait when that provider is unavailable on this machine. Planning runs on the premium tier by default.
- One Skills screen (Settings → Labs): create a skill, edit its instructions with a preview, see where it is used, set its provider and its output template. Output shapes now live as templates instead of inside the skill text.
- "Workflow" replaces "pipeline" across the product: a workflow is the board's sequence of stages, and one unattended run is a workflow run.
- Design review per platform. A design spanning several platforms shows one tab per platform (web, desktop, iOS, Android), a map of which screens exist where, and links between screens on different platforms follow across tabs. Desktop is its own design track, never a wide web page.
- Native desktop and iOS apps on the Launch tab: a Mac app target is detected and built with an Open app button, and a native iOS build installs and launches on the simulator. Desktop apps can have end-to-end tests through an XCUITest driver (Project settings → Testing).
- Trying a task's worktree now runs an isolated stack: its own containers, ports and a fresh seeded database, so it never serves or overwrites the main stack's data.
- Launch detection picks your real dev stack, never the test stack, and a proxy-fronted compose stack opens in the browser from the launch button.
- Auto mode chooses per approval step: ask me, let the automatic reviewer decide, or skip. The reviewer judges the approach as well as the documents and can send work back with a better way.
- When a stage hits a host problem it cannot fix itself (a wedged Docker container, a missing device), it proposes the exact command and the task waits for one click from you instead of spending a development round on it.
- The Brain can approve or send back at a human gate from chat.
Fixed
- Work a task wrote could be lost at delivery: an internal git command failed quietly on some repositories, so the stage's changes were never committed and the merged branch came out empty. Delivery now commits everything or refuses, and never deletes a worktree that still holds uncommitted work.
- A task whose base branch moved between its final checks and its merge re-integrates and re-checks instead of blocking.
- Project setup interrupted by the Mac sleeping restarts the step instead of hanging; the wizard's proposal step is bounded on silence and says why when it falls back to a preset.
- Your own staged deletions in the project root are no longer "restored" by the post-run cleanup.
- A reviewer can no longer end a stage without a verdict: the engine tells each evaluator the verdict is required and re-runs it once if it forgets, instead of failing the stage.
- Times in messages written by the server (retry times, cooldowns) show in your local time zone.
- Desktop sign-in works on both 127.0.0.1 and localhost, and re-linking an account no longer leaves an orphaned device holding a slot.
- Timeline durations add up again: a done task's bar ends, and a failed stage keeps its end time.
- A Claude Code error reply is treated as a refusal, not as an answer.
v0.1.472026-09-15
Fixed
- A sign-in loop that could lock you out for good: if a team removed you (or a workspace you once connected went away), the dead membership could stay in charge of your session — every fresh sign-in succeeded and was immediately overruled, showing "sign in to continue" forever, with each attempt reading as "another device is active". A dead session can no longer outrank the account you just signed in with, and one stuck workspace no longer blocks the whole app.
- Reading git status no longer mis-parses the first modified file on some repositories.
v0.1.462026-09-15
New
- One sign-in connects everything. Approving the device once signs you into every team you belong to — no more picking a "workspace" or switching between them. Billing resolves itself to the right plan (a team that covers you wins over a personal free account), shared projects from all your teams appear in one list, and the workspace switcher is gone because there is nothing left to switch.
- Plans are managed in one place: the account console on the web. The desktop app shows your plan and a scheduled cancellation (with a "Keep plan" button to undo it before it lands), and sends you to the console for any change — so the app and the site can never disagree about what you are paying for.
- Claude Code can run through your own cloud. An Anthropic provider can now point at AWS Bedrock, Google Vertex AI or Azure AI Foundry — billed on the contract you already have, authenticated by your cloud's own IAM, with no Anthropic key or login on the machine.
- Account management cleaned up: billing actions live with the owner, devices show their names, and the sandbox card reflects any team that covers you — not just the account that happens to be primary.
- Updated End User License Agreement (version 2026-09-15), reviewed by legal counsel.
Fixed
- Personal CLI logins (Claude Code, Codex, Copilot) are no longer offered on hosted sandboxes, where they cannot legitimately live — API keys and Morning credits work there as before.
- Adopting a project a teammate shared no longer walls you behind the new-project AI onboarding — it appears immediately and opens onto its setup progress, and a failed setup now says why and how to retry on every surface.
- A missing test device (deleted emulator, no simulator) is reported as exactly that, instead of being mistaken for a build failure and sending the pipeline to "fix" code that compiles fine.
- A blocked task always leaves you a button: when no automatic fix is identified, the Brain proposes an attested retry instead of going silent.
- Back on menu pages returns where you came from, not always to the board.
v0.1.452026-09-14
New
- Early access to updates. Settings → Labs has an "Early access updates (beta)" switch: turn it on to receive prerelease builds before everyone else. One setting serves both the desktop app and a team sandbox; turning it off returns you to stable at the next stable release.
- Prefer remote workers. A project setting that sends tasks to an online worker machine whenever one is capable of running them — checked against each task's platforms, so an iPhone task never lands on a machine without the iOS tools. If no capable worker is online, or nobody picks the task up in time, it quietly runs locally instead. Pressing Play always runs here.
- Reviews can no longer end without a verdict. An evaluator's report must state pass or fail in a form the pipeline can read — prose like "READY" used to route nothing, and a task once reached approval with unfinished work described only in words.
- The pipeline finishes the plan. A task can no longer arrive at your approval gate with plan items "deferred" just because they were big — unfinished work goes back to Development until every plan task is done or carries a real, named blocker. Cutting scope is your decision, made by rejecting with a plan change; a per-project setting restores the old warn-only behaviour.
- Every acceptance criterion is accounted for. The testing phase must now either cover each of the plan's criteria with an executed test or state, per criterion, why it is not covered — "at least one test ran" stops being enough.
- The app version now lives in the account menu (click your avatar) — and clicking it copies the version for a bug report. The tray menu slims down to actions only.
Improved
- Shared projects carry their setup with them. A task synced to a teammate or a worker now brings its type, its design flag, its target platforms and its tracker link — and binds to the same pipeline its author configured, including the review-and-return rules. Before this, an adopted task could silently run a different, weaker pipeline than the one you set up.
Fixed
- A team sandbox no longer offers to create a sandbox from inside itself.
- Every submodule branch Morning pushes now gets its merge request, or a log line saying exactly why not.
- An aborted merge no longer leaves the working tree dirty and the resulting block invisible.
- A machine with no terminal configured can still install an AI provider.
- Morning's stored git token is now used for pushing, not only reading — a push no longer fails on machines whose own git credential is read-only.
v0.1.442026-09-13
New
- Stop sharing a project with your team. One action removes it for everyone, where before you could only unlink it from your own machine or untick people one at a time — and there was no way at all to take it back from a teammate who is an owner or an admin. Everyone's local board stays where it is, and no repository is touched.
- Accepting a shared project now tells you, before it starts, what it is about to clone and whether this machine can. If there is no git credential for that repository's host it says so and takes one right there — which is the only way it could be added on a machine that has no projects yet. Importing a project by git address asks the same way.
- A "This machine" panel in Settings → Integrations: what this copy of Morning can actually run — Docker, the headless browser, the iOS and Android toolchains, Maestro — and which git hosts it holds a credential for.
- The docs and demos switches moved out of project settings onto the Docs and Demos tabs, next to the thing they control.
Improved
- The browser Morning uses for screenshots, demo recordings and page checks now installs itself the first time something needs it. Missing it used to quietly switch off page-render checks, so a blank screen or a crashed page went unnoticed while the project still looked reachable.
- Accepting a shared project is a decision again. Nothing adopts itself on a machine you are sitting at; a dedicated worker, which has nobody to ask, still does it on its own.
- Task source rows open on a click anywhere in the row. The clickable area used to be the label text alone, so the arrow, the space beside it and the row's own padding all did nothing — which read as the card being slow rather than as a missed click.
Fixed
- Removing someone's access to a shared project now actually stops their copy syncing, and tells them why. Their Morning used to keep retrying, log an error every minute, and eventually lock the project as if its data were corrupt.
- The demos switch never saved. It showed whatever you last clicked and every project stayed off.
- When accepting a shared project fails, the message now says what went wrong — usually a missing git credential, named with the host — instead of a code.
- Number fields, date pickers, dropdown menus and autofill are drawn in dark chrome on the dark theme instead of staying light.
- A test that ended before it produced any verdict, because the harness or its connection died, is re-run rather than counted as a failure. Bounded, and it stops itself if the machine is short of disk or memory.
- The Skills section in Settings now follows the Labs switch that governs it instead of staying behind after you turn it off.
v0.1.432026-09-12
Fixed
- Accepting a project shared with you works again. The folder is optional — leave it empty and Morning clones the project into your projects folder — but the Accept button was still refusing to act until you typed one.
- Unlinking a project now sticks. It no longer reappears in the list of projects offered to you, and it is not quietly re-added the next time Morning syncs. Sharing it again works as before.
v0.1.422026-09-12
New
- Sharing a project now shares the code, not just the board. A shared project carries its git address, clones itself into your projects folder and is ready to work on — no folder to name, and nothing to accept.
- Delegate a small change in one sentence. It runs end to end on its own branch and comes back as a diff you review and approve, with the usual checks already run against it.
- Sign in to a hosted workspace with one click. No code to copy, and only your own organisation's workspace opens.
- Single sign-on and an exportable audit trail, for teams that need them. Your identity provider can also add and remove members directly.
- An organisation-wide autonomy ceiling. An owner sets the limit once; individual projects can tighten it, never loosen it.
- Linked tickets now hear back — an issue learns when its work started, was delivered and finished.
- Choose which quality checks run at which stage, and let testing turn a repeated manual check into a permanent automatic one.
- Hosted and self-hosted instances can now run on PostgreSQL.
- The pipeline editor shows more of what actually happens — a full-width stage form, a preview of the context each step receives, and which step a reviewer's findings are sent back to.
- New projects can be set up with or without the approval and accept steps, so the pipeline matches how you like to work.
Improved
- Long jobs no longer look stuck: waiting for a busy stage, restarting after a crash, and slow steps all say what they are doing and how long they have been doing it.
- Closing your laptop mid-run no longer counts as work time, and no longer fails a healthy job that was simply asleep.
- Before starting an expensive test run, Morning checks the machine has the disk and memory to finish it, and says so plainly instead of failing halfway.
- Test runs clean up only what they started. Another project's simulators, emulators and servers are left alone.
- Work delivered as a pull request now arrives as a single clean commit instead of the full working history.
- Documentation can live in its own repository, and screenshots are captured behind a sign-in.
- Task chats keep working in the background when you close the panel, and can be stopped from anywhere.
Fixed
- Sign-in: the code stays on screen. It survives a connection hiccup, says "that code expired" instead of vanishing, and the window to use it is long enough to register and verify your email.
- Chat no longer gets stuck on "thinking…" after a reply has already arrived.
- Syncing a project no longer runs when nothing has changed, which cuts a lot of needless traffic during a run.
- Total time on a task now adds up to what the stages actually show, and the timeline and progress views use one format.
- Approving a task into Done now finishes and merges it immediately instead of waiting for the next cycle.
- Mobile: a task's app is built and launched from that task's own code, and Android tools are found from your SDK rather than the server's environment.
- Your own bug reports are visible whether or not you are an administrator.
v0.1.412026-08-31
New
- Screen wireframes. Sketch a screen's structure first, in grayscale, then paint the design from it. One Studio session holds both the sketch and the design with a switcher; a structural complaint at review patches the sketch and repaints instead of redesigning. The questions Morning asks before drafting moved to the sketch step, where the answers are decided.
- The Tune editor became a real builder on wireframes: insert and remove blocks from a catalog beside the canvas, drag to re-order or nest, select several and wrap them, arrange a container's children (row, column, alignment, gap), resize with eight grips, and attach a note to any region saying what it must do. Designs keep the lighter nudge mode. Saves keep your place, and a save never loses a screen.
- Shared projects. See the projects your teammates shared and adopt one onto this machine; each shared project can now be paused, resumed, synced now or unlinked on its own, and the project list shows its sync state live.
- One People list. Team seats, instance accounts and open invitations in one table, with a role you choose at invite time. Invitations now reach the app itself, not only your inbox.
- Hosted and self-hosted instances can run on PostgreSQL; the desktop app stays on SQLite.
- Deploy from Morning. Connect Vercel, App Store Connect and Google Play once in Settings → Integrations, describe the targets in a config file in the repo, and ship the main branch to Vercel, TestFlight or the Play internal track with one click. A screenshot gallery keeps store assets with an approval state.
- Fast mode per task. Mark a task as fast and it skips the research and documentation steps, caps the review loop at one round and runs on the economy tier.
- The plan is finished, not trimmed. An approved plan's tasks are accounted for at every review; work that was quietly deferred is sent back or shown to you at Accept with the reason.
- Unfinished work fails the review gate: an empty method body with a comment in it, or a new public method nothing calls, is reported instead of slipping through with green tests.
- Providers: log in to Claude Code, Codex or Copilot from the provider card; a green Test re-enables a provider that was switched off after an expired login, and such a provider re-checks itself when a run needs it; each provider row can name its own CLI home directory, so two accounts of one vendor coexist.
- Morning's git-hosting token is used for git itself, not only for opening pull requests. Task branches are named after the task even for titles in Cyrillic.
- Release notes are drafted from the commits, translated, and shown in the app after an update and on the site's changelog.
- Diagrams in plans and reports render as pictures. Skills can carry a per-skill format note, and artifacts can be edited in place.
- Settings gained Notifications and MCP tabs, memories can be edited, and new items arrive as a preview in the top-right corner instead of a floating inbox.
Fixed
- The desktop app's first run no longer flashes the licence and then a sign-in; sign-in works on both 127.0.0.1 and localhost; a refused window says whether its cookie was absent or stale.
- A board that cannot load says so, with a Retry where it can help, instead of spinning forever.
- The web test gate runs its own suite instead of trusting a report another step left behind, and announces a pass in the task log.
- Re-authenticating a paused cloud session shows the device code again.
- A task's skill session can no longer hand work to a background agent that outlives it, which used to end a stage with no output.
- Review no longer trips on the word "placeholder" in ordinary code, and the retry budget no longer resets itself, so a stuck task blocks after its budget instead of looping.
- Housekeeping never deletes a remote task branch, which once closed an open pull request.
- Morning's scratch files inside a submodule are no longer mistaken for your uncommitted work.
- A failed dev-client build on an iOS launch is caught, retried through the simulator signing path, and reported to the Brain.
- The onboarding checklist updates when you add a provider or create a project, without a reload.
- A build or module-resolution failure in a running app is investigated at the first occurrence instead of after three.
- The idea you typed at project creation now reaches design-direction generation, and your creative brief binds the candidates.
v0.1.402026-08-25
New
- An explicit Standard tier override is honored, never dead config
- DeepSeek-harness adoption — loop guard, env scrub, retention, conjunction verdicts, prompt goldens
Fixed
- One shared platform-chip toggle rule (mobile family is exclusive)
- Prompt-cache tokens reach usage accounting from the SDK adapters
v0.1.392026-08-24
New
- App side — seat picker, invites UI, company name, onboarding
- Dependency drift install + the Brain's bash fixes count again
Fixed
- Review fixes — resend stale-closure bug, AC-14 jump links, band pin
- Plan changes reach the app without a restart
v0.1.382026-08-22
New
- Platform glyphs on the delivery-platform chips (project settings + new-project wizard); mobile shows the Apple+Android pair
- Outlined Apple + full-body Android platform marks (Tabler paths, lucide-matched strokes)
- Apple/Android marks on the card; mobile always shows both; gate results per platform on hover
- Task card redesign — header strip, gate results on platform glyphs, scrollable tags, flush progress
- Dedicated flat Settings step after create; first screen stays minimal
- One screen, many states — data-screen-state panels, state-aware data-nav, registry states
- Remember the last e2e gate results on the task card; reset when a new run starts
- Follow-ups are fixed where they were found — in-task by default, explicit deferrals proposed, symptom-only titles
- Execution concurrency is one honest number; per-stage limits always on
Fixed
- Stat cards use the theme border token (bare `border` rendered near-black)
- The top-panel palette glow is server-stated, never inferred
- Hide the tag-row scrollbar (global ::-webkit-scrollbar styling beat the arbitrary variant)
- Keep the last e2e results visible until the new window's first settle
- Remembered e2e results keep every leg of the window
- Per-attempt durations add up to the headline total
- Ceiling aborts are run errors (blocked, not 'paused by user'); sweep the worktree's processes on abnormal run end; deny whole-suite runs in skill sessions
v0.1.372026-08-20
New
- Domain-research locates official protocol specs; Morning snapshots them as reference docs
- PDF text extraction for attachments (unpdf); reference extraction keeps whole documents
- Reference grounding — attachments + links reach every skill; plan pins protocol contracts
- Guided next steps, project-creation tracks (362), first-task friction (364), live gate-leg progress
- Richer generation context, coverage tiers, diagnosable runs
- Sweep stale simulator XCTest attachments (testmanagerd debris)
- Design-system rethink: token gallery, usage notes, library + creation flow, clickable prototypes
- Claude SDK model discovery — the account's real roster, not a hardcode
- Copilot model catalog → current roster; retired-model defaults fixed
- Codex model catalog → current GPT-5 lineup
Improved
- Remove the per-flow design-system select from the toolbar
- Drop the Plan and Design cards; attachments list collapsed by default
Fixed
- The provider row's Claude account reaches every caller; a 401 names the failing credential
- Maestro post-run housekeeping crash is not a verdict; re-run gate failures are counted rounds
- Superproject wip snapshot excludes Morning-internal dirs by pathspec
- Review shows every task-owned submodule, diffs three-way; wip snapshot never commits on a stale clone
- Cross-phase parallelism — skip-and-continue barriers + per-stage limits default on
- Settle-vs-flush race, off-screen spotlight, provision misroute, import URL redaction
- Review r2 — spawn-error guard, scoped project row on jobId GET, scaffold commit identity
- ONE palette-key convention — bare names, normalized at every boundary
- Embedded setup-tokens are live-validated, not presence-trusted
- Connection probes honor the ROW's embedded credential
- QA triage clusters same-reason failures; non-e2e skills can't drive devices
- Harvested screenshots supersede prior rounds — never accumulate
- Expired-OAuth CLI notice is an ERROR that fails over — never output
- Instant previews degrade to default scenes; drop the per-group Feedback button
- Route discovery settles hydration before capturing affordances
- Submodule touch-detection tests its OWN base, not the parent's name
- Diff from the tip-most base branch point, not a stale local base
- Exit handler picks the stream carrying the CAUSE
- CHROME RULE — app chrome is never a value moment in another scenario's story
v0.1.362026-08-18
New
- Pipeline editor — full-width canvas, hint below, Discard/Save footer
- Pipeline tab renders the flow editor inline + user-facing flow→Pipeline rename
- Brain FAB unread — rainbow wave radiating through the icon strokes
- Activity signal, cost/expiry surface, consented destroy
- Full-footage demo authoring rules + Flow gets its own project tab
- One card per scenario — list + results grid unified
- Flows land visible results; regen is breakage-aware and failure-keyed
- Interaction probes reveal the dialog layer; flows must be distinct
- Flows script only observed affordances; small sets grow
- Narration sells from the first word; from-scratch drafts cover 4-7 value moments
- Deleted scenarios never resurrect + Morning ships its own demo seed
- Head-trim, honest video removal, voice-over toggle
- Seeded credentials make demo auth fully autonomous
- Scaffold auth.demo.yaml from the login form when walled
- The Brain fixes demo breakage itself, never via pipeline tasks
- Off-camera auth — auth.demo.yaml signs in once, footage starts authenticated
- Auth-wall fail-fast, true per-row re-record, informative Brain aborts
- Live status bar, active-row spinner + stop, locked rows
- Cold-compile grace for app-network noise + persisted last-words tail
- Restart guard, Stop-triage, and non-empty reasons
- Gate flags host Metro-proxy / IPv4↔IPv6 forwarders
- Progress bars, running-flow timer, project-list QA/docs chips, top-panel edit
- Opt-in parallel iOS+Android comprehensive runs (Metro lanes)
- Live per-flow status table for comprehensive runs
- Docs + Demos become top-level project tabs; Automation is flat Autopilots
- Sibling sweep for shared-shape config fixes
- Parallel mobile gate legs on isolated lanes (opt-in)
- Deterministic post-generation corrections
- Task-launch prebuild + generation teaches freshness
- Per-service launch targets (main vs task)
- Call out task launches of prebuilt-image stacks
- Read ALL warnings/errors; classify; fix or MASK
- QA runs narrate their phase on the card
- Workspace-wide chat binds itself to a project
- Generic connectivity-distress tier in log ingest
- Commit selector + resizable sidebar
- Kafka broker-unreachable visibility (wms incident)
- Docker-daemon-hang fast path + one-click restart
- Notes as its own task-detail tab (after Chat, before Cost)
- Two-pane wizard shell — left-rail steps + roomy content
- Task Notes replace steer, injected into every skill
- Move Language picker from Account to Settings → Appearance
- Per-scenario re-render/remove + per-video download/remove
Improved
- Brain FAB unread = neurons-firing animation, not a corner dot
- Fade the project-edit pencil, divider before launch button
Fixed
- Last test-suite stragglers + interrupted tool chips on completed messages
- Sweep stale test pins green + scenario gen never returns empty
- Cross-bundle in-memory registries ride globalThis
- Abort mid-gate must not seal the phase row (Spec 344b completion half)
- Probe destructive-click guard + recovering-indicator flash
- Project-scoped WS routing + residuals
- Durability review — reconnect recovery, half-open watchdog, event coverage
- Cards keep their videos while a run is in flight
- Generation applies the demo seed before discovery
- The WS port follows the app port — PORT+1
- Remove video scrubs the run snapshot too
- Route discovery uses the off-camera session too
- Webpack engine + isolated dist dir for npm run dev
- Simulator.app opens always name their device
- Base-integration lines never read as a stage claim
- An aborted run never settles gate steps
- Unqueue lands PAUSED on auto-mode boards
- Harness-lib v19 — Metro lives in its own process group
- Off-flow columns rejoin the flow, never silently end it
- Project-side subprocesses never inherit Morning's NODE_ENV
- Retry the Gatekeeper build race up to 3×
- Top-panel project-edit button uses the Pencil icon (matches the list)
- Comprehensive run pins the iOS bundler itself (No script URL red screen)
- Fallback iOS boot never boots another project's pinned sim
- IOS launch never fires on another project's booted sim
- Expo rebuild uses `run:<platform>`, not the bogus `launch:` (the --device error)
- Runner opens the INSTALLED iOS app instead of rebuilding (bundle-id divergence)
- IOS render probe detects the "No script URL" red box (no more green over it)
- Durable iOS bundle-id pin (the (null) "No script URL" red screen) + QA parallel progress line
- Render probe stops flagging login/form screens as blank
- WaitForPorts reconciles to the run-root compose
- Readiness never errors a busy process (soft verdict)
- In-place live-row updates feed the stall sweep
- Fold gate sub-phase lines into the leg's live row
- DependsOn wait extends while upstream is starting
- Removing the first manual config no longer reverts
- Drop duplicate generated services (same cwd+command)
- Generation invariant: no empty launch target
- Make the target filter visible
- Warn on readiness-blind dependency roots
- Rejection voids every worker column from destination to gate
- Single-commit view hides base-inherited submodules
- The web leg carries real evidence
- Gatekeeper race on fresh checkouts: xattr + retry once
- Stable base-worktree dir + free prepare classification
- Commit list excludes base-integration commits
- Prepare failures surface the real error, not npm warn
- Kafka missing-topic errors get an identity
- QA legs run maestro-prepare + gate-parity env + fail-fast
- Batch investigation relaunches after file fixes
- Last wizard step reads "Finish", not "Continue"
- Diff panel stuck at "0/0 files" + cycling git logs
Faster
- Keep the .next cache between builds; skip the duplicate build-time lint
v0.1.352026-08-12
New
- Merge-failure Brain recovery; delivery failures never mangle main
- Add scenario by prompt — author one exact scenario from a free-text brief
- Scenario-gen UX — spinner, visible list, per-scenario regenerate, boot-degrade
- Grouped nav rail + flat sectioned card for app & project settings
- Autonomy — stop-what-you-booted, persist starters, one-shot scenario reconcile
- Run docs/demos even when empty (route auto-discovery)
- Loud untrack hygiene — suspicious untracks and ignored-source divergence are named with the responsible ignore rule
- Seed/ skill contract, capture.json compat, structure lint wiring
- Project layout contract + demo-runs spec docs
- Demo scenario lifecycle (M3c of Spec 315)
- Demo render + multilingual voice (M3a of Spec 315)
- Web demo runs on the docs-run chassis (M2 of Spec 315)
- Project layout contract (M1 of the Spec 315 conception)
Fixed
- Ban generic narration filler; ground scripts in PRODUCT.md + name
- Register demos_automation_enabled as a Labs gate
- 1 — restore develop's gate file; re-apply the lint surgically
- Never sweep build-output dirs (target/, node_modules/, …) into commits
v0.1.342026-08-11
New
- "Push changes to the remote" now gates ALL pushes (incl. wip)
- Full-width left sidebar with icons + selected-tab header
- Delivery is its own tab; project tabs move to a left rail
- Per-project "push merged changes to the remote" setting
- Fix-first triage — investigate, fix, verified retry; inline session-end facts
- Improve the code review screen ISSUES (#6)
Fixed
- Classify Higgsfield's underscore not_enough_credits as the typed credits error
- Total task time counts every phase VISIT, not one per phase name
- No background-task wake exists — never notification-wait or end a turn with backgrounded work
- Make the merge-mode control an explicit two-option choice
v0.1.332026-08-10
Fixed
- Precise ENOENT message in mixed build results + boot PATH-patch source pin
- Resolve the GUI-launch PATH once onto process.env — every spawn inherits it
- Toolchain PATH resolution everywhere Morning spawns project tools; bare Brain retries are propose-only
v0.1.322026-08-10
New
- Layered, submodule-aware repo map for every codebase scan
- Add IDE-style code review screen with sidebar and evidence
Fixed
- Loop guards hold — one failure signature per root error, Brain retries grant one round instead of resetting the budget
- Pre-review gate resolves the node toolchain and never fabricates a lint error from its own crash
- Keep the model's buffered doc when a repo scan exhausts turns
v0.1.312026-08-10
New
- RBAC-gated cross-project reads+writes for global/ops threads
- Multi-round text-protocol loop for narrating models
- Context-fit fuse — warn when a local model will truncate the prompt
- Make narrated-tool-call handling visible in the task timeline
- Recover tool calls a model narrates as text instead of calling them
Fixed
- Wip snapshot no longer skips a submodule when .morning-artifacts leaks in
- Scope hint reflects the OPEN thread, not the selected project
- Salvage reads the saved review, not narration; no false FAIL on "no blocking issues"
- Recover bare-name narrated calls; auto-commit only this turn's edits
- Execute tool calls a model narrates as text, then answer
- Persist a whole assistant turn as one ai_output card, not per line
- Coalesce streamed ai_output into whole lines (no more one-token cards)
v0.1.302026-08-08
New
- Discover models for custom/local (Ollama) endpoints
- Dedicated Ollama card + lock provider/transport on type-specific add
- Add a project language selector to the new-project wizard
- Move the Flows editor from the board top panel into project Settings
- View back in bar, blocked→menu, grouped dividers, leaner rows
- Reorder to Auto·Path·Design·Provider + Appearance menu group
- Remove the "System logs" execution dock
- "View diff" on the task panel → inline comments → backtracking
- Show pipeline progress on every card (backlog + needs-you)
- Localize account usage-window labels
- Localize settings copy, user menu, greeting + compact board-view settings
- Migrate final 71 surfaces via workflow (Tier 2 batch 3)
- Migrate 54 more surfaces via workflow (Tier 2 batch 2)
- Migrate 28 high-traffic surfaces via workflow (Tier 2 batch 1)
- Migrate first-contact surfaces — auth + errors (Tier 1)
- App-UI i18n engine — next-intl cookie-only (Phase B)
- Per-project artifact output language (Phase A)
- Settings clarity for non-technical users
- Image-generation install card + CLI-auth save fix
- Component-library-aware implement across frameworks
- Split task-card progress — phase steps inline, overall on the bottom border
Fixed
- Log tool COMPLETION on the API path so calls don't look stuck/duplicated
- Accept `cmd` alias for run_command + give repair the exact schema
- Discovered custom models use an obvious dropdown, not a datalist
- Don't mislabel every 404 as a missing /v1 (surface the real cause)
- Don't double-count one account under Ollama AND Custom / Local
- Default tool taskId to the current task (stops get_task_info({}) crash)
- Repair malformed tool calls instead of crashing the run (helps Ollama)
- Real health check + clearer setup for custom/local (Ollama) endpoints
- Raise max-images-per-design cap to 12 so saving 7 no longer 400s
- Scroll the add/edit-provider dialog body so the footer stays reachable
- No "Checking AI" flash when toggling Simple/Pro
- Show the real filename for deleted files (not /dev/null)
- Phase-count progress, hide 0%, show phase name when paused
- Match the diff button to the copy button (small, muted, left of branch)
- Three-dot diff — show only the task's own changes
- Omit only individually-huge files; smaller, lighter diff
- Show all files, readable diff colors, icon-only trigger
- Review pass — wrong-file comments, shell injection, gate misrouting
- Recover from a dead refresh token instead of retrying forever
- Keep "QA" verbatim + compact eye-icon board settings in full board
- Lock checkout on unverified accounts + honest expired-session error
- Indeterminate phase bar slides instead of filling full-width
- Guard avg cycle time against corrupt timestamps (389026.4d)
- Tapping a "Needs you" row opens the task
- Phase-steps bar shows only the OPEN attempt, never a stale prior one
v0.1.262026-08-06
New
- Platform-aware provider install instructions
- Liveness pill, provisioning milestones, ready toast (Spec 302r4)
- Version indicator + 'Update now' in the app (Spec 302r3)
- Loopback drain probe GET /api/internal/busy
- Membership-wide sandbox discovery in the app (Spec 302r2)
- Company sandbox, morning side
- Wire worker-mode boot + docs cross-links
- Remote execution concept + worker boot profile + auto-mode collision guard
Fixed
- Drop /login from history after sign-in (Back no longer returns to signin)
- Restore last project synchronously (no empty-state flash on load / screen switch)
- Theme-adaptive splash + window background (no dark-on-light flash)
- Review fixes — AC-6 tarball verification + no double-shipped migrations
v0.1.222026-08-06
New
- Neurons brand backdrop on the first-run splash
- Instant branded splash on first run (no more blank window)
Faster
- Defer non-critical boot work so the server serves sooner
v0.1.192026-08-06
Fixed
- Migrate provider-install tilde-expand to homeDir() (Windows EPERM)
v0.1.182026-08-06
New
- Real provider brand icons (@lobehub/icons) + card fixes
- Provider & integration marketplace
Fixed
- "signed in" for cloud actions = a linked account, not a live token
v0.1.172026-08-05
Fixed
- Tolerant markdown-block extraction (unclosed / wrapped fences)
- Resilient memory harvest — don't fail the whole init on one verbose memory
v0.1.162026-08-05
New
- Remove the status widget — the tray menu is the one shell surface
Fixed
- Gate on a linked cloud account, not a live access token
v0.1.152026-08-05
Fixed
- Visible tray icon + honest macOS update path
v0.1.142026-08-05
Fixed
- No Brain FAB on auth pages; external links never steal the app window
- Device-flow approve link never rendered — verificationUri field mismatch
v0.1.132026-08-05
New
- Cloud-only sign-in for desktop + cloud-linked web instances
Fixed
- Status banners are session-gated (no realtime/limit noise on the login page)
v0.1.122026-08-05
Fixed
- Real Claude auth check + clear "sign in" errors (not cryptic parse failures)
v0.1.102026-08-05
New
- Compiled-node-child MCP so Codex/Copilot get morning tools in the packaged app
Fixed
- MORNING_BUILD env is the reliable build-phase marker
- Build-phase DB connections go to a throwaway temp file
- Stop @vercel/nft globbing the whole home dir on Windows
- Strip runtime pollution from the standalone before packing (4GB → normal dmg)
- Run chat MCP servers in-process (type:'sdk') in the packaged app
v0.1.92026-08-05
New
- Single-device reaction + fix MCP server address
- Durable auto-update + Spec 298 widget remake
- 3 — remove the local@morning.app auto-admin everywhere; desktop uses normal cookie sessions from the cloud-first shadow identity
- Cloud-first desktop identity: first-run cloud sign-in (device flow + EULA), shadow local admin, token-expiry relogin gate with 14-day offline grace
- Branded auth frame (bg_neurons + wordmark), EULA-at-registration in /setup (server-enforced, audited), modal suppressed on auth pages, beta-aware trial banner
Fixed
- Clear the init lock live on project_init events
- Bundle skills/ into the standalone
- Run the morning MCP as an in-process SDK server
- No dead tools in an unbound floating chat
- SignedIn preview uses the refresh-capable token path
- Pipeline % measured across WORK columns
- Desktop WS client connects to the runtime port
- First-run cloud routes — malformed body cannot bypass consent; admin-exists race recovers with a session (review catches)
- /api/eula GET is public + middleware-allow-listed — the pre-auth /setup page needs it; a true first run was stuck otherwise (review catch)
v0.1.32026-08-04
New
- Declared platform targets as authoritative scope
- In-app "Report a bug" (cloud-proxied)
- Device registry in the account card (list sessions, sign out, this-device flag) via admin proxies
- MorningDEV rebrand (MD identity) + runtime/error hardening
- BETA badge + free-beta banner; beta flag through the sync (trial banners suppressed while beta)
- Desktop surfaces: cloudAiEnabled through sync, credits UI hidden while dormant, desktop-trial copy, banded Team label
- Desktop EULA gate (first-run modal, packaged builds only) + cloud-reported acceptance
- Email-verification state in the account card + send-verification button; human checkout-gate error (Spec 260r2)
- Mid-task AI decisions pause the task + answer in global chat
- Allow authors to select backtracking targets for approval gates
- Phase-aware "Needs you" titles and design-mode glyph in Simple View
- Gate the New Project wizard Documentation option behind docs Labs gate
- Gate experimental features behind Labs + settings/brand cleanups
- Redesign Simple View for clarity and scannability
- Higgsfield CLI image backend
- Labs tab with progressive disclosure
- Design scan sees the code on API + product scan verifies
- Studio first-draft grounding (provenance tool-gate + pre-draft quiz)
- Finish the run→launch domain rename
- Burn-rate coloring — amber/red when spending ahead of pace
- Harness_ensure_metro: per-flow Metro liveness re-check
- Deployment-topology visibility for generation + the Brain
- Add Codex account limits via the app-server JSON-RPC
- Surface reset time on the top-bar usage
- Keychain grant-access hint when the usage read is blocked
- Top-bar usage bars, all-provider fetchers, settings ui-kit, docs
- REAL Claude account meter (5h/weekly), the way Orca does it
- Explicit enable/disable Switch per provider
- Live provider usage under the provider (top bar + Capabilities)
- Idempotent mutation ledger (L1)
- Per-project scope + auto-import scheduler
- Candy UI-kit (feedback primitives + fast-win polish)
- Provider usage meter (F2) + budget threshold meter (F6)
- Capabilities matrix, Labs pane, scope badges
- M5 — settings UI + board import dialog
- Ingest tasks from GitHub/GitLab/Jira/Linear/Redmine
- One research scan (init) shown live in the wizard
- Promote/design-disabled deadlock + global rename directive
- Surface key project settings in the New Project wizard
- Direct-edit a plan iteration (deterministic escape from LLM refine)
- Shared "Add to batch" refine queue (Studio + approval)
- Init-aware new-project wizard + governed-doc codebase scans
- Doc-first Design library (+ in-flight Spec 253 wizard)
- The durable automation loop (journal, visibility, verification, escalation)
- /247 — phantom deferral, chat ownership, destructive gate
- Run config materializes into served worktrees
- Destructive-fix approval lives in the Brain's chat proposal
- Worktree-run state isolation (per-task DB + test data)
- Investigations get working tools; build-cache clears unbanned
- Diagnose classifies the fix LAYER before fixing
- One-click approval for destructive fixes + migration-collision knowledge
- Worktree envFile falls back to the project root; .next-run types; chooser back nav
- Task-card serving indicator + flow-run blocking
- Console-signal cold start replaces the 4-probe ladder
- Direct fixes on task-worktree runs — no intermediary
- Cold-start probe retry ladder + Brain handoff on real failures
- Structured Delivery section — branch, PR/MR, Submodules list
- V2 — paid Desktop ($10/mo, 14-day trial), team-only cloud, ribbon retired
- Transcript-driven diagnose fixes (env field, honest strips)
- Recovery quality: gated inputs, honest outputs, one owner
- Scan an existing repo into the greenfield review
- Per-task base branch (stacked tasks)
- Local base fast-forwards to origin at task start
- Morning.run.yml — the run config becomes a repo-resident file
- Git-hosting tokens move to the project Delivery section
- Settings → Git hosting — custom host mapping + UI-managed tokens
- Auto-create remote repos for NEW service submodules
- UI-first Design chooser — full-screen try-on + full redesigns
- The Design library — generate, store, select, and tune designs
- Onboarding guidance (tips, provider guard, signals)
- Per-iteration time breakdown (AI work vs approval wait)
- AC ids, Decisions table, testable negative scope
- Run a project instance from a task's worktree
- Stop button to cancel a live headless investigation
- One collapsed attachment strip for every chat surface
- Coverage catalog — living e2e/README.md matrix across tasks
- Teams inbound message relay (Spec 220) + remote-execution concept
- Reorder subtabs Timeline·Progress·Graph + per-phase durations on Timeline
- Live "investigating…" placeholder for background triage
- Cold-start the design system from the codebase during init
- Web folder picker for the Path field
- Merge Execution+Flow tab, rename Tokens→Cost, tab icons; messenger "Preview" button
- Per-project policy for pre-existing (drift) failures
- Harness_should_rebuild + gate warning so wrappers stop rebuilding every run
- Auto-commit the Brain's direct edits to a project checkout
- Render-check verdict chip on the message bubble
- Wire render-check into all refine surfaces
- Render-check diagnose+verify in the shared core
- Let the Brain edit small localized fixes directly (incl. visual)
- Prefix downloaded artifact filenames with the task title (length-capped)
- Route pre-existing e2e drift failures to Brain bugfix tasks
- Register new screens from the per-task plan (Spec 209 #2)
- Propose_task — one-click "Create task" button for new work (not prose, not a quiz)
- Copy-paste command hints in the channel dialog
- Command menu + per-messenger formatter + native buttons
- Universal render probe (harness_verify_app_renders)
- Branded Brain avatar on AI messages
- Broadcast user messages; gate logs typed as tools
- Message attribution — every human message names its author
- Visible loops — stopped is never silent; approval attests "I fixed it"
- Gitlink rescue — the machine's bytes are Morning's to publish
- Tested table + signature purity + visible base integration (Spec 198–200)
- Tested table — every testing artifact says what actually ran
- Delay the supporter ribbon to the end of the first week
- Sub-tabs inside the project Settings tab
- Root-failure fail-fast + entry canary — N identical failures are ONE defect
- Design pipeline as a positive switch; stop gating the rest of the tab
- Auto-apply is the default — flip existing 'review' projects to 'auto'
- Input-integrity rule — fail at the keystroke, not three steps later
- In-app skill editor (admin, edit .md on disk) — Spec 193
- Canonical iOS sim resolver + one test-visibility setting (Spec 191 + 192)
- Global rate-limit banner in the app shell
- Per-phase default capacities + drop the "WIP" qualifier
- Generic permission-fault detection — detect by error class, confirm by stat
- A real Stop for chat turns (Spec 186 follow-up)
- Read the SVG, don't inline it (Spec 186 T6+T9)
- Apply-as-a-tool on both transports + routes (Spec 186 T4-5)
- Apply-tool infrastructure — seam, endpoint, MCP server (Spec 186 T1-3)
- Extract shared apply seam (Spec 186 Task 1)
- 1-hour prompt cache for durable warm-start
- Declutter generation UI + delete sessions + relocate screens
- Auto-continue on quiz completion + Back button to revise answers
- Pin the live quiz on every chat surface
- Pin the live quiz above the composer + add a custom answer option
- Daily multi-project Brain ops chat
- Collapse Design document history + move it to the bottom of the tab
- Stoppable capture runs, issues split, Brain fix loop
- Stoppable runs, honest verdicts, Brain fix loop
- Weekly promo cooldown + discoverable lifetime purchase in settings
- Channel-aware inbound chat — the bot can act on what it pinged
- Free-tier supporter promo + lifetime unlock
- Product-defect routing, in-chat quizzes, external MCP servers (Specs 170 + 172 + 175)
- Scoped rejection routing + accumulated per-artifact review (Specs 168+169)
- Screenshots group in task artifacts tab (Spec 164 follow-up)
- Add existing screens mid-session
- Resizable + collapsible sidebar panels in review modals and Design Studio
- Read-only codebase access in chat + refine
- Visual tuning editor (Spec 160) + start from existing screens
- Sync app changes into registered screen designs
- Readability contract — word budgets, revision log, style gate
- Vertical-slice task sizing + per-run coarse granularity knob
- Orphaned-record debug count when nothing matches the phase (Task 13)
- Surface partial-stale approve outcomes (Spec 153 Task 14/15)
- Runner/QA/docs acquire the device lock for a pinned device (Task 4)
- Persist last mobile-launch target/device across restarts (Task 10)
- Native-toolchain remediation + clean escalation
- Surface FTS relevance-recall health in the tab (Spec 153 Task 20)
- Idempotent materialise + crash-recovery to the approval gate (Task 7)
- In-phase provider failover before blocking (Spec 153 Task 8)
- Expire stale design proposals + prune resolved ones (Task 9)
- Age-prune tombstones so deleted-id lists don't grow forever
- Pause on 401/403 with a re-auth banner (Spec 153 Task 5)
- Persistent admin alert on provider sticky-disable (Task 18 §17)
- Show run age + force-finalize a stuck run (Spec 153 Task 18 §10)
- Surface executed-flow-count on a mobile gate PASS
- Global realtime-offline banner on WS drop (Spec 153 Task 18 §20)
- Surface per-channel delivery health in the channel manager
- Persist per-channel delivery health (Spec 153 Task 18/19)
- Per-channel outbound rate cap (notify-storm guard)
- Unify thinking / tool-call UI across every chat surface
- Create tasks with enforced "blocked by" dependencies
- Thinking indicator on every chat surface
- Attach office documents and let the model read them
- Process-audit — systemic flow-by-flow gap + durability review
- Self-service display name + reorder phase-control cycle
- Multiuser completion — membership sync scope, attribution, audit-or-exempt
- Visible task-id chip + ?task= deep-link (worktree correlation)
- Design EDIT paths carry the knowledge fragments (Spec 148 follow-up)
- Behavioral principles fragment wired into plan, design, and review
- 'Run configuration' shortcut in the board run dropdown
- Stale dev-client automation + reserved e2e AVD + sim name disambiguation
- Interactive live build — scroll, device frames, liveness note
- Live generation progress (phase, KB streamed, elapsed)
- Parse, gate, and surface the plan TL;DR
- TL;DR contract — the whole concept in one reviewable paragraph
- Grouped settings surfaces + merged Automation tab
- Per-project device affinity + parallel mobile testing
- Owned-only gate scope + the task-ownership contract
- Left-rail assistant chat + hover-overlay refine visuals + theming
- Architecture onboarding + canonical ARCHITECTURE.md
- Brain-first background QA
- Living user docs (per-task updates + localized capture runs)
- Mid-task AI↔user decision channel + task steering
- Tasks 4-10 — distiller, consolidation, recall tool, staleness, user prefs, tests, docs
- Task 3 (+8 store) — memoryPolicy enum, proposals queue, source-gated writes
- Tasks 1-2 — FTS5 recall index + task-relevant injection
- Per-skill live web-research opt-in via frontmatter
- Web search in all chats, task-chat write parity, plan-chat direct edits
- Icon-only Simple/PRO toggle + theme icon in the top panel
- One chat stack, one tool display, one design-refine engine
- Tasks 13-15 — Design Studio (design-first) + cross-cutting tests + docs
- Tasks 11-12 — real-screen previews + Constructor full-doc apply
- Tasks 8-10 — Design Constructor directions + scenes
- Tasks 5-7 — design_doc_policy + Design-tab history/revert/queue
- Tasks 1-4 — flag hygiene + DESIGN.md as single source of truth
- Design badge is icon-only on the card — label moves to hover
- Native provider web search as the default mechanism
- Domain-research as a parallel Planning skill
- Common web research + grounded critique
- Self-clean committed build artifacts on run start (not just fresh worktrees)
- Run-start integration v2 + worktree lifecycle hardening
- Unified per-attempt context contract for every phase entry
- Drive the burning sparks from REAL token spend
- Flame emits rising "token" embers while LIVE
- Per-task auto-resume ceiling (bounds runaway idle-stall loops)
- "Complete" task menu item for last-column tasks (force → done)
- Auto -wipe-data cold boot when the Android emulator is out of storage
- Phase 3 — provider-form tier inputs + per-project tier overrides
- Phase 2 model tiers — per-provider tier_models + per-skill Settings UI
- Per-skill model tiers (economy/standard/premium), provider-agnostic
- Inject a gate-enforcement directive so the skill validates EVERY target OS
- Task Flow tab — per-task execution graph + timeline
- Skip-unchanged-phase caching (#6) + unified interrupt registry (#3)
- Replay-from-checkpoint (#2) + scoped reducers (#4) + fan-out primitive (#5)
- Erase iOS sims on teardown to reclaim disk (not just shut down)
- Surface a bell alert when the macOS Automation/TCC permission is missing
- Auto-apply dependency/product fixes (raised autonomy)
- Replace equalizer with a calm ember-glow live indicator
- Burn-up curve, runtime-health light, rate-limit cooldown
- Cockpit instruments — live "burning" indicator, pipeline, velocity, ETA
- Show the real phase on queued cards (Clock icon marks queued)
- Dashboard redesign — project card + stat cards + zones
- Exhaustion advances to the NEXT phase (universal) — supersedes replan-to-Planning
- Exhaustion moves forward to replan→Approval instead of hard-blocking
- On a build-gate failure, order the fix phase to RUN the real build
- Phase-advance waits in the queue (like a manual run) when the next phase is at capacity
- Show which task the Brain is actively working on
- Surface Brain messages — unread dot on the Brain button + bell
- Boot+hourly sweep to recover dropped blocker triage
- Use the shared floating Brain button (like PRO), drop the rail
- Auto-launch the mobile app preview on Metro-ready
- Simctl fallback when Expo's Simulator-focus osascript fails (TCC)
- Animate + resize lifted side panels, persist width per-user
- Rename "AI Execution Log" → "System logs"; drop stop button + AI icon
- Per-project Metro port so mobile previews run concurrently
- Detect silent port failures → Brain-visible signal
- Autonomy ON by default + triage ALL error classes (not just product)
- Make the Needs you section collapsible too
- Dynamically detect a FREE Metro port (was a fixed hash)
- Rebrand the chat as "Brain"; hide it behind a side button
- Make the Now and Up next sections collapsible
- Platform icons on feature titles + distinct queued state
- Persist per-project + read the durable truth
- Sort actively-running features above paused ones in NOW
- Open the chat on a fresh empty conversation each time
- Redesign — cockpit layout + chat rail, richer cards
- Embed full chat instead of brain-recents; swap the two lanes
- Simple Mode as the default screen, PRO board behind a toggle
- Per-project Metro port — let native e2e runs run concurrently
- Unify the brain + make autonomy durable
- Run the FULL native suite (smoke + comprehensive), like the web gate
- Force the e2e skill to fix EVERY failing flow, not just its own
- Post-escalation fresh in-place round — full self-heal ladder
- Failure-class escalation ladder — Testing×N → Implementing×1 → block
- Remove the per-task "Human-step skipping" feature
- Gather full prompts — capture the assembled system+user per run
- Layer 2 — board-level token analytics + outcome correlation
- Prompt-token analytics — section manifest + cache metrics per run
- Version documents instead of overwriting; browse history on review
- Surface triage in task log + project chat + bell badge
- Inline cycling icon on the Execution-tab phase list (drop the separate section)
- Per-task phase controls — skip or pause any phase
- Glab/gh setup hint + PR-auth health check under "Merge changes without PR"
- Send the 'done' notification on task completion (it was never fired)
- Per-subagent log division for ensemble candidates
- Canonical e2e-layer structure as the single source of truth (no manifest, no guessing)
- Surface fan-out in the Logs tab (visibility for N subagents)
- Per-column WIP capacity (option B) — serialise Testing via a phase-entry semaphore
- Default the `plan` skill to 2 subagents
- Make Brain config UI-settable (Settings → Morning Brain), not env-only
- Phase 4 — earned autonomy + anti-thrash closure + loop-trigger + channel push
- Phase 4 (part) — tool-equipped READ-ONLY autonomous triage
- Phase 4 (part) — config-override reach via update_task
- Phase 3 — conversational + actionable (the propose→act bridge)
- Phase 2 — runtime-exception fold-in (the second silent runner)
- Phase 1 — blocker triage (propose-mode, chat-surfaced)
- Move the subagent count to the column inspector panel
- Health-check + recycle a wedged device before reuse
- Headless android emulator + iOS sim by default (mobile_emulator_headless)
- Per-skill ensemble count control + API passthrough
- Ensemble skills (best-of-N + auto-synthesizer)
- Envelope the roadmap critique (third debate unified)
- Unify the debates' response-envelope reading
- Unify convergence across in-phase + cross-phase loops
- Surface sandbox denials + warn when the model lacks vision
- Vivid light-theme primary (14 88% 43%)
- Generic evidence manifest + investigate-first loop
- Make the failure CLASSIFIER screen-aware (feed it the screenshot)
- Surface the Maestro failure SCREENSHOT into the fail artifact
- Desktop local-single-user, folder picker, shared apiFetch, in-place New Project, R2 auto-update
- Reflect the cloud trial in the app (banner + Upgrade)
- Event-triggered instant sync push + tombstone derivation test
- Sync tombstones — propagate deletions across the team
- Automatic shared-project sync sweep + "Share to team" UI
- Sync engine (local) — board snapshot serialize/merge + reconcile
- Plan-upgrade checkout from the app (Stripe per-seat)
- Cloud account loop — device-code sign-in, entitlements, routes, UI
- Cloud-as-provider seam + local-only-by-default cloud session
- Warm the light-theme primary toward salmon
- Recolor primary from blue to Dawn coral (icon-matched)
- Add Morning sun app icon
- Package the app as native installers + build them in CI
- Autonomous emulator provisioning in harness lib (v2)
- Shared version-pinned Maestro harness lib (Option A)
- Open PR/MR via gh/glab CLI first, REST as fallback
- Open PR/MR via gh/glab CLI first, REST as fallback
- Move close + download actions to the header (match approval dialog)
- Move close + download actions to the header (match approval dialog)
- Log deployed build version at server startup
- Log deployed build version at server startup
- Stamp deployed code version into the run-start log
- Stamp deployed code version into the run-start log
- Open one MR per touched submodule, not just the parent
- Open one MR per touched submodule, not just the parent
- Persist delivered branch + PR/MR URL so it survives Done
- Persist delivered branch + PR/MR URL so it survives Done
- Autonomous environment remediation + env-class budget split
- Autonomous environment remediation + env-class budget split
- Feed cumulative replan history to the implementer (REPEATED-issue awareness)
- Feed cumulative replan history to the implementer (REPEATED-issue awareness)
- Open every artifact full-screen, like plan/design
- Open every artifact full-screen, like plan/design
- Event ("task") autopilots NOTIFY on a board event — never create a task
- Event ("task") autopilots NOTIFY on a board event — never create a task
- Live runtime-error pipeline — detect, classify, surface, (gated) auto-fix
- Live runtime-error pipeline — detect, classify, surface, (gated) auto-fix
- Human-step skipping on the create + edit forms (not just the detail panel)
- Human-step skipping on the create + edit forms (not just the detail panel)
- Channel selector for task-event autopilots
- Channel selector for task-event autopilots
- "Any column" for Target too + default both column fields to Any
- "Any column" for Target too + default both column fields to Any
- Optional source-column filter for event tasks ("Only when in column")
- Optional source-column filter for event tasks ("Only when in column")
- Auto-detect Telegram Chat ID (no hand-running getUpdates)
- Auto-detect Telegram Chat ID (no hand-running getUpdates)
- Collapse plan + design to one row each (opens the latest)
- Collapse plan + design to one row each (opens the latest)
- Inline "where do I get these" setup steps in the channel form
- Inline "where do I get these" setup steps in the channel form
- Trigger derives from kind (report=schedule, task=event) + "Any" event; fix UUID in notifications
- Trigger derives from kind (report=schedule, task=event) + "Any" event; fix UUID in notifications
- Headless android build primary (symmetric to iOS)
- Headless android build primary (symmetric to iOS)
- Headless iOS build as the PRIMARY path (drop expo run:ios)
- Headless iOS build as the PRIMARY path (drop expo run:ios)
- Read-only design viewer + click-a-design-artifact opens it
- Read-only design viewer + click-a-design-artifact opens it
- Clicking a plan artifact opens the read-only plan viewer
- Clicking a plan artifact opens the read-only plan viewer
- Read-only plan-review viewer (version history, no approve/reject)
- Read-only plan-review viewer (version history, no approve/reject)
- Plan version strip shows kind + "what was required" (2/2)
- Snapshot rejected plan as a version + plan-version-history lib (1/2)
- Plan version strip shows kind + "what was required" (2/2)
- Snapshot rejected plan as a version + plan-version-history lib (1/2)
- Signing-proof iOS fallback ends the code-signing loop
- Signing-proof iOS fallback ends the code-signing loop
- Surface pushed branch + PR/MR (or skip reason) in logs AND summary
- Surface pushed branch + PR/MR (or skip reason) in logs AND summary
- Push-only ALWAYS opens the PR/MR; source token from local authorities
- Push-only ALWAYS opens the PR/MR; source token from local authorities
- A gate-build failure tells the AI to FIX THE BUILD
- A gate-build failure tells the AI to FIX THE BUILD
- Delegate the native build to the project's own maestro script
- Delegate the native build to the project's own maestro script
- Surface the optional account token field for anthropic-sdk
- Surface the optional account token field for anthropic-sdk
- Per-account auth via CLAUDE_CODE_OAUTH_TOKEN (multi-account groundwork)
- Per-account auth via CLAUDE_CODE_OAUTH_TOKEN (multi-account groundwork)
- Stick-to-bottom auto-scroll on chats + task Logs tab
- Stick-to-bottom auto-scroll on chats + task Logs tab
- Per-step human-approval skipping — board list + per-task overrides
- Per-step human-approval skipping — board list + per-task overrides
- Configurable approval reject-target + always-require-human; plan data contracts
- Configurable approval reject-target + always-require-human; plan data contracts
- Bare-Xcode iOS support in the pre-maestro rebuild
- Bare-Xcode iOS support in the pre-maestro rebuild
- Rebuild + reinstall from current source before maestro
- Rebuild + reinstall from current source before maestro
- Label debate-loop evaluator groups as "Debate loop (…)"
- Label debate-loop evaluator groups as "Debate loop (…)"
- Explicit reviser selector for the debate loop + test/e2e evaluators
- Explicit reviser selector for the debate loop + test/e2e evaluators
- Move Export into the tab strip; remove Comments filter
- Move Export into the tab strip; remove Comments filter
- Group phase checkpoints by skill with friendly labels
- Group phase checkpoints by skill with friendly labels
- Run unit/integration in Implementing as a debate loop; Testing = e2e only
- Run unit/integration in Implementing as a debate loop; Testing = e2e only
- Export task execution logs as JSON
- Export task execution logs as JSON
- Universal up.sh provisioning at Testing entry + drift-aware failed-flows-first
- Universal up.sh provisioning at Testing entry + drift-aware failed-flows-first
- Mobile ownership scoping + orchestrator-run up.sh provisioning
- Scope web verdict to task-owned specs + known-red baseline
- Mobile ownership scoping + orchestrator-run up.sh provisioning
- Scope web verdict to task-owned specs + known-red baseline
- Wire Google Gemini as a provider (API adapter)
- Wire Google Gemini as a provider (API adapter)
- Per-autopilot channel — route a report to one channel (not all)
- Per-autopilot channel — route a report to one channel (not all)
- Note inbound/public-URL support under the channel type select
- Note inbound/public-URL support under the channel type select
- Slack (Socket Mode) + Discord (Gateway) inbound chat — no public URL
- Slack (Socket Mode) + Discord (Gateway) inbound chat — no public URL
- Telegram inbound via long-polling — chat WITHOUT a public URL
- Telegram inbound via long-polling — chat WITHOUT a public URL
- Inbound chat — reply to the bot in Telegram (epic phase 4, Telegram-first)
- Inbound chat — reply to the bot in Telegram (epic phase 4, Telegram-first)
- Event-driven triggers (epic phase 3)
- Event-driven triggers (epic phase 3)
- Report kind — scheduled stats digests to channels (epic phase 2)
- Report kind — scheduled stats digests to channels (epic phase 2)
- UI-configurable channels (epic phase 1 — foundation for report autopilots + inbound chat)
- UI-configurable channels (epic phase 1 — foundation for report autopilots + inbound chat)
- Settings UI — manage scheduled recurring tasks
- Settings UI — manage scheduled recurring tasks
- Scheduled recurring agent work (multica-inspired)
- Scheduled recurring agent work (multica-inspired)
- Agent-CLI auto-detection — scan the host for known runtimes (multica-inspired)
- Agent-CLI auto-detection — scan the host for known runtimes (multica-inspired)
- Multi-channel "come look" pings — Slack/Discord/Telegram/MS Teams (comparison-#10)
- Auto-open a PR/MR when a task reaches Done in push-only mode (comparison-#9, opt-in)
- Visual-regression gate — coarse design-vs-build drift (comparison-#6, opt-in)
- Multi-channel "come look" pings — Slack/Discord/Telegram/MS Teams (comparison-#10)
- Auto-open a PR/MR when a task reaches Done in push-only mode (comparison-#9, opt-in)
- Visual-regression gate — coarse design-vs-build drift (comparison-#6, opt-in)
- Per-project token & cost budgets with auto-pause/resume
- Apply the Anthropic prompt-cache breakpoint on the api path (was a no-op)
- Per-project token & cost budgets with auto-pause/resume
- Apply the Anthropic prompt-cache breakpoint on the api path (was a no-op)
- Exact rate-limit resets from the SDK's rate_limit_event (aif-handoff idea)
- Exact rate-limit resets from the SDK's rate_limit_event (aif-handoff idea)
- Convergence-aware backtrack — stop early when not converging (comparison-#A)
- Authenticated OUTWARD MCP surface for external clients (comparison-#B, opt-in)
- Convergence-aware backtrack — stop early when not converging (comparison-#A)
- Wire the HTTP MCP server behind KANBAN_MCP_HTTP (#1, stage 4 — cutover)
- Long-lived HTTP kanban MCP server (#1, stage 3, opt-in)
- Session-context registry — keystone for the long-lived HTTP MCP server (#1, stage 1)
- Wire the HTTP MCP server behind KANBAN_MCP_HTTP (#1, stage 4 — cutover)
- Long-lived HTTP kanban MCP server (#1, stage 3, opt-in)
- Session-context registry — keystone for the long-lived HTTP MCP server (#1, stage 1)
- Salvage an evaluator verdict from output when none was recorded (#3)
- Salvage an evaluator verdict from output when none was recorded (#3)
- Per-stage model override (KANBAN_STAGE_MODELS) — first cut of per-stage profiles
- Per-stage model override (KANBAN_STAGE_MODELS) — first cut of per-stage profiles
- Light/dark theme switcher + full color-token migration
- Capture device screen for the summary when Maestro emits none
- Capture device screen for the summary when Maestro emits none
- Persist unsent refinements + pre-fill them into the reject message
- Image card backgrounds + WYSIWYG preview (card matches the picker)
- Customizable project card background + settings entry + layout polish
- Back-to-projects icon + delete task in card overflow menu
- Completion forecast on project cards — est. days to finish all uncompleted tasks
- Post-login projects landing with per-project stat cards + quick switch
- Harvest e2e screenshots into the Implementation Summary + portable export
- Harvest e2e screenshots into the Implementation Summary + portable export
- Class-aware Testing gates — hybrid failure classifier
- Class-aware Testing gates — hybrid failure classifier
- Failure-class routing for the Testing↔Implementing loop (Spec 121 a+b)
- Failure-class routing for the Testing↔Implementing loop (Spec 121 a+b)
- Collapse topbar account cluster into one avatar menu
- WS auth + membership gating + presence (Spec 120 T15)
- Scope notification reads to assigned projects (Spec 120 T16)
- Audit log + attribution (Spec 120 T13)
- Optimistic concurrency on task edits (Spec 120 T17)
- Account security — pw-change, lockout, sessions, forced rotation (Spec 120 T14)
- Session-aware UI + user management (Spec 120 T9-10)
- Guard all API routes — capability + scope (Spec 120 T6-7)
- Login + setup UI; fix Edge build for middleware (Spec 120 T8)
- Auth routes + coarse middleware (Spec 120 T4-5)
- Foundation — schema, auth service, guards (Spec 120 T1-3)
- Upgrade to zod 4 + claude-agent-sdk 0.3.161 (matches CLI 2.1.161)
- Accumulate planner envelopes across tool calls under per-turn cap
- Show tool-call details in task logs
- Auto-generate downloadable "What Was Built + User Flows" artifact on Done
- Gate owns the authoritative Testing verdict (conclusion + run-tests)
- Detect app-didn't-render (build/bundle) failures, surface root cause
- Per-flow fixes, authoring carry-forward, visible Fix stages, android auto-boot
- Tag-scoped auto mode + TagEditor component
- Stalled-task watchdog catches Docker/sim/zombie hangs
- Emit per-platform task log entries so UI shows gate activity
- Per-skill carry-forward continue + retry-from-scratch
- Auto-scaffold i18n language-switch helpers + negative-assertion guidance
- Native-mobile testing track (Maestro) + engine-routing safety net
- Surface patch-apply failures + audit-pass for plan refines
- Plan review surface with cross-phase feedback routing
- Fold design/UI/UX review into review-plan and add faithfulness check in review
- Make feature-research cap configurable from the dialog
- Two-layer memory schema + internal-architecture capture
- Support submodule monorepos in worktrees
- Emit Tool complete events so UI shows tool-call duration badges
- File-based contentPath to bypass Copilot 16k per-turn cap
- Reasoning-effort plumbing across all three SDK adapters
- Render AI Output logs as markdown
- Forward images through all SDK adapters instead of stripping
- WEB/MOBILE/IOS/ANDROID/DESKTOP platform model + bundle fix
- Cross-platform framework awareness + mobile e2e foundation
- Roadmap feature-approval gate
- Design refine UX overhaul + roadmap/plan-run/infra specs WIP
- Full-screen review surface; compact approval dialog
- Chat-applied design patches (phase 3)
- Bidirectional design chat (phase 2)
- Figma-style in-iframe comment bubbles
- Refine batches, in-iframe pulse, side chat, history
- UX overhaul (Spec 099) — picker, kill switch, AI fallback, direction merge
- DependsOn + waitForPorts for service ordering
- Reconcile port from runtime when service binds elsewhere
- Maven/gradle stack addendums with Spring profile hint
- Detect nested frontends + per-service envFile
- Per-service enable/disable toggle
- UX in plan + native design tracks + multi-screen carousel
- Per-scene carousel in live preview surface
- Manual-trigger jobs run before older auto-queued FIFO
- Snapshot projectRoot WIP into a tagged commit before worktree
- Configurable concurrency cap + parallel-skill toggles
- Stop button to abort in-flight AI responses
- Preserve done checkpoints + add Fix tasks on backtrack
- Untrack already-committed files when adding gitignore baseline
- Warn when SDK chat hallucinates kanban actions
- Regroup Project Settings tab into headed sections
- Image attachments, copy button, drop stale-scope threadId
- Require homepage smoke test, fail on scaffolding stubs
- Pair tool logs + per-task implement checkpoints
- Floating AI chat with full app + bash access
- Write stack-aware .gitignore baseline on first worktree
- Fold Attachments into Details, reorder tabs
- Add tailwind-utility-clobbered rule
- Computed-style render-verifier catches cross-file rot
- Close implementation pipeline + make planner design-aware
- Add tailwind-default-suffix fix-pattern
- Add tailwind-v4-opacity-syntax fix-pattern
- Add vite-process-undefined fix-pattern
- Runtime render analyzer + pattern-matched auto-fix
- Refine UX — autoscroll, disclosure, selection bridge
- Flow-driven design positioning
- Surface design stages in planning when designMode=on
- Phase 5 + 6 — drop designSystemId load-bearing paths, docs
- Phase 4 — composer drawer surfaces brand layers
- Phase 3 — prompt-builder + verifier read brand_layers
- Phase 2 — preview route reads brand_layers (?projectId + ?layers)
- Phase 1 — pristine flag + brand_layers backfill
- Per-product-type multi-scene live preview
- Editable productType + platforms in Settings
- Layer-aware composer with preset picker (Spec 077 Phase 8)
- Design as first-class Planning step (Spec 077 Phase 7)
- Project-context awareness (Spec 077 Phase 6)
- Unified design skill + ios/android/desktop platforms (Spec 077 Phase 5)
- Platform targets + multi-platform design reference (Spec 077 Phase 4)
- Scan existing codebases for layers (Spec 077 Phase 3)
- Product type + platforms + per-platform styles (Spec 077 Phase 2)
- Composable layers + presets API (Spec 077 Phase 1)
- Bash audit + end-of-run stray recovery
- No-code authoring + user catalog (Spec 076 Phase 7)
- Catalog distinctness gate + picker UX (Spec 076 Phase 6)
- App→project→task scope chain (Spec 076 Phase 5)
- Post-generation verifier + alignment panel (Spec 076 Phase 4)
- Emit System contract block in prompts (Spec 076 Phase 3)
- Structured DesignSystemSpec parser (Spec 076 Phase 2)
- System primary, composition secondary (Spec 076 Phase 1)
- Phase 4 — visual polish on picker tiles
- Phase 3 — adopt drawer in NewTaskForm + ProjectDesignTab
- Phase 2 — DesignComposerDrawer foundation
- Phase 1 — tokenize thumbnails + route ?tokensFrom
- Teach NN to set designMode on UI tasks
- Retire dead surfaces + docs (Spec 073 Phase 6)
- Inject design + components into implement skills (Spec 073 Phase 5)
- Multi-platform design support (Spec 073 Phase 4)
- Visual template picker + skill thumbnails (Spec 073 Phase 3)
- Per-task toggle + visual system picker (Spec 073 Phase 1+2)
- Manual vs auto trigger; relax pre-merge gate for manual reruns
- Integrated kanban UX
- DesignOnly flag, runbook, seed script (Spec 071 Phase 4)
- Design→implement skill + apply-to-repo (Spec 071 Phase 3)
- Edit-in-place (Spec 071 Phase 2)
- Iteration strip + Refine button
- Project-level direction pin
- Add /skills catalog browser
- Export tokens as CSS / JSON from the catalog
- Phase 3 — catalog browser + comprehensive E2E
- Phase 2 — discovery form, flow override, device frames
- Adopt open-design engine into ai-kanban
- Add skills
- Model dropdown sourced from /api/ai/models
- Multi-provider chain with per-project pin + automatic failover
- Human-wait count chip alongside inFlightCount
- Coordinate task-Done merges + fix board freeze
- Multiple backtrack rules per column (one per evaluator skill)
- Stop recovery — terminate the loop without skipping through
- Opt-in Bash for the diagnose AI
- Copy-able shell commands inside diagnosis cards
- See polyglot monorepo children, not just Node
- Type-aware icons + mobile simulator/emulator launch
- Persist logs, observable+interruptible recovery, hardened lifecycle
- Server-side detection + recovery state survives navigation
- Visible multi-attempt AI recovery loop
- Pattern library + stack-aware AI diagnose
- SDK transport now applies the fix, not just describes it
- Zero-config UX — press Run, project starts, URL opens
- Post-test build gate runs only affected projects
- Post-test build gate — universal compile check after Testing skill
- Add universal run/restart button for any project type
- Retry inherits prior phase time so the timeline stays continuous
- Surface reviewer verdicts + separate dialog into labeled sections
- Preserve foundation, accumulate review feedback, stop shipping rejected plans
- Show in-flight task indicator per project
- Observability + Write tool in inline conflict resolver
- Redirect to Testing after AI conflict resolution to revalidate
- Inline AI conflict resolver between integration and block
- Integrate base into worktree at Done — under the hood, no skill, no reroute
- Universal task-progress contract + project-init blocking workflow
- Edit page replaces modal + new projects ship with default board
- Blocking init bg-process + slim project list + header cleanup
- Rename Approval column model + add humanStepKind variants
- Visual blueprint editor with xyflow + column CRUD/merge backend
- Per-skill telemetry, in-loop typecheck, diff-aware review, pre-review gate
- Ground decomposition in real project shape (digest + read-only tools)
- Fix kanban state and background workers on project selection
- Adoption + coverage + UX + quality overhaul (v2)
- Skills json output
- Move blockers and execution progress to dedicated tab
- Make task description expandable
- Add attachments to creation form and fix dialog layout overflow
- Add full-page settings and statistics view
- Consolidate dual plan/roadmap UI into single roadmap mode
- Add inline phase indicators and comprehensive tests
- Implement backtrack tracking - phases, time accumulation, and API routes
- Wire upload and preview into task detail panel
- Roadmap review
- Add ability to unqueue tasks from execution queue
- Merge conflicts resolution manager
- Git clean behaviour
- Markdown view
- Add markdown viewer with proper spacing and table support
- Implement task attachments with upload, validation, and storage
- Retest on failed test result
- Enable parallel skill execution within phases
- Add comprehensive project statistics
- Copilot sdk + openai fix
- Add task blocker tracking system
- Add manual review for tasks in approval columns
- Persist execution logs across page refreshes
- Add project editing and improve dropdown layout
- Autoexecute phase on task drag-and-drop
- Auto-expand active phase in progress view
- Auto-expand active phase in progress view
- Add branch visibility to task detail panel
- Add branch visibility to task detail panel
- Show pause button directly on running task cards
- Display durations in digital clock format
- Add drag-and-drop task reordering within columns
- Add autostart on move and approval
- Display artifacts in approval dialog for review
- Redesign with spent time display and improved controls
- Recover stuck tasks on startup and resume from completed phases
- Unify artifact viewing with modal dialog
- Add total task time display in detail panel
- Make description textarea expandable with 10 rows
- Show only current phase progress on task cards
- Add phase duration display in task details
- Add interactive AI chat for task Q&A
- Add task editing functionality
- Add filter buttons for log types
- Add multi-provider adapter architecture
- Drive tasks through flow on Run AI, simplify auto-mode
- Add delete task button with confirmation dialog
- Add project path display, persist last selected project
- Initial implementation of AI Kanban app
Improved
- Session selector header, background-generation glow + notification, layout polish
- Rail hierarchy, preview padding, New design action
- Design / Plan / Flows become icon-only with hover tooltips
- Widen task + chat side panels
- Widen the artifact dialog for HTML designs
- Move 'AI shell access' into the AI Diagnosis panel as a Switch
Fixed
- Stamp canonical verdicts on single-flow runs; nudge harness_ensure_metro adoption
- Render the server-reported cap, never a hardcoded 5 (review catch)
- 10s timeout on the acceptance report — never hang sign-in (review catch)
- Unclassified orchestrator errors park blocked with a blocker row, never idle
- Batch refine leaves a chat trail (was silent)
- Based-on-task worktrees keep the parent's submodule gitlink
- Board Palette glows for the pending-quiz state too
- A redesign brief reads the current implementation
- Live chat completion + don't resurface the quiz
- Reconcile brandLayers before studio generation reads it
- Self-heal the brandLayers cache when DESIGN.md drifts
- Headingless §2 palette parses; preview accent picks the brand
- Scanned shadcn palette shows its real colors, not a near-white strip
- Scanned "Current" card renders its palette swatches
- Don't deliver base-inherited submodule advances as the task's MR
- Refresh the studio "Current" system when DESIGN.md changes
- Preview loads fonts + the "Current" card shows swatches
- The preview scene renders colors — normalize HSL channels in token CSS
- Parse plain (no-backtick) DESIGN.md palette tables — the "no colors" bug
- Rename the Design card's "Scan codebase" → "Redraft…"
- Render Tailwind HSL-channel colors as swatches (Style Lab + palette)
- Pin the pre-draft quiz above the composer, not at the top
- Pre-draft quiz saves progress mid-quiz (survives navigation)
- Cool down until the account window actually resets
- Pre-draft quiz reuses the shared QuizCard (one quiz UI)
- A running app's button says "Restart", not "Launch"
- Follow-up — rename leftover user-facing Run/Runner strings
- The pre-draft quiz survives navigation (was strandable)
- Surface the real cause of a GET /screens failure
- Per-account usage — each provider row uses its OWN credential
- Brighter bar fills so meters aren't muddy on the light theme
- Recover usage bars fast on a fresh run
- 2 — scaffold lint for unscoped kills of shared test-infra process families
- Port-in-use on a compose stack routes to the Brain
- Recovery-loop coherence (pending-approval gate, two-tier consent, health-is-truth)
- Scaffold sabotage fails the leg pre-flow; render probe targets the resolved sim
- 2 — guard also denies inline Metro starts on the shared 8081/8082 family
- BashGuard denies inline cross-project device sabotage from skill sessions
- Run-tab config save targets the active run's worktree
- Sticky account-usage cache — stop the meter flickering off
- Hold "Stopping…" until the refresh actually ends
- Stoppable context refresh + cancelled ≠ failed
- Keep the top-bar reset time in the tooltip, not inline
- Project-scoped provider for chats + visible blocked-provider notice
- Carry the precise ALGORITHM, not just the structure
- Reconcile ACs against the shipped app (AC is intent, not proof)
- Reconcile plan concepts against the diff (code is truth)
- Show the re-planned version + reject label after a rejection
- Concept-faithful, plan-complete implementation summaries
- A4 — heal promoted sessions orphaned before the delete-hook
- 2 — pin/residue logic uses the pbxproj+app.json bundle-id union; pin the installed id
- Render tables that lack a blank line before them
- Reject a patch that unbalances code fences
- Version refines by the group key so they supersede the base
- Jump to the refined iteration + hide composer while refining
- Move projectForSet out of the route module (Next route-export rule)
- Tolerate smart-punctuation drift + auto-retry a failed plan patch
- Silent auto-continue on quiz complete — no "please continue" bubble
- Move the refine-batch action onto the preview toolbar
- Grounded codebase re-scan from the Design tab
- New EVIDENCE re-arms one fresh investigation
- Exit-0 before first-ready is a failure; dead ≠ stalled
- Cheap runtime digests fixed at all three layers
- Serving affordance follows card conventions
- The handoff line reports what the Brain actually DID
- Stale log-file discipline — kills the same-fix recovery loop
- TaskCard drops its QueryClient dependency — refresh rides the context
- Cold-start retry no longer clobbered by the drift timer
- Last page diagnosis stays visible after the service stops/fails
- Probe verdict is NEVER suppressed; ladder is additive
- Self-run copy gets its own Next dist dir (.next-run)
- Dev script port is env-driven — ${PORT:-4000}
- Run menu (incl. Run configuration) available in every phase
- Port bump refuses when provably useless — escalates to the Brain
- Port bump rewrites the COMMAND's hard-coded port, not just the field
- Port-in-use-self auto-bumps the runConfig port
- Fatal patterns without an auto-fix reach the Brain diagnosis
- Global chat opens at the tail + auto-selects the newest conversation; drop single-platform hint card
- Review opens the CLICKED artifact, not the stage's first
- Undeclared CSS vars: verifier rule + Style Lab token block
- Follow-up turns re-attach the newest prior image
- Placeholder-marker gate stops flagging legitimate UI placeholder idioms
- Approved actions confirm visibly — ops-chat + task-log line
- Drop the redundant Progress heading inside the Progress execution mode
- 5 — a late "fixed" verifier never silences live work
- 3/r4 — cascades aren't errors; limits reach the banner
- 2 — diagnose fix survives restart; ONE banner mechanism
- Investigation lights the pinned banner; Stop reaches the loop
- Review — manual Done path honors override push+MR routing
- No-progress stops are final; activity rides the banner
- Coherent run-recovery narrative, in-flight owns the story
- Diverged submodule gitlinks advance to the submodule's own base
- Headless probe classifies default-port emulator boots
- Spec docs stay TRACKED on self-runs; exact-path redirect elsewhere
- Port-block isolation for concurrent instances
- Submodule auto-merge targets the SUBMODULE's own base branch
- Keep pipeline paperwork + ALL scratch dirs out of the user's repo
- Evict sibling-project apps from the pinned sim; forbid the skill from borrowing foreign devices
- Shared-Metro-port squats are lint-flagged; harness gains the canonical iOS bundler pin
- Task duration total no longer banks sleep dead-time on completed phases
- Worktree-targeted runs investigate and diagnose the tree they execute from
- Specs 231+232 — pin the iOS bundle URL per run; lint wrappers that ignore the iOS session pin
- Plural runtime errors become ONE consolidated incident
- Cross-project parallelism (web port affinity, probe TTL, lock owners)
- Visibility setting can no longer lie in either direction
- Total time counts AI work only, even after live WS updates
- Recovery/comms wedge fixes (resolution resets the guard)
- Generate-from-context works on SDK/CLI providers, not just API
- Approval gate surfaces ALL of a selected stage's artifacts, not just the newest
- Stop no longer duplicates the last reply (skip full save after abort)
- Brain reply no longer disappears on tab switch (non-destructive refetch)
- Generate Roadmap dialog footer stays inside the dialog
- Total time matches the Progress tab (canonical, sleep-bounded)
- Ops-thread replies answer WITH the project's context, not morning's repo
- Render inline markdown in setup-dialog candidate previews
- Render ARCHITECTURE.md as markdown, not raw <pre>
- Total time excludes human-gate waits inside cycle groups
- "needs you" recovery chat carries the PROPOSED FIX, not just the cause
- Recovery chat never sticks on "investigating" — resolve on every terminal state
- Diverged local base no longer over-owns sibling specs
- Auto-gitignore the e2e scratch so it stops polluting the repo
- Give the runtime-error triage the crashed backend's own logs
- Collapse per-endpoint network errors + surface recovery in chat
- One coherent story — runtime-error triage defers to active recovery
- Schedule crons in local machine time, not UTC
- Don't bank a slept phase's idle wall-clock across retries
- Task-logs download uses the same task-title prefix as artifacts
- Tune color pickers no longer clipped/hidden
- Copy-message works over non-localhost http (insecure context)
- Show a chat-style working banner, not a Tool-call stack
- AVD provisioning must produce consistent, installed-image AVDs
- Pin-less projects get a per-project AVD, not the shared morning_e2e
- Brain task proposals show without a manual reload
- Surface the reason a blocked task won't run (Spec 149 follow-up)
- Mobile designs bucket into the iOS track, not Web (Spec 090/117)
- Carousel per-screen render fills the frame, not center-bleed
- Lock the plan on a design-scoped rejection
- Reconcile screenIds BEFORE the fan-out/screen-count gate
- Reconcile task.screenIds from the plan at design fan-out entry
- Reject data-screen on a non-<section> element
- Reject invented screen attributes at the plan gate (Spec 209 #1)
- Hide empty streaming bubble until first token; global chat routes new work to tasks
- Reject screen-attribute lookalikes (one-column stack)
- Pass visibility intent through to the affinity boot (recycle-windowed)
- Collapse the setup-hang killed message to the stable family too
- Killed-run message gives facts + investigate tools, not a flow-blaming guess
- Stable identity + environment class for 0-owned-executed coverage fail
- Min-h-0 on task + floating chat scroll areas so autoscroll works
- Pre-merge snapshot of dirty base — stop merges blocking on uncommitted tracked files
- Exclude human-step gate time from task work duration
- Idempotent manual Run — a repeat press is never weaker
- Re-pump the board when a mobile task leaves the device phase
- Whitespace-tolerant fallback for rendered-DOM SEARCH drift
- Count thinking as activity in the task refine chat
- Scroll the Edit channel dialog body
- Reachable run config + honest label during recovery
- One-way Android visibility recycle — stale headless emulators re-boot windowed
- Proposals remember they were acted on
- Don't mistake emulator -list-avds warnings for AVD names
- Morning-prefixed AVD name in the bootstrap (create + boot)
- Narrow android categorical block — respect AVD creation
- Blocked-task sweep idempotence per blocker, not per task
- Block fast on categorical android absence, don't loop the budget
- Audit-cover the skill editor PUT; pin keepMessageOnAbort semantics
- Nudge for a missing entry canary on app-not-rendering
- Route app-not-rendering to Development, not Testing-in-place
- Stable identity for app-not-rendering gate loop
- The gitlink guard verifies the task's delta, not the repo's standing debt
- One blocked message — the triage IS the notification
- Done-push failures create a git_error blocker
- Dedup identical ⚠ git: warnings at the log choke point
- Drop the wrong legend on the Types pie
- Dangling-gitlink guard fetches before refusing
- Content-aware submodule delivery — housekeeping merges never become MRs
- Scope newline preservation to paragraphs (kill the big gaps)
- Auto-scroll every chat to the bottom on open
- False "no flows found" on all-green runs — digit-bound the count probe
- Move model-tiers note, fix ?tab deep link, chart colors, drop design note
- Preserve newlines in chat messages
- Unique click selector for siblings sharing a class
- Profile avatar matches the task-card avatar
- Collapse the task-details Screenshots section by default
- Direction-aware drag reorder + no misleading drop animation
- Fit the lightbox image to the dialog, don't overflow
- Fallback never steals a foreign-pinned device
- Flip to 'running' on any non-running status at run start
- Fix test-surfaced follow-up defects in-task, never defer
- Unwritable-tree blocking + evidence freshness
- No queued-status flash on Run when a slot is free
- Rate-limit banner counts only limit blockers, not any retryAt
- AI status reads "Not available" instead of "none — not connected"
- Pin approval footer + add 'one per feature' granularity
- Surface ask_user_quiz so the Brain asks batches as a quiz, not prose
- Composer textarea auto-grows with its content
- Stop no longer removes the sent question
- Flaky-gate loop defenses — honest rebuilds, stable signatures, honest verdicts
- Thread project device pins into skill/wrapper runs
- Setup-abort false positive + verdict-parser hardening
- App-shell flex column so top banners don't break scrollable screens
- Show the thinking banner on switch-back (Spec 186 follow-up)
- Restore switched-away chat message + live "writing" bar
- Cache stable prefix + ride SVG as source, not vision
- Ceiling is a sanity backstop, not a budget — inactivity governs
- Restore generous ceiling — don't cut off a producing run
- Cap reasoning effort — design is generative, not a thinking puzzle
- SVG attachments ride as a reference, not inlined into the brief
- Stop agentic repo-crawl hangs — generate from injected context
- Accept SVG attachments (route as text source, not a vision image)
- Watchdog fires only on true INACTIVITY (thinking counts as active)
- No-PROGRESS watchdog + honest phase/banner (studio generation)
- Vacuous-green diagnosis says fix-the-RUNNER when owned flows exist
- Idle-based generation watchdog + live tool/thinking banner in studio
- Tool-call crawl re-arms the first-output watchdog
- Phase-aware watchdog + cancel — stop 30-min "reading your project" runaways
- Morning heals the retired global Maestro lock itself
- Quiz/side-channel broadcasts must append, not hijack the temp bubble
- Manual restart cancels the in-flight blocker triage
- Instance-scope the mobile gate boot pre-flight
- Count model thinking as activity + always clear the chat spinner
- Extract the LAST complete HTML document from model output (blank preview)
- Bare 'budget' regex made not-converging blockers invisible to triage
- Flow attribution on the wrapper path + in-attempt suite verify
- Auto-create per-project Android emulator on first run
- Render pasted images as small zoomable thumbnails
- Render data:image attachments in messages (were showing as filename text)
- Enforce feature coverage per declared platform — close vacuous-green gates
- Address Spec 175 code-review findings
- Dynamic base-branch resolution — follow the checkout, no env, no hardcodes
- Approval-dialog design chat sees the full design — unify on the studio core
- Platform picker preselects the project default regardless of the design switch
- E2e gate integrity — pin-aware scoping, mobile QA baseline, serial-guarded adb
- Default platformTargets to the project platform so tasks aren't shipped without e2e
- Promote a refined design as the screen head (refines no longer leave heads pinned to the pre-refine design)
- Post-test build gate writes a superseding PASS so a resolved failure stops looping the task
- Task chat can reach every artifact — full content via temp files + relevance order
- Design refines collapse to one latest version (implement no longer anchors on stale design)
- Verification claim gate supersedes its prior FAIL on recovery (Spec 169 r10)
- Stream tool activity to the studio chat banner (Spec 166 follow-up)
- Chat sees the full document + liveness-based turn guard
- Review suggestions — approval resets debate seed history
- Never truncate consumed artifacts + planning-loop guards
- Resolve screen heads across mixed platform keys
- Gate refresh-from-app on SDK transport (Spec 161 review #2)
- Surface SDK tool activity on task chat + unify the thinking banner
- Guard against dangling submodule gitlinks on delivery
- Atomic cross-process run claim (Spec 153 Task 16 review)
- Heartbeat lease so a live run isn't reclaimed cross-process (Spec 153 Task 16)
- Register launch timers/siblings for stop + sync first-track PID flush (Task 10)
- Don't cross-swap concurrent asks on WS broadcast (Spec 153 Task 11)
- Retry decomposition on an unparseable response (Spec 153 Task 7)
- Route native-toolchain build failures as environment (Spec 155 Phase 1)
- Hydrate FloatingChat unread badge on mount + WS reconnect (Task 3)
- Resync plan-run state on WS reconnect (Spec 153 Task 7)
- Supersede pending questions on abort/error, not just finalize
- CAS approve/cancel so a race can't clobber a terminal state (Task 7)
- Verify app still installed before native build-skip
- Rowid watermark eliminates same-second boundary loss
- Boot-reconcile orphaned 'generating' sessions to error
- Guard constructor apply against lost updates (base-hash precondition)
- Make promoteDesignSession race-safe (one task per session)
- Make applySnapshot atomic — one transaction, no torn board
- Default to no-patch banner on a change-framed ask
- Record collapsed cron ticks after downtime instead of silent skip
- Distinct user-facing warning when a configured backend is broken
- Completion gate trusts emit-time taskDone, not a post-hoc re-read
- Explicit disabled channel skips, never broadcasts to all
- Surface reconcile failure as a distinct needs-review notification
- Surface an unhonored model-tier request on the task (Spec 153 Task 20)
- Quarantine a poisoned project after repeated apply failures (Spec 153 Task 5)
- Reset everyN counter only on a real fire (Spec 153 Task 16)
- Boot-prune SIGKILL-orphaned QA/docs run worktrees (Spec 153 Task 16)
- Prune merged docs-refresh branches (Spec 153 Task 16)
- Synchronous FTS index write closes the crash-drift window (Spec 153 Task 14)
- Atomic interrupt claim prevents double-apply — Task 17 complete
- Sweep skips in-flight triage to prevent double-triage (Spec 153 Task 17)
- Expire open brain-action interrupts on task finalize (Spec 153 Task 9)
- Label degraded-classification fix tasks (Spec 153 Task 17)
- Surface a note when triage investigation fails (Spec 153 Task 17)
- Flag savedPartial after the real save, not before (Spec 153 Task 11)
- Persist inbound offset per-update to stop replay double-replies (Spec 153 Task 11)
- Preserve screen gates on blockedBy replace; drop redundant strip
- Guard blockedBy against dependency cycles
- Inbound routing always replies, even on total failure (Spec 153 Task 11)
- Explicit budget-exhaustion signal on propose (Spec 153 Task 17)
- Null worktreeInfo on teardown so leaked plan-run worktrees sweep (Spec 153 Task 7, partial)
- TOCTOU run-launch guard + budget-gate auto fix-tasks (Spec 153 Task 16, partial)
- Distiller in-flight guard prevents double-distill (Spec 153 Task 14, partial)
- Wall-clock safety timeout on both /ask routes (Spec 153 Task 11, partial)
- Reconnect resync + wildcard replay RBAC + backoff (Spec 153 Task 3, mostly)
- Durability cluster — recovery, re-entrancy, rebuild-cancel, sibling kill (Spec 153 Task 10)
- Clean up sync state + tags on delete/sign-out (Spec 153 Task 6, mostly)
- Bounded retry on outbound sends (Spec 153 Task 19, partial)
- Per-project in-flight latch coalesces concurrent syncs (Spec 153 Task 5, partial)
- Deterministic same-second log ordering via rowid (Spec 153 Task 13, partial)
- Resolve incidents when a task's blockers clear (Spec 153 Task 17, partial)
- Atomic + serialized DOMAIN.md writes (Spec 153 Task 20, partial)
- Expired cloud token no longer sticky-disables the provider (Spec 153 Task 8, partial)
- Supersede pending questions above the project guard (Spec 153 Task 9, partial)
- Suspend-tolerant watchdog deadlines (Spec 153 Task 12)
- Device lock on the direct Maestro path — closes the audit P0 (Spec 153 Task 4)
- Ephemeral-table prune sweep + fault-isolated boot recovery (Spec 153 Task 1)
- Atomic read-modify-write hardening (Spec 153 Task 2)
- Declutter the header — meta line + separate labels row
- Task-card actor avatar leads the footer on the left
- Render markdown in the TL;DR card (was showing literal * / **)
- Verify a stale build-gate verdict before escalating
- Self-healing base integration
- RCT_METRO_PORT is compile-time — bare-RN launches use RCT_jsLocation (device-verified)
- Bare-RN apps launch with SIMCTL_CHILD_RCT_METRO_PORT — deep links can't carry a port
- End expo at install-complete — its stray :8081 deep link can hit ANOTHER project
- Expo rejects --port with --no-bundler — heal the deep-link race by detection instead
- Kill the 'No script URL provided' race — expo deep-links to the right Metro
- Backlog-aware dependency blocking
- Make the expected osascript/TCC note actionable
- Stop Chrome autofill from ghosting the settings page
- The rejecting evaluator must re-judge — never carry its checkpoints
- Plan-debate feedback must not repaint a studio-approved design
- Show the TL;DR once — strip the markdown copy under the hero card
- Never stream raw patch code into the chat bubble
- Repeated components fan out to all instances instead of failing
- Selector-aware HTML windowing — the model could not see end-of-body elements
- Make the stream real — the idle watchdog starved on a silent SDK
- Stop killing healthy generations at the wall-clock (the 'aborted by user' error)
- Sleep-aware stall handling — keep-awake assertion + loop-guard exemption
- Apply ?tab= deep link post-hydration (review finding)
- Narrate the instant iOS launch path
- Verify iOS launches against the NATIVE project's bundle id
- Annotate expo's expected osascript failure in the run log
- Heal the "No script URL provided" cold dev-client launch
- Self-finishing expo run:ios — survive the osascript step, target the runner's Metro
- Reverse ALL service ports on android launch + Brain owns diagnose bash
- Expo run:android device naming + late adb reverse + iOS launch narration
- Visible pinned-emulator boots + honest picker + sticky iOS pin
- Spirit-guide launch fixes — verified iOS deep-links + per-project AVD creation
- Review of 6aff5de — unblock the event loop on sim boot + idempotent tests
- Durable multi-emulator mobile runs (spirit-guide Metro conflicts)
- Stop Done↔Testing livelock — ignore merged-in base commits in the unvalidated-changes check
- Honor 'skip' phase-control at backtrack re-entry too (parity with pause)
- Honor a 'pause' phase-control at backtrack re-entry, not only fresh forward entry
- Build the app from base before mobile docs capture
- Mobile designs fill the viewport — stop self-drawing a phone shell (phone-in-a-phone)
- Manual drag-to-Done runs the Done column skills (commit + summarize) before merging
- Structured Maestro verdicts, device leases, per-platform locks, build skip
- Remove env knobs — same-error cap hardcoded at 3, lifetime backstop deleted
- Signature-scoped loop guard + Brain-first blocker triage
- Visible selection on click + pending sweep through background refines
- Model runs survive client disconnects in all four review-chat routes
- Auto-resume the latest session per platform on load
- Tabs mirror the project's declared platforms; route renamed to /design-studio
- Project-scoped, background generation, attachments, per-platform sessions, full project context
- Honor designDisabled in the planner's design-context builders
- Auto-resolve native/config churn to base — mobile Done merges stop blocking on the same 9 files
- Make the build-artifact untrack async so it doesn't freeze the UI
- Make the untrack-on-add step work on large repos (85k tracked build files)
- Recover from a timed-out base merge instead of blocking Done
- Post-delivery race guard + RN/Expo build-output gitignore baseline
- Log merge success on the manual move-to-Done path
- Never force-sync a DIRTY submodule (preserve mid-phase work)
- Honor the e2e skill's failureClass verdict (stop the Testing loop)
- Don't block backend-diagnostic curl/docker (URL & API routes aren't file paths)
- Route a failed server-address setup flow as environment
- Don't claim "Patch applied — re-queued" when the patch wasn't applied
- Don't run the generic JVM `classes` task against Android/RN projects
- Don't freeze the UI while merging the default branch (async git fetch/merge)
- Stop duplicate "Skill stalled" blockers + pause orphaned runs
- Match pipeline legend colors to segments; liven the equalizer
- Publish task_updated on every status change (instant board updates)
- End the "Brain working" indicator when the Brain produces its answer
- Also delete the empty WIP-pushed remote branch on no-changes
- Don't push + open a PR when the branch has no real changes vs base
- Unqueue clears a stranded queued task; revert the stranding hybrid
- Resilient migrate — record + skip a drifted migration instead of stopping
- Free Android storage harder — clear staged APKs + logcat before evicting apps
- Capacity-waiting tasks bypass the stalled-skill detector
- Clear message when a mobile port can't be bumped (was "no free port")
- Scope the floating chat to the project on screen (URL ?project=)
- Open installed dev client via deep link instead of rebuilding
- Reflect Brain autonomy ON-by-default in the UI
- Reap orphaned worktree Metro/dev-servers at startup
- Never port-bump a mobile/Metro service (breaks the device app)
- Give the e2e SKILL the dynamic Metro port too (stop 8081 squat)
- Reap stale worktree Metro before provisioning (no 8081 squat)
- Actually kill running Android emulators in cleanup (emu kill)
- Clean up stranded mobile devices on boot + shutdown
- A full column no longer head-of-line-blocks the queue
- Inject RCT_METRO_PORT for Expo/Metro services (not just PORT)
- Manual runs bypass per-column capacity — run separately, not queued
- Idle-stalled skill auto-resumes instead of freezing the board
- Requeue in-flight tasks on shutdown so auto-mode resumes them
- Overwrite the stale FAIL with a PASS so a fixed task recovers
- Retry RESETS the backtrack budget so a fixed task can recover
- Surface a NEW thread (Brain triage) live, not only after reload
- Don't let a passing gate mask a failing sibling gate
- Distinguish paused from running in Happening now
- Kill a task's e2e daemons (Metro / dev-servers) before teardown
- Sort platform icons by a fixed canonical order
- Surface the e2e-skip cause + show platform icons on cards
- Scope the escalation to Implementing as PRODUCT-ONLY (no test edits)
- Name the off-screen/scroll cause of an assertVisible failure
- A phase reset must clear carry-forward so skills actually re-run
- Scan the WORKING TREE, not stale committed HEAD
- DnD into a column (and Retry-from-scratch) restarts the phase
- A phase-control pause runs through backtrack loops, not re-pause
- Manual resume of a blocked task resets the backtrack budget
- HEAD-independent hard loop cap — stop infinite retry on a recurring blocker
- Durable per-phase WIP snapshot — stop losing worktree (esp. submodule) work
- Route a real app-compile failure to Implementing, not Testing
- Group the synthesized plan into Planning, not OTHER
- Dedup the live artifact list + hide ensemble candidates
- A candidate's set_phase_conclusion must not abort its siblings
- Target the repo's real default branch + integrate submodule base
- Make the anti-loop guard worktree-state-aware
- Run the production build in Testing, not just the test suite
- Force-end the turn after set_phase_conclusion even when busy
- Resolve task flowId from type label on create/edit/run
- Manual resume clears provider limit-cooldown
- Detect touched submodules by committed gitlink, not drifted clone HEAD
- Drop the skip/pause badge from the task card
- Two loops on the same task — phantom tsconfig claim + e2e re-scaffold
- Manual move-to-Done now opens the PR/MR (parent + submodules), like the auto flow
- Enforce per-column capacity at DISPATCH (incl. manual resume)
- Handle a vanished worktree — restore from branch, else block clearly
- Cooldown pause must be 'paused', not 'blocked', or it never resumes
- Give the chat each task's WORKTREE path so it stops inspecting `main`
- Classify prepare/expo INSTALL failures as 'install' (env), not 'build'
- Move 🧩 subagent filter to its own sub-row under the main filters
- "Task run complete" fires only at genuine task-done, not every phase
- "Task run complete" no longer fires on blocked tasks
- Race-free stick-to-bottom (scrollbar-drag up no longer yanks)
- Stick-to-bottom no longer yanks the view when you scroll up
- Per-column mode relaxes the global cap; show count/max in column header
- Fan out on resume when the skill hasn't produced its artifact yet
- Re-run the producer on a backtrack — Fix items were skipped + rubber-stamped
- Persist subagent count — patch schema dropped flow.ensembleConfig
- Fail-fast + environment-route on a setup/backend abort (API not healthy)
- Auto-resumable blocker when all providers are in limit-cooldown
- Patch-id-aware unvalidated-commit check + clear stale gate conclusion
- Signature fingerprint for issue-less failures (silent-loop hardening)
- Prebuild native dirs in the worktree (Expo CNG gitignores ios/)
- Expo prebuild before iOS build — fixes "xcworkspace does not exist" loop
- Remove CI=1 from the gate env — it broke native expo builds
- Scope the tag picker to NOT-done tasks only
- Show the LATEST phase row's conclusion, not a stale early attempt
- Don't salvage a verdict for the gate-evaluated e2e skill
- Tell the agent the evidence/toolchain dirs ARE readable on a read-deny
- Drop non-null assertion (unknown eslint rule broke build)
- Confine outputFileTracingRoot (fixes Windows EPERM Application Data)
- Create-project silently failing + dashboard New Project card
- Reliably skip migrations during desktop build + retry on lock
- Ad-hoc re-sign macOS bundle after afterPack (fixes 'damaged')
- Skip DB migrations + backfills during next build (SQLITE_BUSY race)
- App icon + empty-window (better-sqlite3 ABI) + free ports
- Use Morning sun mark on the dashboard header
- Replace Bot header icon with Morning sun mark
- Apply Windows .cmd/shell fix to after-pack hook too
- Shell:true for Windows .cmd spawns (Node 20 EINVAL)
- Unsigned mac build, Linux homepage, Windows rebuild bin
- Harness lib v3 — clean the DRIVER, not the whole emulator
- Allow mobile/web test toolchain paths outside the worktree (cross-OS)
- Allow mobile/web test toolchain paths outside the worktree (cross-OS)
- Route server/runtime infra failures to environment, not product
- Close budget-leak under interruption + add maxBacktracks floor
- Route server/runtime infra failures to environment, not product
- Close budget-leak under interruption + add maxBacktracks floor
- Branch new tasks from origin/<base>, not the drifting local base
- Branch new tasks from origin/<base>, not the drifting local base
- Exclude desktop/ from root tsconfig type-checking
- Exclude desktop/ from root tsconfig type-checking
- Logs autoscroll no longer fights scroll-up under a fast feed
- Logs autoscroll no longer fights scroll-up under a fast feed
- Keep e2e harness/connection failures in Testing, not Implementing
- Keep e2e harness/connection failures in Testing, not Implementing
- Settle on process exit, not stdio close (orphaned-daemon root fix)
- Settle on process exit, not stdio close (orphaned-daemon root fix)
- Name the orphaned-daemon-FD cause so the e2e skill can fix it
- Name the orphaned-daemon-FD cause so the e2e skill can fix it
- Show the live Done-phase progress on the card (was a blank Done card)
- Show the live Done-phase progress on the card (was a blank Done card)
- Don't silently complete a broken delivery — block on PR auth-fail + un-pushable submodule
- Don't silently complete a broken delivery — block on PR auth-fail + un-pushable submodule
- Resolve planRun scope by run_id + scope asset upload by board
- Resolve planRun scope by run_id + scope asset upload by board
- Defer to PRODUCT-class skill FAIL only — don't loop the finished feature on unrelated drift
- Defer to PRODUCT-class skill FAIL only — don't loop the finished feature on unrelated drift
- Gate fails when the e2e skill fails — defer to ANY skill FAIL, not just product
- Gate fails when the e2e skill fails — defer to ANY skill FAIL, not just product
- Don't override a PRODUCT-class skill FAIL with an owned-spec pass
- Don't override a PRODUCT-class skill FAIL with an owned-spec pass
- Don't flag/recover intentional project-root writes (DESIGN.md, .gitignore)
- Don't flag/recover intentional project-root writes (DESIGN.md, .gitignore)
- KeyFailureLine ignores optional-step WARNED noise, prefers the hard assertion
- KeyFailureLine ignores optional-step WARNED noise, prefers the hard assertion
- Lead the failure reason with the REAL engine error, not Spec-115 boilerplate
- Lead the failure reason with the REAL engine error, not Spec-115 boilerplate
- A build killed for silence is a timeout, not a compile error
- A build killed for silence is a timeout, not a compile error
- Drive single-flow maestro-run.sh wrappers per-flow
- Drive single-flow maestro-run.sh wrappers per-flow
- Re-wiring an edge moves the column's board position too
- Re-wiring an edge moves the column's board position too
- Make scroll-up actually stick — gate follow on a ref, not racy state
- Make scroll-up actually stick — gate follow on a ref, not racy state
- A column inserted mid-pipeline renders at that spot on the board, not last
- A column inserted mid-pipeline renders at that spot on the board, not last
- Stick-to-bottom autoscroll instead of always force-scrolling
- Stick-to-bottom autoscroll instead of always force-scrolling
- Block on non-convergence instead of advancing to the (already-passed) Approval gate
- Block on non-convergence instead of advancing to the (already-passed) Approval gate
- Delegate native build to maestro-prepare (correct binary) + aggressive android storage self-heal
- Delegate native build to maestro-prepare (correct binary) + aggressive android storage self-heal
- Freeform plan reject snapshots a version too (history/diff after any reject)
- Freeform plan reject snapshots a version too (history/diff after any reject)
- Provision the runtime (Metro + seed) via maestro-prepare before native flows
- Provision the runtime (Metro + seed) via maestro-prepare before native flows
- Dedup event fires per source task — stop the per-event task flood
- Dedup event fires per source task — stop the per-event task flood
- Keep raw markdown in acceptance criteria so bold/code render
- Keep raw markdown in acceptance criteria so bold/code render
- Scope the Statistics route as a board, not a project (RBAC 404)
- Scope the Statistics route as a board, not a project (RBAC 404)
- Universal install-failure self-heal + classify by step, not error string
- Universal install-failure self-heal + classify by step, not error string
- One column field (filter = destination) + actually RUN the created task
- One column field (filter = destination) + actually RUN the created task
- Trim task card title right padding (pr-16 → pr-14)
- GenerateId works in non-secure contexts (LAN HTTP)
- Trim task card title right padding (pr-16 → pr-14)
- GenerateId works in non-secure contexts (LAN HTTP)
- Signing-proof iOS fallback must build the APP scheme, not a pod
- Signing-proof iOS fallback must build the APP scheme, not a pod
- Caller's MAESTRO_DEVICE must win; flag wrong-platform builds
- Caller's MAESTRO_DEVICE must win; flag wrong-platform builds
- Integrate latest base into the task branch before push
- Integrate latest base into the task branch before push
- Seconds-free compact duration that doesn't overlap the task title
- Seconds-free compact duration that doesn't overlap the task title
- A post-conclusion hang COMPLETES the phase, not a timeout-fail
- A post-conclusion hang COMPLETES the phase, not a timeout-fail
- Expo build must use --no-bundler (+ pin sim) so it terminates
- Expo build must use --no-bundler (+ pin sim) so it terminates
- Recoverable blocked tasks; iOS signing is environment not product
- Recoverable blocked tasks; iOS signing is environment not product
- A skill may only complete checkpoints it OWNS (no cross-skill ticking)
- A skill may only complete checkpoints it OWNS (no cross-skill ticking)
- Treat set_phase_conclusion as authoritative done — short grace then force wrap-up
- Treat set_phase_conclusion as authoritative done — short grace then force wrap-up
- Break the stream at the terminal `result` (root cause of stuck-on-conclusion)
- Break the stream at the terminal `result` (root cause of stuck-on-conclusion)
- Tool-aware idle watchdog — catch the SDK turn that never ends
- Tool-aware idle watchdog — catch the SDK turn that never ends
- On resume, scope the implement prompt to its OWN checkpoints
- On resume, scope the implement prompt to its OWN checkpoints
- On resume, skip worker skills whose checkpoints are already done
- On resume, skip worker skills whose checkpoints are already done
- MCP emit can't hang a tool; honest stall labels + low-disk fast-fail
- MCP emit can't hang a tool; honest stall labels + low-disk fast-fail
- Don't re-execute an already-finished phase on resume
- Don't re-execute an already-finished phase on resume
- Stop run events redirecting to dashboard-picked project + unclickable dialog X
- Stop run events redirecting to dashboard-picked project + unclickable dialog X
- Harvest mobile screenshots from the worktree + require flows to test to the result
- Harvest mobile screenshots from the worktree + require flows to test to the result
- A gate PASS must not override the skill's explicit FAIL
- A gate PASS must not override the skill's explicit FAIL
- Don't block a "no device booted" skip as a creds failure
- Don't block a "no device booted" skip as a creds failure
- Reject marks Approval phase failed, not completed (no 100%-green)
- Reject marks Approval phase failed, not completed (no 100%-green)
- Durable backtrack budget + a failed fix can't reach Done
- Durable backtrack budget + a failed fix can't reach Done
- Self-seed evaluator checkpoints so the loop is never invisible
- Self-seed evaluator checkpoints so the loop is never invisible
- Web gate PASS overrides the e2e skill's self-fail (stop the drift backtrack loop)
- Web gate PASS overrides the e2e skill's self-fail (stop the drift backtrack loop)
- Fold backtrack Fix checkpoints into the worker group (not a "Steps" group)
- Fold backtrack Fix checkpoints into the worker group (not a "Steps" group)
- Place Export in the Logs sub-filter row, not the main tab strip
- Place Export in the Logs sub-filter row, not the main tab strip
- Re-run the reviser's DOWNSTREAM worker skills too, not just the reviser
- Re-run the reviser's DOWNSTREAM worker skills too, not just the reviser
- Catch Maestro false-green + stop embedding failure-frame screenshots
- Catch Maestro false-green + stop embedding failure-frame screenshots
- Treat a transient unhealthy stack as environment-retry, not a block
- Treat a transient unhealthy stack as environment-retry, not a block
- Apply ownership scoping to the forced-reporter fallback too
- Apply ownership scoping to the forced-reporter fallback too
- MobileLaunchButtons — drop redundant target !== 'web' that broke the build
- MobileLaunchButtons — drop redundant target !== 'web' that broke the build
- Soften user chat bubble + float mobile device picker as overlay
- Soften user chat bubble + float mobile device picker as overlay
- Teams — support Power Automate Workflows webhooks (MessageCard was rejected)
- Teams — support Power Automate Workflows webhooks (MessageCard was rejected)
- Hide Target column for event-driven tasks (event fires regardless of column)
- Hide Target column for event-driven tasks (event fires regardless of column)
- Friendly schedule, derived report period, drop Flow, scrollable dialog
- Friendly schedule, derived report period, drop Flow, scrollable dialog
- Require screenshots on MOBILE + for any UI task (was web/design-only) — empty summaries
- Require screenshots on MOBILE + for any UI task (was web/design-only) — empty summaries
- Block on alwaysLoad + fast-boot the kanban MCP child to stop status='pending'
- Block on alwaysLoad + fast-boot the kanban MCP child to stop status='pending'
- Legible dark log console in both themes + eye-comfort palette pass
- Make ThemeToggle segments share width so they fit the account menu
- Widen the MCP-connect retry (startupTimeout is inert) to span the post-gate load spike
- Widen the MCP-connect retry (startupTimeout is inert) to span the post-gate load spike
- Clear stale FAIL artifact on pass; drop misleading final-state device screenshot
- Clear stale FAIL artifact on pass; drop misleading final-state device screenshot
- Stop re-running e2e authoring on backtrack — sync prompt skip-markers with the row
- Stop re-running e2e authoring on backtrack — sync prompt skip-markers with the row
- Give the kanban MCP server a generous startup timeout (fragile-timeout root cause)
- Give the kanban MCP server a generous startup timeout (fragile-timeout root cause)
- Re-running tests must also reset the downstream results checkpoints
- Re-running tests must also reset the downstream results checkpoints
- Retry the SDK session on intermittent kanban-MCP connect failure
- Retry the SDK session on intermittent kanban-MCP connect failure
- Hide harvested e2e screenshots from attachments + export only referenced
- Hide harvested e2e screenshots from attachments + export only referenced
- Force JSON reporter when no report found, instead of hard-block
- Force JSON reporter when no report found, instead of hard-block
- Deterministically rebuild native app on missing-module crash
- Deterministically rebuild native app on missing-module crash
- Launch tsx MCP servers from local bin, declare tsx as dependency
- Launch tsx MCP servers from local bin, declare tsx as dependency
- Startup Claude CLI↔SDK version-skew check + skip wrong-platform pinned CLI
- Block the task when the kanban MCP server never connects (don't silently stall)
- Block the task when the kanban MCP server never connects (don't silently stall)
- Clean cover-band card background (drop the muddy scrim)
- Project page "Back" returns to wherever you came from (router.back)
- Project page "back" goes to the dashboard, not a hardcoded board
- Back-to-projects is icon-only with a tooltip (no inline text in the topbar)
- See test files inside git submodules (stop the false "claimed-but-missing" loop)
- See test files inside git submodules (stop the false "claimed-but-missing" loop)
- Bound stalled/hung phases in project statistics duration (no more "duration stack")
- Bound stalled/hung phases in project statistics duration (no more "duration stack")
- App sign-in 401 / E2E_PASSWORD is the e2e skill's to fix (retryable), not a human block
- App sign-in 401 / E2E_PASSWORD is the e2e skill's to fix (retryable), not a human block
- Never carry the "Run tests" checkpoint forward — re-run + re-validate every entry
- Never carry the "Run tests" checkpoint forward — re-run + re-validate every entry
- Mobile gate resolves Android SDK env itself (adb/emulator/maestro on PATH)
- Mobile gate resolves Android SDK env itself (adb/emulator/maestro on PATH)
- Mobile gate blocks-for-human only on USER-OWNED failures; device/toolchain stays retryable
- Mobile gate blocks-for-human only on USER-OWNED failures; device/toolchain stays retryable
- Drop single-skill idle timeout (it killed long e2e); rely on a generous wall-clock cap
- Drop single-skill idle timeout (it killed long e2e); rely on a generous wall-clock cap
- Clamp idle tails on terminal phase rows closed long after work stopped
- Clamp idle tails on terminal phase rows closed long after work stopped
- Auto-boot an iOS simulator in the mobile gate + fuller iOS setup in e2e skill
- Auto-boot an iOS simulator in the mobile gate + fuller iOS setup in e2e skill
- Mobile coverage gate blocks (non-retryable) when nothing ran
- Mobile coverage gate blocks (non-retryable) when nothing ran
- Stop counting pause/idle tails in per-task Total Time
- Stop counting pause/idle tails in per-task Total Time
- Integrate origin/<base> too + make run-start base refresh visible
- Integrate origin/<base> too + make run-start base refresh visible
- Refresh task worktree from base at run start (stale-infra fix)
- Cap single-skill execution with an absolute wall-clock ceiling
- Refresh task worktree from base at run start (stale-infra fix)
- Cap single-skill execution with an absolute wall-clock ceiling
- Web e2e gate catches "pre-feature frontend image" feature-UI skips
- Web e2e gate catches "pre-feature frontend image" feature-UI skips
- Block non-retryable infra/credential e2e failures immediately instead of looping
- Durable web e2e gate — never ship a web task whose browser suite didn't run
- Block non-retryable infra/credential e2e failures immediately instead of looping
- Durable web e2e gate — never ship a web task whose browser suite didn't run
- Block task when backtrack budget exhausted instead of advancing
- Block task when backtrack budget exhausted instead of advancing
- Setup/login form feedback + Secure-cookie over HTTP
- Make e2e tests required and block if was not executed
- Mcp server calls update
- Don't count pause→resume idle time as phase duration
- Pin the Claude CLI to the SDK's version so MCP works
- Claude path resolution fails loudly + health check matches runner
- Extract resume hints from provider limit errors for auto-resume
- Debate-loop evaluator dropdown can show/select non-column critics
- Repair stale Planning debate evaluator ('plan' → 'review-plan')
- Relabel pre-merge WIP commits into proper conventional commits
- Make mobile-singleton barrier symmetric + device-phase-only
- Scope mobile-singleton barrier to the device phase, not task start
- Guard the APP-facing backend URL + enforce app==harness isolation
- Reset a partially-run skill's checkpoints on resume of a paused phase
- Deterministic backend-isolation guard + hardened discovery
- Enforce isolated backend + speed up Android Maestro input
- Stop false 'missing artifacts' on '|' separators, design, prose plans
- Scope authoring carry-forward to intent='auto' only
- Mobile-coverage gate owns the run-tests checkpoint
- Durable Maestro/Playwright runner — stream, time-out, decide
- Un-block event loop during mobile-coverage Maestro runs
- Break continue-loop + unblock Maestro test reads
- Continue short-circuit on 100% carry + move overflow to top-right
- Filter test-tooling noise from Implementation Fix: items
- Maestro-run.sh auto-recovers from iOS driver-session flakiness
- Test-plan review on re-entry + bound stalled-phase idle in durations
- Make maestro-run.sh a UDID-resolving translator; align gate args
- Use --platform not --device for Maestro + skip unavailable platforms
- Drop invalid --format=json from Maestro gate invocation
- Stop Maestro orphan/watchdog leak + bound the hang vectors
- Serialize mobile-native tasks (simulator is a singleton)
- Serialize + timeout-guard Maestro runs (sim-singleton + hang prevention)
- Mobile target runs BOTH iOS and Android (drop Spec-117 collapse)
- Close brandLayers race + reject duplicate screen ids
- Preserve submodule commits across Done-stage merge
- Include .kanban-artifacts/ in managed gitignore entries
- Emit per-tool task_log rows (not a bulk \"Tool calls: a, b, c\")
- Force contentPath in user-prompt + tail-anchor (system-prompt update wasn't reaching the model)
- Carry over plan prep checkpoints when a prior plan exists (iteration mode)
- Mandate contentPath for design save (refine + fresh modes)
- Restrict project-init reads to user workspace
- Pre-complete design prep checkpoints when refine mode applies
- Harvest memorySuggestions only when task is accepted
- Tolerate bulleted \`- **Screens:**\` marker so uxFingerprint isn't empty
- Stop re-running design on prose-only plan edits + preserve original checkpoint timestamps
- Structured tool-call output replaces text-JSON parsing (research phase)
- Strip per-screen scaffolding so previous-frame label doesn't bleed in fluid viewport
- Broaden fan-out gate + model-driven screen discovery
- Orchestrator-level per-screen fan-out, drop chunked-save
- Actually populate execInput.reasoningEffort from skill defaults
- Jsonrepair fallback rescues AI JSON with trailing commas + unescaped newlines
- Neutral diagnostic when save_artifact never fired
- RFC 6266 Content-Disposition for non-ASCII filenames
- Land paused status synchronously, don't overwrite
- Per-provider maxSteps ceiling for streamText
- Run scripts in carousel + scope refine to active screen
- Trim irrelevant artifacts, fix silent SDK hangs
- Suppress legacy refine side panel in lifted mode
- Tighten compact card; raise surface above the dialog
- Unblock production build for refine-state guard + design-chat helpers
- Plan + design were running in the same parallel group
- Render image thumbnails inline + dark-mode contrast
- Remove direction list from catalog page
- Reset recovery budget on manual re-diagnose
- Cancel stale queued jobs on server restart
- Show configured provider label in top-bar AI status
- Keep task.status=running across phase boundaries
- Correct bypass flag, terminal-row mutation, retry re-entry
- Honor Stop button on the Done phase
- Refuse SIGKILL when child.pid is stale or recycled
- Break paragraphs between AI step text and tool calls
- Rely solely on .gitignore + surface gate events
- Auto-take base for ignored / build-output paths
- Never paint stale 100% over a bypassed audit row
- Translate Claude tool names so model isn't stranded with zero tools
- Chat MCP — auto-resolve board/column + workspace overview tool
- Retry pass 1 without digest when agent calls zero tools
- Floating chat no longer creates a thread on every open
- Commit our .gitignore writes so the next worktree passes assertBaseClean
- Attribute review-plan to Planning, not Other
- SDK chat gets MCP kanban + artifacts grouped by phase
- Refuse global chat under SDK transport instead of lying
- Cancel in-flight run on manual move so new column gets fresh execution
- Real checkpoints + update_progress nudge
- Pause source phase during backtrack detour
- Clear stale gate artifact, self-heal task_moved cache
- Normalize Next.js → 'web', seed e2e in Testing
- Detect sibling frontend when root has package.json
- Pre-review gate filters warnings + keeps row open on FAIL
- Smaller FAB, lift above log dock
- Stream tool-call events + strip system_notification leaks
- Floating chat — left FAB, thread naming, real tool calls
- Dedup floating-chat race between WS broadcast and SSE stream
- Collapse per-card 1Hz timers + drop /progress refetch storm
- Unref poll timer + WS-driven UI to stop hang and fix Ctrl+C
- Block set_phase_conclusion when canonical artifact missing
- Auto-complete "Save X" checkpoint after save_artifact succeeds
- Force canonical name/type in save_artifact to keep upserts stable
- Always inject stack profile into plan-skill prompts
- Stop NODE_ENV=production leak into spawned dev servers
- Set type="button" so toggling never submits the parent form
- Exempt .kanban-worktrees/ from base-clean check
- Sync save_artifact enum with schema artifactType
- Use existing design as spec; never produce design artifacts
- Make brand-layers contract directive on radii + border
- Pristine projects auto-refresh brand_layers on boot
- Strip post-</html> trailer; parse markdown-bold radii + border
- Render HTML artifacts in iframe, not raw source
- Render HTML artifacts in iframe even when type='other'
- -muted is a surface, not the chroma `muted` token
- Canonical swatches + colon-label normalization
- Live preview thumbnail for unified design skill
- Description-update regression + parity with NewTaskForm
- Escape JSX entities to satisfy react/no-unescaped-entities
- ResolveDesignChain — read task fields through unknown bag
- Recover from "untracked working tree files would be overwritten"
- Modernize seed + runbook for Spec 073 per-task model
- Paste + drop file capture
- Plug auto-mode leaks that hung the page and blocked Ctrl+C
- Escape apostrophe in skill-folder hint
- Align dropdown items with trigger text
- Align dropdown items with trigger text
- No-creds SDKs + per-provider setup hints + readable errors
- Break blocked-by retry loop, rework prerequisite UX, unstick recovery panel after Stop
- Trim replan-context to current round + chunked log flush with backpressure
- Static-site shape + test-script annotation so AI doesn't propose `npm run e2e` as the run target
- Cancel orphan poll loops, drop stale auto-mode entries, debounce PID flush
- Render every backtrack rule, not just the last one
- Normalize backtrackConfig in getBoardFullState too (was bypassing the service-layer normaliser)
- Reset reviser+evaluator checkpoints between iterations
- Drop '|' delimiter, carry over preserved-artifact skills, add evaluator checkpoints
- Debate-non-convergence advances to next column for human decision
- Show 'needs revision' phases as amber, not solid green
- Detect kLSApplicationNotFoundErr + show full err message
- Allow enable_diagnose_bash through PATCH /api/settings
- Kill orphan children on fresh Start + surface real xcrun/adb errors
- Pre-flight resolves ./mvnw against service cwd, not kanban-app cwd
- Reflect server-side detection + recovery state on the header button
- Unstick the 'Restarting' phase + emit exhausted at limit
- Help AI return valid updatedRunConfig + recover when it doesn't
- Launch mobile via xcrun/adb deep link, not stdin keystroke
- Infer mobileStack from command when AI omits it
- Validate AI-proposed runConfig before applying
- Unstick on interactive prompts + auto-apply fixes from quick-death
- Logs no longer flash + lines no longer cut off
- Config gen sandbox + sub-dir manifest detection + UI gates
- Pre-flight failures defer to post-loop, always fire recovery
- Pre-flight + quick-death failures route through recovery loop
- Config edits stick — auto-save on remove, dirty-flag inputs
- Keep recovery panel visible across mid-cycle restart
- Trigger AI diagnose on readiness timeout + include stdout
- Readiness probe waits as long as logs are flowing
- "Running" now waits for the port to actually answer
- Give header its own stacking context so Run dropdown wins
- Pass configured port through PORT env var to spawn
- Resolve ESLint and Next.js build warnings
- Pre-flight port check + self-run detection
- Suppress auto-recover on quick-death + show stderr
- SIGTERM the whole process group, not just the shell
- Handle agent transcripts + leaked tool-call markup
- Strip markdown fences before parsing AI diagnosis
- Actually invoke SDK executor (was awaiting handle)
- Tolerate fenced/prose-wrapped/empty AI output
- Also swallow ADD-COLUMN race on multi-worker migration startup
- Collapse shared ScrollArea on Chat tab to remove the empty bottom strip
- Show actual progress on cancelled/paused phase rows + prefer active over completed
- Resolve conflict markers in mcp-kanban.ts + boardId-fix test that shipped via 3e074b1
- Refuse to auto-commit a worktree that contains conflict markers or has MERGE_HEAD
- Redirect to Testing when worktree has unvalidated commits since Testing completed
- Discard local generated-artifact mods in projectRoot before mergeAndCleanup checkout
- JSON-output fallback when both tool calls and markdown bash blocks fail
- Pre-merge gate also runs on advanceAfterResume's terminal-column path
- Wildcard subscribers receive per-board broadcasts
- Auto-commit dirty worktree before pre-merge integration
- Support SDK transport (Copilot SDK / Claude Code) — branch executeChat by transport
- Pin prose elements to --foreground so headings/strong/code stay readable
- Execute markdown bash blocks as fallback when SDK didn't fire tool calls
- Conflict resolver via Bash heredoc + flip queued→running at run start
- Pre-feed file contents + Write-only directive prompt
- Integrate against LOCAL base, not origin/<base>, in worktree integration
- On-demand artifact tools + surface empty-response and real errors
- Feed merge-conflict context to the implement skill via rejection-feedback channel
- Retire phase rows on merge-conflict reroute so they actually re-execute
- Defer to debate verdict instead of fighting it with the artifact gate
- Retire terminal rows on superseded close so reject re-executes
- Pick the configured evaluator's artifact, not newest of type
- Self-heal projects without a board on first GET (kanban switch was a no-op)
- Drop hardcoded HIGHLIGHTS allowlist for universal deps
- Bias output toward implement-feature tasks (no research verbs)
- Wire allowedTools so planner text-only contract applies on Copilot
- Whitelist tsconfig.tsbuildinfo + specs/** intentional writes
- Pass boardId through completeCheckpoint/failCheckpoint/setPhaseConclusion
- Cap merge-conflict recovery to stop infinite reroute loop
- Pass boardId to progress service functions for real-time SSE events
- Forbid AI from asking clarifying questions; better parse error
- Run roadmap inside the worktree + forbid AI from writing files
- Cap vitest forks at 4 + per-worker DB to prevent memory blowup
- Make git lifecycle fail-loud instead of silent wipe
- Worktrees sandbox
- Enable image paste from clipboard in task creation
- Enable image paste via clipboard in upload component
- Filter E2E test projects from list
- Missed plan added
- Prevent roadmap artifacts from being merged into main branch
- Resolve checkpoint name mismatch in phase init
- Handle 'need revision' routing and prevent infinite replan loop
- Replace per-hook EventSource with shared WebSocket subscriptions
- Fix tool call formatting in execution logs
- E2e tests port
- Abort stale board fetch on project change
- Project clear on tests
- Build issue
- Task performing issues
- Git clean
- Inherit theme foreground color in prose
- Pass KANBAN_API_URL to child process for event emission
- Correct statistics endpoint board validation
- Task dialog reset
- Solve conflicts on merging branch
- Copilot integration
- Add missing SSE event types for blocker events
- Resolve type issues in chat implementation
- Eliminate message duplication, empty responses, and add AI greeting
- Build issues
- Build issues
- Sync EditTaskDialog form state when dialog reopens
- Prevent test files from being created in project root
- Tasks flow and worktrees
- Worktree
- Revive debate + backtrack loops for existing flows
- Pin MCP child DATABASE_URL so artifacts reach the main DB
- Resolve unknown tool names in execution log
- Updating ui issue
- Handle unknown tool calls in execution log
- Claude path recognition
- Add INTERNAL_SECRET to styles webServer env to fix TS build error
- Assign task flow from planned type for branch lifecycle
- Fix phase progress stale closure and add diagnostic logging
- Prevent paused tasks from auto-advancing to next phase
- Prevent position collisions when tasks complete concurrently
- Prevent card jump animation during drag-and-drop reorder
- Preserve task progress when stopping or resuming execution
- Add artifact gate to single-column (no-flow) execution path
- Show duration in common M:SS / H:MM:SS format
- Reduce noise from status checks, model resolution, and SDK output
- Load historical logs on page refresh
- Retry phases with missing artifacts + externalize ws from bundle
- Fix layout by moving chat tab outside ScrollArea
- Ensure progress bar updates reliably during AI execution
- Ensure tasks reach 'done' status and emit SSE events
- Prevent AI from calling move_task during phase execution
- Resume rejected/approved tasks in auto-mode and clean up phases
- Prevent date serialization error in phase duration display
- Correct message cleanup on error
- Bridge SSE events from MCP child process to parent via HTTP
- Handle stuck running tasks after server restart
- Persist activeExecutions map across HMR re-imports
- SSE disconnect, phase checkpoints & log verbosity
- Resolve skill folder from project path instead of process.cwd()
- Add status indicator, enable Run AI without flow, wire auto-mode execution
- Auto-mode toggle, status endpoint, backlog task query
- Refresh board instantly after task deletion
- Add cross-browser scrollbar consistency + e2e tests
- Prevent main view scrollbars when tasks overflow columns
- Align ExecutionPanel SSE parsing with server protocol
- Portal dialog and register select options when closed
- Project creation dialog and select display name
- Wire Tailwind v4 for production builds
- Resolve 5 failing comprehensive tests
Faster
- Warm once per gate leg — harness_once memoizes idempotent setup
- Skip the rebuild + maestro run when the skill already failed
- Skip the rebuild + maestro run when the skill already failed
- Defer kanban MCP server's heavy module graph until first tool call
- Defer kanban MCP server's heavy module graph until first tool call
- Scoped in-place re-run — re-run only the failing skill, not the whole phase (Spec 121 d)
- Scoped in-place re-run — re-run only the failing skill, not the whole phase (Spec 121 d)
- Canonical single-flow invocation + deterministic RESULT block (cut output-capture re-runs)
- Canonical single-flow invocation + deterministic RESULT block (cut output-capture re-runs)
- Stop the mobile suite running 2-3× (skill over-run + authoritative gate re-run)
- Stop the mobile suite running 2-3× (skill over-run + authoritative gate re-run)
- Pre-flight device skip + failed-first re-run (both engines)
v0.1.22026-06-20
New
- Desktop local-single-user, folder picker, shared apiFetch, in-place New Project, R2 auto-update
- Reflect the cloud trial in the app (banner + Upgrade)
- Event-triggered instant sync push + tombstone derivation test
- Sync tombstones — propagate deletions across the team
- Automatic shared-project sync sweep + "Share to team" UI
- Sync engine (local) — board snapshot serialize/merge + reconcile
- Plan-upgrade checkout from the app (Stripe per-seat)
- Cloud account loop — device-code sign-in, entitlements, routes, UI
- Cloud-as-provider seam + local-only-by-default cloud session
Fixed
- Confine outputFileTracingRoot (fixes Windows EPERM Application Data)
- Create-project silently failing + dashboard New Project card
v0.1.12026-06-20
New
- Warm the light-theme primary toward salmon
- Recolor primary from blue to Dawn coral (icon-matched)
- Add Morning sun app icon
- Package the app as native installers + build them in CI
Fixed
- Reliably skip migrations during desktop build + retry on lock
- Ad-hoc re-sign macOS bundle after afterPack (fixes 'damaged')
- Skip DB migrations + backfills during next build (SQLITE_BUSY race)
- App icon + empty-window (better-sqlite3 ABI) + free ports
- Use Morning sun mark on the dashboard header
- Replace Bot header icon with Morning sun mark
- Apply Windows .cmd/shell fix to after-pack hook too
- Shell:true for Windows .cmd spawns (Node 20 EINVAL)
- Unsigned mac build, Linux homepage, Windows rebuild bin