Cuukit Lonca 9.1.0

Lonca now reads a pull request's comments, takes your own brief for a task, and finds the criteria an issue asks for outside its list.

• PR comments. The task tab lists the pull request's line comments, review bodies and conversation comments. Lonca judges each comment before acting on it: it applies the ones it agrees with and replies on GitHub, drafts a reply when it disagrees, and asks you when it is unsure. How much happens without you follows the orchestrator's autonomy level, and a comment that waits for you sends a notification.
• A brief for a handed-off task. Tell the brief agent what the issue does not say: a method, a constraint, a screenshot, a Figma selection. It writes the task's brief, which every stage reads and the Documents tab shows. A change after the hand-off sends the task back only as far as it reaches. Figma selections in the brief take the place of the issue's.
• Criteria the issue asks for. Background AI adds the criteria the issue asks for outside its criteria section (the description, comments, attachments), marked Generated with where each came from. The author's criteria are never changed. The planner and the scenario writer may object to a generated one; the orchestrator or you keep, reword or drop it, and the PR lists what was dropped.
• Scenarios follow the criteria. A dropped criterion's E2E scenario no longer runs, the scenario writer is told what changed in the criteria, and a criterion left without a scenario is caught when the scenarios are written.
• Question cards are coloured by the kind of stop.
• Settings → Plan. A new tab shows your plan. Every feature stays on for everyone in this release. Usage measurement is there too, off by default: switched on, Lonca counts a few numbers on your machine (decisions kept, tasks at once, advisor decisions), never content, and you can see and export them.
• No going back to an older version. From this release on, an older Lonca does not open a profile a newer one has used, so it never reads data it does not understand.

Your configuration needs no changes for this release.

Cuukit Lonca 9.0.0

Lonca now works with OpenAI's Codex CLI beside Claude Code.

• Codex as a stage agent. Each stage agent runs on Claude Code or Codex, in its terminal or headless, with its own model and effort (Settings → Orchestration). Codex runs with Lonca's hooks: its state is shown and it is held to a repository's tests: policy, as Claude Code is. A headless Codex run is read-only.
• The advisor and Background AI on Codex too. Background AI (drafts, summaries) and the advisor can run on either CLI.
• Models from the CLI. Settings lists the models each CLI offers your account, and the effort slider shows only the levels the chosen model takes.
• Every run measured and priced. Tokens and cost for every run, terminal or headless, whichever CLI ran it, at Anthropic's and OpenAI's list prices, refreshed weekly. Haiku 5.5 is priced by its prompt-length tiers.
• Settings. The tabs are in a new order; "Orchestrator" is now "Orchestration", "AI command" is "Background AI".

Your configuration is brought up to date by the configuration agent after the update. A Claude Code alias (opus, sonnet, haiku) on a stage agent that runs on Codex is replaced with a Codex model.

Cuukit Lonca 8.4.0

What's new

The orchestrator keeps going while work moves

• Past max_rounds, a loop goes on by itself while each round brings new findings. It stops and asks when a finding comes back, or at three times the limit. Each added round is on the timeline.
• A scenario the E2E runner corrects runs in its next e2e-fix, even one the plan left out.
• The reviewer writes the PR's description: in the repository's pull request template when there is one, short and concrete otherwise. The PR lists the acceptance criteria met.
• The UI step's visual checks show their match ratio on the timeline.

Conversations

• What you write while an agent works shows when you write it, not later or never.
• A sent message shows at once.
• A Claude Code dialog waiting in an agent's terminal shows in the conversation, with keys to answer it.
• Stage agents and the orchestrator start with Claude Code's auto-mode setup offer off.
• The orchestrator in the menu talks about the task you have open.

Interface

• Task tabs: Overview, Pipeline, Worktree, AC, Documents.
• The Pipeline tab links the PR beside its step, and offers Skip beside a running gate.
• Calmer cards for questions and decisions; Running is purple.
• A Jira update is a small dot beside the status.
• Clicking a notification opens its task and tab, also from the Orchestrator view.
• Run App opens the app in the browser once it answers.
• A new repository comes with its E2E routine on.
• Settings: Save works again when the E2E routine's options sit under it.
• Everything Lonca writes is in English; files from earlier versions still read, and show in English.

Cuukit Lonca 8.3.0

Figma designs: a UI step before the implementer

• A task whose issue or story links a Figma selection ("Copy link to selection") gets a UI step first, in repositories with figma: true and a Figma token saved in the new Settings → Figma. The link can be in the description, a comment or a web link, in any Jira smart-link view.
• The UI implementer (Opus by default) builds the presentational components with the repository's own tokens and component library. Lonca checks each one against the Figma render in its own browser (pixel diff, text, style and geometry findings), and the step ends only when every component passed its check or has a stated reason why not.
• Screens behind a sign-in are shown in a small preview kit: the global styles, the theme and the component with mock data, served by the repository's own bundler. Lonca keeps the kit per repository, outside the PR, checks its health without an agent, and asks you once before it runs its command.
• The implementer then wires the components to real data, routes and state without changing how they look. Tasks without Figma links run as before.
• The Figma token stays in Lonca, encrypted; no agent sees it, and agents can inspect only the task's own links.

Images and folders

• Draft with AI reads the images pasted into an idea, and a task gets copies of its story's images.
• When an agent needs a folder outside its task, Lonca asks you with the macOS permission prompt or a folder picker, and copies what you allow into the task.

Sessions

• Every agent session that has both views has a Terminal / Conversation switch.

Configuration

• The configuration agent applies the 8.3.0 notes after the update: {{readDirs}} in ai.args, figma: true for repositories with a browser UI, and the optional ui_implementer stage agent.
• A Jira smart link shown as a card or an embed is now kept in the task's snapshot instead of being dropped.

Cuukit Lonca 8.2.0

Starting a task

• Setup and hand-off errors can be dismissed, and each offers one next step: Retry setup, Set up from scratch, or Try again. The Overview card names the step that stopped.
• If Claude Code does not trust the repository, Lonca says so at hand-off and before you start your own session.
• A hand-off no longer stops on a branch that already exists on the remote. The Pipeline tab asks whether to use that branch or open a new revision. A remote that cannot be read is set up from the local state, with a notice.
• A setup that stopped names its step in the notification.
• A network or GitHub blink is tried again after 20 seconds. A lost tmux connection and a PR that went missing are retried by themselves. Questions wait for an advisor that is still opening.

Cheaper rounds

• Settings → Orchestrator shows what each stage agent's runs cost, resumed and fresh, and lonca.log has a line per run.
• A passing E2E round runs without the E2E runner agent. The agent starts only when a scenario fails.
• The reviewer and final check do not run again on code that is unchanged since their last clean pass.
• The implementer kept between rounds is sent only what is new. Its conversation starts fresh once its context passes 120k tokens.
• The repository's MCP servers go only to the stage agents that work on the app.
• One setup reads its Jira issue once, and a round fetches its base once.
• New, opt-in: gate.scope_after_light runs gate commands on the changed files after light rounds. The configuration agent sets it up where your test runner supports it.

Fixes

• A push that landed before its PR failed counts as the branch's last push, so the next push after a rebase is no longer refused.
• Transcripts under CLAUDE_CONFIG_DIR are read.
• A plan line that says only "None." no longer stops the task with a question.
• A tmux server that ends while Lonca reads a session no longer stops the step.

Cuukit Lonca 8.1.0

Cuukit Lonca 8.1.0

Everything Jira knows, in the task's context. Lonca now reads every comment, the history, every field (custom fields under their Jira names), linked issues, subtasks, web links and every attachment, the ones in comments included. Confluence pages linked from comments are read too. Large issues are held to caps that keep the newest comments and history.

When the issue changes, the orchestrator decides what it means. A task whose issue changed in Jira since Lonca last read it shows "Updated in Jira" with how long ago. Refresh reads it again, and for a task in the pipeline a read-only reassessor weighs the change: continue (with a note to the agent at work), plan again, fix only the implementation, rewrite the test scenarios, or ask you. Acceptance criteria follow the change. The task moves under your autonomy setting; under assistant the move waits for you. Lonca's own status moves never count as a change.

Your own session is yours. The Claude Code session you open on a task now has its own conversation, kept per task, instead of resuming whatever ran last in the worktree, which after a pipeline step was a stage agent's. After the update the configuration agent changes --continue in your agent's command to {{conversation}}.

Orchestrator

• Only the implementer and the test runner wait for a test slot, and the label says which one waits.
• Moving a task back to an earlier step stops the agent still working on the step it left.
• A new task's worktree starts from a freshly fetched base branch.

Cuukit Lonca 8.0.1

Cuukit Lonca 8.0.1

Settings folders

• Moving the Jira or local stories folder no longer stops when a folder of the same name is already at the new place: the two are joined. Nothing is overwritten; a file both have with different contents is kept twice, the moved one named for the old folder. New subtasks of those stories then go to the new place too.
• A file or link standing at the new place under a folder's name keeps only that folder where it is; the others move.

Window

• Dialogs now cover the title and status bars; their buttons no longer stay bright and clickable over a dialog.
• The Orchestrator drawer no longer dims the window behind it.
• The task list keeps the order you chose. Before, any task with a notification, read or not, was ordered by it first, so while the orchestrator worked no sort seemed to apply. Unread notifications still come first.

Cuukit Lonca 8.0.0

Cuukit Lonca 8.0.0

Sign-in per repository. Each repository that needs a sign-in has its own sign-in script, which the configuration agent writes and tries with your credentials before keeping it. Credentials are entered per repository under Settings → Workspaces and encrypted with the macOS Keychain; a saved browser account moves there by itself. E2E scenarios and their API calls run signed in, and a session about to run out is renewed before the next scenario.

Lonca's runner always runs E2E. e2e.runner is gone; an app on a fixed port says so with e2e.fixed_port. After the update, the configuration agent brings your config up to date.

The configuration agent is a conversation. In Settings, talk to it to change the configuration. It asks with buttons when it needs an answer, and the conversation is kept across a restart.

Settings folders. Changing a folder asks before anything moves, and moves it at the next start; a folder in use is never touched.

Orchestrator

• A task that used up its runs before a step goes on to that step when you give an extra round, instead of being sent back to implementation.
• A round that corrected only the task's E2E scenarios counts as progress and goes on to the E2E run.
• Under consultant and autopilot, a round that changed nothing goes to the advisor before it stops the task.
• The task files a role reads in most runs come in its prompt.
• The conversation view shows what an agent asks in its terminal.

Cuukit Lonca 7.4.0

Load control (experimental, off by default). Turn it on in Settings → Orchestrator. Lonca then decides from the machine's CPU and memory how many tasks and tests run at once. It freezes or pauses low-priority work while the machine is busy and gives high-priority tasks the machine first. Nothing stays frozen longer than 10 minutes or paused longer than 20. New config key: orchestrator.test_concurrency (default 4).

Cuukit Lonca 7.3.0

• Stage agents get a role-specific system prompt from Lonca; the task goes alone as the first message. The person's system_prompt_file no longer applies to stage agents (the configuration agent removes it where it is only used by them).
• Where the organization turned Claude Code's file-loading flags off, the stage system prompt is passed inline.
• The advisor's CLAUDE.md explains what Lonca does and asks for fewer calls.
• Enter sends in the conversation box; Shift+Enter and Option+Enter open a new line.
• The workspace follows the sidebar handle with no delay.