Docs

Mission board

One board for everything your bots work on: planned missions, hand-offs, routine runs and loose ends you take on. A task counts as done only when a check you can see has passed, never because a bot said so.

On from the start

The board is on from the start. To hide it, turn off Settings → All spaces → Mission board → Show the Mission board. Turning it off hides the board and stops new tasks; nothing is deleted.

The columns

The Needs you now tile counts the same cards as the Needs you column, so the two always match.

What "Verified" means

A task becomes Verified only when its checks pass. A check can be:

• a command that has to succeed, such as running the tests
• a web address that has to answer, optionally containing some text
• a file that has to exist or contain some text
• a review by a teammate who isn't the task's owner, pointing at evidence
• you, pressing Verify

A bot can't verify its own work. A check a bot adds to its own task waits for your OK before it counts, so "the file I just wrote exists" isn't enough.

Optional checks are shown but never hold a task back. A Verified task with an optional check still waiting says "Check passed" and "1 optional check waiting".

Hand-offs: the lead reviews

When a bot hands work to a teammate with an @mention, the hand-off gets its own card. When the teammate finishes, the card goes to Review and the bot that handed the work out reviews the result in its next reply:

• A pass makes the card Verified.
• A fail sends the work back to the teammate with the reason. It stays one card the whole time.
• No review: the lead gets one reminder, which lists every card still waiting. After that the card says Waiting for <lead> to review and nothing asks again. You can still press Verify yourself.
• One-way notices, where nobody is asked about the result, never wait in Review. They're shown as Closed, "Notice, no review needed", not Verified.
• A card that was blocked when its hand-off finished, or left waiting by a restart, gets its own review turn.

Only the lead's review counts: the teammate who did the work still can't approve its own card. Cards made before 0.4.0 have no result attached for the lead to review, so verify them yourself or leave them.

Missions

A mission groups tasks toward one goal. A task waits for the tasks it depends on. When a lead bot plans a mission, it waits for your approval before anything runs.

Token limits

Since 0.3.0 nothing has a token limit by default: tasks, missions and requests that hand work around run until they're done. A task or mission shows plain usage, such as "64.9k tokens used".

• Set a limit next to Cost on a task, or under a mission's cost, offers 50k, 200k, 1M or a custom amount. A task stopped by its limit offers Remove limit and continue and Raise limit.
• Settings → All spaces → Usage sets defaults: Token limits for new tasks and missions, and Hand-offs per request for the tokens one request may spend on bots handing work to each other. Both start at No limit. With a limit set, a request pauses there and offers Continue.
• Updating to 0.4.0 switches off a hand-off limit still at the old 200k default, once. A number you set yourself stays.

What stops runaway loops are the hand-off caps: one bot can be asked up to 4 times per message, and one message can start up to 16 hand-offs. Past that, the chat shows "Paused after 16 hand-offs so it doesn't loop. Continue?" with a Continue button, or you can type "continue".

"May be invalid"

If the file a verified task points to changes or disappears, or you redo it, that task and the tasks built on it get a May be invalid badge and wait under Needs you. They stay Verified until you pick Still valid, Redo or Cancel. Redo makes a copy that runs again, so the verified history stays as it was. How tasks connect shows a mission's tasks step by step, with a Check files now button.

Checkpoints and restore

kwant takes a checkpoint of a task's folder before and after each run, and whenever you ask for one. From the task's panel you can preview a checkpoint and restore it.

• Only the files that task's runs changed are written back. Files another task or you changed are listed, and left alone unless you tick them.
• kwant takes one more checkpoint just before restoring, so you can undo the restore.
• What the bot did outside its folder, such as commands, web requests or messages, is listed as not undone.
• A restored task waits on hold until you press Run. Pick Restore and run again to start it straight away.

Checkpoints stay on your server.

AutoFix

When a task fails, its panel says why in plain words and what to fix. AutoFix is on from 0.6.5, under Settings → All spaces → Mission board → Retry failed tasks with AutoFix, and it plans a retry that fits the failure:

• A temporary error, like a timeout or an overloaded service: it waits 1, then 5, then 25 minutes and tries again.
• A failed check: it runs the task again and passes along what the check said.
• A stalled run: it starts again from the last good checkpoint.

A failed task shows its tries, what retries have cost, and Retry now and Don't retry. Choose Ask me first to approve each retry yourself, or Retry on its own. Retries stay within the task's attempts and a retry budget, a quarter of the task's token limit when it has one.

Proposed changes

A bot can propose changing a task's check or acceptance, or a routine, for example when a check can't pass for a reason outside the task. The proposal appears under Needs you with the before and after, and nothing changes until you approve it.

Insights: verified results and their cost

New in 0.6. At the top of the board, under the week's numbers, kwant shows two measures:

• Verified per active week: how many tasks were verified each week over the last 8 weeks, and the average over the weeks that had any.
• Cost per verified result: what this week's settled tasks cost, failed runs included, divided by the verified results that reported a cost.
• Never a fake $0. When some runs didn't report a cost, it says so ("Based on 6 verified results that reported a cost; 2 runs reported none"). When none did, it says Not reported.

The same "Needs you" area holds fact conflicts to settle (with the Truth Ledger on) and Flight School requests from bots asking for a higher tier.

Performance

Under the board, on All work, a Performance section shows how each bot's finished work goes, per kind of work and per bot, or By model:

• how many tasks finished, and the verified rate with its likely range, such as "56% (27–81%)";
• median time and median tokens, cost per task, and how often you stepped in.

Filter by kind of work, bot and period (30 days, 90 days or all time). A group with fewer than 5 tasks says "few tasks", so a lucky streak doesn't look like a trend. It's report-only: the numbers never change which bot gets work, and only you can see them. Tasks finished before the update are counted too.

Swarm Auditor and open risks

The Auditor is on from 0.6.5, under Settings → All spaces → Mission board → Challenge important work with an audit. A task gets one challenge when it reaches Review if it's high or urgent priority, or if no outside check will judge it:

1. A teammate from outside the owner's lead loop lists concrete risks, each with evidence.
2. The owner answers once.
3. What stays unresolved is listed on the task as open risks until you Resolve or Dismiss each one. Board cards show how many are still open.

An audit never verifies, fails or moves a task, and "no risks raised" isn't a pass. Both turns ask before changing anything; while one waits on your approval card, the task says "Audit waiting for your answer" with a link to it. By default there are at most 5 audits a day, and they run on the engine's fast, cheap model. Change both under the switch: Audit model offers Fast, Balanced, Deep or the bot's own model.

On a phone

The board works on a phone: it shows one column at a time, picked from tabs that show how many tasks each holds.