Tasks

A task is a goal with the criteria that say it is done, kept on this computer, that Vults can hand to an agent working in a clone of its own. The Nest’s Tasks tab lists them (The Nest).

Keeping a task

#

/task in the chat keeps a task in one message, sent to no model: the goal on the first line, then a line for each criterion that says it is done, starting with -:

/task Fix the flaky login test
- passes 20 runs in a row
- no sleep added

On one line, done when: starts the criteria, split by ;. Its roost is the repository of the conversation’s folder. The Nest’s Tasks tab lists them, the ones still to do first, with each criterion to tick, who works on it (you, Zeca or an agent), where it stands (To do, Doing, Blocked, Done, Dropped) and what it cost once Vults starts sessions for it. New task makes one there too. With the GitHub connector on, a task in a GitHub repository gets Open an issue: on your click, your gh opens an issue with its goal and criteria, and the task links to it. Any task can link one you already have: #12 (in its GitHub repository) or the issue’s address, then Link. Tasks stay in vults.sqlite, on this computer.

Delegating

#

/delegate makes a task the same way and starts it at once on the agent the router picks, and the task’s card says why. A project’s own rule picks first (/delegate default codex, or claude, typed in a conversation in that project; /delegate default none clears it); without one, Claude Code. If that agent’s plan is above 85% and the other’s is known to be under it, the task goes to the other. Once a project has 10 finished tasks of one kind (a fix, tests, docs, a feature), the agent that finished more of them there replaces the rule for that kind, though the 85% still holds. A task counts for the agent that owns it when it is marked done or dropped.

Starting an agent on a task

#

Start with Claude Code works on a task in a repository. Vults makes the task a clone of its own, on a branch of its own (vults/task-12), under Vults’ data folder: a --shared clone, which reads your repository’s objects but never writes them. It starts Claude Code there on the task’s goal and criteria, told to commit on that branch and never push. It runs in Claude Code’s sandbox (bubblewrap): it can’t start without it, every command still asks, and it writes only in its clone and the temp folders, never your repository, its branches, hooks or config. Commands you excluded from the sandbox in your own Claude Code settings still run outside it, each asking first. It reads your own settings, so it shows in the flock like any session and its permissions are island cards, answered by you, your rules or your policies. Two tasks run side by side without touching each other’s files. When its turn ends, Vults fetches the task’s branch, and only that branch, into your repository, without running anything of the clone’s; What it said last under the task keeps the end of its reply, and Go on continues the same session. Start fresh hands the task to a new session instead, when the old one has grown long: the old session writes a context pack (Objective, Constraints, Done, In progress, Decisions, Files, Failures, Next), Vults puts in what it knows for certain (the goal, which criteria are ticked, the files changed on the branch), and the new session starts from that pack alone. The pack stays under the task as Handoff notes; if the old session writes none, nothing is handed off and it keeps its session. Stop asks its turn to stop, waits, then ends it; only a session Vults started can be stopped, and a running task can’t be deleted. What it cost joins the task. Remove its folder brings the branch back and removes the clone; work never committed in it goes with it, so it asks twice.

Codex

#

Start with Codex does the same with Codex, in its own sandbox (workspace-write: its clone and the temp folders). Codex asks for approvals on its own channel: Vults turns each into the island card its hook would send (the whole command, or an edit’s files and line counts; a call it can’t name whole is a no), so your rules, policies and phone answer it the same way, and a stopped run takes its waiting card away. On a task with a session, Hand off to Codex (or to Claude Code) is Start fresh on the other agent: the session writes its pack, and the other one starts from it.

Other agents

#

Start with another agent… lists the agents installed here that speak the Agent Client Protocol: OpenCode, Qwen Code, Copilot CLI and Kimi Code. One works in the task’s clone the same way, logged in with its own account. Vults offers it no files and no terminal of its own: it uses its own tools and asks before each one that writes, runs or fetches, and each ask is an island card, answered only for that once, never with the agent’s own “always”. A change to a file outside the clone, or in its .git, a command run elsewhere, a move or a delete is a no without a card, and so is a file reached through a link that points out of the clone. OpenCode runs every tool unasked by default, and Qwen Code and Kimi Code may by your own settings: Vults starts each with a setting, for that process or session alone, that makes it ask (never written to its config), and a write or command that still runs unasked stops the turn. Their own hooks ask nothing in these sessions, so a call never shows two cards. Stop, Go on, Start fresh and Review work as with Codex; the review is Codex’s. Checked with OpenCode, and with Kimi Code 2.1.1 against a stand-in model; Qwen Code and Copilot CLI follow the same protocol but were not tried with a real account yet.

Review

#

Review with Codex (or with Claude Code) has the other agent review a task’s work, in its clone, as a reviewer that wrote none of it, in a sandbox where it writes nothing: it runs the tests, reads the diff against where the branch began, checks each criterion, and lists what is wrong. Codex uses its own reviewer. When the review asks for changes, its findings go to the task’s session, which fixes them and commits, and the next round reviews again, up to three rounds. A pass ends it; a review with no verdict, or a third round still asking for changes, is yours to read. The last review stays under the task, and you still mark the task done yourself.

Cost and a daily budget

#

Vults counts what the agents report they cost: each turn of a session it started for a task or a /continue (Claude Code’s dollars; Codex reports tokens only, so its tasks show their tokens), and each Claude Code session it watches, from its status line, counted from the first time Vults sees it (what it cost before stays out of today). Nothing is guessed from a price list. Zeca’s own chat, summaries and sleep are not counted here, and the budget does not stop them. On a subscription the dollars are what the tokens would cost at API prices, which your plan does not charge; Vults says so when it can tell (a plan that reports its usage windows). Each task shows its cost and tokens, each roost in the Nest’s Activity → By project its cost today and in all, and the top of the Tasks tab today’s.

The Nest’s Activity → Cost sets a daily budget in US dollars. At 80% a notification says so; once it is spent another one does, and Vults starts no new session until the next day: no Start, handoff, review, /delegate or /continue. That is all it does. It never stops a session that is running, and never blocks or slows an agent you started yourself. Costs stay in vults.sqlite, on this computer; Clear history leaves them.