The cniska-task plugin, in plugin/ and the cniska marketplace, tells every session where queued and running work lives, puts the running and in-review tasks in its context, and enforces a halt and the check gate. It reads the store only through the CLI.
Whenever a session starts, resumes, clears or compacts, it adds a short note, plugin/hooks/note.md: queued and running work lives in agent-task, not in a ledger, a notes file or a plan, add, start, review and done queue, start, hand back and close a task, and every check runs through agent-task check.
When the working directory holds the brief of a running or in-review task, it adds the running and in-review tasks after the note as list --started prints them, briefChanged included, so the session sees a task whose brief no longer has the briefSha recorded when it started.
The session hook asks agent-task with list --started --brief-under the session's directory, which runs no queued task's check, and puts what it prints in the context as it is, under a heading. It prints the note alone when that command is refused, no_brief_under included; it never exits with an error. Without jq, which reads the session's directory from the hook's input, it prints the note alone and leaves a line on stderr.
While a halt is raised, as agent-task halted reports it, the plugin stops the session's work:
- Session start: a heading saying all agent work is halted, then the refusal
haltedprints with its reason, come before the note. - Tool calls: a
PreToolUsehook refusesEdit,Write,NotebookEdit,Bash,PowerShell,Monitor,Agent,WorkflowandEnterWorktree, naming the reason. Every other tool works, so a halted agent can still read and report. - agent-task commands: a
BashorPowerShellcommand that runs onlyagent-task, with no shell operator, expansion or control character, goes through, sohaltandresumestill work. - Unreadable input: while halted, a call whose input the hook cannot read, without
jqor not JSON, is refused.
The hooks decide on halted's exit status alone, and let calls through when they cannot ask it. When agent-task is missing or fails --version, session start prints only a line saying the halt could not be checked; when it is too old to know halted, that line comes before the note.
With no halt raised, the PreToolUse hook passes each Bash, PowerShell and Monitor command to agent-task gate. When the command runs a test, check, verify, lint, typecheck or build other than through agent-task check, it refuses the call, naming the agent-task check command to run instead. So every agent's checks run one at a time across the machine, and only while it is under its limits.
- Every session: the gate holds in every session the plugin runs in, a coordinator's included.
- The owner's shell: a terminal the owner runs has no hook, so it is never held.
- Lets through when it cannot ask: when
agent-taskis missing or too old to knowgate, or the hook cannot read the command withoutjq, the call goes through.
The hooks export AGENT_TASK_NO_UPDATE_CHECK, so agent-task's update line never enters what they capture and show, a halt refusal included.
The plugin brings neither Bun nor the CLI. Install the CLI first, as the README says; the plugin checks nothing until agent-task is on the PATH Claude Code starts with. Its hooks also need jq.
The hooks see the PATH Claude Code started with, which can differ from a shell's, for example when the desktop app is opened from the Dock, or when the line the curl install adds to a shell file does not reach it. When session start says the halt could not be checked, or prints its note but no running task from a directory that holds one's brief, run agent-task doctor under that PATH.