Tasks and plans

The tasks snippet puts agents inside a structured work tracking system. Agents create tasks for real work, bind their sessions to them, log progress as comments, and mark work done with a summary of what they verified. A human monitors all of it through the dashboard — status flows through the system, not through chat messages that scroll away.

Installed via the snippet system; it rides along with memory by default:

cast task install

The objects

Tasks are work items — features, bugs, chores — with priorities, dependencies, and a status workflow (draft → open → in_progress → in_review → done). Plans group tasks under a goal and acceptance criteria for work with multiple distinct parts. Docs hold the prose — specs, investigations, handoffs. All three have short IDs (ct-4102, pl-88, doc:…) that render as live reference cards when written in prose anywhere in codecast.

The rules the snippet sets

The snippet is mostly judgment, not commands. Its rules:

• Create a task when the work will change code or produce a deliverable and will run long enough for someone to check on it. Skip it for questions, quick lookups, and small changes that finish in minutes.
• Create a plan only for work with multiple distinct parts. Single-task work gets a task.
• Bind before you build. cast task start ct-4102 claims the task and binds the session to it; cast plan bind pl-88 attaches to a plan. Sizable work done unbound is invisible to the human tracking it.
• Check existing work first. Search before creating: cast task ls -q "auth", cast plan ls -q "auth", cast task ready for unclaimed work. Claim rather than duplicate.
• Escalate explicitly. BLOCKED: <reason>, NEEDS_CONTEXT: <what>, and DONE_WITH_CONCERNS: <concern> are recognized markers that flag the session for human attention.

The working loop

cast task start ct-4102                          # claim + bind the session
cast task comment ct-4102 "reproduced; fix in progress" -t progress
cast task done ct-4102 -m "fix + regression test, verified e2e"

Plan-bound work adds coordination duties: record directional decisions with cast plan comment pl-88 "decision" -d -r "rationale", flag dependencies, and suggest splitting tasks that grew too large. Decisions logged this way become part of the plan's permanent timeline, visible to every future session that binds to it.

Context recovery

Long sessions get compacted. The snippet tells agents to reground from the system rather than trust compacted memory:

cast task context --current    # everything about the session's current task
cast plan context --current    # the plan: goal, tasks, decisions, discoveries

These print the full work item — description, comments, linked sessions — so a compacted agent recovers exactly the state it needs.

Where this leads

Tasks and plans are the substrate for the heavier machinery: workflows bind execution graphs to them, and orchestration decomposes a plan into tasks and runs them in parallel across agents. Triggers handle the time dimension — work that should happen after the session ends.