Agents and models
Choose the Coding agent and Model per Workspace, set the Reasoning effort, and read the agent's work as it streams in.
Each Workspace is owned by one Coding agent at a time: Claude Code, Codex, or Cursor. The agent is the vendor of access, not the brain. It determines which Models you can reach through your subscription for that agent; the Model is what actually runs. Both are chosen from the prompt input.
Models and Reasoning effort
Switching agent re-lists the Models available through it and snaps to that agent's flagship default. Each Model carries its own capabilities:
- Reasoning effort is the per-Model thinking budget, set from a compact gauge in the prompt input. The exact rungs are the Model's own, and the gauge hides for a Model that exposes no effort axis.
- Fast mode, faster output at the same level of thinking, appears only on Models that support it.
Cohalen never proxies your Model access: the agent runs under your own subscription or keys, on your machine.
Turns
One prompt, the agent's work, and the streamed reply make a Turn. The Workspace is busy from send until the reply completes, and a running Turn can be cancelled; cancel is the stop control for a running agent. Turns are continuous: each one resumes the Workspace's durable Session, so the agent keeps its context across Turns and across app restarts.
The agent's reply renders as ordered parts, in the order the agent produced them: prose, a collapsible Thinking trace, and each Tool call (Read, Edit, Bash, and the rest) inline with its status and result. Everything the agent does shows in the Chat; nothing about the running agent hides in the Terminal.
Skills
A Skill is a reusable SKILL.md playbook each agent loads natively from
its own directories. Cohalen stays zero-touch on Skill content: it never
stores or injects a Skill, because re-injecting one the agent already loads
would double it into the context. What Cohalen does add is the / picker in
the prompt input, which autocompletes Skill names from the active
Workspace's agent's own roots; the agent then resolves the name itself.
Separate from Skills, a user-global custom instruction in Settings is Cohalen-owned text carried into every Turn, for preferences you want every agent to follow on this machine.
Prompting for a clean Changeset
The Changeset you review is everything the Workspace changed, so scope the prompt to one task and state the acceptance criteria in it (what to build, what to test). When the reply needs correction, prefer a Note on the diff line over a fresh prose prompt: the Note carries the exact file and line to the agent's next Turn. See Reviewing changes.
Next: Reviewing changes, reading what the agent did.