Cairn Documentation
AI Pipeline
Every ⌘K command flows through a 5-stage pipeline before reaching your review screen.
Models
The router is Claude Haiku 5.5, chosen for speed: it only decides which specialist answers and what context it needs. Specialists run on Claude Sonnet 5.5 by default. With your own Anthropic key, Profile → AI Providers → Default Model can switch them to Claude Opus 5.5 for heavier decompositions; high-quality 3D models always use it.
Each call says how much the model should think before it answers (its effort) rather than setting a temperature, which the current models don't take:
Thinking counts toward a reply's length limit, so Cairn leaves room for it on top of what the reply needs. Cairn reads only the reply's text; the thinking itself isn't shown or stored.
When Claude Declines
Claude's safeguards can decline a request, most often one that reads like weapons, malware or biological work. On Sonnet 5.5 and Opus 5.5, Cairn turns on Anthropic's server-side fallback: a declined request is retried on the model Anthropic recommends for it, the reply continues where the first one stopped, and the proposal's pipeline steps in Change Review name the model that answered. If the router declines, Cairn picks the specialist by keywords instead and the specialist gets its own turn.
If the request is still declined, nothing is proposed and the palette says which safeguard declined it, for example “Claude’s cybersecurity safeguards declined this request. Rephrase it, or say more about what the system is for.” A safety-critical system described plainly (a battery's thermal cutoff, a vehicle's braking) is ordinary engineering; saying what the system is for usually settles it.
Usage and cost count every attempt, each at its own model's rates: after a fallback, the declined attempt is billed as well as the one that answered. A decline before any output is billed only in some categories, so Cairn's estimate can run a little high there.
The Validator
The validator renames ids the model already has (an AI numbers from 001), completes what a specialist left out that Cairn can fill in (a data type's category, a unit it names), and then runs the same guard that decides Apply. Anything the guard would refuse is an error on its operation in review, and the specialist is asked once more with the errors named. A plan of several steps is renamed step by step, so two steps that took the same new id each keep their own references. The trace above the review names each step, its model and its time.
When a Request Fails
A failed request is titled for what failed: the AI couldn't be reached, it declined, its reply couldn't be read, or no step of a plan produced a change. A plan whose steps all failed ends there, with each step named, rather than open an empty review. A plan that reaches your monthly allowance midway stops and shows the allowance, with when it resets. While a request runs, the AI buttons in the lenses say Cairn is still working instead of queuing another.
Specialists
The router selects the right specialist based on your request: