Skip to content
cairn

cairn ask

cairn ask turns intent into governed plans and answers questions about the platform’s state. It is safe by construction, not by prompt: the agent’s entire tool belt is the read-only API surface — there is no run, resume, cancel, install, or admin tool for it to call, so it literally cannot mutate anything. When you ask it to do something, it resolves the right workflow and hands back a ready-to-run plan for a human to execute.

CLI

Section titled “CLI”
Terminal window
cairn ask                                   # interactive REPL
cairn ask plan training/finetune -i model=...   # deterministic plan, no LLM; exit 0 iff runnable
cairn ask run "fine-tune a model on my support tickets"

ask plan assembles the plan for a known template deterministically. ask run uses an LLM only to resolve intent → template, then runs the same deterministic planner and prints the plan plus the exact cairn run command — it does not execute. Flags: --profile, --model (default gpt-4o-mini), --json.

The server agent — POST /ask

Section titled “The server agent — POST /ask”

The server exposes a ReAct read agent over the same governed core:

POST /ask            {"query": "...", "thread_id": "..."}   (auth required)
GET  /ask/thread/{thread_id}
  • Read tools only — query runs, costs, audit, remote jobs, packs and contracts, grounded knowledge retrieval.
  • Grounded answers — the system prompt is fail-closed: it answers from tool results, not world knowledge.
  • Tenant-scoped server-side — account and workspace come from the authenticated request, never from the body; thread ids are namespaced per account, so a guessed id yields an empty transcript.
  • Cost-bounded and audited — each turn runs under a hard budget cap through cairn’s own governed model gateway, and every answer emits an ask.answered audit event (query, tools used, recommended workflow).
  • Durable conversations — with OBS_CHECKPOINTER=postgres, ask threads survive restarts.

When the answer is “run this workflow”, the response carries a structured action that surfaces the run form — the human clicks, policy applies as always.