jj.bettong.agent — org.clojars.jj/bettong 0.1.0-SNAPSHOT

The agent loop: prompt in, tool calls executed, final answer out.

The agent starts with no abilities at all. What it can do is exactly what you pass in :capabilities - see jj.bettong.capability.

(require '[jj.bettong.agent :as agent] '[jj.bettong.capabilities :as caps])

(def assistant (agent/deepseek {:capabilities [(caps/filesystem)]})) (agent/run assistant "Create file /tmp/foo.txt with content 'foo'")

Agent clj protocol

Something that can run a prompt to completion.

Implement it yourself when you want a stand-in - a canned agent for your own tests, one that caches, one that routes between models.

run clj

(run this prompt)

(run this prompt opts)

Runs prompt and returns {:output <final assistant text> :steps <number of model calls> :tool-calls [{:tool .. :capability .. :args .. :result ..} ...] :messages <full transcript> :usage {:requests .. :prompt-tokens .. :completion-tokens .. :total-tokens :cache-hit-tokens .. :cache-miss-tokens .. :reasoning-tokens ..} :done? <false if it hit :max-steps>}

opts overrides the agent's own configuration for this one call.

approve-all clj

(approve-all _)

The default approval hook: every tool call is allowed.

ask-in-terminal clj

(ask-in-terminal {:keys [tool args]})

An :approve-fn that prints the tool call and waits for y/n on stdin. Anything other than y/yes denies the call.

base-system-prompt clj

clear! clj

(clear! session)

Forgets the conversation, keeping the agent. The next turn starts fresh.

config-keys clj

Everything an agent accepts, at construction or as a per-call override.

deepseek clj

(deepseek)

(deepseek config)

An agent backed by DeepSeek's API. Configure it once, run it many times:

(def assistant (agent/deepseek {:api-key (System/getenv "DEEPSEEK_API_KEY") :model "deepseek-v4-flash" :capabilities [(caps/filesystem {:root "/srv/data"})]}))

(agent/run assistant "Summarise report.txt") (agent/run assistant "And again in French" {:max-steps 3}) ; per-call override

Options: :capabilities what the agent may do. A seq of anything satisfying jj.bettong.capability/Capability, including plain maps. Empty by default: with no capabilities it has no tools and can only talk back. :api-key defaults to the DEEPSEEK_API_KEY environment variable :model defaults to deepseek-v4-flash :base-url defaults to https://api.deepseek.com - point it at any OpenAI-compatible endpoint :max-steps model calls before giving up (default 10) :system replace the base system prompt; capability instructions are still appended to it :approve-fn (fn [{:keys [tool args capability]}]) -> anything truthy allows the call; false/nil, or {:allowed? false :reason <string>}, denies it and hands the reason to the model :on-event fn called with progress maps :temperature passed through to the model :chat-fn override the LLM call itself, for testing

default-max-steps clj

parse-arguments clj

(parse-arguments arguments)

Tool-call arguments arrive as a JSON string; decode to a keyword map.

session clj

(session agent)

An agent that remembers. Each turn continues the transcript instead of starting one, so follow-ups work and the provider's cache keeps paying:

(def chat (agent/session assistant)) (agent/run chat "Describe Lithuania using wikipedia") (agent/run chat "Now in one sentence") ; knows what 'it' is

A session is itself an Agent, so anything that takes one takes this.

session-usage clj

(session-usage session)

What the conversation has cost in total, across every turn.

system-prompt clj

(system-prompt capabilities override)

The base prompt plus whatever the granted capabilities want the model to know.

transcript clj

(transcript session)

Everything said so far, as the messages the API sees.