Behaviour for injecting dynamic content into an agent's system prompt.

Implement this behaviour in a sidecar module to inject per-session context (e.g. memory, project state) into the system prompt before every LLM turn.

Behaviour

Use use Planck.Agent.Hooks.Prompt and override either or both callbacks. Both default to returning nil (no injection).

defmodule MySidecar.Hooks.Memory do
  use Planck.Agent.Hooks.Prompt

  @impl true
  def after_prompt(session_id) do
    case :ets.lookup(:memory, session_id) do
      [{^session_id, content}] -> content
      [] -> nil
    end
  end
end

Dispatch

Planck.Agent stores the resolved hook module atom in state and calls before_prompt/3 and after_prompt/3 before every LLM turn:

Hooks.Prompt.before_prompt(state.prompt_hook, state.session_id, state.sidecar_node)
Hooks.Prompt.after_prompt(state.prompt_hook, state.session_id, state.sidecar_node)

• module: nil — returns nil (no injection).
• sidecar_node: nil — calls module.prepend/1 / module.append/1 in-process.
• sidecar_node set — calls via RPC; returns nil on :badrpc.

The default RPC timeout is 5000 ms; override hook_timeout/0 to declare a custom expected latency.