Documentation
¶
Overview ¶
Package ollama provides a clikit Act backend that talks to a local ollama daemon over HTTP. Importing it is the opt-in: the clikit core never pulls in a model runner on its own, and nothing here installs or launches a daemon — consumers bring their own ollama (and pull their own weights).
Index ¶
- Constants
- type Commander
- func (o Commander) Load(ctx context.Context) error
- func (o Commander) Loaded(ctx context.Context) (bool, error)
- func (o Commander) Parse(output string) ([]clikit.Action, error)
- func (o Commander) Propose(ctx context.Context, prompt string) (<-chan string, error)
- func (o Commander) Unload(ctx context.Context) error
Constants ¶
const DefaultEndpoint = "http://127.0.0.1:11434"
DefaultEndpoint is the local ollama HTTP API. Loopback by design — nothing here reaches the network.
const DefaultModel = "qwen2.5:3b"
DefaultModel is the model a Commander uses unless overridden. A 3B instruct model is the sweet spot for facet classification: 0.5–1.5B models are too weak (they return uniform, undifferentiated picks), while 3B differentiates well and stays small enough for a resident daemon to hold cheaply. Warm latency on a short prompt is ~1–2s CPU-only, sub-second on a GPU; keep the classifier's output terse, since on CPU each generated token costs ~60ms. Configure per host via Commander.Model.
Variables ¶
This section is empty.
Functions ¶
This section is empty.
Types ¶
type Commander ¶
type Commander struct {
Endpoint string // ollama base URL; defaults to DefaultEndpoint
Model string // local model tag; defaults to DefaultModel
System clikit.DocCorpus // system prompt: the classifier's role/identity
// KeepAlive controls how long ollama holds the model resident after this
// call (e.g. "30m", or "-1" to never unload). Empty uses the daemon default.
KeepAlive string
// Wrap, if set, turns the raw user prompt into the message actually sent —
// hosts use it to frame the task as classification and embed the prompt as
// inert data (see the omp backend's Commander.Wrap).
Wrap func(prompt string) string
// Client, if set, is used for the HTTP call (tests inject a stub); otherwise
// http.DefaultClient. The request carries the Propose context, so cancelling
// it aborts the call.
Client *http.Client
// contains filtered or unexported fields
}
Commander is a local-model Act backend: it POSTs to a resident ollama daemon and streams the model's output so the box shows it working, then Parse turns the completed output into a validated Action set. Unlike the omp backend it makes no subprocess and needs no auth — the model runs locally and stays warm, so there is no process-spawn or auth latency per call. It depends only on net/http.
The classification prompt is neutral by design: it scopes a task to what the task objectively needs, never to any one operator's habits. Deliberate over-provisioning ("max everything for this one") stays a manual pick in the host tool, so the model's suggestion is always a sensible baseline.
func NewCommander ¶
NewCommander builds a Commander with the default endpoint, local model, and the given system prompt (which must instruct the model to answer with a JSON object of settings). The model starts unloaded — the user loads it explicitly (see clikit.Loadable).
func (Commander) Load ¶
Load pins the model resident (keep_alive -1) so subsequent suggestions are warm; the first load pulls the weights into RAM and can take a while on CPU.
func (Commander) Loaded ¶
Loaded reports whether the model is resident right now, per the daemon's live process list (/api/ps) — the source of truth, independent of the intent flag.
func (Commander) Parse ¶
Parse turns the model's completed output into proposed changes, reusing the same tolerant JSON extraction as the omp backend.
Source Files