Skip to content

Every agent's model. One place. Codex on DeepSeek, Claude Code on Kimi, from the menu bar.

provider: Ollama's and Ollama Cloud's decision models classify routing groups through System One (Nanka on Discord)

yetone9588501c0266
diff --git a/docs/subsystems/providers-accounts.md b/docs/subsystems/providers-accounts.md
index 4038f6581..f95a3eda0 100644
--- a/docs/subsystems/providers-accounts.md
+++ b/docs/subsystems/providers-accounts.md
@@ -104,9 +104,19 @@ apart, so its decision models are the ones its `/v1/models/status` gives
[`omlx_decide.go`](../../internal/provider/omlx_decide.go)). MLX-Serve marks
them in its `/v1/models` with `capabilities: ["decisions"]`, kept under
`<id>.decisions` (`mlxServeDecides`, `listedMLXServeDecide` in
-[`mlxserve_decide.go`](../../internal/provider/mlxserve_decide.go)). A server
-with none lists none, and no Jev is made up for it. An oMLX or MLX-Serve
-provider saved before the preset had a decision base gets its chat base as one.
+[`mlxserve_decide.go`](../../internal/provider/mlxserve_decide.go)). Ollama
+and Ollama Cloud (Nanka on Discord; Ollama 0.35.0+, 0.40.5+ for cloud models)
+serve System One at `/v1/systemone` beside their chat API, and list decision
+models (Tev, Nimble, Clef) in `/v1/models` like any other. Their `/api/tags`,
+asked with the same key, gives each model `capabilities`, `decision` for
+these and `vision` for one that reads images. Those are kept under
+`<id>.decisions`, with the cloud's 65,536-token window when the list gives
+none (`ollamaDecides`, `listedOllamaDecide` in
+[`ollama_decide.go`](../../internal/provider/ollama_decide.go)). An Ollama
+older than its System One lists no capabilities, so none. A server
+with none lists none, and no Jev is made up for it. An oMLX, MLX-Serve or
+Ollama provider saved before the preset had a decision base gets its chat
+base as one.
The CLI and TUI report model-list fetch errors for remote and other mixed providers
(such as OpenRouter, serving conversation and decision models). Only a
diff --git a/internal/gui/local_address_test.go b/internal/gui/local_address_test.go
index 5e56516d7..77113bf82 100644
--- a/internal/gui/local_address_test.go
+++ b/internal/gui/local_address_test.go
@@ -30,7 +30,7 @@ func TestProviderSaveLocalPresetAtItsURLs(t *testing.T) {
}
}
for _, c := range []struct{ id, name, chat, responses, anthropic, decide string }{
- {"ollama", "Ollama", "http://127.0.0.1:1/v1", "", "http://127.0.0.1:1", ""},
+ {"ollama", "Ollama", "http://127.0.0.1:1/v1", "", "http://127.0.0.1:1", "http://127.0.0.1:1/v1"},
{"omlx", "oMLX", "http://127.0.0.1:2/v1", "http://127.0.0.1:2/v1", "http://127.0.0.1:2", "http://127.0.0.1:2/v1"},
{"lmstudio", "LM Studio", "http://127.0.0.1:3/v1", "", "", ""},
} {
@@ -48,7 +48,8 @@ func TestProviderSaveLocalPresetAtItsURLs(t *testing.T) {
if moved.Chat != c.chat || moved.Responses != c.responses || moved.Anthropic != c.anthropic {
t.Fatalf("%s at %q %q %q", c.id, moved.Chat, moved.Responses, moved.Anthropic)
}
- // oMLX's decision API is the same server's, so it moves too (#1641)
+ // oMLX's and Ollama's decision API is the same server's, so it moves
+ // too (#1641; Nanka on Discord)
if moved.Decide != c.decide {
t.Fatalf("%s decides at %q, want %q", c.id, moved.Decide, c.decide)
}
diff --git a/internal/provider/decide.go b/internal/provider/decide.go
index 3485b2be4..04334b0c8 100644
--- a/internal/provider/decide.go
+++ b/internal/provider/decide.go
@@ -150,10 +150,18 @@ func openRouterDecisions(base string) string {
}
// listsDecisions reports whether p's decision models are a list of their
-// own, OpenRouter's, oMLX's (omlxDecides) or MLX-Serve's (mlxServeDecides),
+// own, OpenRouter's or a server's that marks them (serverMarksDecisions),
// rather than Jev's ids among its chat models.
func (p Provider) listsDecisions() bool {
- return openRouterDecisions(p.Decide) != "" || p.omlxDecides() || p.mlxServeDecides()
+ return openRouterDecisions(p.Decide) != "" || p.serverMarksDecisions()
+}
+
+// serverMarksDecisions reports whether p's server says which of the models
+// it serves are decision models, oMLX's (omlxDecides), MLX-Serve's
+// (mlxServeDecides) or Ollama's (ollamaDecides): it has no Jev of its own,
+// and lists none while it has none.
+func (p Provider) serverMarksDecisions() bool {
+ return p.omlxDecides() || p.mlxServeDecides() || p.ollamaDecides()
}
// decisionsID is where a provider that serves conversations too keeps the
@@ -195,7 +203,7 @@ func (p Provider) decideListDue() bool {
return false
}
live, _, ok := catalog.Live(id)
- if p.omlxDecides() || p.mlxServeDecides() {
+ if p.serverMarksDecisions() {
// a server with no decision model loaded lists none, for good
return !ok
}
@@ -281,8 +289,8 @@ func (p Provider) Jev() string {
}
// OpenRouter's Jev is ~typesafe/jev-latest, in its list of decision
// models; its Jev Router is a chat model
- if p.omlxDecides() || p.mlxServeDecides() {
- // oMLX and MLX-Serve have no Jev of their own: their first decision
+ if p.serverMarksDecisions() {
+ // oMLX, MLX-Serve and Ollama have no Jev of their own: their first decision
// model, if any
if ms := p.decideListed(); len(ms) > 0 {
return ms[0].ID
@@ -389,7 +397,7 @@ func (p Provider) decideListed() []catalog.Model {
if live, _, ok := catalog.Live(decisionsID(p.ID)); ok && len(live) > 0 {
return live
}
- if p.omlxDecides() || p.mlxServeDecides() {
+ if p.serverMarksDecisions() {
return nil // none until its server says which
}
return []catalog.Model{{ID: OpenRouterJev, Name: "Jev"}}
@@ -818,6 +826,8 @@ func (p Provider) fetchDecide(ctx context.Context) ([]catalog.Model, error) {
list = omlxDecisions(p.Decide)
} else if p.mlxServeDecides() {
list = mlxServeDecisions(p.Decide)
+ } else if p.ollamaDecides() {
+ list = ollamaDecisions(p.Decide)
} else if bailianDecides(p.Decide) {
// Bailian lists its chat models only: its decision model is asked
return p.decideAsked(ctx)
@@ -860,6 +870,14 @@ func (p Provider) fetchDecide(ctx context.Context) ([]catalog.Model, error) {
}
return ms, catalog.SaveLive(decisionsID(p.ID), p.Decide, ms)
}
+ if p.ollamaDecides() {
+ // none is an answer too: no model of the server's is one
+ ms, err := listedOllamaDecide(p.Decide, b)
+ if err != nil {
+ return nil, fmt.Errorf("%s: %s", p.Name, err)
+ }
+ return ms, catalog.SaveLive(decisionsID(p.ID), p.Decide, ms)
+ }
ms, err := listedDecide(b, all)
if (err != nil || len(ms) == 0) && own {
return p.decideAsked(ctx)
diff --git a/internal/provider/ollama_decide.go b/internal/provider/ollama_decide.go
new file mode 100644
index 000000000..465d1791a
--- /dev/null
+++ b/internal/provider/ollama_decide.go
@@ -0,0 +1,74 @@
+package provider
+
+import (
+ "encoding/json"
+ "fmt"
+ "slices"
+ "strings"
+
+ "github.com/yetone/magpie/internal/catalog"
+)
+
+// Ollama (0.35.0 and later, 0.40.5 for its cloud models) serves decision
+// models (Nimble, Clef, Tev …) at POST /v1/systemone, System One's
+// questions and answers, on the local server and at ollama.com with an API
+// key alike (docs.ollama.com/api/systemone; Nanka on Discord). Its
+// /v1/models lists them among its chat models with nothing to tell them
+// apart; its /api/tags, asked with the same key, gives each model its
+// capabilities, "decision" for these (and "vision" for one that reads
+// images, as Clef does). That list is the provider's decision models, as
+// oMLX's and MLX-Serve's are (listsDecisions), and a model it doesn't mark
+// stays a chat model. An Ollama older than its System One lists no
+// capabilities, so none.
+func (p Provider) ollamaDecides() bool {
+ return (p.Preset == "ollama" || p.Preset == "ollama-cloud") && p.Decide != "" && !p.DecideOnly()
+}
+
+// ollamaCloudContext is the window Ollama's docs give a System One model
+// in its cloud.
+const ollamaCloudContext = 65_536
+
+// ollamaDecisions is where an Ollama server lists its models with their
+// capabilities: /api/tags at the root its /v1 is under.
+func ollamaDecisions(base string) string {
+ return strings.TrimSuffix(strings.TrimRight(base, "/"), "/v1") + "/api/tags"
+}
+
+// listedOllamaDecide are the decision models in Ollama's /api/tags, by the
+// name /v1/models gives them too, with the window the list gives (a GGUF's
+// details.context_length), else, in Ollama's cloud (at ollama.com, or a
+// local server's cloud model, which names its remote_host), the cloud's.
+func listedOllamaDecide(base string, b []byte) ([]catalog.Model, error) {
+ var out struct {
+ Models *[]struct {
+ Name string `json:"name"`
+ Model string `json:"model"`
+ RemoteHost string `json:"remote_host"`
+ Capabilities []string `json:"capabilities"`
+ Details struct {
+ ContextLength int `json:"context_length"`
+ } `json:"details"`
+ } `json:"models"`
+ }
+ if err := json.Unmarshal(b, &out); err != nil || out.Models == nil {
+ return nil, fmt.Errorf("not Ollama's model list")
+ }
+ cloud := HostOf(base) == "ollama.com"
+ ms := []catalog.Model{}
+ for _, m := range *out.Models {
+ id := m.Model
+ if id == "" {
+ id = m.Name
+ }
+ if id == "" || !slices.Contains(m.Capabilities, "decision") {
+ continue
+ }
+ images := slices.Contains(m.Capabilities, "vision")
+ ctx := m.Details.ContextLength
+ if ctx <= 0 && (cloud || m.RemoteHost != "") {
+ ctx = ollamaCloudContext
+ }
+ ms = append(ms, catalog.Model{ID: id, Name: id, Context: ctx, ImageInput: &images, Images: images})
+ }
+ return ms, nil
+}
diff --git a/internal/provider/ollama_decide_test.go b/internal/provider/ollama_decide_test.go
new file mode 100644
index 000000000..692f5940a
--- /dev/null
+++ b/internal/provider/ollama_decide_test.go
@@ -0,0 +1,160 @@
+package provider
+
+import (
+ "context"
+ "encoding/json"
+ "net/http"
+ "net/http/httptest"
+ "slices"
+ "sync"
+ "testing"
+)
+
+// The bytes below were saved from Ollama 0.40.6 (a local server with
+// tev1:0.8b and smollm:135m pulled) and from ollama.com/api/tags on
+// 2026-10-19, trimmed to the rows a test needs.
+const (
+ ollamaLocalTags = `{"models":[{"name":"smollm:135m","model":"smollm:135m","modified_at":"2026-10-19T23:15:35.885392239+08:00","size":91739413,"digest":"b0b2a46174385c0adcaa77ff245ffeced5fc4a61447b6f221b2beb5c5a760133","details":{"parent_model":"","format":"gguf","family":"llama","families":["llama"],"parameter_size":"134.52M","quantization_level":"Q4_0","context_length":2048,"embedding_length":576,"runner":"ggml"},"capabilities":["completion"]},{"name":"tev1:0.8b","model":"tev1:0.8b","modified_at":"2026-10-19T23:14:31.147707454+08:00","size":1524372929,"digest":"b661d56508169d2b6c4cab5a047409f50bd9694a215bc232469d44291f4bdbf1","details":{"parent_model":"","format":"safetensors","family":"qwen3_5","families":["qwen3_5"],"parameter_size":"752M","quantization_level":"BF16","runner":"mlx"},"capabilities":["decision"]}]}`
+ ollamaLocalModels = `{"object":"list","data":[{"id":"smollm:135m","object":"model","created":1792422935,"owned_by":"library"},{"id":"tev1:0.8b","object":"model","created":1792422871,"owned_by":"library"}]}`
+ ollamaSystemOne = `{"model":"tev1:0.8b","answers":{"ok":{"type":"noul","noul":0.2234409293546222}},"usage":{"input_tokens":107,"output_tokens":0},"prompt_eval_cached_count":0}`
+ ollamaNoChat = `{"error":{"message":"\"tev1:0.8b\" does not support chat","type":"invalid_request_error","param":null,"code":null}}`
+ ollamaCloudTags = `{"models":[{"name":"glm-5.3","model":"glm-5.3","modified_at":"2026-08-28T08:00:00-07:00","size":755433728000,"digest":"632dfda18c6d","details":{"parent_model":"","format":"","family":"","families":null,"parameter_size":"","quantization_level":""},"capabilities":["completion","thinking","tools"]},{"name":"clef:27b","model":"clef:27b","modified_at":"2026-10-04T09:00:00-07:00","size":54969731792,"digest":"490a326b58a0","details":{"parent_model":"","format":"","family":"","families":null,"parameter_size":"","quantization_level":""},"capabilities":["decision","vision"]}]}`
+ // Ollama before System One (0.6.8) lists no capabilities
+ ollamaOldTags = `{"models":[{"name":"qwen3:8b","model":"qwen3:8b","modified_at":"2025-05-01T10:00:00+08:00","size":5225388164,"digest":"500a1f067a9f","details":{"parent_model":"","format":"gguf","family":"qwen3","families":["qwen3"],"parameter_size":"8.2B","quantization_level":"Q4_K_M"}}]}`
+)
+
+// fakeOllama answers as Ollama 0.40.6 does: /api/tags and /v1/models list
+// a chat model and a decision model, /v1/systemone answers System One, and
+// chat completions turn the decision model away.
+func fakeOllama(t *testing.T) (base string, asked func() []string) {
+ t.Helper()
+ var mu sync.Mutex
+ var models []string
+ srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+ w.Header().Set("Content-Type", "application/json")
+ switch r.URL.Path {
+ case "/api/tags":
+ w.Write([]byte(ollamaLocalTags))
+ case "/v1/models":
+ w.Write([]byte(ollamaLocalModels))
+ case "/v1/systemone":
+ var q struct{ Model string }
+ json.NewDecoder(r.Body).Decode(&q)
+ mu.Lock()
+ models = append(models, q.Model)
+ mu.Unlock()
+ w.Write([]byte(ollamaSystemOne))
+ case "/v1/chat/completions":
+ var q struct{ Model string }
+ json.NewDecoder(r.Body).Decode(&q)
+ if q.Model == "tev1:0.8b" {
+ w.WriteHeader(http.StatusBadRequest)
+ w.Write([]byte(ollamaNoChat))
+ return
+ }
+ w.Write([]byte(`{"id":"chatcmpl-382","object":"chat.completion","created":1792422957,"model":"smollm:135m","system_fingerprint":"fp_ollama","choices":[{"index":0,"message":{"role":"assistant","content":"Hello! How can I"},"finish_reason":"length"}],"usage":{"prompt_tokens":10,"completion_tokens":5,"total_tokens":15}}`))
+ default:
+ http.NotFound(w, r)
+ }
+ }))
+ t.Cleanup(srv.Close)
+ return srv.URL, func() []string {
+ mu.Lock()
+ defer mu.Unlock()
+ return slices.Clone(models)
+ }
+}
+
+// An Ollama provider saved before its preset had System One (Nanka on
+// Discord): the models its /api/tags marks "decision" classify routing
+// groups through /v1/systemone, and agents no longer see them as chat
+// models.
+func TestOllamaDecisionModels(t *testing.T) {
+ h := t.TempDir()
+ t.Setenv("HOME", h)
+ t.Setenv("USERPROFILE", h)
+ t.Setenv("XDG_CONFIG_HOME", t.TempDir())
+ t.Setenv("XDG_CACHE_HOME", t.TempDir())
+ base, asked := fakeOllama(t)
+
+ if err := Save(Provider{ID: "ollama", Name: "Ollama", Preset: "ollama", Chat: base + "/v1", Anthropic: base}); err != nil {
+ t.Fatal(err)
+ }
+ p, err := Find("ollama")
+ if err != nil || !p.Decides() || p.DecideOnly() || p.Decide != base+"/v1" {
+ t.Fatalf("Ollama has no decision API at its own server: %+v", p)
+ }
+ if _, _, _, err := p.refetch(context.Background()); err != nil {
+ t.Fatal(err)
+ }
+ p, _ = Find("ollama")
+
+ if !p.DecidesModel("tev1:0.8b") || p.DecidesModel("smollm:135m") {
+ t.Fatal("decision models are not the ones Ollama's capabilities say are")
+ }
+ var deciders []string
+ for _, e := range Deciders() {
+ deciders = append(deciders, e.ID)
+ }
+ if !slices.Equal(deciders, []string{"ollama/tev1:0.8b"}) {
+ t.Fatalf("deciders: %v", deciders)
+ }
+ var agents []string
+ for _, e := range providerEntries() {
+ agents = append(agents, e.Model)
+ }
+ if !slices.Equal(agents, []string{"smollm:135m"}) {
+ t.Fatalf("agents' models: %v", agents)
+ }
+
+ q, m, err := RouteDecider("ollama/tev1:0.8b")
+ if err != nil || q.ID != "ollama" || m != "tev1:0.8b" {
+ t.Fatalf("routed to %s %s: %v", q.ID, m, err)
+ }
+ if err := q.AskSystemOne(context.Background(), m); err != nil {
+ t.Fatal(err)
+ }
+ if _, m, _ := RouteDecider("ollama"); m != "tev1:0.8b" {
+ t.Fatalf("Ollama's default decision model: %q", m)
+ }
+ if got := asked(); !slices.Equal(got, []string{"tev1:0.8b"}) {
+ t.Fatalf("asked: %v", got)
+ }
+ results := p.TestModels(context.Background(), []string{"tev1:0.8b"})
+ if len(results) != 1 || !results[0].OK || results[0].Protocol != "decide" {
+ t.Fatalf("TestModels tev1:0.8b: %+v", results)
+ }
+}
+
+// Ollama Cloud's list marks Clef a decision model that reads images, with
+// the cloud's window; a chat model, and every model of an Ollama from
+// before System One, is none.
+func TestListedOllamaDecide(t *testing.T) {
+ ms, err := listedOllamaDecide("https://ollama.com/v1", []byte(ollamaCloudTags))
+ if err != nil || len(ms) != 1 || ms[0].ID != "clef:27b" || ms[0].Context != 65_536 || ms[0].ImageInput == nil || !*ms[0].ImageInput {
+ t.Fatalf("cloud: %+v %v", ms, err)
+ }
+ ms, err = listedOllamaDecide("http://localhost:11434/v1", []byte(ollamaLocalTags))
+ if err != nil || len(ms) != 1 || ms[0].ID != "tev1:0.8b" || ms[0].Context != 0 || ms[0].ImageInput == nil || *ms[0].ImageInput {
+ t.Fatalf("local: %+v %v", ms, err)
+ }
+ if ms, err := listedOllamaDecide("http://localhost:11434/v1", []byte(ollamaOldTags)); err != nil || len(ms) != 0 {
+ t.Fatalf("old Ollama: %+v %v", ms, err)
+ }
+ if _, err := listedOllamaDecide("http://localhost:11434/v1", []byte(`404 page not found`)); err == nil {
+ t.Fatal("a 404's body read as a list")
+ }
+ if got := ollamaDecisions("https://ollama.com/v1"); got != "https://ollama.com/api/tags" {
+ t.Fatalf("list at %s", got)
+ }
+ for _, id := range []string{"ollama", "ollama-cloud"} {
+ pr := Preset(id)
+ if pr == nil || pr.Decide == "" || pr.Decide != pr.Chat {
+ t.Fatalf("%s's preset has no System One beside its chat: %+v", id, pr)
+ }
+ }
+ p := Provider{ID: "ollama-x", Preset: "ollama-cloud", Chat: "http://127.0.0.1:1/v1", Decide: "http://127.0.0.1:1/v1"}
+ if got := p.decideListed(); len(got) != 0 || p.Jev() != "" {
+ t.Fatalf("made up: %v %q", got, p.Jev())
+ }
+}
diff --git a/internal/provider/presets.go b/internal/provider/presets.go
index 25ca6b3a4..79e7609cf 100644
--- a/internal/provider/presets.go
+++ b/internal/provider/presets.go
@@ -415,9 +415,12 @@ var presets = []PresetDef{
Endpoint: "https://<resource>.openai.azure.com",
EndpointHint: "Your resource's endpoint, from Keys and Endpoint in the Azure portal. Once added, magpie asks the resource for its deployments with this key, over its v1 API; the deployments' names are the model ids.",
Website: "https://ai.azure.com", KeysURL: "https://portal.azure.com/#view/Microsoft_Azure_ProjectOxford/CognitiveServicesHub/~/OpenAI"},
- // Ollama's own hosted models: the local server's API, at ollama.com with a key
+ // Ollama's own hosted models: the local server's API, at ollama.com with
+ // a key. Its decision models (Clef), served at /v1/systemone, are the
+ // ones its /api/tags gives the "decision" capability (ollama_decide.go)
{ID: "ollama-cloud", Name: "Ollama Cloud", Icon: "ollama", Kind: KindVendor, Catalog: "ollama-cloud", Hosts: true,
Chat: "https://ollama.com/v1", Anthropic: "https://ollama.com",
+ Decide: "https://ollama.com/v1",
Note: "cloud models, with an API key",
Website: "https://docs.ollama.com/cloud", KeysURL: "https://ollama.com/settings/keys"},
// Nous Research's Nous Portal (#392): its own Hermes models and other
@@ -569,9 +572,13 @@ var presets = []PresetDef{
Decide: "https://maas.qianwenaiapi.com/compatible-mode/v1",
Note: "routes groups · picks model and effort",
Website: "https://www.qianwenai.com/models/decision-model-preview", KeysURL: "https://platform.qianwenai.com/home/api-keys"},
+ // its decision models (Nimble, Clef, Tev …), served at /v1/systemone
+ // since 0.35.0, are the ones its /api/tags gives the "decision"
+ // capability (ollama_decide.go)
{ID: "ollama", Name: "Ollama", Icon: "ollama", Kind: KindLocal, NoKey: true,
Chat: "http://localhost:11434/v1", Anthropic: "http://localhost:11434",
- Note: "your local models", Website: "https://ollama.com"},
+ Decide: "http://localhost:11434/v1",
+ Note: "your local models", Website: "https://ollama.com"},
{ID: "lmstudio", Name: "LM Studio", Icon: "lmstudio", Kind: KindLocal, NoKey: true,
Chat: "http://localhost:1234/v1",
Note: "local server, port 1234 by default", Website: "https://lmstudio.ai"},
diff --git a/internal/provider/provider.go b/internal/provider/provider.go
index 4f30fb1b8..dc4241bf3 100644
--- a/internal/provider/provider.go
+++ b/internal/provider/provider.go
@@ -1275,9 +1275,10 @@ func normalize(p Provider) Provider {
p.Decide = pr.Decide
}
}
- // an oMLX or MLX-Serve provider saved before its preset had System One (#1641)
- // asks it where its chat completions are, at whatever port it was given
- if (p.Preset == OMLXPreset || p.Preset == MLXServePreset) && p.Decide == "" && p.Chat != "" {
+ // an oMLX, MLX-Serve or Ollama provider saved before its preset had
+ // System One (#1641; Nanka on Discord) asks it where its chat
+ // completions are, at whatever address it was given
+ if (p.Preset == OMLXPreset || p.Preset == MLXServePreset || p.Preset == "ollama" || p.Preset == "ollama-cloud") && p.Decide == "" && p.Chat != "" {
p.Decide = p.Chat
}
if p.Routing != Ordered && p.Routing != Rotate && p.Routing != LeastUsed && p.Routing != Pace && p.Routing != Weighted && !PluginRouting(p.Routing) {