Compare commits

...

7 Commits

Author SHA1 Message Date
Jesse Gross
14e57e6192 server: use tiered VRAM-based default context length
Replace binary low VRAM mode with tiered VRAM thresholds that set
default context lengths for all models:

- < 24 GiB VRAM: 4,096 context
- 24-48 GiB VRAM: 32,768 context
- >= 48 GiB VRAM: 262,144 context
2026-01-27 17:38:29 -08:00
Jesse Gross
05359f383a server: fix ollama ps showing configured instead of actual context length
When context length is clamped to the model's trained context length,
ollama ps now shows the actual clamped value instead of the originally
configured value.
2026-01-27 16:32:08 -08:00
Gabe Goodhart
7b62c41060 cmd/config: use envconfig.Host() for base API in launch config packages (#13937) 2026-01-27 13:30:00 -08:00
Parth Sareen
26acab64b7 docs: add clawdbot (#13925) 2026-01-26 18:32:54 -08:00
Gyungrai Wang
e0f03790b1 parsers/ministral: fix nested tool call parsing by counting brace nesting (#13905)
* parsers/ministral: fix nested tool call parsing by counting brace nesting

* fix lint error

* parsers: refactor ministral parser

The old one was very tied to expecting to see only one token at a time,
which I don't like to assume (who knows what the future might hold wrt
speculative decoding, etc). This new one follows a similar structure to
qwen3-coder's parser, which incidentally makes it easier to test as well
(since we can test the individual events that come out when given
particular inputs).

---------

Co-authored-by: Devon Rifkin <drifkin@drifkin.net>
2026-01-26 15:03:43 -08:00
Parth Sareen
3ab842b0f5 cmd: clawdbot config fixes (#13922) 2026-01-26 14:34:29 -08:00
Parth Sareen
b8e8ef8929 cmd: ollama launch clawdbot (#13921) 2026-01-26 13:40:59 -08:00
23 changed files with 1880 additions and 114 deletions

View File

@@ -6,6 +6,8 @@ import (
"os/exec"
"path/filepath"
"runtime"
"github.com/ollama/ollama/envconfig"
)
// Claude implements Runner for Claude Code integration
@@ -50,7 +52,7 @@ func (c *Claude) Run(model string) error {
cmd.Stdout = os.Stdout
cmd.Stderr = os.Stderr
cmd.Env = append(os.Environ(),
"ANTHROPIC_BASE_URL=http://localhost:11434",
"ANTHROPIC_BASE_URL="+envconfig.Host().String(),
"ANTHROPIC_API_KEY=",
"ANTHROPIC_AUTH_TOKEN=ollama",
)

195
cmd/config/clawdbot.go Normal file
View File

@@ -0,0 +1,195 @@
package config
import (
"bytes"
"encoding/json"
"fmt"
"io"
"os"
"os/exec"
"path/filepath"
"strings"
"github.com/ollama/ollama/envconfig"
)
type Clawdbot struct{}
func (c *Clawdbot) String() string { return "Clawdbot" }
const ansiGreen = "\033[32m"
func (c *Clawdbot) Run(model string) error {
if _, err := exec.LookPath("clawdbot"); err != nil {
return fmt.Errorf("clawdbot is not installed, install from https://docs.clawd.bot")
}
models := []string{model}
if config, err := loadIntegration("clawdbot"); err == nil && len(config.Models) > 0 {
models = config.Models
}
if err := c.Edit(models); err != nil {
return fmt.Errorf("setup failed: %w", err)
}
cmd := exec.Command("clawdbot", "gateway")
cmd.Stdin = os.Stdin
// Capture output to detect "already running" message
var outputBuf bytes.Buffer
cmd.Stdout = io.MultiWriter(os.Stdout, &outputBuf)
cmd.Stderr = io.MultiWriter(os.Stderr, &outputBuf)
err := cmd.Run()
if err != nil && strings.Contains(outputBuf.String(), "Gateway already running") {
fmt.Fprintf(os.Stderr, "%sClawdbot has been configured with Ollama. Gateway is already running.%s\n", ansiGreen, ansiReset)
return nil
}
return err
}
func (c *Clawdbot) Paths() []string {
home, _ := os.UserHomeDir()
p := filepath.Join(home, ".clawdbot", "clawdbot.json")
if _, err := os.Stat(p); err == nil {
return []string{p}
}
return nil
}
func (c *Clawdbot) Edit(models []string) error {
if len(models) == 0 {
return nil
}
home, err := os.UserHomeDir()
if err != nil {
return err
}
configPath := filepath.Join(home, ".clawdbot", "clawdbot.json")
if err := os.MkdirAll(filepath.Dir(configPath), 0o755); err != nil {
return err
}
// Read into map[string]any to preserve unknown fields
config := make(map[string]any)
if data, err := os.ReadFile(configPath); err == nil {
_ = json.Unmarshal(data, &config)
}
// Navigate/create: models.providers.ollama (preserving other providers)
modelsSection, _ := config["models"].(map[string]any)
if modelsSection == nil {
modelsSection = make(map[string]any)
}
providers, _ := modelsSection["providers"].(map[string]any)
if providers == nil {
providers = make(map[string]any)
}
ollama, _ := providers["ollama"].(map[string]any)
if ollama == nil {
ollama = make(map[string]any)
}
ollama["baseUrl"] = envconfig.Host().String() + "/v1"
// needed to register provider
ollama["apiKey"] = "ollama-local"
// TODO(parthsareen): potentially move to responses
ollama["api"] = "openai-completions"
// Build map of existing models to preserve user customizations
existingModels, _ := ollama["models"].([]any)
existingByID := make(map[string]map[string]any)
for _, m := range existingModels {
if entry, ok := m.(map[string]any); ok {
if id, ok := entry["id"].(string); ok {
existingByID[id] = entry
}
}
}
var newModels []any
for _, model := range models {
entry := map[string]any{
"id": model,
"name": model,
"reasoning": false,
"input": []any{"text"},
"cost": map[string]any{
"input": 0,
"output": 0,
"cacheRead": 0,
"cacheWrite": 0,
},
// TODO(parthsareen): get these values from API
"contextWindow": 131072,
"maxTokens": 16384,
}
// Merge existing fields (user customizations)
if existing, ok := existingByID[model]; ok {
for k, v := range existing {
if _, isNew := entry[k]; !isNew {
entry[k] = v
}
}
}
newModels = append(newModels, entry)
}
ollama["models"] = newModels
providers["ollama"] = ollama
modelsSection["providers"] = providers
config["models"] = modelsSection
// Update agents.defaults.model.primary (preserving other agent settings)
agents, _ := config["agents"].(map[string]any)
if agents == nil {
agents = make(map[string]any)
}
defaults, _ := agents["defaults"].(map[string]any)
if defaults == nil {
defaults = make(map[string]any)
}
modelConfig, _ := defaults["model"].(map[string]any)
if modelConfig == nil {
modelConfig = make(map[string]any)
}
modelConfig["primary"] = "ollama/" + models[0]
defaults["model"] = modelConfig
agents["defaults"] = defaults
config["agents"] = agents
data, err := json.MarshalIndent(config, "", " ")
if err != nil {
return err
}
return writeWithBackup(configPath, data)
}
func (c *Clawdbot) Models() []string {
home, err := os.UserHomeDir()
if err != nil {
return nil
}
config, err := readJSONFile(filepath.Join(home, ".clawdbot", "clawdbot.json"))
if err != nil {
return nil
}
modelsSection, _ := config["models"].(map[string]any)
providers, _ := modelsSection["providers"].(map[string]any)
ollama, _ := providers["ollama"].(map[string]any)
modelList, _ := ollama["models"].([]any)
var result []string
for _, m := range modelList {
if entry, ok := m.(map[string]any); ok {
if id, ok := entry["id"].(string); ok {
result = append(result, id)
}
}
}
return result
}

625
cmd/config/clawdbot_test.go Normal file
View File

@@ -0,0 +1,625 @@
package config
import (
"encoding/json"
"fmt"
"os"
"path/filepath"
"testing"
)
func TestClawdbotIntegration(t *testing.T) {
c := &Clawdbot{}
t.Run("String", func(t *testing.T) {
if got := c.String(); got != "Clawdbot" {
t.Errorf("String() = %q, want %q", got, "Clawdbot")
}
})
t.Run("implements Runner", func(t *testing.T) {
var _ Runner = c
})
t.Run("implements Editor", func(t *testing.T) {
var _ Editor = c
})
}
func TestClawdbotEdit(t *testing.T) {
c := &Clawdbot{}
tmpDir := t.TempDir()
setTestHome(t, tmpDir)
configDir := filepath.Join(tmpDir, ".clawdbot")
configPath := filepath.Join(configDir, "clawdbot.json")
cleanup := func() { os.RemoveAll(configDir) }
t.Run("fresh install", func(t *testing.T) {
cleanup()
if err := c.Edit([]string{"llama3.2"}); err != nil {
t.Fatal(err)
}
assertClawdbotModelExists(t, configPath, "llama3.2")
assertClawdbotPrimaryModel(t, configPath, "ollama/llama3.2")
})
t.Run("multiple models - first is primary", func(t *testing.T) {
cleanup()
if err := c.Edit([]string{"llama3.2", "mistral"}); err != nil {
t.Fatal(err)
}
assertClawdbotModelExists(t, configPath, "llama3.2")
assertClawdbotModelExists(t, configPath, "mistral")
assertClawdbotPrimaryModel(t, configPath, "ollama/llama3.2")
})
t.Run("preserve other providers", func(t *testing.T) {
cleanup()
os.MkdirAll(configDir, 0o755)
os.WriteFile(configPath, []byte(`{"models":{"providers":{"anthropic":{"apiKey":"xxx"}}}}`), 0o644)
if err := c.Edit([]string{"llama3.2"}); err != nil {
t.Fatal(err)
}
data, _ := os.ReadFile(configPath)
var cfg map[string]any
json.Unmarshal(data, &cfg)
models := cfg["models"].(map[string]any)
providers := models["providers"].(map[string]any)
if providers["anthropic"] == nil {
t.Error("anthropic provider was removed")
}
})
t.Run("preserve top-level keys", func(t *testing.T) {
cleanup()
os.MkdirAll(configDir, 0o755)
os.WriteFile(configPath, []byte(`{"theme":"dark","mcp":{"servers":{}}}`), 0o644)
if err := c.Edit([]string{"llama3.2"}); err != nil {
t.Fatal(err)
}
data, _ := os.ReadFile(configPath)
var cfg map[string]any
json.Unmarshal(data, &cfg)
if cfg["theme"] != "dark" {
t.Error("theme was removed")
}
if cfg["mcp"] == nil {
t.Error("mcp was removed")
}
})
t.Run("preserve user customizations on models", func(t *testing.T) {
cleanup()
c.Edit([]string{"llama3.2"})
// User adds custom field
data, _ := os.ReadFile(configPath)
var cfg map[string]any
json.Unmarshal(data, &cfg)
models := cfg["models"].(map[string]any)
providers := models["providers"].(map[string]any)
ollama := providers["ollama"].(map[string]any)
modelList := ollama["models"].([]any)
entry := modelList[0].(map[string]any)
entry["customField"] = "user-value"
configData, _ := json.MarshalIndent(cfg, "", " ")
os.WriteFile(configPath, configData, 0o644)
// Re-run Edit
c.Edit([]string{"llama3.2"})
data, _ = os.ReadFile(configPath)
json.Unmarshal(data, &cfg)
models = cfg["models"].(map[string]any)
providers = models["providers"].(map[string]any)
ollama = providers["ollama"].(map[string]any)
modelList = ollama["models"].([]any)
entry = modelList[0].(map[string]any)
if entry["customField"] != "user-value" {
t.Error("custom field was lost")
}
})
t.Run("edit replaces models list", func(t *testing.T) {
cleanup()
c.Edit([]string{"llama3.2", "mistral"})
c.Edit([]string{"llama3.2"})
assertClawdbotModelExists(t, configPath, "llama3.2")
assertClawdbotModelNotExists(t, configPath, "mistral")
})
t.Run("empty models is no-op", func(t *testing.T) {
cleanup()
os.MkdirAll(configDir, 0o755)
original := `{"existing":"data"}`
os.WriteFile(configPath, []byte(original), 0o644)
c.Edit([]string{})
data, _ := os.ReadFile(configPath)
if string(data) != original {
t.Error("empty models should not modify file")
}
})
t.Run("corrupted JSON treated as empty", func(t *testing.T) {
cleanup()
os.MkdirAll(configDir, 0o755)
os.WriteFile(configPath, []byte(`{corrupted`), 0o644)
if err := c.Edit([]string{"llama3.2"}); err != nil {
t.Fatal(err)
}
data, _ := os.ReadFile(configPath)
var cfg map[string]any
if err := json.Unmarshal(data, &cfg); err != nil {
t.Error("result should be valid JSON")
}
})
t.Run("wrong type models section", func(t *testing.T) {
cleanup()
os.MkdirAll(configDir, 0o755)
os.WriteFile(configPath, []byte(`{"models":"not a map"}`), 0o644)
if err := c.Edit([]string{"llama3.2"}); err != nil {
t.Fatal(err)
}
assertClawdbotModelExists(t, configPath, "llama3.2")
})
}
func TestClawdbotModels(t *testing.T) {
c := &Clawdbot{}
tmpDir := t.TempDir()
setTestHome(t, tmpDir)
t.Run("no config returns nil", func(t *testing.T) {
if models := c.Models(); len(models) > 0 {
t.Errorf("expected nil/empty, got %v", models)
}
})
t.Run("returns all ollama models", func(t *testing.T) {
configDir := filepath.Join(tmpDir, ".clawdbot")
os.MkdirAll(configDir, 0o755)
os.WriteFile(filepath.Join(configDir, "clawdbot.json"), []byte(`{
"models":{"providers":{"ollama":{"models":[
{"id":"llama3.2"},
{"id":"mistral"}
]}}}
}`), 0o644)
models := c.Models()
if len(models) != 2 {
t.Errorf("expected 2 models, got %v", models)
}
})
}
// Helper functions
func assertClawdbotModelExists(t *testing.T, path, model string) {
t.Helper()
data, _ := os.ReadFile(path)
var cfg map[string]any
json.Unmarshal(data, &cfg)
models := cfg["models"].(map[string]any)
providers := models["providers"].(map[string]any)
ollama := providers["ollama"].(map[string]any)
modelList := ollama["models"].([]any)
for _, m := range modelList {
if entry, ok := m.(map[string]any); ok {
if entry["id"] == model {
return
}
}
}
t.Errorf("model %s not found", model)
}
func assertClawdbotModelNotExists(t *testing.T, path, model string) {
t.Helper()
data, _ := os.ReadFile(path)
var cfg map[string]any
json.Unmarshal(data, &cfg)
models, _ := cfg["models"].(map[string]any)
providers, _ := models["providers"].(map[string]any)
ollama, _ := providers["ollama"].(map[string]any)
modelList, _ := ollama["models"].([]any)
for _, m := range modelList {
if entry, ok := m.(map[string]any); ok {
if entry["id"] == model {
t.Errorf("model %s should not exist", model)
}
}
}
}
func assertClawdbotPrimaryModel(t *testing.T, path, expected string) {
t.Helper()
data, _ := os.ReadFile(path)
var cfg map[string]any
json.Unmarshal(data, &cfg)
agents := cfg["agents"].(map[string]any)
defaults := agents["defaults"].(map[string]any)
model := defaults["model"].(map[string]any)
if model["primary"] != expected {
t.Errorf("primary model = %v, want %v", model["primary"], expected)
}
}
func TestClawdbotPaths(t *testing.T) {
c := &Clawdbot{}
t.Run("returns path when config exists", func(t *testing.T) {
tmpDir := t.TempDir()
setTestHome(t, tmpDir)
configDir := filepath.Join(tmpDir, ".clawdbot")
os.MkdirAll(configDir, 0o755)
os.WriteFile(filepath.Join(configDir, "clawdbot.json"), []byte(`{}`), 0o644)
paths := c.Paths()
if len(paths) != 1 {
t.Errorf("expected 1 path, got %d", len(paths))
}
})
t.Run("returns nil when config missing", func(t *testing.T) {
tmpDir := t.TempDir()
setTestHome(t, tmpDir)
if paths := c.Paths(); paths != nil {
t.Errorf("expected nil, got %v", paths)
}
})
}
func TestClawdbotModelsEdgeCases(t *testing.T) {
c := &Clawdbot{}
tmpDir := t.TempDir()
setTestHome(t, tmpDir)
configDir := filepath.Join(tmpDir, ".clawdbot")
configPath := filepath.Join(configDir, "clawdbot.json")
cleanup := func() { os.RemoveAll(configDir) }
t.Run("corrupted JSON returns nil", func(t *testing.T) {
cleanup()
os.MkdirAll(configDir, 0o755)
os.WriteFile(configPath, []byte(`{corrupted`), 0o644)
if models := c.Models(); models != nil {
t.Errorf("expected nil, got %v", models)
}
})
t.Run("wrong type at models level", func(t *testing.T) {
cleanup()
os.MkdirAll(configDir, 0o755)
os.WriteFile(configPath, []byte(`{"models":"string"}`), 0o644)
if models := c.Models(); models != nil {
t.Errorf("expected nil, got %v", models)
}
})
t.Run("wrong type at providers level", func(t *testing.T) {
cleanup()
os.MkdirAll(configDir, 0o755)
os.WriteFile(configPath, []byte(`{"models":{"providers":"string"}}`), 0o644)
if models := c.Models(); models != nil {
t.Errorf("expected nil, got %v", models)
}
})
t.Run("wrong type at ollama level", func(t *testing.T) {
cleanup()
os.MkdirAll(configDir, 0o755)
os.WriteFile(configPath, []byte(`{"models":{"providers":{"ollama":"string"}}}`), 0o644)
if models := c.Models(); models != nil {
t.Errorf("expected nil, got %v", models)
}
})
t.Run("model entry missing id", func(t *testing.T) {
cleanup()
os.MkdirAll(configDir, 0o755)
os.WriteFile(configPath, []byte(`{"models":{"providers":{"ollama":{"models":[{"name":"test"}]}}}}`), 0o644)
if len(c.Models()) != 0 {
t.Error("expected empty for missing id")
}
})
t.Run("model id is not string", func(t *testing.T) {
cleanup()
os.MkdirAll(configDir, 0o755)
os.WriteFile(configPath, []byte(`{"models":{"providers":{"ollama":{"models":[{"id":123}]}}}}`), 0o644)
if len(c.Models()) != 0 {
t.Error("expected empty for non-string id")
}
})
}
func TestClawdbotEditSchemaFields(t *testing.T) {
c := &Clawdbot{}
tmpDir := t.TempDir()
setTestHome(t, tmpDir)
configPath := filepath.Join(tmpDir, ".clawdbot", "clawdbot.json")
if err := c.Edit([]string{"llama3.2"}); err != nil {
t.Fatal(err)
}
data, _ := os.ReadFile(configPath)
var cfg map[string]any
json.Unmarshal(data, &cfg)
models := cfg["models"].(map[string]any)
providers := models["providers"].(map[string]any)
ollama := providers["ollama"].(map[string]any)
modelList := ollama["models"].([]any)
entry := modelList[0].(map[string]any)
// Verify required schema fields
if entry["reasoning"] != false {
t.Error("reasoning should be false")
}
if entry["input"] == nil {
t.Error("input should be set")
}
if entry["contextWindow"] == nil {
t.Error("contextWindow should be set")
}
if entry["maxTokens"] == nil {
t.Error("maxTokens should be set")
}
cost := entry["cost"].(map[string]any)
if cost["cacheRead"] == nil {
t.Error("cost.cacheRead should be set")
}
if cost["cacheWrite"] == nil {
t.Error("cost.cacheWrite should be set")
}
}
func TestClawdbotEditModelNames(t *testing.T) {
c := &Clawdbot{}
tmpDir := t.TempDir()
setTestHome(t, tmpDir)
configPath := filepath.Join(tmpDir, ".clawdbot", "clawdbot.json")
cleanup := func() { os.RemoveAll(filepath.Join(tmpDir, ".clawdbot")) }
t.Run("model with colon tag", func(t *testing.T) {
cleanup()
if err := c.Edit([]string{"llama3.2:70b"}); err != nil {
t.Fatal(err)
}
assertClawdbotModelExists(t, configPath, "llama3.2:70b")
assertClawdbotPrimaryModel(t, configPath, "ollama/llama3.2:70b")
})
t.Run("model with slash", func(t *testing.T) {
cleanup()
if err := c.Edit([]string{"library/model:tag"}); err != nil {
t.Fatal(err)
}
assertClawdbotModelExists(t, configPath, "library/model:tag")
assertClawdbotPrimaryModel(t, configPath, "ollama/library/model:tag")
})
t.Run("model with hyphen", func(t *testing.T) {
cleanup()
if err := c.Edit([]string{"test-model"}); err != nil {
t.Fatal(err)
}
assertClawdbotModelExists(t, configPath, "test-model")
})
}
func TestClawdbotEditAgentsPreservation(t *testing.T) {
c := &Clawdbot{}
tmpDir := t.TempDir()
setTestHome(t, tmpDir)
configDir := filepath.Join(tmpDir, ".clawdbot")
configPath := filepath.Join(configDir, "clawdbot.json")
cleanup := func() { os.RemoveAll(configDir) }
t.Run("preserve other agent defaults", func(t *testing.T) {
cleanup()
os.MkdirAll(configDir, 0o755)
os.WriteFile(configPath, []byte(`{"agents":{"defaults":{"model":{"primary":"old"},"temperature":0.7}}}`), 0o644)
c.Edit([]string{"llama3.2"})
data, _ := os.ReadFile(configPath)
var cfg map[string]any
json.Unmarshal(data, &cfg)
agents := cfg["agents"].(map[string]any)
defaults := agents["defaults"].(map[string]any)
if defaults["temperature"] != 0.7 {
t.Error("temperature setting was lost")
}
})
t.Run("preserve other agents besides defaults", func(t *testing.T) {
cleanup()
os.MkdirAll(configDir, 0o755)
os.WriteFile(configPath, []byte(`{"agents":{"defaults":{},"custom-agent":{"foo":"bar"}}}`), 0o644)
c.Edit([]string{"llama3.2"})
data, _ := os.ReadFile(configPath)
var cfg map[string]any
json.Unmarshal(data, &cfg)
agents := cfg["agents"].(map[string]any)
if agents["custom-agent"] == nil {
t.Error("custom-agent was lost")
}
})
}
const testClawdbotFixture = `{
"theme": "dark",
"mcp": {"servers": {"custom": {"enabled": true}}},
"models": {
"providers": {
"anthropic": {"apiKey": "xxx"},
"ollama": {
"baseUrl": "http://127.0.0.1:11434/v1",
"models": [{"id": "old-model", "customField": "preserved"}]
}
}
},
"agents": {
"defaults": {"model": {"primary": "old"}, "temperature": 0.7},
"custom-agent": {"foo": "bar"}
}
}`
func TestClawdbotEdit_RoundTrip(t *testing.T) {
c := &Clawdbot{}
tmpDir := t.TempDir()
setTestHome(t, tmpDir)
configDir := filepath.Join(tmpDir, ".clawdbot")
configPath := filepath.Join(configDir, "clawdbot.json")
os.MkdirAll(configDir, 0o755)
os.WriteFile(configPath, []byte(testClawdbotFixture), 0o644)
if err := c.Edit([]string{"llama3.2", "mistral"}); err != nil {
t.Fatal(err)
}
data, _ := os.ReadFile(configPath)
var cfg map[string]any
json.Unmarshal(data, &cfg)
// Verify top-level preserved
if cfg["theme"] != "dark" {
t.Error("theme not preserved")
}
mcp := cfg["mcp"].(map[string]any)
servers := mcp["servers"].(map[string]any)
if servers["custom"] == nil {
t.Error("mcp.servers.custom not preserved")
}
// Verify other providers preserved
models := cfg["models"].(map[string]any)
providers := models["providers"].(map[string]any)
if providers["anthropic"] == nil {
t.Error("anthropic provider not preserved")
}
// Verify agents preserved
agents := cfg["agents"].(map[string]any)
if agents["custom-agent"] == nil {
t.Error("custom-agent not preserved")
}
defaults := agents["defaults"].(map[string]any)
if defaults["temperature"] != 0.7 {
t.Error("temperature not preserved")
}
}
func TestClawdbotEdit_Idempotent(t *testing.T) {
c := &Clawdbot{}
tmpDir := t.TempDir()
setTestHome(t, tmpDir)
configDir := filepath.Join(tmpDir, ".clawdbot")
configPath := filepath.Join(configDir, "clawdbot.json")
os.MkdirAll(configDir, 0o755)
os.WriteFile(configPath, []byte(testClawdbotFixture), 0o644)
c.Edit([]string{"llama3.2", "mistral"})
firstData, _ := os.ReadFile(configPath)
c.Edit([]string{"llama3.2", "mistral"})
secondData, _ := os.ReadFile(configPath)
if string(firstData) != string(secondData) {
t.Error("repeated edits with same models produced different results")
}
}
func TestClawdbotEdit_MultipleConsecutiveEdits(t *testing.T) {
c := &Clawdbot{}
tmpDir := t.TempDir()
setTestHome(t, tmpDir)
configDir := filepath.Join(tmpDir, ".clawdbot")
configPath := filepath.Join(configDir, "clawdbot.json")
os.MkdirAll(configDir, 0o755)
os.WriteFile(configPath, []byte(testClawdbotFixture), 0o644)
for i := range 10 {
models := []string{"model-a", "model-b"}
if i%2 == 0 {
models = []string{"model-x", "model-y", "model-z"}
}
if err := c.Edit(models); err != nil {
t.Fatalf("edit %d failed: %v", i, err)
}
}
data, _ := os.ReadFile(configPath)
var cfg map[string]any
if err := json.Unmarshal(data, &cfg); err != nil {
t.Fatalf("file is not valid JSON after multiple edits: %v", err)
}
if cfg["theme"] != "dark" {
t.Error("theme lost after multiple edits")
}
}
func TestClawdbotEdit_BackupCreated(t *testing.T) {
c := &Clawdbot{}
tmpDir := t.TempDir()
setTestHome(t, tmpDir)
configDir := filepath.Join(tmpDir, ".clawdbot")
configPath := filepath.Join(configDir, "clawdbot.json")
backupDir := filepath.Join(os.TempDir(), "ollama-backups")
os.MkdirAll(configDir, 0o755)
uniqueMarker := fmt.Sprintf("test-marker-%d", os.Getpid())
original := fmt.Sprintf(`{"theme": "%s"}`, uniqueMarker)
os.WriteFile(configPath, []byte(original), 0o644)
if err := c.Edit([]string{"model-a"}); err != nil {
t.Fatal(err)
}
backups, _ := filepath.Glob(filepath.Join(backupDir, "clawdbot.json.*"))
foundBackup := false
for _, backup := range backups {
data, _ := os.ReadFile(backup)
if string(data) == original {
foundBackup = true
break
}
}
if !foundBackup {
t.Error("backup with original content not found")
}
}
func TestClawdbotEdit_CreatesDirectoryIfMissing(t *testing.T) {
c := &Clawdbot{}
tmpDir := t.TempDir()
setTestHome(t, tmpDir)
configDir := filepath.Join(tmpDir, ".clawdbot")
if _, err := os.Stat(configDir); !os.IsNotExist(err) {
t.Fatal("directory should not exist before test")
}
if err := c.Edit([]string{"model-a"}); err != nil {
t.Fatal(err)
}
if _, err := os.Stat(configDir); os.IsNotExist(err) {
t.Fatal("directory was not created")
}
}

View File

@@ -7,6 +7,8 @@ import (
"os/exec"
"path/filepath"
"slices"
"github.com/ollama/ollama/envconfig"
)
// Droid implements Runner and Editor for Droid integration
@@ -117,7 +119,7 @@ func (d *Droid) Edit(models []string) error {
newModels = append(newModels, modelEntry{
Model: model,
DisplayName: model,
BaseURL: "http://localhost:11434/v1",
BaseURL: envconfig.Host().String() + "/v1",
APIKey: "ollama",
Provider: "generic-chat-completion-api",
MaxOutputTokens: 64000,

View File

@@ -218,7 +218,7 @@ func TestDroidEdit(t *testing.T) {
}
}
if model["baseUrl"] != "http://localhost:11434/v1" {
if model["baseUrl"] != "http://127.0.0.1:11434/v1" {
t.Errorf("unexpected baseUrl: %s", model["baseUrl"])
}
if model["apiKey"] != "ollama" {
@@ -447,7 +447,7 @@ const testDroidSettingsFixture = `{
{
"model": "existing-ollama-model",
"displayName": "existing-ollama-model",
"baseUrl": "http://localhost:11434/v1",
"baseUrl": "http://127.0.0.1:11434/v1",
"apiKey": "ollama",
"provider": "generic-chat-completion-api",
"maxOutputTokens": 64000,

View File

@@ -41,6 +41,7 @@ type Editor interface {
// integrations is the registry of available integrations.
var integrations = map[string]Runner{
"claude": &Claude{},
"clawdbot": &Clawdbot{},
"codex": &Codex{},
"droid": &Droid{},
"opencode": &OpenCode{},
@@ -242,6 +243,7 @@ func LaunchCmd(checkServerHeartbeat func(cmd *cobra.Command, args []string) erro
Supported integrations:
claude Claude Code
clawdbot Clawdbot
codex Codex
droid Droid
opencode OpenCode

View File

@@ -9,6 +9,8 @@ import (
"path/filepath"
"slices"
"strings"
"github.com/ollama/ollama/envconfig"
)
// OpenCode implements Runner and Editor for OpenCode integration
@@ -88,7 +90,7 @@ func (o *OpenCode) Edit(modelList []string) error {
"npm": "@ai-sdk/openai-compatible",
"name": "Ollama (local)",
"options": map[string]any{
"baseURL": "http://localhost:11434/v1",
"baseURL": envconfig.Host().String() + "/v1",
},
}
}

View File

@@ -102,6 +102,7 @@
"group": "Integrations",
"pages": [
"/integrations/claude-code",
"/integrations/clawdbot",
"/integrations/cline",
"/integrations/codex",
"/integrations/droid",

View File

@@ -0,0 +1,48 @@
---
title: Clawdbot
---
Clawdbot is a personal AI assistant that runs on your own devices. It bridges messaging services (WhatsApp, Telegram, Slack, Discord, iMessage, and more) to AI coding agents through a centralized gateway.
## Install
Install [Clawdbot](https://clawd.bot/)
```bash
npm install -g clawdbot@latest
```
Then run the onboarding wizard:
```bash
clawdbot onboard --install-daemon
```
<Note>Clawdbot requires a larger context window. It is recommended to use a context window of at least 64k tokens. See [Context length](/context-length) for more information.</Note>
## Usage with Ollama
### Quick setup
```bash
ollama launch clawdbot
```
This configures Clawdbot to use Ollama and starts the gateway.
If the gateway is already running, no changes need to be made as the gateway will auto-reload the changes.
To configure without launching:
```shell
ollama launch clawdbot --config
```
## Recommended Models
- `qwen3-coder`
- `glm-4.7`
- `gpt-oss:20b`
- `gpt-oss:120b`
Cloud models are also available at [ollama.com/search?c=cloud](https://ollama.com/search?c=cloud).

View File

@@ -201,7 +201,7 @@ var (
// Enable the new Ollama engine
NewEngine = Bool("OLLAMA_NEW_ENGINE")
// ContextLength sets the default context length
ContextLength = Uint("OLLAMA_CONTEXT_LENGTH", 4096)
ContextLength = Uint("OLLAMA_CONTEXT_LENGTH", 0)
// Auth enables authentication between the Ollama client and server
UseAuth = Bool("OLLAMA_AUTH")
// Enable Vulkan backend

View File

@@ -282,7 +282,7 @@ func TestVar(t *testing.T) {
func TestContextLength(t *testing.T) {
cases := map[string]uint{
"": 4096,
"": 0,
"2048": 2048,
}

View File

@@ -80,6 +80,7 @@ type LlamaServer interface {
GetPort() int
GetDeviceInfos(ctx context.Context) []ml.DeviceInfo
HasExited() bool
ContextLength() int
}
// llmServer is an instance of a runner hosting a single model
@@ -1901,6 +1902,10 @@ func (s *llmServer) VRAMByGPU(id ml.DeviceID) uint64 {
return 0
}
func (s *llmServer) ContextLength() int {
return s.options.NumCtx
}
func (s *ollamaServer) GetDeviceInfos(ctx context.Context) []ml.DeviceInfo {
devices, err := ml.GetDevicesFromRunner(ctx, s)
if err != nil {

View File

@@ -4,6 +4,7 @@ import (
"encoding/json"
"fmt"
"strings"
"unicode"
"github.com/ollama/ollama/api"
)
@@ -17,12 +18,34 @@ const (
ministralCollectingToolArgs
)
// ministralEvent represents an event emitted during parsing
type ministralEvent interface {
isMinistralEvent()
}
type ministralEventContent struct {
content string
}
type ministralEventThinking struct {
thinking string
}
type ministralEventToolCall struct {
name string
args string // raw JSON string
}
func (ministralEventContent) isMinistralEvent() {}
func (ministralEventThinking) isMinistralEvent() {}
func (ministralEventToolCall) isMinistralEvent() {}
type MinistralParser struct {
state ministralParserState
buffer strings.Builder
tools []api.Tool
hasThinkingSupport bool
currentTool *api.Tool
pendingToolName string // stores tool name while collecting args
}
func (p *MinistralParser) HasToolSupport() bool {
@@ -63,74 +86,251 @@ func toolByName(tools []api.Tool, n string) (*api.Tool, error) {
return nil, fmt.Errorf("tool '%s' not found", n)
}
func (p *MinistralParser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
p.buffer.WriteString(s)
const (
ministralToolCallsTag = "[TOOL_CALLS]"
ministralThinkTag = "[THINK]"
ministralThinkEndTag = "[/THINK]"
ministralArgsTag = "[ARGS]"
)
// eat consumes the parser's buffer, and returns a list of any unambiguous
// events from the current parser state. The second return value indicates
// whether to keep looping (true when state transitions, false when waiting
// for more data).
func (p *MinistralParser) eat() ([]ministralEvent, bool) {
var events []ministralEvent
switch p.state {
case ministralCollectingContent:
if strings.Contains(p.buffer.String(), "[TOOL_CALLS]") {
before, _ := splitAtTag(&p.buffer, "[TOOL_CALLS]", false)
if before != "" {
return before, "", calls, nil
bufStr := p.buffer.String()
// Check for [TOOL_CALLS] tag
if strings.Contains(bufStr, ministralToolCallsTag) {
split := strings.SplitN(bufStr, ministralToolCallsTag, 2)
before := strings.TrimRightFunc(split[0], unicode.IsSpace)
if len(before) > 0 {
events = append(events, ministralEventContent{content: before})
}
after := split[1]
p.buffer.Reset()
p.buffer.WriteString(after)
p.state = ministralCollectingToolName
} else if strings.Contains(p.buffer.String(), "[THINK]") {
return events, true
}
// Check for [THINK] tag
if strings.Contains(bufStr, ministralThinkTag) {
split := strings.SplitN(bufStr, ministralThinkTag, 2)
before := strings.TrimRightFunc(split[0], unicode.IsSpace)
if len(before) > 0 {
events = append(events, ministralEventContent{content: before})
}
after := split[1]
p.buffer.Reset()
p.buffer.WriteString(after)
p.state = ministralCollectingThinkingContent
return "", "", calls, nil
} else {
p.buffer.Reset()
return s, "", calls, nil
return events, true
}
// Check for partial tag overlap with [TOOL_CALLS] or [THINK]
overlapToolCalls := overlap(bufStr, ministralToolCallsTag)
overlapThink := overlap(bufStr, ministralThinkTag)
maxOverlap := max(overlapToolCalls, overlapThink)
if maxOverlap > 0 {
// Withhold the potential partial tag
beforePartialTag := bufStr[:len(bufStr)-maxOverlap]
trailingWS := trailingWhitespaceLen(beforePartialTag)
ambiguousStart := len(beforePartialTag) - trailingWS
unambiguous := bufStr[:ambiguousStart]
ambiguous := bufStr[ambiguousStart:]
p.buffer.Reset()
p.buffer.WriteString(ambiguous)
if len(unambiguous) > 0 {
events = append(events, ministralEventContent{content: unambiguous})
}
return events, false
}
// No tag found: emit content but withhold trailing whitespace
whitespaceLen := trailingWhitespaceLen(bufStr)
ambiguousStart := len(bufStr) - whitespaceLen
unambiguous := bufStr[:ambiguousStart]
ambiguous := bufStr[ambiguousStart:]
p.buffer.Reset()
p.buffer.WriteString(ambiguous)
if len(unambiguous) > 0 {
events = append(events, ministralEventContent{content: unambiguous})
}
return events, false
case ministralCollectingThinkingContent:
if strings.Contains(p.buffer.String(), "[/THINK]") {
thinkingContent, after := splitAtTag(&p.buffer, "[/THINK]", true)
p.state = ministralCollectingContent
if after != "" {
p.buffer.Reset()
return after, thinkingContent, calls, nil
}
return "", thinkingContent, calls, nil
} else {
bufStr := p.buffer.String()
if strings.Contains(bufStr, ministralThinkEndTag) {
split := strings.SplitN(bufStr, ministralThinkEndTag, 2)
thinkingContent := split[0]
after := strings.TrimLeftFunc(split[1], unicode.IsSpace)
p.buffer.Reset()
return "", s, calls, nil
}
case ministralCollectingToolName:
if strings.Contains(p.buffer.String(), "[ARGS]") {
name, _ := splitAtTag(&p.buffer, "[ARGS]", false)
t, err := toolByName(p.tools, name)
if err != nil {
return "", "", calls, err
p.buffer.WriteString(after)
if len(thinkingContent) > 0 {
events = append(events, ministralEventThinking{thinking: thinkingContent})
}
p.currentTool = t
p.state = ministralCollectingToolArgs
return "", "", calls, nil
}
return "", "", calls, nil
case ministralCollectingToolArgs:
if strings.Contains(p.buffer.String(), "}") {
before, _ := splitAtTag(&p.buffer, "}", false)
before += "}"
var args api.ToolCallFunctionArguments
if err := json.Unmarshal([]byte(before), &args); err != nil {
// todo - throw a better error
return "", "", calls, err
}
p.state = ministralCollectingContent
return events, true
}
call := api.ToolCall{
// Check for partial overlap with [/THINK]
if overlapLen := overlap(bufStr, ministralThinkEndTag); overlapLen > 0 {
unambiguous := bufStr[:len(bufStr)-overlapLen]
ambiguous := bufStr[len(bufStr)-overlapLen:]
p.buffer.Reset()
p.buffer.WriteString(ambiguous)
if len(unambiguous) > 0 {
events = append(events, ministralEventThinking{thinking: unambiguous})
}
return events, false
}
// No tag found: emit all thinking content
p.buffer.Reset()
if len(bufStr) > 0 {
events = append(events, ministralEventThinking{thinking: bufStr})
}
return events, false
case ministralCollectingToolName:
bufStr := p.buffer.String()
if strings.Contains(bufStr, ministralArgsTag) {
split := strings.SplitN(bufStr, ministralArgsTag, 2)
toolName := split[0]
after := split[1]
p.pendingToolName = toolName
p.buffer.Reset()
p.buffer.WriteString(after)
p.state = ministralCollectingToolArgs
return events, true
}
// Wait for more data
return events, false
case ministralCollectingToolArgs:
bufStr := p.buffer.String()
jsonEnd := findJSONEnd(bufStr)
if jsonEnd != -1 {
jsonStr := bufStr[:jsonEnd+1]
remaining := bufStr[jsonEnd+1:]
events = append(events, ministralEventToolCall{
name: p.pendingToolName,
args: jsonStr,
})
p.pendingToolName = ""
p.buffer.Reset()
p.buffer.WriteString(remaining)
p.state = ministralCollectingContent
return events, true
}
// Wait for more data
return events, false
default:
panic("unexpected ministral event")
}
}
// parseEvents loops calling eat() until it returns false
func (p *MinistralParser) parseEvents() []ministralEvent {
var all []ministralEvent
keepLooping := true
for keepLooping {
var events []ministralEvent
events, keepLooping = p.eat()
all = append(all, events...)
}
return all
}
func (p *MinistralParser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
p.buffer.WriteString(s)
events := p.parseEvents()
var contentBuilder, thinkingBuilder strings.Builder
var toolCalls []api.ToolCall
for _, event := range events {
switch e := event.(type) {
case ministralEventContent:
contentBuilder.WriteString(e.content)
case ministralEventThinking:
thinkingBuilder.WriteString(e.thinking)
case ministralEventToolCall:
// Validate tool exists
tool, toolErr := toolByName(p.tools, e.name)
if toolErr != nil {
return contentBuilder.String(), thinkingBuilder.String(), toolCalls, toolErr
}
// Parse JSON arguments
var args api.ToolCallFunctionArguments
if jsonErr := json.Unmarshal([]byte(e.args), &args); jsonErr != nil {
return contentBuilder.String(), thinkingBuilder.String(), toolCalls, jsonErr
}
toolCalls = append(toolCalls, api.ToolCall{
Function: api.ToolCallFunction{
Name: p.currentTool.Function.Name,
Name: tool.Function.Name,
Arguments: args,
},
}
calls = append(calls, call)
return "", "", calls, nil
})
}
return "", "", calls, nil
}
return p.buffer.String(), thinking, calls, nil
return contentBuilder.String(), thinkingBuilder.String(), toolCalls, nil
}
// findJSONEnd finds the index of the closing brace that completes a JSON object.
// It properly handles nested objects, arrays, and strings (including escaped characters).
// Returns -1 if the JSON is not yet complete.
func findJSONEnd(s string) int {
depth := 0
inString := false
escaped := false
for i, r := range s {
if inString {
switch {
case escaped:
// If the previous character was a backslash, skip this character
escaped = false
case r == '\\':
// Mark the next character as escaped
escaped = true
case r == '"':
// End of string literal
inString = false
}
continue
}
switch r {
case '"':
// Start of string literal
inString = true
case '{', '[':
// Increase nesting level for objects and arrays
depth++
case '}', ']':
// Decrease nesting level
depth--
if depth == 0 {
// Reached the end of the root JSON structure
return i
}
}
}
return -1
}

View File

@@ -0,0 +1,545 @@
package parsers
import (
"reflect"
"testing"
"github.com/ollama/ollama/api"
)
func TestMinistralParserStreaming(t *testing.T) {
type step struct {
input string
wantEvents []ministralEvent
}
cases := []struct {
desc string
tools []api.Tool
steps []step
think bool // whether to enable thinking support
}{
// Content streaming
{
desc: "simple content",
steps: []step{
{input: "Hello, how can I help you?", wantEvents: []ministralEvent{
ministralEventContent{content: "Hello, how can I help you?"},
}},
},
},
{
desc: "streaming content word by word",
steps: []step{
{input: "Hello,", wantEvents: []ministralEvent{ministralEventContent{content: "Hello,"}}},
{input: " how", wantEvents: []ministralEvent{ministralEventContent{content: " how"}}},
{input: " can I help?", wantEvents: []ministralEvent{ministralEventContent{content: " can I help?"}}},
},
},
// Simple tool calls
{
desc: "simple tool call",
tools: []api.Tool{{Function: api.ToolFunction{Name: "get_weather"}}},
steps: []step{
{input: `[TOOL_CALLS]get_weather[ARGS]{"location": "San Francisco"}`, wantEvents: []ministralEvent{
ministralEventToolCall{name: "get_weather", args: `{"location": "San Francisco"}`},
}},
},
},
{
desc: "tool call with nested object",
tools: []api.Tool{{Function: api.ToolFunction{Name: "create_entities"}}},
steps: []step{
{input: `[TOOL_CALLS]create_entities[ARGS]{"entities": [{"entityType": "Person", "name": "Jack", "observations": ["Works as a baker"]}]}`, wantEvents: []ministralEvent{
ministralEventToolCall{name: "create_entities", args: `{"entities": [{"entityType": "Person", "name": "Jack", "observations": ["Works as a baker"]}]}`},
}},
},
},
{
desc: "tool call with deeply nested objects",
tools: []api.Tool{{Function: api.ToolFunction{Name: "update_config"}}},
steps: []step{
{input: `[TOOL_CALLS]update_config[ARGS]{"settings": {"user": {"profile": {"name": "John", "age": 30}}, "theme": "dark"}}`, wantEvents: []ministralEvent{
ministralEventToolCall{name: "update_config", args: `{"settings": {"user": {"profile": {"name": "John", "age": 30}}, "theme": "dark"}}`},
}},
},
},
{
desc: "tool call with array of objects",
tools: []api.Tool{{Function: api.ToolFunction{Name: "process_items"}}},
steps: []step{
{input: `[TOOL_CALLS]process_items[ARGS]{"items": [{"id": 1}, {"id": 2}, {"id": 3}]}`, wantEvents: []ministralEvent{
ministralEventToolCall{name: "process_items", args: `{"items": [{"id": 1}, {"id": 2}, {"id": 3}]}`},
}},
},
},
{
desc: "tool call with escaped quotes in string",
tools: []api.Tool{{Function: api.ToolFunction{Name: "search"}}},
steps: []step{
{input: `[TOOL_CALLS]search[ARGS]{"query": "say \"hello\""}`, wantEvents: []ministralEvent{
ministralEventToolCall{name: "search", args: `{"query": "say \"hello\""}`},
}},
},
},
{
desc: "tool call with braces inside string",
tools: []api.Tool{{Function: api.ToolFunction{Name: "format"}}},
steps: []step{
{input: `[TOOL_CALLS]format[ARGS]{"template": "Hello {name}!"}`, wantEvents: []ministralEvent{
ministralEventToolCall{name: "format", args: `{"template": "Hello {name}!"}`},
}},
},
},
{
desc: "empty JSON object",
tools: []api.Tool{{Function: api.ToolFunction{Name: "no_args"}}},
steps: []step{
{input: `[TOOL_CALLS]no_args[ARGS]{}`, wantEvents: []ministralEvent{
ministralEventToolCall{name: "no_args", args: `{}`},
}},
},
},
{
desc: "JSON with newlines in string",
tools: []api.Tool{{Function: api.ToolFunction{Name: "write"}}},
steps: []step{
{input: `[TOOL_CALLS]write[ARGS]{"content": "line1\nline2\nline3"}`, wantEvents: []ministralEvent{
ministralEventToolCall{name: "write", args: `{"content": "line1\nline2\nline3"}`},
}},
},
},
{
desc: "backslash in string value",
tools: []api.Tool{{Function: api.ToolFunction{Name: "path"}}},
steps: []step{
{input: `[TOOL_CALLS]path[ARGS]{"dir": "C:\\Users\\test"}`, wantEvents: []ministralEvent{
ministralEventToolCall{name: "path", args: `{"dir": "C:\\Users\\test"}`},
}},
},
},
// Content after tool call
{
desc: "content after tool call",
tools: []api.Tool{{Function: api.ToolFunction{Name: "test"}}},
steps: []step{
// NOTE: It's unclear if this is valid Ministral output, but the parser
// currently treats text after a tool call as regular content. This test
// documents that behavior so we notice if it changes.
{input: `[TOOL_CALLS]test[ARGS]{"a": 1}some content after`, wantEvents: []ministralEvent{
ministralEventToolCall{name: "test", args: `{"a": 1}`},
ministralEventContent{content: "some content after"},
}},
},
},
// Multiple tool calls
{
desc: "multiple tool calls in sequence",
tools: []api.Tool{
{Function: api.ToolFunction{Name: "get_weather"}},
{Function: api.ToolFunction{Name: "get_time"}},
},
steps: []step{
{input: `[TOOL_CALLS]get_weather[ARGS]{"location": "NYC"}[TOOL_CALLS]get_time[ARGS]{"timezone": "EST"}`, wantEvents: []ministralEvent{
ministralEventToolCall{name: "get_weather", args: `{"location": "NYC"}`},
ministralEventToolCall{name: "get_time", args: `{"timezone": "EST"}`},
}},
},
},
{
desc: "multiple tool calls streamed separately",
tools: []api.Tool{
{Function: api.ToolFunction{Name: "tool_a"}},
{Function: api.ToolFunction{Name: "tool_b"}},
},
steps: []step{
{input: `[TOOL_CALLS]tool_a[ARGS]{"x": 1}`, wantEvents: []ministralEvent{
ministralEventToolCall{name: "tool_a", args: `{"x": 1}`},
}},
{input: `[TOOL_CALLS]tool_b[ARGS]{"y": 2}`, wantEvents: []ministralEvent{
ministralEventToolCall{name: "tool_b", args: `{"y": 2}`},
}},
},
},
// Streaming tool calls
{
desc: "streaming tool call with nested objects",
tools: []api.Tool{{Function: api.ToolFunction{Name: "create_entities"}}},
steps: []step{
{input: "[TOOL_CALLS]create_entities[ARGS]", wantEvents: []ministralEvent{}},
{input: `{"entities": [{"entityType": "Person",`, wantEvents: []ministralEvent{}},
{input: ` "name": "Jack",`, wantEvents: []ministralEvent{}},
{input: ` "observations": ["Works`, wantEvents: []ministralEvent{}},
{input: ` as a baker"]}`, wantEvents: []ministralEvent{}},
{input: `]}`, wantEvents: []ministralEvent{
ministralEventToolCall{name: "create_entities", args: `{"entities": [{"entityType": "Person", "name": "Jack", "observations": ["Works as a baker"]}]}`},
}},
},
},
{
desc: "streaming with incomplete JSON waits for completion",
tools: []api.Tool{{Function: api.ToolFunction{Name: "test"}}},
steps: []step{
{input: "[TOOL_CALLS]test[ARGS]{", wantEvents: []ministralEvent{}},
{input: `"a": {`, wantEvents: []ministralEvent{}},
{input: `"b": 1`, wantEvents: []ministralEvent{}},
{input: `}`, wantEvents: []ministralEvent{}},
{input: `}`, wantEvents: []ministralEvent{
ministralEventToolCall{name: "test", args: `{"a": {"b": 1}}`},
}},
},
},
// Partial tag handling
{
desc: "partial tool tag fakeout",
steps: []step{
{input: "abc[TOOL", wantEvents: []ministralEvent{ministralEventContent{content: "abc"}}},
{input: " not a tag", wantEvents: []ministralEvent{ministralEventContent{content: "[TOOL not a tag"}}},
},
},
{
desc: "tool call tag split across chunks",
tools: []api.Tool{{Function: api.ToolFunction{Name: "test"}}},
steps: []step{
{input: "[TOOL_", wantEvents: []ministralEvent{}},
{input: "CALLS]test[ARGS]{}", wantEvents: []ministralEvent{
ministralEventToolCall{name: "test", args: `{}`},
}},
},
},
{
desc: "content before tool call",
tools: []api.Tool{{Function: api.ToolFunction{Name: "get_weather"}}},
steps: []step{
{input: "hello [TOOL_CALLS]get_weather[ARGS]{}", wantEvents: []ministralEvent{
ministralEventContent{content: "hello"},
ministralEventToolCall{name: "get_weather", args: `{}`},
}},
},
},
{
desc: "whitespace between content and tool call is trimmed",
tools: []api.Tool{{Function: api.ToolFunction{Name: "test"}}},
steps: []step{
{input: "content \n [TOOL_CALLS]test[ARGS]{}", wantEvents: []ministralEvent{
ministralEventContent{content: "content"},
ministralEventToolCall{name: "test", args: `{}`},
}},
},
},
{
desc: "tabs and newlines before tool call are trimmed",
tools: []api.Tool{{Function: api.ToolFunction{Name: "test"}}},
steps: []step{
{input: "content\t\n\t[TOOL_CALLS]test[ARGS]{}", wantEvents: []ministralEvent{
ministralEventContent{content: "content"},
ministralEventToolCall{name: "test", args: `{}`},
}},
},
},
{
desc: "non-breaking space before tool call is trimmed",
tools: []api.Tool{{Function: api.ToolFunction{Name: "test"}}},
steps: []step{
// \u00a0 is non-breaking space, which unicode.IsSpace considers whitespace
{input: "content\u00a0[TOOL_CALLS]test[ARGS]{}", wantEvents: []ministralEvent{
ministralEventContent{content: "content"},
ministralEventToolCall{name: "test", args: `{}`},
}},
},
},
{
desc: "whitespace before THINK tag is trimmed",
steps: []step{
{input: "content \n [THINK]thinking[/THINK]after", wantEvents: []ministralEvent{
ministralEventContent{content: "content"},
ministralEventThinking{thinking: "thinking"},
ministralEventContent{content: "after"},
}},
},
},
{
desc: "trailing whitespace withheld then emitted",
steps: []step{
{input: "Hello ", wantEvents: []ministralEvent{ministralEventContent{content: "Hello"}}},
{input: "world", wantEvents: []ministralEvent{ministralEventContent{content: " world"}}},
},
},
{
desc: "trailing newline withheld then emitted",
steps: []step{
{input: "Hello\n", wantEvents: []ministralEvent{ministralEventContent{content: "Hello"}}},
{input: "world", wantEvents: []ministralEvent{ministralEventContent{content: "\nworld"}}},
},
},
// Thinking support
{
desc: "thinking content",
think: true,
steps: []step{
{input: "thinking here[/THINK]", wantEvents: []ministralEvent{
ministralEventThinking{thinking: "thinking here"},
}},
{input: "content after", wantEvents: []ministralEvent{
ministralEventContent{content: "content after"},
}},
},
},
{
desc: "thinking with whitespace after end tag",
think: true,
steps: []step{
{input: "my thoughts[/THINK] \n response", wantEvents: []ministralEvent{
ministralEventThinking{thinking: "my thoughts"},
ministralEventContent{content: "response"},
}},
},
},
{
desc: "non-breaking space after think end tag is trimmed",
think: true,
steps: []step{
// \u00a0 is non-breaking space
{input: "thinking[/THINK]\u00a0response", wantEvents: []ministralEvent{
ministralEventThinking{thinking: "thinking"},
ministralEventContent{content: "response"},
}},
},
},
{
desc: "partial think end tag",
think: true,
steps: []step{
{input: "thinking[/THI", wantEvents: []ministralEvent{ministralEventThinking{thinking: "thinking"}}},
{input: "NK]after", wantEvents: []ministralEvent{ministralEventContent{content: "after"}}},
},
},
{
desc: "think tag fakeout",
think: true,
steps: []step{
{input: "thinking[/THI", wantEvents: []ministralEvent{ministralEventThinking{thinking: "thinking"}}},
{input: "not end tag", wantEvents: []ministralEvent{ministralEventThinking{thinking: "[/THInot end tag"}}},
},
},
{
desc: "thinking then tool call",
think: true,
tools: []api.Tool{{Function: api.ToolFunction{Name: "test"}}},
steps: []step{
{input: "let me think[/THINK][TOOL_CALLS]test[ARGS]{}", wantEvents: []ministralEvent{
ministralEventThinking{thinking: "let me think"},
ministralEventToolCall{name: "test", args: `{}`},
}},
},
},
// Content then THINK tag transition
{
desc: "content then think tag",
steps: []step{
{input: "content[THINK]thinking[/THINK]more", wantEvents: []ministralEvent{
ministralEventContent{content: "content"},
ministralEventThinking{thinking: "thinking"},
ministralEventContent{content: "more"},
}},
},
},
// Unicode handling
{
desc: "unicode content",
steps: []step{
{input: "你好 🌍 مرحبا", wantEvents: []ministralEvent{
ministralEventContent{content: "你好 🌍 مرحبا"},
}},
},
},
{
desc: "unicode in tool args",
tools: []api.Tool{{Function: api.ToolFunction{Name: "greet"}}},
steps: []step{
{input: `[TOOL_CALLS]greet[ARGS]{"message": "你好 🌍"}`, wantEvents: []ministralEvent{
ministralEventToolCall{name: "greet", args: `{"message": "你好 🌍"}`},
}},
},
},
}
for _, tc := range cases {
t.Run(tc.desc, func(t *testing.T) {
parser := MinistralParser{}
parser.hasThinkingSupport = tc.think
parser.Init(tc.tools, nil, nil)
for i, step := range tc.steps {
parser.buffer.WriteString(step.input)
gotEvents := parser.parseEvents()
if len(gotEvents) == 0 && len(step.wantEvents) == 0 {
// avoid deep equal on empty vs. nil slices
continue
}
if !reflect.DeepEqual(gotEvents, step.wantEvents) {
t.Errorf("step %d: input %q: got events %#v, want %#v", i, step.input, gotEvents, step.wantEvents)
}
}
})
}
}
func TestMinistralParser_Errors(t *testing.T) {
t.Run("unknown tool returns error", func(t *testing.T) {
p := &MinistralParser{}
p.Init([]api.Tool{{Function: api.ToolFunction{Name: "known_tool"}}}, nil, nil)
_, _, _, err := p.Add(`[TOOL_CALLS]unknown_tool[ARGS]{"a": 1}`, true)
if err == nil {
t.Fatal("expected error for unknown tool")
}
})
t.Run("invalid JSON returns error", func(t *testing.T) {
p := &MinistralParser{}
p.Init([]api.Tool{{Function: api.ToolFunction{Name: "test"}}}, nil, nil)
_, _, _, err := p.Add(`[TOOL_CALLS]test[ARGS]{invalid json}`, true)
if err == nil {
t.Fatal("expected error for invalid JSON")
}
})
}
func TestFindJSONEnd(t *testing.T) {
tests := []struct {
name string
input string
expected int
}{
{
name: "simple object",
input: `{"a": 1}`,
expected: 7,
},
{
name: "nested object",
input: `{"a": {"b": 2}}`,
expected: 14,
},
{
name: "array inside object",
input: `{"items": [1, 2, 3]}`,
expected: 19,
},
{
name: "braces in string",
input: `{"template": "Hello {name}!"}`,
expected: 28,
},
{
name: "escaped quotes",
input: `{"msg": "say \"hi\""}`,
expected: 20,
},
{
name: "incomplete object",
input: `{"a": {"b": 1}`,
expected: -1,
},
{
name: "deeply nested",
input: `{"a": {"b": {"c": {"d": 1}}}}`,
expected: 28,
},
{
name: "object with trailing content",
input: `{"a": 1} extra`,
expected: 7,
},
{
name: "array",
input: `[{"a": 1}, {"b": 2}]`,
expected: 19,
},
{
name: "escaped backslash before quote",
input: `{"path": "C:\\"}`,
expected: 15,
},
{
name: "empty string",
input: "",
expected: -1,
},
{
name: "no opening brace",
input: "hello world",
expected: -1,
},
{
name: "only opening brace",
input: "{",
expected: -1,
},
{
name: "unclosed string",
input: `{"key": "unclosed`,
expected: -1,
},
{
name: "double escaped backslash then quote",
input: `{"path": "C:\\\\"}`,
expected: 17,
},
{
name: "unicode in key and value",
input: `{"키": "값"}`,
expected: 13,
},
{
name: "nested arrays",
input: `{"matrix": [[1, 2], [3, 4]]}`,
expected: 27,
},
{
name: "mixed nesting",
input: `{"a": [{"b": {"c": [1, 2, 3]}}]}`,
expected: 31,
},
}
for _, tt := range tests {
t.Run(tt.name, func(t *testing.T) {
result := findJSONEnd(tt.input)
if result != tt.expected {
t.Errorf("findJSONEnd(%q) = %d, want %d", tt.input, result, tt.expected)
}
})
}
}
func TestMinistralParser_HasToolSupport(t *testing.T) {
p := &MinistralParser{}
if !p.HasToolSupport() {
t.Error("expected HasToolSupport to return true")
}
}
func TestMinistralParser_HasThinkingSupport(t *testing.T) {
p := &MinistralParser{hasThinkingSupport: false}
if p.HasThinkingSupport() {
t.Error("expected HasThinkingSupport to return false")
}
p = &MinistralParser{hasThinkingSupport: true}
if !p.HasThinkingSupport() {
t.Error("expected HasThinkingSupport to return true")
}
}

View File

@@ -3,6 +3,7 @@ package parsers
import (
"strings"
"unicode"
"unicode/utf8"
"github.com/ollama/ollama/api"
"github.com/ollama/ollama/harmony"
@@ -114,3 +115,33 @@ func splitAtTag(sb *strings.Builder, tag string, trimAfter bool) (string, string
sb.WriteString(after)
return before, after // return events
}
// overlap returns the longest overlap between the suffix of s and the prefix of delim
func overlap(s, delim string) int {
max := min(len(delim), len(s))
for i := max; i > 0; i-- {
if strings.HasSuffix(s, delim[:i]) {
return i
}
}
return 0
}
// trailingWhitespaceLen returns the length in bytes of trailing whitespace in s
func trailingWhitespaceLen(s string) int {
remaining := s
total := 0
for len(remaining) > 0 {
r, size := utf8.DecodeLastRuneInString(remaining)
// if it's an invalid utf8 rune, assume it isn't whitespace
if r == utf8.RuneError && size == 1 {
break
}
if !unicode.IsSpace(r) {
break
}
total += size
remaining = remaining[:len(remaining)-size]
}
return total
}

View File

@@ -11,7 +11,6 @@ import (
"strconv"
"strings"
"unicode"
"unicode/utf8"
"github.com/ollama/ollama/api"
"github.com/ollama/ollama/logutil"
@@ -194,36 +193,6 @@ func eat(p *Qwen3CoderParser) ([]qwenEvent, bool) {
}
}
// TODO(drifkin): move this to a shared location
// longest overlap between suffix of s and prefix of delim
func overlap(s, delim string) int {
max := min(len(delim), len(s))
for i := max; i > 0; i-- {
if strings.HasSuffix(s, delim[:i]) {
return i
}
}
return 0
}
func trailingWhitespaceLen(s string) int {
remaining := s
total := 0
for len(remaining) > 0 {
r, size := utf8.DecodeLastRuneInString(remaining)
// if it's an invalid utf8 rune, assume it isn't whitespace
if r == utf8.RuneError && size == 1 {
break
}
if !unicode.IsSpace(r) {
break
}
total += size
remaining = remaining[:len(remaining)-size]
}
return total
}
type XMLFunctionCall struct {
XMLName xml.Name `xml:"function"`
Name string `xml:"name,attr"`

View File

@@ -75,16 +75,12 @@ func experimentEnabled(name string) bool {
var useClient2 = experimentEnabled("client2")
// Low VRAM mode is based on the sum of total VRAM (not free) and triggers
// reduced context length on some models
var lowVRAMThreshold uint64 = 20 * format.GibiByte
var mode string = gin.DebugMode
type Server struct {
addr net.Addr
sched *Scheduler
lowVRAM bool
addr net.Addr
sched *Scheduler
defaultNumCtx int
}
func init() {
@@ -107,8 +103,12 @@ var (
errBadTemplate = errors.New("template error")
)
func modelOptions(model *Model, requestOpts map[string]any) (api.Options, error) {
func (s *Server) modelOptions(model *Model, requestOpts map[string]any) (api.Options, error) {
opts := api.DefaultOptions()
if opts.NumCtx == 0 {
opts.NumCtx = s.defaultNumCtx
}
if err := opts.FromMap(model.Options); err != nil {
return api.Options{}, err
}
@@ -140,20 +140,11 @@ func (s *Server) scheduleRunner(ctx context.Context, name string, caps []model.C
return nil, nil, nil, fmt.Errorf("%s %w", name, err)
}
opts, err := modelOptions(model, requestOpts)
opts, err := s.modelOptions(model, requestOpts)
if err != nil {
return nil, nil, nil, err
}
// This model is much more capable with a larger context, so set that
// unless it would penalize performance too much
if !s.lowVRAM && slices.Contains([]string{
"gptoss", "gpt-oss",
"qwen3vl", "qwen3vlmoe",
}, model.Config.ModelFamily) {
opts.NumCtx = max(opts.NumCtx, 8192)
}
runnerCh, errCh := s.sched.GetRunner(ctx, model, opts, keepAlive)
var runner *runnerRef
select {
@@ -1720,10 +1711,18 @@ func Serve(ln net.Listener) error {
for _, gpu := range gpus {
totalVRAM += gpu.TotalMemory - envconfig.GpuOverhead()
}
if totalVRAM < lowVRAMThreshold {
s.lowVRAM = true
slog.Info("entering low vram mode", "total vram", format.HumanBytes2(totalVRAM), "threshold", format.HumanBytes2(lowVRAMThreshold))
// Set default context based on VRAM tier
// Use slightly lower thresholds (47/23 GiB vs. 48/24 GiB) to account for small differences in the exact value
switch {
case totalVRAM >= 47*format.GibiByte:
s.defaultNumCtx = 262144
case totalVRAM >= 23*format.GibiByte:
s.defaultNumCtx = 32768
default:
s.defaultNumCtx = 4096
}
slog.Info("vram-based default context", "total_vram", format.HumanBytes2(totalVRAM), "default_num_ctx", s.defaultNumCtx)
err = srvr.Serve(ln)
// If server is closed from the signal handler, wait for the ctx to be done
@@ -1897,8 +1896,8 @@ func (s *Server) PsHandler(c *gin.Context) {
Details: modelDetails,
ExpiresAt: v.expiresAt,
}
if v.Options != nil {
mr.ContextLength = v.Options.NumCtx
if v.llama != nil {
mr.ContextLength = v.llama.ContextLength()
}
// The scheduler waits to set expiresAt, so if a model is loading it's
// possible that it will be set to the unix epoch. For those cases, just

View File

@@ -15,6 +15,7 @@ import (
)
func TestGenerateDebugRenderOnly(t *testing.T) {
t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
gin.SetMode(gin.TestMode)
mock := mockRunner{
@@ -208,6 +209,7 @@ func TestGenerateDebugRenderOnly(t *testing.T) {
}
func TestChatDebugRenderOnly(t *testing.T) {
t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
gin.SetMode(gin.TestMode)
mock := mockRunner{

View File

@@ -20,6 +20,7 @@ import (
// TestGenerateWithBuiltinRenderer tests that api/generate uses built-in renderers
// when in chat-like flow (messages present, no suffix, no template)
func TestGenerateWithBuiltinRenderer(t *testing.T) {
t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
gin.SetMode(gin.TestMode)
mock := mockRunner{
@@ -204,6 +205,7 @@ func TestGenerateWithBuiltinRenderer(t *testing.T) {
// TestGenerateWithDebugRenderOnly tests that debug_render_only works with built-in renderers
func TestGenerateWithDebugRenderOnly(t *testing.T) {
t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
gin.SetMode(gin.TestMode)
mock := mockRunner{

View File

@@ -162,6 +162,7 @@ func TestGenerateChatRemote(t *testing.T) {
}
func TestGenerateChat(t *testing.T) {
t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
gin.SetMode(gin.TestMode)
mock := mockRunner{
@@ -878,6 +879,7 @@ func TestGenerateChat(t *testing.T) {
}
func TestGenerate(t *testing.T) {
t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
gin.SetMode(gin.TestMode)
mock := mockRunner{
@@ -2355,6 +2357,7 @@ func TestGenerateWithImages(t *testing.T) {
// TestImageGenerateStreamFalse tests that image generation respects stream=false
// and returns a single JSON response instead of streaming ndjson.
func TestImageGenerateStreamFalse(t *testing.T) {
t.Setenv("OLLAMA_CONTEXT_LENGTH", "4096")
gin.SetMode(gin.TestMode)
p := t.TempDir()

View File

@@ -0,0 +1,127 @@
package server
import (
"testing"
)
func TestModelOptionsNumCtxPriority(t *testing.T) {
tests := []struct {
name string
envContextLen string // empty means not set (uses 0 sentinel)
defaultNumCtx int // VRAM-based default
modelNumCtx int // 0 means not set in model
requestNumCtx int // 0 means not set in request
expectedNumCtx int
}{
{
name: "vram default when nothing else set",
envContextLen: "",
defaultNumCtx: 32768,
modelNumCtx: 0,
requestNumCtx: 0,
expectedNumCtx: 32768,
},
{
name: "env var overrides vram default",
envContextLen: "8192",
defaultNumCtx: 32768,
modelNumCtx: 0,
requestNumCtx: 0,
expectedNumCtx: 8192,
},
{
name: "model overrides vram default",
envContextLen: "",
defaultNumCtx: 32768,
modelNumCtx: 16384,
requestNumCtx: 0,
expectedNumCtx: 16384,
},
{
name: "model overrides env var",
envContextLen: "8192",
defaultNumCtx: 32768,
modelNumCtx: 16384,
requestNumCtx: 0,
expectedNumCtx: 16384,
},
{
name: "request overrides everything",
envContextLen: "8192",
defaultNumCtx: 32768,
modelNumCtx: 16384,
requestNumCtx: 4096,
expectedNumCtx: 4096,
},
{
name: "request overrides vram default",
envContextLen: "",
defaultNumCtx: 32768,
modelNumCtx: 0,
requestNumCtx: 4096,
expectedNumCtx: 4096,
},
{
name: "request overrides model",
envContextLen: "",
defaultNumCtx: 32768,
modelNumCtx: 16384,
requestNumCtx: 4096,
expectedNumCtx: 4096,
},
{
name: "low vram tier default",
envContextLen: "",
defaultNumCtx: 4096,
modelNumCtx: 0,
requestNumCtx: 0,
expectedNumCtx: 4096,
},
{
name: "high vram tier default",
envContextLen: "",
defaultNumCtx: 262144,
modelNumCtx: 0,
requestNumCtx: 0,
expectedNumCtx: 262144,
},
}
for _, tt := range tests {
t.Run(tt.name, func(t *testing.T) {
// Set or clear environment variable
if tt.envContextLen != "" {
t.Setenv("OLLAMA_CONTEXT_LENGTH", tt.envContextLen)
}
// Create server with VRAM-based default
s := &Server{
defaultNumCtx: tt.defaultNumCtx,
}
// Create model options (use float64 as FromMap expects JSON-style numbers)
var modelOpts map[string]any
if tt.modelNumCtx != 0 {
modelOpts = map[string]any{"num_ctx": float64(tt.modelNumCtx)}
}
model := &Model{
Options: modelOpts,
}
// Create request options (use float64 as FromMap expects JSON-style numbers)
var requestOpts map[string]any
if tt.requestNumCtx != 0 {
requestOpts = map[string]any{"num_ctx": float64(tt.requestNumCtx)}
}
opts, err := s.modelOptions(model, requestOpts)
if err != nil {
t.Fatalf("modelOptions failed: %v", err)
}
if opts.NumCtx != tt.expectedNumCtx {
t.Errorf("NumCtx = %d, want %d", opts.NumCtx, tt.expectedNumCtx)
}
})
}
}

View File

@@ -804,6 +804,7 @@ func (s *mockLlm) GetPort() int { return -
func (s *mockLlm) GetDeviceInfos(ctx context.Context) []ml.DeviceInfo { return nil }
func (s *mockLlm) HasExited() bool { return false }
func (s *mockLlm) GetActiveDeviceIDs() []ml.DeviceID { return nil }
func (s *mockLlm) ContextLength() int { return 0 }
// TestImageGenRunnerCanBeEvicted verifies that an image generation model
// loaded in the scheduler can be evicted when idle.

View File

@@ -347,6 +347,11 @@ func (s *Server) VRAMByGPU(id ml.DeviceID) uint64 {
return s.vramSize
}
// Context length is not applicable for image generation.
func (s *Server) ContextLength() int {
return 0
}
func (s *Server) Embedding(ctx context.Context, input string) ([]float32, int, error) {
return nil, 0, errors.New("not supported")
}