feat(llm): add Ollama provider (local/self-hosted) (#22)

internal/llm/ollama: OllamaProvider implements llm.Provider via
github.com/ollama/ollama/api, with JSON-schema structured output (chat Format)
matching the other adapters. Local/self-hosted, so EXEMPT from the allow_external
BAA gate (no third-party disclosure) — the strongest data-control option; the
pipeline sanitizer gate still applies. No API key; host defaults to
http://localhost:11434. Selectable as primary or router fallback.

Review fixes: surface truncated (done_reason=length) and empty structured-output
responses as terminal errors so the fallback router advances (parity with
openai/anthropic). config.yaml documents that the ollama host MUST stay within
the HIPAA control boundary (a third-party 'Ollama-compatible' endpoint would
reintroduce external disclosure without the gate).

Verified with ADC off (hermetic httptest): go build/vet clean, golangci-lint 0
issues, go test -race green, internal coverage 90.0%, linux static cross-compile OK.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: avivl

cmd/sre-agent/main.go

@@ -27,6 +27,7 @@ import (
 	"github.com/avivl/cloud-sre-agent/internal/llm"
 	"github.com/avivl/cloud-sre-agent/internal/llm/anthropic"
 	"github.com/avivl/cloud-sre-agent/internal/llm/gemini"
+	"github.com/avivl/cloud-sre-agent/internal/llm/ollama"
 	"github.com/avivl/cloud-sre-agent/internal/llm/openai"
 	"github.com/avivl/cloud-sre-agent/internal/llm/router"
 	"github.com/avivl/cloud-sre-agent/internal/obs"
@@ -228,6 +229,13 @@ func buildOneProvider(ctx context.Context, e config.ProviderConfig) (llm.Provide
 			APIKey:  apiKey,
 			BaseURL: e.BaseURL,
 		})
+	case config.KindOllama:
+		// Ollama is local/self-hosted: no API key, host defaults inside the
+		// adapter when empty.
+		return ollama.New(ollama.Config{
+			Model: e.Model,
+			Host:  e.Host,
+		})
 	default:
 		return nil, fmt.Errorf("unsupported llm provider kind %q", e.Kind)
 	}

config.yaml

@@ -24,6 +24,14 @@ sources:
 #               consumer Developer API (gated, see allow_non_baa).
 #   openai    — OpenAI (external third party; gated, see allow_external).
 #   anthropic — Anthropic (external third party; gated, see allow_external).
+#   ollama    — local/self-hosted model server. NOT an external third party:
+#               prompt content stays on infrastructure you control, so it is
+#               EXEMPT from the allow_external gate and needs no API key. Set a
+#               `host` (defaults to http://localhost:11434).
+#               IMPORTANT: the `host` MUST resolve to infrastructure within your
+#               HIPAA control boundary. Pointing it at a third-party
+#               "Ollama-compatible" endpoint reintroduces external disclosure
+#               WITHOUT the allow_external gate — the code cannot detect this.
 #
 # HIPAA gate: openai and anthropic are EXTERNAL third parties NOT covered by
 # Google's BAA. Selecting either (as primary or fallback) is refused at startup
@@ -74,6 +82,9 @@ llm:
   #     # base_url: https://my-compatible-gateway.example.com   # optional host override
   #   - kind: anthropic
   #     model: claude-opus-4-8
+  #   - kind: ollama          # local/self-hosted; no allow_external, no API key
+  #     model: llama3.1
+  #     host: http://localhost:11434   # optional; this is the default
 
 output:
   dir: ./out

go.mod

@@ -14,6 +14,7 @@ require (
 	github.com/knadh/koanf/providers/env v1.1.0
 	github.com/knadh/koanf/providers/file v1.2.1
 	github.com/knadh/koanf/v2 v2.3.5
+	github.com/ollama/ollama v0.30.10
 	github.com/openai/openai-go/v3 v3.41.0
 	github.com/spf13/cobra v1.10.2
 	github.com/stretchr/testify v1.11.1
@@ -57,6 +58,7 @@ require (
 	github.com/hashicorp/go-retryablehttp v0.7.8 // indirect
 	github.com/inconshreveable/mousetrap v1.1.0 // indirect
 	github.com/knadh/koanf/maps v0.1.2 // indirect
+	github.com/mailru/easyjson v0.7.7 // indirect
 	github.com/mitchellh/copystructure v1.2.0 // indirect
 	github.com/mitchellh/reflectwalk v1.0.2 // indirect
 	github.com/pb33f/ordered-map/v2 v2.3.1 // indirect
@@ -67,6 +69,7 @@ require (
 	github.com/tidwall/match v1.1.1 // indirect
 	github.com/tidwall/pretty v1.2.1 // indirect
 	github.com/tidwall/sjson v1.2.5 // indirect
+	github.com/wk8/go-ordered-map/v2 v2.1.8 // indirect
 	go.einride.tech/aip v0.83.0 // indirect
 	go.opencensus.io v0.24.0 // indirect
 	go.opentelemetry.io/auto/sdk v1.2.1 // indirect

go.sum

@@ -123,6 +123,7 @@ github.com/influxdata/tdigest v0.0.1 h1:XpFptwYmnEKUqmkcDjrzffswZ3nvNeevbUSLPP/Z
 github.com/influxdata/tdigest v0.0.1/go.mod h1:Z0kXnxzbTC2qrx4NaIzYkE1k66+6oEDQTvL95hQFh5Y=
 github.com/invopop/jsonschema v0.14.0 h1:MHQqLhvpNUZfw+hM3AZDYK7jxO8FZoQeQM77g8iyZjg=
 github.com/invopop/jsonschema v0.14.0/go.mod h1:ygm6C2EaVNMBDPpaPlnOA2pFAxBnxGjFlMZABxm9n2I=
+github.com/josharian/intern v1.0.0/go.mod h1:5DoeVV0s6jJacbCEi61lwdGj/aVlrQvzHFFd8Hwg//Y=
 github.com/knadh/koanf/maps v0.1.2 h1:RBfmAW5CnZT+PJ1CVc1QSJKf4Xu9kxfQgYVQSu8hpbo=
 github.com/knadh/koanf/maps v0.1.2/go.mod h1:npD/QZY3V6ghQDdcQzl1W4ICNVTkohC8E73eI2xW4yI=
 github.com/knadh/koanf/parsers/yaml v1.1.0 h1:3ltfm9ljprAHt4jxgeYLlFPmUaunuCgu1yILuTXRdM4=
@@ -139,6 +140,8 @@ github.com/kr/pretty v0.3.1 h1:flRD4NNwYAUpkphVc1HcthR4KEIFJ65n8Mw5qdRn3LE=
 github.com/kr/pretty v0.3.1/go.mod h1:hoEshYVHaxMs3cyo3Yncou5ZscifuDolrwPKZanG3xk=
 github.com/kr/text v0.2.0 h1:5Nx0Ya0ZqY2ygV366QzturHI13Jq95ApcVaJBhpS+AY=
 github.com/kr/text v0.2.0/go.mod h1:eLer722TekiGuMkidMxC/pM04lWEeraHUUmBw8l2grE=
+github.com/mailru/easyjson v0.7.7 h1:UGYAvKxe3sBsEDzO8ZeWOSlIQfWFlxbzLZe7hwFURr0=
+github.com/mailru/easyjson v0.7.7/go.mod h1:xzfreul335JAWq5oZzymOObrkdz5UnU4kGfJJLY9Nlc=
 github.com/mattn/go-colorable v0.1.13 h1:fFA4WZxdEF4tXPZVKMLwD8oUnCTTo08duU7wxecdEvA=
 github.com/mattn/go-colorable v0.1.13/go.mod h1:7S9/ev0klgBDR4GtXTXX8a3vIGJpMovkB8vQcUbaXHg=
 github.com/mattn/go-isatty v0.0.20 h1:xfD0iDuEKnDkl03q4limB+vH+GxLEtL/jb4xVJSWWEY=
@@ -147,6 +150,8 @@ github.com/mitchellh/copystructure v1.2.0 h1:vpKXTN4ewci03Vljg/q9QvCGUDttBOGBIa1
 github.com/mitchellh/copystructure v1.2.0/go.mod h1:qLl+cE2AmVv+CoeAwDPye/v+N2HKCj9FbZEVFJRxO9s=
 github.com/mitchellh/reflectwalk v1.0.2 h1:G2LzWKi524PWgd3mLHV8Y5k7s6XUvT0Gef6zxSIeXaQ=
 github.com/mitchellh/reflectwalk v1.0.2/go.mod h1:mSTlrgnPZtwu0c4WaC2kGObEpuNDbx0jmZXqmk4esnw=
+github.com/ollama/ollama v0.30.10 h1:xeW6KBUqB2g88jSHBoxFiaP1gsLbRIB7LcjTreAFHn8=
+github.com/ollama/ollama v0.30.10/go.mod h1:TjwyryJftKpcf7ByoIuZWso/Wx2Jr2AcGubxadv13dY=
 github.com/openai/openai-go/v3 v3.41.0 h1:9GkxcN02U5NG0WGdQjZ0cTSu/pMXEyzL2LfF0ruZCck=
 github.com/openai/openai-go/v3 v3.41.0/go.mod h1:cdufnVK14cWcT9qA1rRtrXx4FTRsgbDPW7Ia7SS5cZo=
 github.com/pb33f/ordered-map/v2 v2.3.1 h1:5319HDO0aw4DA4gzi+zv4FXU9UlSs3xGZ40wcP1nBjY=
@@ -183,6 +188,8 @@ github.com/tidwall/pretty v1.2.1 h1:qjsOFOWWQl+N3RsoF5/ssm1pHmJJwhjlSbZ51I6wMl4=
 github.com/tidwall/pretty v1.2.1/go.mod h1:ITEVvHYasfjBbM0u2Pg8T2nJnzm8xPwvNhhsoaGGjNU=
 github.com/tidwall/sjson v1.2.5 h1:kLy8mja+1c9jlljvWTlSazM7cKDRfJuR/bOJhcY5NcY=
 github.com/tidwall/sjson v1.2.5/go.mod h1:Fvgq9kS/6ociJEDnK0Fk1cpYF4FIW6ZF7LAe+6jwd28=
+github.com/wk8/go-ordered-map/v2 v2.1.8 h1:5h/BUHu93oj4gIdvHHHGsScSTMijfx5PeYkE/fJgbpc=
+github.com/wk8/go-ordered-map/v2 v2.1.8/go.mod h1:5nJHM5DyteebpVlHnWMV0rPz6Zp7+xBAnxjb1X5vnTw=
 gitlab.com/gitlab-org/api/client-go/v2 v2.43.0 h1:CxvWrDmW6/NAmnFeC4if5SGSP/0X54RIYox4sMo7pH0=
 gitlab.com/gitlab-org/api/client-go/v2 v2.43.0/go.mod h1:pTbeBowtVA+0/ZExWEZYUGOrpu5qlRN5ZyOUf27BnVY=
 go.einride.tech/aip v0.83.0 h1:TI21IdeOnLTwZEJ3BxtImIZk6bsN2Q+sd0x99SLiQ+M=
@@ -243,6 +250,8 @@ golang.org/x/sys v0.0.0-20190412213103-97732733099d/go.mod h1:h1NjWce9XRLGQEsW7w
 golang.org/x/sys v0.0.0-20200930185726-fdedc70b468f/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
 golang.org/x/sys v0.45.0 h1:dO4czNzziLiiXplLQgBCEpCvXQ3dnkn0SdaZSYdQ+FY=
 golang.org/x/sys v0.45.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+golang.org/x/term v0.43.0 h1:S4RLU2sB31O/NCl+zFN9Aru9A/Cq2aqKpTZJ6B+DwT4=
+golang.org/x/term v0.43.0/go.mod h1:lrhlHNdQJHO+1qVYiHfFKVuVioJIheAc3fBSMFYEIsk=
 golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=
 golang.org/x/text v0.3.3/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.37.0 h1:Cqjiwd9eSg8e0QAkyCaQTNHFIIzWtidPahFWR83rTrc=

internal/config/config.go

@@ -65,13 +65,19 @@ type TracingConfig struct {
 }
 
 // LLM provider kinds. "gemini" is Google's Gemini (via Vertex AI or the
-// Developer API); "openai" and "anthropic" are external third-party services.
+// Developer API); "openai" and "anthropic" are external third-party services;
+// "ollama" is a local/self-hosted model server (no external disclosure).
 const (
 	KindGemini    = "gemini"
 	KindOpenAI    = "openai"
 	KindAnthropic = "anthropic"
+	KindOllama    = "ollama"
 )
 
+// DefaultOllamaHost is the Ollama daemon address used when an ollama provider's
+// host is left empty.
+const DefaultOllamaHost = "http://localhost:11434"
+
 // LLM backend identifiers (gemini kind only). Vertex AI is the only BAA-eligible
 // Google backend and is the default; the consumer Gemini Developer API is not
 // covered by Google's BAA and is gated behind an explicit opt-in.
@@ -96,7 +102,7 @@ const AllowExternalLLMEnv = "SRE_ALLOW_EXTERNAL_LLM"
 // the gemini kind, BaseURL applies to openai/anthropic. API keys are never
 // stored here — they are read from the environment at wire time.
 type ProviderConfig struct {
-	// Kind selects the adapter: "gemini", "openai", or "anthropic".
+	// Kind selects the adapter: "gemini", "openai", "anthropic", or "ollama".
 	Kind string `koanf:"kind"`
 	// Model is the model name (e.g. "gemini-2.5-flash", "gpt-4o-mini",
 	// "claude-opus-4-8"). Required.
@@ -124,6 +130,13 @@ type ProviderConfig struct {
 	// BaseURL optionally overrides the provider API host (e.g. a compatible
 	// gateway). Empty uses the SDK default. Applies to openai/anthropic.
 	BaseURL string `koanf:"base_url"`
+
+	// --- ollama fields ---
+
+	// Host is the Ollama daemon base URL (e.g. "http://localhost:11434"). Empty
+	// defaults to DefaultOllamaHost. Applies to the ollama kind only; Ollama is
+	// local/self-hosted and uses no API key.
+	Host string `koanf:"host"`
 }
 
 // LLMConfig selects the primary LLM provider plus an ordered list of fallbacks.
@@ -154,6 +167,9 @@ type LLMConfig struct {
 	// BaseURL optionally overrides the primary provider API host when Provider is
 	// openai/anthropic. Ignored for gemini.
 	BaseURL string `koanf:"base_url"`
+	// Host is the Ollama daemon base URL when Provider is ollama. Empty defaults
+	// to DefaultOllamaHost. Ignored for other kinds.
+	Host string `koanf:"host"`
 
 	// Fallbacks is the ordered list of fallback providers, tried in turn when the
 	// primary (and preceding fallbacks) fail.
@@ -178,6 +194,7 @@ func (l LLMConfig) Primary() ProviderConfig {
 		APIKeyEnv:   l.APIKeyEnv,
 		AllowNonBAA: l.AllowNonBAA,
 		BaseURL:     l.BaseURL,
+		Host:        l.Host,
 	}
 }
 
@@ -471,6 +488,12 @@ func (l LLMConfig) validateProvider(where string, p ProviderConfig) error {
 	switch p.Kind {
 	case KindGemini:
 		return l.validateGeminiBackend(where, p)
+	case KindOllama:
+		// Ollama is local/self-hosted: prompt content stays on infrastructure
+		// the operator controls, so it is EXEMPT from the external-disclosure
+		// (BAA) gate. The model is already validated above; host defaults to
+		// DefaultOllamaHost when empty, so there is nothing further to check.
+		return nil
 	case KindOpenAI, KindAnthropic:
 		if !l.AllowsExternal() {
 			return fmt.Errorf("config: %s: provider %q is an external third-party service; prompt content "+
@@ -482,7 +505,7 @@ func (l LLMConfig) validateProvider(where string, p ProviderConfig) error {
 		}
 		return nil
 	default:
-		return fmt.Errorf("config: %s: provider kind %q must be %q, %q, or %q", where, p.Kind, KindGemini, KindOpenAI, KindAnthropic)
+		return fmt.Errorf("config: %s: provider kind %q must be %q, %q, %q, or %q", where, p.Kind, KindGemini, KindOpenAI, KindAnthropic, KindOllama)
 	}
 }
 

internal/config/config_test.go

@@ -228,6 +228,81 @@ func TestValidate_ExternalProviders(t *testing.T) {
 	})
 }
 
+// TestValidate_Ollama covers the local/self-hosted ollama provider, which is
+// EXEMPT from the external-disclosure (BAA) gate: it must be selectable as
+// primary and as a fallback WITHOUT allow_external.
+func TestValidate_Ollama(t *testing.T) {
+	base := func() Config {
+		return Config{
+			Sources:   []SourceConfig{{Type: "file", Path: "x.log"}},
+			Output:    OutputConfig{Dir: "./out"},
+			Target:    TargetLocal,
+			Validator: ValidatorNone,
+			Log:       LogConfig{Format: "json"},
+		}
+	}
+
+	t.Run("ollama as primary without opt-in is allowed", func(t *testing.T) {
+		c := base()
+		c.LLM = LLMConfig{Provider: KindOllama, Model: "llama3.1"}
+		require.NoError(t, c.Validate())
+	})
+
+	t.Run("ollama as primary with explicit host is allowed", func(t *testing.T) {
+		c := base()
+		c.LLM = LLMConfig{Provider: KindOllama, Model: "llama3.1", Host: "http://ollama.internal:11434"}
+		require.NoError(t, c.Validate())
+	})
+
+	t.Run("ollama as fallback without opt-in is allowed", func(t *testing.T) {
+		c := base()
+		c.LLM = LLMConfig{
+			Provider: KindGemini, Model: "gemini-2.5-flash",
+			Backend: BackendVertex, Project: "p", Location: "us-central1",
+			Fallbacks: []ProviderConfig{{Kind: KindOllama, Model: "llama3.1"}},
+		}
+		// No AllowExternal set: ollama is exempt from the external gate.
+		require.False(t, c.LLM.AllowsExternal())
+		require.NoError(t, c.Validate())
+	})
+
+	t.Run("ollama missing model fails", func(t *testing.T) {
+		c := base()
+		c.LLM = LLMConfig{Provider: KindOllama, Model: ""}
+		err := c.Validate()
+		require.Error(t, err)
+		require.Contains(t, err.Error(), "model is required")
+	})
+}
+
+// TestLoad_OllamaProvider loads an ollama config from YAML to confirm the host
+// field round-trips and the provider is accepted without allow_external.
+func TestLoad_OllamaProvider(t *testing.T) {
+	p := writeConfig(t, `
+sources:
+  - type: file
+    path: ./x.log
+llm:
+  provider: ollama
+  model: llama3.1
+  host: http://localhost:11434
+  fallbacks:
+    - kind: ollama
+      model: llama3.2
+      host: http://ollama.internal:11434
+`)
+	cfg, err := Load(p)
+	require.NoError(t, err)
+	assert.Equal(t, KindOllama, cfg.LLM.Provider)
+	assert.Equal(t, "http://localhost:11434", cfg.LLM.Host)
+	require.Len(t, cfg.LLM.Fallbacks, 1)
+	assert.Equal(t, KindOllama, cfg.LLM.Fallbacks[0].Kind)
+	assert.Equal(t, "http://ollama.internal:11434", cfg.LLM.Fallbacks[0].Host)
+
+	// Primary projection carries the host through.
+	assert.Equal(t, "http://localhost:11434", cfg.LLM.Primary().Host)
+}
+
 func TestLoad_PrimaryAndFallbacks(t *testing.T) {
 	p := writeConfig(t, `
 sources: