feat(llm): add config-driven LLM provider package [AG-000]

[ShawkyZ]

What

Adds pkg/llm, a vendor-agnostic LLM provider package (built on langchaingo), so LLM access can be shared across Snyk tools instead of living in the remy extension.

• Contract: Provider interface + ChatRequest/ChatResponse/Message/ToolCall/… types.
• Adapters: anthropic, openai, litellm, ollama, vertex (Gemini via genai + Claude via anthropic-sdk-go), bedrock. Retry/backoff, temperature body-strip, base-URL normalization all preserved.
• Registry (config.go + registry.go): Resolve / New / NewFromConfig read provider selection, API keys, base URLs, and gateway headers from a configuration.Configuration.
• Config: named env-var constants + canonical snyk_llm_* keys wired via SetSupportedEnvVars + AddAlternativeKeys. Existing env vars keep working (ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_CLOUD_*, AWS_REGION, LITELLM_*). SNYK_LLM_EXTRA_HEADERS is the new primary name; REMY_EXTRA_HEADERS retained as an alternative key. API keys are read on demand, never persisted to storage.
• Networking: optional WithNetworkAccess threads the unauthorized HTTP transport (proxy/CA/FIPS, no Snyk auth headers) under provider clients.
• De-branding: WithBedrockAppID sets the AWS User-Agent app id; the package ships unbranded (the SNYK_AGENTIC_FIX id is supplied by the caller).
• Generated pkg/mocks/llm.go; /pkg/llm/ assigned to @snyk/agentic-code-security in CODEOWNERS.

Ports the full provider test suite and adds a config-driven registry test + a networking-compose test. go build/vet/test ./pkg/llm/... clean; go mod tidy stable.

⚠️ Dependency weight — needs owner sign-off

This pulls langchaingo, aws-sdk-go-v2/{config,bedrockruntime}, google.golang.org/genai, anthropics/anthropic-sdk-go, and golang.org/x/oauth2/google into GAF's graph — inherited by snyk/cli and every GAF consumer. Please confirm this is acceptable, or whether pkg/llm should be isolated (separate go.mod / build tags) before merge.

Consumer

The remy extension is migrating off its internal/provider to this package (companion PR in snyk/remy-cli-extension). CODEOWNERS routes /pkg/llm/ review to Agentic Code Security.

🤖 Generated with Claude Code

---

Note

High Risk
Large new surface area with third-party LLM credentials, gateway auth stripping, and heavy dependencies inherited by all GAF consumers; misconfiguration could leak keys or send code to the wrong endpoint.

Overview
Introduces pkg/llm, a shared LLM layer for GAF: a Provider contract (ChatCompletion, tools, normalized stop reasons) plus Resolve / New / NewFromConfig that read configuration.Configuration and familiar env vars (ANTHROPIC_*, OPENAI_*, GOOGLE_CLOUD_*, LITELLM_*, AWS_REGION, optional SNYK_LLM_EXTRA_HEADERS with legacy REMY_EXTRA_HEADERS).

Backends are wired through langchaingo and small custom llms.Model shims: Anthropic, OpenAI, LiteLLM (OpenAI-compatible proxy), Ollama (/api/chat with tools), Vertex (Gemini via genai, Claude on Model Garden via anthropic-sdk-go), and Bedrock. Shared behavior includes gateway header injection, placeholder keys stripped on the wire when unset, OpenAI/LiteLLM HTTPS base-URL normalization, temperature stripped from JSON bodies, per-call timeouts, and 429/529 backoff. Optional WithNetworkAccess uses the unauthorized HTTP transport (proxy/CA without Snyk auth); WithBedrockAppID brands Bedrock traffic.

Also adds FakeProvider, generated pkg/mocks/llm.go, CODEOWNERS for Agentic Code Security, and new direct deps (langchaingo, AWS Bedrock SDK, google.golang.org/genai, anthropic-sdk-go, etc.) on the root module.

^{Reviewed by Cursor Bugbot for commit ebf453a. Bugbot is set up for automated code reviews on this repo. Configure here.}

Usage

Runnable godoc examples are in pkg/llm/example_test.go. In short:

import (
    "github.com/snyk/go-application-framework/pkg/configuration"
    "github.com/snyk/go-application-framework/pkg/llm"
)

// One-shot: resolve the provider from config (env vars / flags / Set) and build it.
config := configuration.New() // in a workflow: ictx.GetConfiguration()
provider, err := llm.NewFromConfig(config)
if err != nil { /* no provider configured */ }

resp, err := provider.ChatCompletion(ctx, &llm.ChatRequest{
    SystemPrompt: "You are a helpful assistant.",
    MaxTokens:    1024,
    Messages:     []llm.Message{{Role: llm.RoleUser, Content: "Say hello."}},
})
_ = resp.Content

Inside a GAF workflow — thread the shared transport and (for Bedrock) the app id:

provider, err := llm.NewFromConfig(ictx.GetConfiguration(),
    llm.WithNetworkAccess(ictx.GetNetworkAccess()), // proxy/CA/FIPS, no Snyk auth headers
    llm.WithLogger(ictx.GetEnhancedLogger()),
    llm.WithBedrockAppID("MY_APP"),                 // AWS User-Agent app/<id>
)

Resolve then build (when you need the provider name/model for telemetry first):

res, err := llm.Resolve(config)                 // res.Provider, res.Model
provider, err := llm.New(config, res, llm.WithNetworkAccess(na))

Tool calling — pass Tools, then echo results back as ToolResult:

resp, _ := provider.ChatCompletion(ctx, &llm.ChatRequest{
    Model:    "gpt-4o",
    Messages: []llm.Message{{Role: llm.RoleUser, Content: "Scan ./api."}},
    Tools: []llm.ToolDefinition{{
        Name: "scan", Description: "Run a Snyk scan on a path.",
        InputSchema: []byte(`{"type":"object","properties":{"path":{"type":"string"}},"required":["path"]}`),
    }},
})
if resp.StopReason == llm.StopToolUse {
    for _, call := range resp.ToolCalls { /* run call.Name(call.Input); reply with llm.ToolResult{ToolCallID: call.ID, ...} */ }
}

Provider selection — explicit via config.Set(llm.CONFIG_PROVIDER, "anthropic") (or --provider); otherwise auto-detected from ANTHROPIC_API_KEY (or ANTHROPIC_BASE_URL) then OPENAI_API_KEY. Other providers read GOOGLE_CLOUD_PROJECT/GOOGLE_CLOUD_LOCATION, AWS_REGION, LITELLM_BASE_URL/LITELLM_API_KEY, and gateway headers from SNYK_LLM_EXTRA_HEADERS.

[snyk-io]

⛔ Snyk checks have failed. 5 issues have been found so far.

Status	Scan Engine	Critical	High	Medium	Low	Total (5)
⛔	Open Source Security	1	4	0	0	5 issues
✅	Licenses	0	0	0	0	0 issues
✅	Code Security	0	0	0	0	0 issues
✅	Secrets	0	0	0	0	0 issues

💻 Catch issues earlier using the plugins for VS Code, JetBrains IDEs, Visual Studio, and Eclipse.

[snyk-io]

⛔ Snyk checks have failed. 5 issues have been found so far.

Status	Scan Engine	Critical	High	Medium	Low	Total (5)
⛔	Open Source Security	1	4	0	0	5 issues
✅	Licenses	0	0	0	0	0 issues
✅	Code Security	0	0	0	0	0 issues

💻 Catch issues earlier using the plugins for VS Code, JetBrains IDEs, Visual Studio, and Eclipse.

[snyk-pr-review-bot]

PR Reviewer Guide 🔍

🧪 PR contains tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Brittle Body Manipulation 🟡 [minor] The stripJSONBodyFields function performs a full Unmarshal into map[string]json.RawMessage to remove the temperature field. While safe for current small chat payloads, this bypasses langchaingo's internal state management. If future SDK versions start using streaming or multi-part bodies, this logic will silently pass through the unparseable body, potentially re-introducing the 'temperature is deprecated' error from providers. func stripJSONBodyFields(req *http.Request, fields []string) error { raw, err := io.ReadAll(req.Body) _ = req.Body.Close() if err != nil { return err } out := raw var m map[string]json.RawMessage if json.Unmarshal(raw, &m) == nil { changed := false for _, f := range fields { if _, ok := m[f]; ok { delete(m, f) changed = true } } if changed { if b, mErr := json.Marshal(m); mErr == nil { out = b } } } req.Body = io.NopCloser(bytes.NewReader(out)) req.ContentLength = int64(len(out)) req.GetBody = func() (io.ReadCloser, error) { return io.NopCloser(bytes.NewReader(out)), nil } return nil } Implicit Provider Priority 🟡 [minor] The Resolve function defaults to anthropic if both ANTHROPIC_API_KEY and OPENAI_API_KEY are present. While documented, this implicit priority can surprise users who set multiple keys for different tools and expect GAF to require explicit selection when ambiguous. for _, n := range []string{"anthropic", "openai"} { def := providers[n] if !def.autoDetect { continue } if def.apiKeyKey != "" && config.GetString(def.apiKeyKey) != "" { return Resolution{Provider: n, Model: modelFor(config, def)}, nil } // A configured base URL is enough to pick the provider even without a // key — the gateway it points at may supply auth. if def.baseURLKey != "" && config.GetString(def.baseURLKey) != "" { return Resolution{Provider: n, Model: modelFor(config, def)}, nil } }

📚 Repository Context Analyzed This review considered 63 relevant code sections from 15 files (average relevance: 1.00) pkg/mocks/workflow.go (lines 1-238) pkg/mocks/configuration.go (lines 1-224) internal/mocks/api.go (lines 1-170) pkg/apiclients/feature_flag_gateway/mocks/feature_flag_gateway.go (lines 1-194) pkg/apiclients/ldx_sync_config/mocks/ldx_sync.go (lines 1-180) pkg/apiclients/ldx_sync_config/ldx_sync/2026-05-07/ldx_sync.go (lines 1-230) internal/api/fileupload/uploadrevision/fake_client.go (lines 1-147) ... and 8 more file(s)

🤖 Repository instructions applied (from AGENTS.md)

[cursor]

Cursor Bugbot has reviewed your changes and found 4 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit ebf453a. Configure here.}

[cursor]

Vertex cross-family model override

High Severity

vertexModel builds either a Gemini (genai) or Claude (anthropic-sdk-go) client from the default model at construction, but GenerateContent picks the backend from each call’s WithModel override. A per-request model in the other family hits generateGemini or generateClaude with a nil client and fails, even though the type documents per-call model overrides.

 

^{Reviewed by Cursor Bugbot for commit ebf453a. Configure here.}

[cursor]

Tool error flag ignored

Medium Severity

ToolResult exposes IsError, but mapping into provider messages never uses it: toLCMessage only copies ToolCallID and Content, and the Vertex Claude path always passes false into NewToolResultBlock. Failed tool runs are sent as successful results to the model.

Additional Locations (1)

• pkg/llm/vertex_model.go#L495-L499

 

^{Reviewed by Cursor Bugbot for commit ebf453a. Configure here.}

[cursor]

Bedrock skips temperature strip

Medium Severity

Anthropic and OpenAI adapters strip langchaingo’s hardcoded temperature field via headerTransport, but NewBedrockAdapter wraps langchaingo Bedrock with no equivalent HTTP middleware. Claude-on-Bedrock requests can still send deprecated temperature:0 and be rejected by the model.

 

^{Reviewed by Cursor Bugbot for commit ebf453a. Configure here.}

[cursor]

Bedrock ignores network transport

Medium Severity

WithNetworkAccess threads the unauthorized proxy/CA transport into other providers, but the Bedrock registry build function and NewBedrockAdapter never accept or apply that RoundTripper. Bedrock traffic uses the AWS SDK default HTTP client and bypasses corporate proxy/TLS settings used elsewhere.

Additional Locations (1)

• pkg/llm/bedrock.go#L54-L63

 

^{Reviewed by Cursor Bugbot for commit ebf453a. Configure here.}