Pass --openai-api-base-path / --anthropic-api-base-path to AWF when URL contains path

[lpcox]

Summary

When OPENAI_BASE_URL or ANTHROPIC_BASE_URL in engine.env contains a path prefix (e.g., https://host.com/serving-endpoints), the compiler strips the path and only passes the hostname via --openai-api-target. This causes API requests to go to https://host/v1/chat/completions instead of https://host/serving-endpoints/v1/chat/completions, returning 404.

The AWF firewall already supports --openai-api-base-path and --anthropic-api-base-path flags (added in a recent release). The compiler just needs to extract the path component and pass it.

Upstream tracking: /issues/21320#issuecomment-4131507565

Impact

Any OPENAI_BASE_URL or ANTHROPIC_BASE_URL with a required path prefix silently fails:

• Databricks serving endpoints (/serving-endpoints)
• Azure OpenAI deployments (/openai/deployments/<name>)
• Corporate LLM routers with path-based routing

Root Cause

extractAPITargetHost() in pkg/workflow/awf_helpers.go:344-347 intentionally strips the path:

// Remove path suffix if present (everything after first /)
if idx := strings.Index(host, "/"); idx != -1 {
    host = host[:idx]
}

This only returns the hostname for --openai-api-target. The path component is discarded and never passed to AWF.

Solution

1. Add extractAPIBasePath() function (pkg/workflow/awf_helpers.go)

Add a new function that extracts the path component from the URL:

// extractAPIBasePath extracts the path component from a custom API base URL in engine.env.
// Returns the path prefix (e.g., "/serving-endpoints") or empty string if no path is present.
func extractAPIBasePath(workflowData *WorkflowData, envVar string) string {
    if workflowData == nil || workflowData.EngineConfig == nil || workflowData.EngineConfig.Env == nil {
        return ""
    }

    baseURL, exists := workflowData.EngineConfig.Env[envVar]
    if !exists || baseURL == "" {
        return ""
    }

    // Remove protocol prefix
    host := baseURL
    if idx := strings.Index(host, "://"); idx != -1 {
        host = host[idx+3:]
    }

    // Extract path (everything after first /)
    if idx := strings.Index(host, "/"); idx != -1 {
        path := host[idx:] // e.g., "/serving-endpoints"
        // Remove trailing slash if present
        path = strings.TrimRight(path, "/")
        if path != "" {
            return path
        }
    }

    return ""
}

2. Pass base path flags in BuildAWFArgs() (pkg/workflow/awf_helpers.go)

After the existing --openai-api-target and --anthropic-api-target lines (~L213-223), add:

// Pass base path if URL contains a path component
openaiBasePath := extractAPIBasePath(config.WorkflowData, "OPENAI_BASE_URL")
if openaiBasePath != "" {
    awfArgs = append(awfArgs, "--openai-api-base-path", openaiBasePath)
    awfHelpersLog.Printf("Added --openai-api-base-path=%s", openaiBasePath)
}

anthropicBasePath := extractAPIBasePath(config.WorkflowData, "ANTHROPIC_BASE_URL")
if anthropicBasePath != "" {
    awfArgs = append(awfArgs, "--anthropic-api-base-path", anthropicBasePath)
    awfHelpersLog.Printf("Added --anthropic-api-base-path=%s", anthropicBasePath)
}

3. Add tests (pkg/workflow/awf_helpers_test.go)

Add tests for extractAPIBasePath:

func TestExtractAPIBasePath(t *testing.T) {
    tests := []struct {
        name     string
        url      string
        expected string
    }{
        {"databricks serving endpoint", "https://host.com/serving-endpoints", "/serving-endpoints"},
        {"azure openai deployment", "https://host.com/openai/deployments/gpt-4", "/openai/deployments/gpt-4"},
        {"simple path", "https://host.com/v1", "/v1"},
        {"trailing slash stripped", "https://host.com/api/", "/api"},
        {"no path", "https://host.com", ""},
        {"bare hostname", "host.com", ""},
        {"root path only", "https://host.com/", ""},
    }
    // ... table-driven test body
}

Add integration test for BuildAWFArgs:

t.Run("includes openai-api-base-path when URL has path component", func(t *testing.T) {
    // Set OPENAI_BASE_URL to "https://host.com/serving-endpoints"
    // Verify args contain both --openai-api-target and --openai-api-base-path
})

4. Recompile workflows

make build && make recompile && make agent-finish

No new frontmatter needed

This fix uses the existing engine.env frontmatter. Workflow authors already configure custom endpoints via env vars:

engine:
  id: codex
  env:
    OPENAI_BASE_URL: "https://stone-dataplatform-production.cloud.databricks.com/serving-endpoints"
    OPENAI_API_KEY: ${{ secrets.DATABRICKS_KEY }}

The compiler will automatically extract and pass the path component — no new frontmatter fields required.

Checklist

• Add extractAPIBasePath() function to awf_helpers.go
• Call it in BuildAWFArgs() for both OpenAI and Anthropic URLs
• Add unit tests for extractAPIBasePath()
• Add integration test for BuildAWFArgs with path-containing URLs
• make build && make recompile && make agent-finish

[github-actions]

🏷️ Issue Triaged

Hi @lpcox! I've categorized this issue as bug based on the following analysis:

Reasoning: The compiler incorrectly strips the path component from OPENAI_BASE_URL/ANTHROPIC_BASE_URL, causing API requests to hit the wrong endpoint and return 404 errors. This is a regression in correct behavior.

View Triage Details

Analysis

• Keywords detected: "strips the path", "404", "base-path", "URL contains path prefix"
• Issue type indicators: Existing functionality produces incorrect output (wrong URL constructed), leads to runtime 404 errors — classic bug pattern
• Confidence: High

Recommended Next Steps

• Investigate the URL parsing logic in the compiler where --openai-api-target / --anthropic-api-target are set
• Preserve the path component when constructing the AWF flags (--openai-api-base-path / --anthropic-api-base-path)
• Add a test case covering base URLs with path prefixes (e.g. `(host.com/redacted)

References: Triage run §23600217702

Note

🔒 Integrity filter blocked 33 items

The following items were blocked because they don't meet the GitHub integrity level.

• Threat detection taking 5m+ #23130 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• PR creation fails if there are merge conflicts with intervening updates to target branch #23129 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• BUG: 'GH_HOST' set with GITHUB_OUTPUT (not env or env-threaded), breaking Process Safe Outputs on GHES #23093 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• BUG: Downstream 'GH_AW_SAFE_OUTPUTS_CONFIG_PATH' and 'GH_AW_SAFE_OUTPUTS_TOOLS_PATH' variables not available as env vars due to GITHUB_OUTPUT-only write #23092 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• BUG: 'setup-cli' action referenced by mutable tag instead of commit SHA in maintenance workflow #23088 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Agentic Plan: Improve gh aw upgrade warning for github/gh-aw-actions/setup when no GitHub Releases exist #23083 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• AWF allowlist not enforced for plain-HTTP connections to numeric IPs; web-fetch may bypass proxy enforcement #23079 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• push_repo_memory fails with ENOBUFS on large repos after disabling sparse checkout #23024 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• BUG: AWF agent cannot reach Postgres started via services: #22939 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• bug: agentic workflows fail with token limit overflow when repository has comprehensive custom instructions #22929 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• gh-aw compiler must reject env.* expressions in markdown per documented safety policy #22914 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Failed to register MCP tools | HTTP 400: Bad Request #22913 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• MCP gateway should enforce tool allowlist at the gateway layer, not only at the Claude client layer #22908 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• cache-memory restore must clear or validate prior-run artifacts before agent launch #22902 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Assign to agent seems to be limited for 1 assignment per issue, I would like to extend it to multiple assignments so agent can work on multiple repos. #22897 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Anthropic API proxy must enforce per-request credential validation #22894 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• ... and 17 more items

To allow these resources, lower min-integrity in your GitHub frontmatter:

tools:
  github:
    min-integrity: approved  # merged | approved | unapproved | none

Generated by Issue Triage Agent · ◷

[copilot-swe-agent]

The agent encountered an error and was unable to start working on this issue: This may be caused by a repository ruleset violation. See granting bypass permissions for the agent, or please contact support if the issue persists.