fix(ci): remove API keys from agent env when api-proxy is enabled#803
Conversation
When api-proxy is enabled (indicated by BASE_URL environment variables), API keys should NOT be exposed to the agent container for security. The api-proxy sidecar holds the credentials and injects auth headers. Previously, the workflow was passing both: - CODEX_API_KEY and OPENAI_API_KEY (should NOT be in agent env) - OPENAI_BASE_URL and ANTHROPIC_BASE_URL (should be in agent env) This defeated the security isolation provided by api-proxy. Changes: - Removed CODEX_API_KEY and OPENAI_API_KEY from agent environment block - Kept OPENAI_BASE_URL and ANTHROPIC_BASE_URL for routing to api-proxy - The awf CLI still receives keys via `sudo -E` and `--env-all` - awf passes keys only to api-proxy container, not agent container Security model: - awf reads keys from host environment (process.env) - awf passes keys only to api-proxy sidecar (src/docker-manager.ts:908-909) - Agent only receives BASE_URL variables (src/docker-manager.ts:948-955) - api-proxy injects auth headers and routes through Squid 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com>
ebfb173
into
claude/fix-github-actions-workflow-again
There was a problem hiding this comment.
Pull request overview
This PR updates the Codex smoke-test GitHub Actions workflow to avoid exposing LLM API credentials to the agent container when the API proxy sidecar is used, aligning CI with the intended credential-isolation model.
Changes:
- Removes
CODEX_API_KEYandOPENAI_API_KEYfrom the Codex execution step environment. - Keeps
OPENAI_BASE_URL/ANTHROPIC_BASE_URLset to route requests via theapi-proxysidecar.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| GH_AW_PROMPT: /tmp/gh-aw/aw-prompts/prompt.txt | ||
| GH_AW_SAFE_OUTPUTS: ${{ env.GH_AW_SAFE_OUTPUTS }} | ||
| GH_DEBUG: 1 | ||
| GITHUB_STEP_SUMMARY: ${{ env.GITHUB_STEP_SUMMARY }} |
There was a problem hiding this comment.
Removing both CODEX_API_KEY and OPENAI_API_KEY here means the awf process no longer has an API key in its environment to pass into the api-proxy sidecar (awf only reads OpenAI credentials from process.env.OPENAI_API_KEY). Since this step still forces OPENAI_BASE_URL to http://api-proxy:10000, Codex/OpenAI calls will likely hit the sidecar without auth headers and fail (e.g., 401). Consider keeping OPENAI_API_KEY available to the awf invocation while preventing it from being forwarded into the agent container (e.g., stop using --env-all and pass only the needed env vars explicitly, or add filtering so OPENAI_API_KEY/CODEX_API_KEY are not propagated to the agent when api-proxy is enabled).
| GITHUB_STEP_SUMMARY: ${{ env.GITHUB_STEP_SUMMARY }} | |
| GITHUB_STEP_SUMMARY: ${{ env.GITHUB_STEP_SUMMARY }} | |
| OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }} | |
| CODEX_API_KEY: ${{ secrets.CODEX_API_KEY }} |
…ng (#802) * Initial plan * fix(ci): add BASE_URL environment variables for CODEX api-proxy routing CODEX was not being directed to use the api-proxy because the OPENAI_BASE_URL and ANTHROPIC_BASE_URL environment variables were not explicitly set in the smoke-codex workflow. While AWF automatically sets these variables when generating the docker-compose configuration (if API keys are present), explicitly setting them in the workflow env ensures they are available to the CODEX agent for routing API calls through the api-proxy sidecar. This fix adds: - OPENAI_BASE_URL=http://api-proxy:10000 - ANTHROPIC_BASE_URL=http://api-proxy:10001 to the 'Run Codex' step environment variables. Fixes job failure in run 63483600453. * fix(ci): remove API keys from agent env when api-proxy is enabled (#803) * Initial plan * fix(ci): remove API keys from agent env when api-proxy is enabled When api-proxy is enabled (indicated by BASE_URL environment variables), API keys should NOT be exposed to the agent container for security. The api-proxy sidecar holds the credentials and injects auth headers. Previously, the workflow was passing both: - CODEX_API_KEY and OPENAI_API_KEY (should NOT be in agent env) - OPENAI_BASE_URL and ANTHROPIC_BASE_URL (should be in agent env) This defeated the security isolation provided by api-proxy. Changes: - Removed CODEX_API_KEY and OPENAI_API_KEY from agent environment block - Kept OPENAI_BASE_URL and ANTHROPIC_BASE_URL for routing to api-proxy - The awf CLI still receives keys via `sudo -E` and `--env-all` - awf passes keys only to api-proxy container, not agent container Security model: - awf reads keys from host environment (process.env) - awf passes keys only to api-proxy sidecar (src/docker-manager.ts:908-909) - Agent only receives BASE_URL variables (src/docker-manager.ts:948-955) - api-proxy injects auth headers and routes through Squid 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> --------- Co-authored-by: anthropic-code-agent[bot] <242468646+Claude@users.noreply.github.com> Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> * fix(firewall): add api-proxy to allowed domains when enabled (#804) * Initial plan * fix(firewall): add api-proxy to allowed domains when enabled Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> --------- Co-authored-by: anthropic-code-agent[bot] <242468646+Claude@users.noreply.github.com> Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> * fix(squid): handle bare hostnames without leading dot for api-proxy (#805) * Initial plan * fix(squid): handle bare hostnames without leading dot for api-proxy Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> --------- Co-authored-by: anthropic-code-agent[bot] <242468646+Claude@users.noreply.github.com> Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> * fix(agent): enable direct api-proxy access and remove api key from env (#807) * Initial plan * fix(agent): enable direct api-proxy access and remove api key from env - Add iptables bypass for api-proxy in setup-iptables.sh - Pass AWF_ENABLE_API_PROXY env var from docker-manager.ts - Remove ANTHROPIC_API_KEY from agent env in workflows - Update postprocess script to strip API key from compiled workflows Fixes agent->api-proxy connectivity and security vulnerability where API key was exposed to agent container instead of isolated to api-proxy. --------- Co-authored-by: anthropic-code-agent[bot] <242468646+Claude@users.noreply.github.com> * fix: pass ANTHROPIC_API_KEY to validation in all Claude workflows (#808) * Initial plan * fix: pass ANTHROPIC_API_KEY to validation step in smoke-claude workflow Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> * fix: pass ANTHROPIC_API_KEY to validation in all Claude workflows Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> --------- Co-authored-by: anthropic-code-agent[bot] <242468646+Claude@users.noreply.github.com> Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> * fix: only enable api-proxy when API keys are provided (#810) * Initial plan * fix: only enable api-proxy when API keys are provided Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> * feat(workflow): enable api-proxy for smoke-claude workflow - Add --enable-api-proxy flag to awf command - Add --anthropic-api-key flag to pass API key to api-proxy sidecar - Add ANTHROPIC_API_KEY to env block for agent step - API key is shared with api-proxy, kept out of agent container - Agent uses ANTHROPIC_BASE_URL to direct requests to api-proxy Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> * refactor(workflow): remove redundant --enable-api-proxy flag The --enable-api-proxy flag defaults to true (src/cli.ts:724), so it doesn't need to be explicitly specified. The api-proxy sidecar will automatically deploy when API keys are present. See docs/api-proxy-sidecar.md:51 which states "The API proxy is enabled by default and automatically deploys when API keys are present" Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> --------- Co-authored-by: anthropic-code-agent[bot] <242468646+Claude@users.noreply.github.com> Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> * fix: remove unknown --anthropic-api-key flag from smoke-claude workflow (#811) * Initial plan * fix: remove unknown --anthropic-api-key flag from smoke-claude workflow Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> --------- Co-authored-by: anthropic-code-agent[bot] <242468646+Claude@users.noreply.github.com> Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> * fix: use api-proxy IP address instead of hostname for BASE_URL (#813) * Initial plan * fix: use api-proxy IP address instead of hostname for BASE_URL Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> --------- Co-authored-by: anthropic-code-agent[bot] <242468646+Claude@users.noreply.github.com> Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> * fix: use api-proxy IP address instead of hostname for BASE_URL (#813) (#815) * Initial plan * fix: use api-proxy IP address instead of hostname for BASE_URL (#813) Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> --------- Co-authored-by: anthropic-code-agent[bot] <242468646+Claude@users.noreply.github.com> Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> * fix(api-proxy): use IP address for API base URLs to avoid DNS issues In chroot mode, Docker container hostname resolution can fail because the DNS resolver may not properly reach Docker's embedded DNS. Use the api-proxy IP address directly (e.g., http://172.30.0.30:10001) instead of the hostname (http://api-proxy:10001) to eliminate DNS resolution as a failure point. Also add test coverage for the host-iptables api-proxy ACCEPT rule. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: exclude API keys from agent when api-proxy is enabled (#814) * Initial plan * fix: exclude API keys from agent when api-proxy is enabled --------- Co-authored-by: anthropic-code-agent[bot] <242468646+Claude@users.noreply.github.com> * fix(agent): use AWF_API_PROXY_IP env var for api-proxy iptables rules (#817) * Initial plan * fix(agent): use AWF_API_PROXY_IP env var for api-proxy iptables Move api-proxy iptables rules to use pre-set AWF_API_PROXY_IP environment variable instead of dynamic hostname resolution, and place FILTER rules in the correct position (after NAT setup, before final DROP rule). Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> --------- Co-authored-by: anthropic-code-agent[bot] <242468646+Claude@users.noreply.github.com> Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> * Add debug logging for BASE_URL environment variables in agent container (#816) * Initial plan * feat: add debug logging for BASE_URL environment variables Add logging to display ANTHROPIC_BASE_URL and OPENAI_BASE_URL values that are set for the agent container. This helps debug configuration issues when running Claude Code CLI in the workflow. The logging is added after the Docker Compose config is generated, showing whether BASE_URL variables are set or using defaults. Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> --------- Co-authored-by: anthropic-code-agent[bot] <242468646+Claude@users.noreply.github.com> Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> * fix: simplify api-proxy iptables bypass condition (#819) * Initial plan * fix: simplify api-proxy iptables bypass condition The NAT bypass for api-proxy traffic was checking both AWF_ENABLE_API_PROXY and AWF_API_PROXY_IP, but this was too strict. When the HTTP_PROXY environment variable is set, HTTP clients will send requests through the proxy unless NO_PROXY is configured or iptables rules prevent it. The issue was that the NAT bypass required both flags, causing traffic to 172.30.0.30 (api-proxy) to be sent through Squid when it should go directly. Squid then blocked these requests because the IP address wasn't in the domain whitelist. The fix simplifies the condition to only check AWF_API_PROXY_IP, matching the pattern used for the OUTPUT FILTER ACCEPT rules (lines 285-289). This ensures that when AWF_API_PROXY_IP is set, traffic to that IP bypasses Squid at the NAT level, preventing HTTP clients from routing through the proxy regardless of HTTP_PROXY settings. Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> --------- Co-authored-by: anthropic-code-agent[bot] <242468646+Claude@users.noreply.github.com> Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> * Initial plan (#818) Co-authored-by: anthropic-code-agent[bot] <242468646+Claude@users.noreply.github.com> * fix: add api-proxy IP to squid allowlist (#820) * Initial plan * fix: add api-proxy IP to squid allowlist Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> --------- Co-authored-by: anthropic-code-agent[bot] <242468646+Claude@users.noreply.github.com> Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> --------- Co-authored-by: anthropic-code-agent[bot] <242468646+Claude@users.noreply.github.com> Co-authored-by: lpcox <15877973+lpcox@users.noreply.github.com> Co-authored-by: Landon Cox <landon.cox@microsoft.com> Co-authored-by: Jiaxiao (mossaka) Zhou <duibao55328@gmail.com> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When api-proxy is enabled, API keys were being passed to both the agent container and the api-proxy sidecar, defeating the security isolation model where credentials should only exist in the sidecar.
Changes
CODEX_API_KEYandOPENAI_API_KEYfrom agent environment block insmoke-codex.lock.ymlOPENAI_BASE_URLandANTHROPIC_BASE_URLto route requests through api-proxyFlow
The awf CLI receives keys from host environment via
sudo -E→ reads fromprocess.env→ passes only to api-proxy container (not agent) → api-proxy injects auth headers.This restores the documented security model: agent cannot access credentials, only routes to proxy.