# `comet eval` creates this complete template automatically on first startup at: # Unix: ~/.comet/eval/.env # Windows: %USERPROFILE%\.comet\eval\.env # Edit that user-level file as needed. Existing files are never overwritten. # Process environment variables take precedence over this file. # Default subject Agent when --agent is not supplied. # BENCH_EVAL_AGENT=claude-code # Common subject (Bench) configuration. # BENCH_API_KEY= # BENCH_BASE_URL= # BENCH_MODEL= # Optional independent LLM-as-judge configuration. # BENCH_LLM_JUDGE=1 # BENCH_JUDGE_AGENT= # BENCH_JUDGE_API_KEY= # BENCH_JUDGE_BASE_URL= # BENCH_JUDGE_MODEL= # Provider-specific variables remain supported and override common fallbacks. # ANTHROPIC_API_KEY= # Optional Claude Code model override. # BENCH_CC_MODEL= # BENCH_CC_VERSION=latest # Codex / Qoder / CodeBuddy credentials when selected with --agent. # OPENAI_API_KEY= # OPENAI_BASE_URL= # OPENAI_MODEL= # CODEX_API_KEY= # CODEX_BASE_URL= # CODEX_MODEL= # BENCH_CODEX_VERSION=latest # QODER_PERSONAL_ACCESS_TOKEN= # QODER_BASE_URL= # QODER_MODEL= # BENCH_QODER_VERSION=latest # CODEBUDDY_API_KEY= # CODEBUDDY_AUTH_TOKEN= # CODEBUDDY_BASE_URL= # CODEBUDDY_MODEL= # CODEBUDDY_SMALL_FAST_MODEL= # CODEBUDDY_BIG_SLOW_MODEL= # CODEBUDDY_CODE_SUBAGENT_MODEL= # CODEBUDDY_CUSTOM_HEADERS= # CODEBUDDY_INTERNET_ENVIRONMENT= # BENCH_CODEBUDDY_VERSION=latest # BENCH_CODEX_MODEL= # BENCH_QODER_MODEL= # BENCH_CODEBUDDY_MODEL= # Optional user-installed custom Agent adapter registry. # Default: ~/.comet/eval/adapters (Windows: %USERPROFILE%\.comet\eval\adapters) # COMET_EVAL_ADAPTERS_DIR= # Custom Agent credential/model/base URL names are declared by its adapter.yaml. # Add those names here only after installing an adapter, for example: # MY_AGENT_API_KEY= # MY_AGENT_MODEL= # MY_AGENT_BASE_URL= # Optional: path to a custom user-simulator prompt for auto_user workflow evals. # Relative paths resolve from eval/. Defaults to eval/simulator-instruction.md. # BENCH_SIMULATOR_PROMPT_FILE=simulator-instruction.md # Optional legacy Judge auth token. The common BENCH_JUDGE_API_KEY above is # mapped to the selected Agent's native credential when this is not set. # BENCH_JUDGE_AUTH_TOKEN= # Claude Code via Anthropic-compatible proxy (BigModel / mimo / OpenRouter etc.). # When ANTHROPIC_API_KEY is unset, claude authenticates with these instead. # ANTHROPIC_AUTH_TOKEN= # ANTHROPIC_BASE_URL= # ANTHROPIC_MODEL= # ANTHROPIC_DEFAULT_HAIKU_MODEL= # ANTHROPIC_DEFAULT_SONNET_MODEL= # ANTHROPIC_DEFAULT_OPUS_MODEL= # ANTHROPIC_DEFAULT_OPUS_MODEL_NAME= # ANTHROPIC_DEFAULT_SONNET_MODEL_NAME= # CLAUDE_CODE_SUBAGENT_MODEL= # Eval never writes API keys into published assets, manifests, reports, or the # workspace. During a Docker run, Agent-specific config roots are mounted as # container-local tmpfs files. Codex config uses env_key, CodeBuddy uses an # apiKeyHelper, and Qoder reads its PAT from the isolated process environment. # Required only for eval/langsmith. # LANGSMITH_API_KEY= # LANGSMITH_TRACING=true # LANGSMITH_ENDPOINT=https://api.smith.langchain.com # LANGSMITH_PROJECT=comet-skill-eval # The Claude Code plugin variables (TRACE_TO_LANGSMITH and CC_LANGSMITH_*) are # derived from the LANGSMITH_* settings by eval/langsmith/tests/conftest.py. # TRACE_TO_LANGSMITH=false # Optional: name the LangSmith experiment (project) per treatment so CONTROL vs # skill-injected runs compare in the experiment view. # LANGSMITH_EXPERIMENT=COMET_FULL_040_BETA # Optional trajectory tracing via the official langsmith-tracing Claude Code # plugin. When LANGSMITH_TRACING=true, the LangSmith suite auto-builds the # plugin into eval/.cache/langsmith-cc-plugin via node:20 and reuses that cache. # Do not mount a Windows-host-built node_modules directory into the Linux eval container. # Unset = rubric + comparison still logged, just without the nested trajectory. # CC_LANGSMITH_PLUGIN_DIR=/abs/path/to/comet/eval/.cache/langsmith-cc-plugin # CC_LANGSMITH_PLUGIN_AUTO_BUILD=true # Required only for `comet eval --suite langfuse` run mode. Credentials stay in # the process environment and are never written to manifests or reports. # LANGFUSE_PUBLIC_KEY= # LANGFUSE_SECRET_KEY= # LANGFUSE_BASE_URL=https://cloud.langfuse.com # LANGFUSE_TRACING_ENVIRONMENT=ci