scaling-group/eve/configure-eve-driver

Configure the EvE Driver

Use the public presets first:

• driver: codex_smoke is for fast runtime validation.
• driver: codex_max is for full non-interactive runs.
• driver: opencode_smoke is the equivalent fast OpenCode path.
• driver: opencode_max is the equivalent full-run OpenCode path.

Do not add a new preset for a one-off variation. Prefer Hydra overrides until the role is stable enough to document as a supported entrypoint.

Public Surface

Driver presets live in configs/eve/driver/. Keep that directory small:

• codex_smoke.yaml: non-interactive Codex backend, small turn budget, web search disabled.
• codex_max.yaml: non-interactive Codex backend, full-run budget, web search enabled.
• opencode_smoke.yaml: non-interactive OpenCode backend with a small native step cap.
• opencode_max.yaml: non-interactive OpenCode backend with a full-run native step cap.

Task smoke configs live beside task configs, but keep the root config surface compact:

• circle_packing.smoke is the single tracked circle packing smoke preset.
• By default it validates the ordinary loop.
• Use explicit Hydra overrides from that same entrypoint for judge evaluation or role-specific system prompt smoke paths.

Task smoke configs use driver: codex_smoke by default; select driver=opencode_smoke to exercise OpenCode. Do not shrink loop.n_workers_phase2 in the default smoke path; the two-worker default covers Phase 2 multi-worker behavior and optimizer sync. Use loop.n_parallel_phase2 when you only need to reduce concurrent solver agents.

Normal Commands

Fast smoke:

uv run python -m scaling_evolve.algorithms.eve.runner --config-name=circle_packing.smoke

OpenCode smoke (use a native provider/model identifier):

uv run python -m scaling_evolve.algorithms.eve.runner --config-name=circle_packing.smoke driver=opencode_smoke driver.model=deepseek/deepseek-chat

Judge evaluation smoke:

uv run python -m scaling_evolve.algorithms.eve.runner \
  --config-name=circle_packing.smoke \
  evaluation=circle_packing.judge \
  label=circle-packing-judge-smoke

Role-specific system prompt smoke:

uv run python -m scaling_evolve.algorithms.eve.runner \
  --config-name=circle_packing.smoke \
  evaluation=circle_packing.judge \
  loop.max_iterations=1 \
  label=circle-packing-codex-prompt-smoke \
  +driver.overrides.solver.system_prompt_file=configs/eve/optimizer/circle_packing/prompt/CODEX_SYSTEM_PROMPT.md

Full run:

uv run python -m scaling_evolve.algorithms.eve.runner --config-name=circle_packing

OpenCode full run:

uv run python -m scaling_evolve.algorithms.eve.runner --config-name=circle_packing driver=opencode_max driver.model=deepseek/deepseek-chat

One-off interactive/debug run:

uv run python -m scaling_evolve.algorithms.eve.runner \
  --config-name=circle_packing.smoke \
  driver.driver=codex_tmux \
  driver.open_iterm2=true

If overriding from a full config, keep the backend switch explicit:

uv run python -m scaling_evolve.algorithms.eve.runner \
  --config-name=circle_packing \
  driver.driver=codex_tmux \
  driver.open_iterm2=true

Do not preserve interactive/debug overrides as new YAML presets unless the team decides they are a supported public entry.

Selection Semantics

The driver config lives under the top-level driver: key.

• driver.driver selects the backend and wins over driver.provider.
• driver.provider is accepted for older configs. Avoid it for new public presets.
• If a config inherits from codex_max, changing only driver.provider will not switch the backend because driver.driver=codex_exec is still set.
• Role overrides are shallow merges: EvE starts from the base driver mapping, then overlays driver.overrides.solver or driver.overrides.eval.
• Do not configure driver.overrides.optimizer; the independent optimizer driver role was removed.

Supported backend names are codex_exec, codex_tmux, and opencode.

Authentication

• Codex uses its native authentication by default; alternate providers may use model_providers.*.env_key.
• OpenCode uses credentials saved by opencode auth login unless an explicit model_providers.*.env_key is configured.
• Explicit environment credentials take precedence. Do not rely on undeclared ambient variables.

Driver Options

Common fields:

• driver.executable: command name to launch; defaults to codex or opencode according to the backend.
• driver.model: backend-native model identifier.
• driver.rollout_max_turns: per-rollout agent turn/step cap.
• driver.timeout_seconds: per-rollout wall-clock timeout.
• driver.system_prompt_file: repo-relative path to a backend system prompt file.
• driver.model_providers: provider config map; if an entry declares env_key, EvE fails fast when that environment variable is missing.
• driver.token_pricing: per-driver pricing metadata override. Prefer shared pricing in configs/pricing.yaml unless a run needs a local override.

Codex-only fields (codex_exec, codex_tmux):

• driver.reasoning_effort: reasoning budget where supported.
• driver.effort_level: accepted as a fallback spelling for Codex-style reasoning effort.
• driver.web_search: disabled, cached, or live.
• driver.budget_prompt: whether EvE injects the budget prompt wrapper.
• driver.enable_multi_agent: optional backend feature override.
• driver.personality: optional backend personality override.
• driver.model_provider: optional alternate provider id for Codex-style backends.

OpenCode-only semantics (opencode):

• driver.model must use OpenCode's native provider/model form when set; omit it to use OpenCode's configured default.
• driver.variant passes the provider-specific native variant unchanged.
• driver.rollout_max_turns is applied through OpenCode's native agent.build.steps setting.
• driver.budget_prompt must be false; Codex hook injection is not translated.
• Codex-only options are rejected instead of ignored or mapped.
• HOME, XDG state, project discovery, and external skills are isolated per EvE workspace.
• EvE-owned AGENTS.md, guidance skills, and helper agents are projected into the isolated

OpenCode configuration.

• Only model_providers.*.env_key is reused; other Codex provider settings are rejected

instead of translated.

• Spawn creates a native OpenCode session; resume uses its exact session ID in the same workspace. Native fork, cross-workspace migration, and snapshots are unsupported.

Tmux-only fields (codex_tmux):

• driver.approval_policy: approval policy passed to the interactive backend.
• driver.sandbox_mode: sandbox mode passed to the interactive backend.
• driver.allow_network: when sandbox_mode is omitted, true maps to danger-full-access; otherwise the default is workspace-write.
• driver.open_iterm2: whether EvE opens the tmux session in a visible terminal window.
• driver.completion_filename: internal tmux completion marker filename.
• driver.instruction_filename: internal tmux instruction filename.

Implementation References

This skill is grounded in the current runtime surfaces:

• src/scaling_evolve/algorithms/eve/runtime/driver.py for backend selection, role overrides, tmux pool creation, sandbox/web-search normalization, provider environment resolution, and pricing config loading.
• src/scaling_evolve/providers/agent/drivers/codex_exec.py for non-interactive Codex fields.
• src/scaling_evolve/providers/agent/drivers/opencode.py for non-interactive OpenCode spawn/resume, JSONL parsing, and native step limits.
• src/scaling_evolve/providers/agent/drivers/codex_tmux.py for interactive Codex/tmux fields.
• src/scaling_evolve/providers/agent/config.py for persistent session provider config validation.