[Deep-Sixed]

Summary

Follow-up to Deep-Sixed/Cerberus#29. A jev alias can now ask Jev, in the same Decisions API call, how much reasoning effort a request needs, as well as which model should serve it.

cerberus/smart:
  mode: jev
  candidates:
    - { provider: google,     credential: free-1,  model: gemini-flash }
    - { provider: openrouter, credential: primary, model: anthropic/claude-sonnet, reasoning_effort: medium }
  jev:
    decider: { provider: openrouter, credential: primary }
    reasoning_efforts: [low, medium, high]   # ask, and choose only among these

Where the answer applies

• It replaces the configured reasoning_effort of every planned candidate that sets one, and is never added to a candidate that sets none. Cerberus already configures effort per route, and some providers reject the parameter (OpenAI rejects it for non-reasoning models), so adding it everywhere could turn a working route into a 400.
• The reasoning_efforts list keeps Jev to values every such route accepts. For example, leave out minimal if one of those routes rejects it.
• The existing rule for callers is unchanged: a caller's own reasoning_effort wins unless the candidate sets reasoning_effort_override.

Each answer is checked on its own

• An effort outside the list, or a missing effort answer, leaves the configured effort in place. The model choice still stands.
• When only one model is left, the call is still made, for the effort question alone. When there's nothing to ask, no call is made and no request text leaves.
• A failed call (timeout, HTTP error, …) records the same reason for both answers. The request runs in configured order with configured efforts.
• DispatchPlan may change the effort of a target that already sets one, and nothing else about it. dispatch() refuses any other difference from a configured target.

Validation

jev.reasoning_efforts is refused when it:

• has fewer than two values
• repeats a value
• contains a value outside minimal / low / medium / high
• has no pool candidate setting reasoning_effort for the answer to replace, since it would never be applied

Wire and telemetry

• The decision adapter now sends a list of questions and returns one answer per question:
  • route: unchanged
  • effort: {"id": "effort", "type": "choice", "question": …, "options": [...]}
• The jev record gains effort, effort_confidence and effort_reason. The reason is one of not_requested, not_applicable, no_choice, choice_outside_options, or the reason the call failed.
• The event's existing reasoning_effort field shows the effort that was applied.
• /admin/routes and the console show which efforts Jev may choose.

Test plan

• ./scripts/ci.sh: ruff clean, 431 passed. 12 new tests in tests/test_jev_dispatch.py and 1 new schema test.
• Mutation checks:
  • applying the effort to every candidate fails 5 tests
  • accepting any effort answer fails 2
  • letting a plan add an effort fails 1
• node --check src/cerberus/ui/static/app.js
• python scripts/check-public-history.py: 0 findings.
• Live probe against the Decisions API. This shares Add jev: Jev decides which model, Cerberus executes it #29's open question about the field names inside a question and an answer.

🤖 Generated with Claude Code

https://claude.ai/code/session_01P4Nm4MVSMmG6JyZi8WumtK

---

Generated by Claude Code