Configurable reasoning effort for ce-plan and ce-brainstorm model elevation

[lechatlisse]

Configurable reasoning effort for ce-plan and ce-brainstorm model elevation

Problem

Compound Engineering 3.22.1 lets a checkout select the elevated planning and brainstorming
model with plan_model and brainstorm_model, but the Claude CLI elevation adapter fixes
reasoning effort at high:

EFFORT="high"   # settled: elevation runs at high effort

By contrast, cross-model document and code review now support persistent
cross_model_effort configuration. A checkout can therefore select Fable and its effort for
review, but cannot make the equivalent choice for an unusually difficult plan without patching
a versioned plugin cache or bypassing the supported elevation route.

Requested behavior

Please expose optional planning and brainstorming effort settings, for example:

plan_model: fable
plan_effort: high
brainstorm_model: fable
brainstorm_effort: high

The exact schema is flexible; a shared elevation_effort would also work if separate settings
are not useful.

Suggested contract

• Preserve high as the default so existing behavior is unchanged.
• Support the effort values honored by the resolved model/adapter, including xhigh for Fable.
• Use the existing configuration precedence: current-task request, local config, tracked config,
  then the default.
• Reject or visibly skip unsupported explicit values rather than silently substituting.
• Record requested and actual effort alongside the existing requested/served model receipt when
  the adapter can verify it.
• Cover the new settings in ce-setup, the config template/reference, and adapter parity tests.

Use case

We run Compound Engineering from the ChatGPT/Codex desktop app with Codex orchestrating and
implementing, Fable authoring the reasoning-heavy plan step, and Fable providing the independent
adversarial review. high is the right standing default. For occasional architecture, security,
tenancy, or migration plans, we would like to request a higher supported effort for the elevated
planning step without changing the rest of the session or maintaining a local plugin patch.

Existing Claude Code routing limitation

This behavior is already tracked upstream in
anthropics/claude-code#79562, an open
bug with a reproduction for /effort xhigh switching a Fable session to Opus. Related reports
also document false-positive Fable safety-classifier fallback and session-level stickiness, for
example
anthropics/claude-code#66657.

A receipt-verified smoke test on Claude CLI 2.1.233 reproduced the underlying routing behavior:

• --model fable --effort high served claude-fable-5 in the observed run.
• A stream receipt for --model fable --effort xhigh explicitly emitted a
  model_refusal_fallback event from claude-fable-5 to claude-opus-4-8, categorized as
  cyber, even though the prompt only requested a fixed harmless reply and tools were disabled.

This points to Claude Code's Fable classifier-fallback path, not a Compound Engineering defect,
and it means effort alone does not prove which model actually served the request. Compound
Engineering should continue to validate the requested and served model and fail closed on a
family substitution. We are keeping both planning and review at the observed
fable/high combination for now, while still validating every receipt.

Environment

• compound-engineering 3.22.1
• ChatGPT/Codex desktop app on macOS
• Claude CLI with Fable available

Related but not duplicate: #1335 added model/effort configuration for cross-model reviews. This
request is limited to ce-plan and ce-brainstorm model elevation.