feat(cua): make reasoning effort declarable, and record what actually ran (#497)

[danielgwilson]

feat(cua): make reasoning effort declarable, and record what actually ran (#497)

Effort was reachable in the provider and unreachable from a lab. Nothing in
cua-actor-lab ever passed it and the schema had no field for it, so every
computer-use run humanish has ever done was the provider default — not by
decision but by omission. That is the kind of silent constant that turns a
finding about a product into a finding about a default nobody chose: a
participant who abandons a flow at medium and finishes it at high did not find
a usability problem, the harness did.

• actors[].reasoningEffort with a per-lane override, the same shape persona
  and stopWhen already have. Two lanes at different efforts is a control that
  runs in ONE study, under one set of conditions.
• The vocabulary is the documented union (none..max) and it validates, because
  OpenAI states support is model-dependent. An unsupported level fails on the
  first turn with the provider's own message rather than being downgraded —
  a silent downgrade would make the trace claim an effort the run did not use.
• ActorTrace.modelSettings (humanish.model-settings.v1): the effort the
  request ACTUALLY carried, including the default when a lab declared none.
  A trace that cannot say what effort produced it cannot be compared with one
  that can. Absent when a provider declares no settings; verify tolerates it.
• The lab screen and the fan-out plan line both show it, and the plan line is
  what you read before spending money. "per-lane" when a roster disagrees,
  rather than picking one lane's answer for all of them.
• labs/effort-contrast-demo.yaml: the sibling of persona-contrast-demo. One
  persona, one mission, two efforts. Both lanes dry-run with identical prompt
  digests, so effort is the only declared difference.

Keeping gpt-5.6-sol as the default is unchanged and deliberate: Artificial
Analysis, who ran OpenAI's pre-release eval, reports Luna and Sol always on the
Pareto frontier ahead of Terra, and OpenAI reports Sol at 62.6% on OSWorld 2.0.

Closes #497.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
humanish	Ignored		Aug 20, 2026 8:34pm