plat 631

← goals / models

PLAT-631: Workflow model card shows execution effort instead of Builder effort and retained turns ignore effort changes

Field	Value
State	deployed
Priority	P2
Product	goals
Area	models
Summary	The workflow model card showed the High execution default beside a Medium Builder chat; effort-only changes also bypassed retained runtime refresh.

What happened

On Excellence, Vaibhav's Timouthy candidate-sourcing workflow follows Claude's managed provider profile. Its Builder default is Sonnet 5.5 with Medium effort; its High execution tier differs. The single Model card displayed the first execution tier while the adjacent chat used Builder. A read-only process check confirmed actual --model claude-sonnet-5-5 --effort medium, and the launch log also recorded Medium. Effort was passed correctly for this managed profile; the displayed default described the wrong role.

A second issue was found in retained workflow delivery: provider, model and account were compared, but options (including reasoning effort) were not. An effort-only edit could keep the existing CLI's launch flags. The continuation request could also keep stale tab options instead of the manifest's actual Builder options.

Code and Crew use their separate single Builder-model panel and persisted profile reasoning setting. Existing effort selection/account preservation and conversation-restart checks pass; this screenshot's session is a workflow, not a Code or Crew chat. No execution-role controls were added to those products.

Fix

• Display Builder's model and effort in the workflow's single card. Preserve provider defaults on mount; choosing a model/effort still applies it to every role with the selected account. Collapsing per-role choices confirms use of Builder's setting.
• Record the actual manifest model, account and options in retained turn context.
• Compare a canonical options fingerprint for warm and saved runtimes. Legacy snapshots reconnect once while preserving native conversation history.
• Queue runtime changes behind a running turn before closing its CLI. Check authority first: changed permissions still revoke the old runtime immediately.

Validation

Frontend panel/role regressions and existing Code/Crew model-panel checks pass; TypeScript build passes. Backend regressions drive the live-input HTTP handler through busy queueing and idle rebuild after an effort-only change; unchanged High effort retains the CLI. Existing permission/provider/account checks pass. A production Builder query assembly regression checks the prepared model and saved continuation use manifest High over stale browser Medium, with its options fingerprint surviving saved runtime serialization. It stops before a model request; external workspace state is mocked.

Deployment

Deployed to Excellence in release agents-b0c18a1f-20261006161016 on 2026-10-06. Public health returns 200; the Linux slot/confinement self-test reports 156 passed, 0 failed, 16 skipped. Verified the deployed frontend contains the Builder-card fix and the backend source includes the retained options comparison. An authenticated read of the live provider manifest returns Builder Sonnet 5.5 with Medium effort, and the workflow's managed profile and selected private account remain intact.

No user configuration was changed and no model message was injected into Vaibhav's conversation. High-effort selection and next-turn refresh are covered by the query-assembly and live-input handler regressions above; they were not exercised with a paid production turn.

Left

None for this fix.