V0.5+: revisit effort picker (Opus 4.8 likely makes effort more impactful)

[yyq1025]

Why

Opus 4.8 / Opus 4.8 1M just landed (89f656d). Effort level is expected to have a larger behavioral impact on Opus 4.8 than on the 4.x models we shipped against, so the per-session effort picker we cut is worth reconsidering.

Background — what we had and why it was cut

A full per-session effort picker existed and was deliberately removed:

• Built: 388fd9c (getModels RPC + model picker), f8f87a2 (model+effort apply), 5af60da (per-model supportedEffortLevels in MODEL_METADATA)
• Cut: 03ba3b4 — "cut effort picker (trust SDK adaptive thinking)" (~512 LOC removed). Chip now shows model only.
• Follow-up: 3f5a950 — daemon still hardcodes effort: "xhigh" in sidecode metadata purely for Desktop local_*.json schema compat (never read at runtime).

Today the daemon never passes --effort; the user's account-level Settings.effortLevel (set via Desktop /effort) is honored implicitly.

Original reasons for cutting (still partly valid — weigh before re-adding)

1. SDK does zero call-time validation of effortLevel — applyFlagSettings/setModel silently accept nonsense (max / 'test' / bad combos) and return undefined. Effort is unobservable at apply time.
2. Models can't self-report their effort, so we can't confirm a switch took.
3. Anthropic Desktop itself doesn't cross-validate model+effort (ships e.g. "Sonnet + xhigh" combos), so there's no canonical mapping to mirror.
4. Only ground truth is result.modelUsage keys after a turn completes.

What re-adding involves

• Adapt/revert 03ba3b4. The atomic mid-session switcher is applyFlagSettings({ model, effortLevel }) (see packages/daemon/src/router.ts).
• Re-surface effort in the model chip — packages/app/src/components/transcript/input-bar.tsx.
• Protocol: modelEntry / getModels / setSessionSelection schemas in packages/protocol/src/index.ts; consider restoring supportedEffortLevels in packages/daemon/src/models-metadata.ts.

Open question to settle first

Before shipping UI, empirically confirm effort actually changes Opus 4.8 output — reuse the probe-script pattern from 2adbc8e against claude-opus-4-8[1m]. If it's still unobservable, the cut reasoning holds and we keep trusting the account default.