fix(models): default thinking level to high when no effort is set

[charlesvien]

Problem

Users who pick claude-fable-5 hit hard failures whenever a generation runs without a configured effort/thinking level (8 users and 142 failed generations on 06-11 alone, invisible to error monitoring since $ai_http_status=None).

fable-5 only supports adaptive thinking and rejects the SDK's default thinking: { type: "disabled" } with invalid_request_error: "thinking.type.disabled" is not supported for this model. The effort level only reaches the SDK when present, so a fable-5 session with no effort set lets the SDK default to disabled thinking, which fable-5 rejects outright.

Same issue as #2618. This PR takes a different approach: instead of scoping a medium default to fable-5 behind a requiresAdaptiveThinking flag, it makes high the default thinking level for every effort-capable model. That keeps thinking enabled by default and removes the disabled-thinking request shape that fable-5 rejects.

Changes

• Add DEFAULT_EFFORT ("high") and a resolveEffortForModel helper in session/models.ts. The helper honors an explicit effort, falls back to high for effort-capable models, and leaves models without effort support (e.g. haiku) unset so they keep disabled thinking.
• In claude-agent.ts createSession, after the gateway modelId is resolved, apply the default effort via applyFlagSettings when the user has not chosen one. Gating on the resolved id (not the raw setting) means the default also applies to the default model and aliases, not just explicit fable-5 picks.
• Fix the mid-session model-switch path (rebuildEffortConfigOption): when switching to an effort-capable model with no explicit effort, apply the default to the live query instead of only updating the display. This closes the same fable-5 failure when the model is changed from the picker, which fix(models): default fable-5 to adaptive thinking when no effort is set #2618 left open.
• Default the effort config display and the buildConfigOptions fallback to high so the UI matches the request shape.

How did you test this?

• Added parameterized unit tests in models.test.ts: resolveEffortForModel (effort-capable models default to high; haiku/opus-4-6 stay unset; explicit effort is honored, including max on fable-5) and DEFAULT_EFFORT === "high".
• Updated the claude-agent.resume-model.test.ts query mock to include applyFlagSettings (the real SDK query exposes it; the mock did not).
• Ran pnpm --filter @posthog/agent exec vitest run src/adapters/claude src/server — 486 tests pass.
• pnpm --filter @posthog/agent typecheck passes; biome check clean on changed files; node scripts/check-host-boundaries.mjs shows no new violations.

Automatic notifications

• Publish to changelog?
• Alert Sales and Marketing teams?

[github-actions]

React Doctor found no issues in the changed files. 🎉

_{Reviewed by React Doctor for commit 7251bfc.}

[greptile-apps]

_{Reviews (1): Last reviewed commit: "default thinking level to high for effor..." | Re-trigger Greptile}

[github-actions]

Well-motivated fix with clear comments, consistent constant usage, and new tests covering the new function. No showstoppers — purely an internal agent behavior default change within a single package.