feat(pi): offer Pi's thinking levels in Ask AI

[josdirksen]

Ask AI's Pi provider now has a thinking-level picker, and the level you pick reaches the Pi session.

What it does

• The levels come from each model, as reported by your installed pi. They are read from each get_available_models row (reasoning + thinkingLevelMap) with pi's own rule (getSupportedThinkingLevels, mirrored by piSupportedThinkingLevels in packages/core/model-catalog.ts):

  • off through high are offered unless the model's map sets that level to null.
  • xhigh and max are offered only when the model's map names them.
  • A non-reasoning model gets no picker.

  The endpoint's effort validation already checks against each model's own list, so a level the selected model does not list is never sent.

• The level is applied after the model. The session sends set_thinking_level after set_model, because the levels a model accepts depend on the model. With nothing picked, the bar reads Auto, nothing is sent, and pi keeps your own default. Pi does not report a default level, so the catalog does not claim one. The code-review Ask AI bar no longer shows the first listed level as if it were active, and it has an Auto item.

• Levels need pi 0.84.3 or newer. Before pi 0.84.3 (pi commit 2ff8ba622), RPC set_thinking_level also rewrote the user's global default thinking level in pi's settings. The provider now runs pi --version alongside model discovery. It offers and sends levels only on pi >= 0.84.3. On an older pi, or when the version cannot be read, there is no picker and nothing is sent. The version is kept as the provider's toolVersion. A session that wants a level before discovery has run uses the same probe, so pi --version runs once. Pre-0.84.3 set_model has the same global-persistence behaviour. A code comment notes it, and this PR does not change it.

• Model labels read Name (provider), like OpenCode's. Two providers that serve the same model stay distinguishable. If two rows would still have the same label, both fall back to provider/id. The review/guide launchers' Pi picker is not affected, because it reads pi --list-models separately.

• The launcher's PI_THINKING is now built from core's PI_THINKING_LEVELS and is no longer a second hand-written list. Its behaviour is unchanged: it still leaves out max, which --thinking accepts only on pi 0.80.6 and newer.

Tests

• packages/ai/providers/pi-sdk-models.test.ts: a fake pi (--version + --mode rpc), run against both the Bun and Node providers. It covers:
  • per-model levels, including max and a model with no off;
  • Name (provider) labels;
  • the level sent after set_model;
  • nothing sent on Auto;
  • on pi 0.84.2, no levels offered and none sent, even when a level arrives.
• packages/core/model-catalog.test.ts: piCatalogFromRpc (level derivation, label collisions) and the version gate.
• bun test packages/ai packages/core apps/pi-extension: all pass. DOM_TESTS=1 runs of review-editor and ui pass, except four DocBadges tests that fail only in the combined run. That file is not changed here, and it passes when run alone. bun run typecheck (re-vendors Pi, including the new pi-version.ts) passes.