Context
Anthropic's Opus 4.6 and Sonnet 4.6 use adaptive thinking — the model decides when and how much to think based on task complexity, rather than being given a fixed token budget. The upstream pi-ai provider (v0.58.1) already supports this: when a 4.6 model is selected, it sends thinking: { type: "adaptive" } instead of thinking: { type: "enabled", budget_tokens: N }, and maps our thinking levels to effort levels (low/medium/high/max).
The plumbing works. But the add-in UI doesn't distinguish between budget-based and adaptive thinking — users see the same level picker (off/low/med/high/max) with no indication that the underlying behavior is different.
Current behavior
- Thinking level picker shows:
off → low → med → high → max
- Labels:
THINKING_LEVEL_LABELS = { xhigh: "Max" }
- Hints:
THINKING_LEVEL_HINTS = { xhigh: "Slowest — deepest reasoning" }
- Status bar:
xhigh → "max"
- Under the hood: pi-ai maps these to
{ type: "adaptive" } + effort levels for 4.6 models
What could improve
1. Thinking level hints should mention adaptive behavior for 4.6 models
Currently "Slowest — deepest reasoning" doesn't tell the user that the model will adaptively choose how much to think. Something like "Adaptive — model chooses reasoning depth" for the default level, and "Adaptive — maximum reasoning effort" for xhigh.
2. Consider making "adaptive" the recommended default for 4.6 models
The current default for reasoning-capable models is "high" (src/taskpane/init.ts:908). For adaptive-thinking models, a dedicated "adaptive" level (or just defaulting to medium and letting the model decide) may produce better cost/quality tradeoffs than always-high.
3. Status bar could show thinking mode
E.g. "adaptive: high" vs "budget: high" — or just an icon/indicator when adaptive mode is active.
4. Thinking duration display
src/compat/thinking-duration.ts shows elapsed thinking time. For adaptive thinking, the duration is more variable — consider showing whether the model chose to think at all (some simple queries may skip thinking entirely in adaptive mode).
Key files
src/taskpane/keyboard-shortcuts.ts — getThinkingLevels(), THINKING_COLORSsrc/taskpane/status-popovers.ts — THINKING_LEVEL_LABELS, THINKING_LEVEL_HINTSsrc/taskpane/status-bar.ts — status bar thinking displaysrc/taskpane/init.ts:908 — default thinking level selectionsrc/compat/thinking-duration.ts — thinking time displaynode_modules/@mariozechner/pi-ai/dist/providers/anthropic.js — supportsAdaptiveThinking(), effort mappingnode_modules/@mariozechner/pi-ai/dist/models.js — supportsXhigh()
Context
Anthropic's Opus 4.6 and Sonnet 4.6 use adaptive thinking — the model decides when and how much to think based on task complexity, rather than being given a fixed token budget. The upstream pi-ai provider (v0.58.1) already supports this: when a 4.6 model is selected, it sends
thinking: { type: "adaptive" }instead ofthinking: { type: "enabled", budget_tokens: N }, and maps our thinking levels to effort levels (low/medium/high/max).The plumbing works. But the add-in UI doesn't distinguish between budget-based and adaptive thinking — users see the same level picker (
off/low/med/high/max) with no indication that the underlying behavior is different.Current behavior
off → low → med → high → maxTHINKING_LEVEL_LABELS = { xhigh: "Max" }THINKING_LEVEL_HINTS = { xhigh: "Slowest — deepest reasoning" }xhigh → "max"{ type: "adaptive" }+ effort levels for 4.6 modelsWhat could improve
1. Thinking level hints should mention adaptive behavior for 4.6 models
Currently "Slowest — deepest reasoning" doesn't tell the user that the model will adaptively choose how much to think. Something like "Adaptive — model chooses reasoning depth" for the default level, and "Adaptive — maximum reasoning effort" for xhigh.
2. Consider making "adaptive" the recommended default for 4.6 models
The current default for reasoning-capable models is
"high"(src/taskpane/init.ts:908). For adaptive-thinking models, a dedicated "adaptive" level (or just defaulting tomediumand letting the model decide) may produce better cost/quality tradeoffs than always-high.3. Status bar could show thinking mode
E.g. "adaptive: high" vs "budget: high" — or just an icon/indicator when adaptive mode is active.
4. Thinking duration display
src/compat/thinking-duration.tsshows elapsed thinking time. For adaptive thinking, the duration is more variable — consider showing whether the model chose to think at all (some simple queries may skip thinking entirely in adaptive mode).Key files
src/taskpane/keyboard-shortcuts.ts—getThinkingLevels(),THINKING_COLORSsrc/taskpane/status-popovers.ts—THINKING_LEVEL_LABELS,THINKING_LEVEL_HINTSsrc/taskpane/status-bar.ts— status bar thinking displaysrc/taskpane/init.ts:908— default thinking level selectionsrc/compat/thinking-duration.ts— thinking time displaynode_modules/@mariozechner/pi-ai/dist/providers/anthropic.js—supportsAdaptiveThinking(), effort mappingnode_modules/@mariozechner/pi-ai/dist/models.js—supportsXhigh()