Fix a configuration bug with Opus 4.7 adaptive thinking#3286
Conversation
|
Could you also re-generate the models list to pick up the AWS Bedrock models? I don't want to do it in this PR as it's noise and unrelated. |
| case "xhigh": | ||
| return modelId.includes("opus-4-6") || modelId.includes("opus-4.6") ? "max" : "high"; | ||
| return modelId.includes("opus-4-6") || | ||
| modelId.includes("opus-4.6") || | ||
| modelId.includes("opus-4-7") || | ||
| modelId.includes("opus-4.7") | ||
| ? "max" | ||
| : "high"; |
There was a problem hiding this comment.
Opus 4.7 supports a real xhigh level now
https://www.anthropic.com/news/claude-opus-4-7#:~:text=More%20effort%20control,or%20xhigh%20effort.
There was a problem hiding this comment.
Pi doesn't expose a value in between high and xhigh. I think semantically Pi's xhigh should map to the highest available value which is max here.
There was a problem hiding this comment.
Pi doesn't expose a value in between
highandxhigh. I think semantically Pi'sxhighshould map to the highest available value which ismaxhere.
For Opus 4.7, xhigh and max are too distinct, so it should probably support both for more flexibility.
There was a problem hiding this comment.
If we have to pick between xhigh and max for the top thinking level available in pi, then I agree that we should choose max. But why not expose both?
There was a problem hiding this comment.
Hmm yeah, it's a bit debatable whether Pi's xhigh should map to xhigh or max. Seems like fairly diminishing returns to be honest going from xhigh to max.
@badlogic or @mitsuhiko any thoughts on this?
There was a problem hiding this comment.
well at the very least xhigh should not map to max given that graph that was posted.
There was a problem hiding this comment.
Yeah, agree on that given the graph. I'll update.
There was a problem hiding this comment.
@mitsuhiko Opus 4.7 now maps xhigh -> xhigh
There was a problem hiding this comment.
yeah, it's a mess. pi-ai package has a ThinkingLevel for everything must map to. providers keep adding the weirdest shit, with partial support in some models.
maybe we can trick the type system into per model thinking levels #3208 somehow, so streamSimple completions work correctly. it still sucks.
There was a problem hiding this comment.
What I did in my local fork is map minimal -> low, low -> medium, medium -> high, high -> xhigh, xhigh -> max
Which is pretty terrible but at least you can access all the levels
|
I looks like commit 6dca0c3 have removed the effort |
Consolidates 62 upstream commits (v0.67.4 -> v0.67.68) into the fork. Preserves all fork-local work (OAuth cloak, antigravity pool, context tiers, OSC 8 file hyperlinks, gemini3 sig skip, codex_cli_rs headers, compaction context hooks, gpt-5.4 models). Conflict resolutions (4 files): - packages/ai/src/providers/google-gemini-cli.ts: layered upstream's onResponse hook call into our pool-aware retry loop; kept our finally-block pool.saveNow() path. - packages/tui/src/utils.ts: adopted upstream's AnsiCodeTracker OSC 8 impl wholesale. It's cleaner than our f16dfd5 version (preserves hyperlink across SGR reset via separate clear(), emits ST-terminated sequences). Dropped our 7d75fee SEGMENT_RESET fix since upstream's design never introduced the bug. - packages/tui/src/components/markdown.ts: kept our resolveHref hook and wrapHyperlink helper but gated OSC 8 emission on getCapabilities().hyperlinks; switched to upstream's hyperlink() helper for consistency. Merges file path linking (ours) with terminal capability detection (upstream earendil-works#3248, #30a8a41f). - packages/tui/test/{markdown,wrap-ansi}.test.ts: updated our OSC 8 assertions to ST-terminated sequences (\x1b\\) to match new emitter; added setCapabilities({hyperlinks:true}) gating where needed; afterEach resetCapabilitiesCache to isolate test state. Dependency upgrades from upstream: - @anthropic-ai/sdk 0.73.0 -> 0.90.0 - @aws-sdk/client-bedrock-runtime 3.983.0 -> 3.1030.0 - @mistralai/mistralai 1.14.1 -> 2.2.0 Upstream features folded in: - fix(ai): Opus 4.7 adaptive thinking + xhigh effort (earendil-works#3286) - feat(ai): thinkingDisplay option (summarized/omitted/raw) - feat(coding-agent): after_provider_response hook (earendil-works#3128) - fix(ai): trust requested Codex service tier (earendil-works#3307) - feat(bedrock): Bearer token auth for Converse API (earendil-works#3125) - feat(agent,coding-agent): per-tool executionMode override (earendil-works#3345) - feat(tui): OSC 8 hyperlinks with terminal capability detection (earendil-works#3248, #30a8a41f) - feat(coding-agent,tui): argument-hint frontmatter in prompts (earendil-works#2780) - Plus ~40 smaller fixes across ai/coding-agent/tui/agent. Verification: - packages/tui: 559/559 tests pass - packages/ai: 144/144 OAuth cloak tests pass; all fork-local suites pass (antigravity-pool, context-tiers, gemini3-unsigned-tool-call, codex-stream, supports-xhigh, etc.). Remaining test failures (13) are pre-existing live-API E2E tests (no credentials in this env). - packages/agent: 39/39 tests pass. - packages/coding-agent: 1020/1021 tests pass; 1 flaky bash timeout test (passes in isolation), unrelated to merge. - All four packages build cleanly.
No description provided.