fix(anthropic): make the SDK compatible with the anthropic 1.x client (temperature crash + strict-endpoint nulls)

[philipph-askui]

Summary

Makes the SDK compatible with the newer anthropic client (1.x). Fixes to the Anthropic messages path (src/askui/models/anthropic/messages_api.py), which both AnthropicVlmProvider and AskUIVlmProvider route Claude through.

The old problem

Our dependency was unbounded (anthropic>=0.86.0), so a fresh install pulls the latest anthropic (1.2.0). But anthropic 1.x removed the temperature parameter from beta.messages.create, and create_message forwarded temperature= to it unconditionally — even the omit sentinel when the user set none. Result: the agent crashes on the first model call, no matter the config:

TypeError: Messages.create() got an unexpected keyword argument 'temperature'

This is why it fires for everyone on a fresh install (e.g. it took down a downstream project that never sets temperature at all).

Relationship to #313: #313 is a one-line stopgap that caps the dependency to anthropic<1, keeping installs on the old (working) 0.x client. This PR is the durable fix — it makes the code work with anthropic 1.x and moves the constraint the other way (anthropic>=1,<2). The two therefore change the same dependency line in opposite directions: this PR supersedes #313, so once it lands #313 can be closed (or, if #313 shipped first as a hotfix, this PR flips the cap forward onto 1.x).

1. temperature crash → gate by model

Root cause (confirmed against the real API): the adaptive-thinking Claude generation deprecated sampling params — from Sonnet 5 / Opus 4.7 onward the API rejects temperature outright (400 "temperature is deprecated for this model."), and even on Sonnet 4.6 / Opus 4.6 a non-default temperature is rejected while adaptive thinking is active. Newer clients also removed the typed temperature param entirely.

Fix: decide per request whether to send temperature, using the SDK's per-model capability check accepts_sampling_params(model_id) plus a request-level adaptive-thinking check:

• Sampling-capable + not adaptive thinking (Sonnet 4/4.5/4.6, Opus 4.1/4.5/4.6, Haiku 4.5, …) → forward temperature in the request body via extra_body (works regardless of client version).
• Sampling-deprecated models (Sonnet 5, Opus 4.7/4.8, Fable 5) or adaptive thinking on → drop it and warn once per model.

This preserves backward compatibility (incl. the Android agent's deterministic temperature=0.0 via make_non_thinking_settings) and avoids per-call log spam, while never crashing.

2. Omit null fields when serializing content blocks

from_content_block used model_dump(), serializing unset optionals as explicit null (cache_control: null, citations: null). Real Anthropic and AskUI tolerate this, but stricter Anthropic-compatible endpoints (e.g. OpenRouter's) reject it with cache_control: expected object, received null. Fixed with exclude_none=True (verified it does not strip Nones inside tool input dicts).

Dependency / CI

• Require anthropic>=1,<2 (plus the bedrock/vertex extras): the declared constraint now matches the 1.x client the code targets, and <2 prevents the next major from silently breaking users the way 1.x did. Lock resolves 1.13.0 (was 1.2.0 before the 2026-10-14 update).
• Bumped anthropic in pdm.lock to 1.13.0 so CI runs against what users actually get on a fresh install (the lock previously pinned 0.116.0, masking the crash).
• Real-signature guard test: binds the SDK's actual create() kwargs (with/without temperature, sampling & non-sampling models) against the real installed client signature.

Update 2026-10-14 — merged main, tested with Claude Haiku 5.5

• Merged origin/main (14 commits, now on top of 0.47.0). Conflicts only in pyproject.toml / pdm.lock: fix(deps): cap anthropic to <1 to prevent temperature-kwarg crash #313's anthropic<1 cap is merged on main, so this PR now flips it to >=1,<2 as planned — fix(deps): cap anthropic to <1 to prevent temperature-kwarg crash #313 is superseded once this lands. The lock was taken from main and re-locked with --update-reuse: only anthropic moves (0.116.0 → 1.13.0, plus its new deps httpcore2, httpx2, httpx2-jsfetch, truststore); every other pin stays on main's versions.
• CI checks locally on anthropic 1.13.0: format --check, lint, typecheck:all, test:unit (830 passed) — all green.
• Live, AskUI API (plain / adaptive thinking / temperature=0.3 / screenshot + tool-use two-turn loop), all PASS:
  • claude-haiku-5-5 on dev and prod; claude-sonnet-5, claude-opus-5-5, claude-sonnet-4-6 on dev.
  • The temperature=0.3 case returned 400 on released 0.47.0 for Haiku 5.5 and Sonnet 5 (temperature is deprecated for this model); with this PR it passes.
• Outgoing request checked (spy on beta.messages.create): Haiku 5.5 + temperature → dropped, one warning per provider instance; Sonnet 4.6 + temperature → sent as extra_body.temperature; Sonnet 4.6 + adaptive thinking → dropped with warning; Android non-thinking settings → Haiku 5.5 sends no temperature, Sonnet 4.6 sends 0.0. Never a typed temperature kwarg.
• OpenRouter (AnthropicVlmProvider / AnthropicImageQAProvider with base_url="https://openrouter.ai/api", auth_token=<key>): plain, agent thinking (effort="medium"), Android non-thinking, temperature=0.3, screenshot + tool-use loop and get() all PASS for claude-haiku-5-5 and anthropic/claude-haiku-5.5; plain/thinking/non-thinking/temperature also pass for anthropic/claude-haiku-4.5, anthropic/claude-sonnet-4.6, anthropic/claude-sonnet-5.
  • On main/0.47.0 every request with an image (tool loop, get()) fails there with 400 invalid_union — the error points at [n].cache_control: expected object, received null, which the exclude_none=True change here fixes.
• Not covered: the direct Claude API (no credits on the test key).
• OpenRouter dotted ids now classified correctly (commit 4b89cfb): thinking.py matched dash prefixes only, so anthropic/claude-haiku-4.5 got adaptive instead of budget thinking and …-sonnet-4.6 / …-haiku-4.5 had temperature dropped. _normalize now maps a dot between two digits to a dash (Bedrock's -v1:0 untouched); 9 new unit cases.

Re-test 2026-10-14 after the dotted-id fix (also contains #318's Haiku 5.5 defaults)

• Locally: format --check, lint, typecheck:all, test:unit (839 passed) — green.
• OpenRouter, 7 scenarios each (plain, agent thinking, Android non-thinking, temperature, screenshot + tool loop, get(), pricing) for the defaults (no model_id), anthropic/claude-haiku-5.5, anthropic/claude-haiku-4.5, anthropic/claude-sonnet-4.6, anthropic/claude-sonnet-5: 35/35 PASS. Dotted Haiku 4.5 now sends budget thinking and temperature: 0.0 on non-thinking runs.
• AskUI API, dev + prod: defaults of all three providers and claude-haiku-5-5 (plain / thinking / temperature / tool loop), plus claude-sonnet-4-6 on dev: 22/22 PASS.

Testing

• Unit tests: temperature via extra_body for sampling-capable models, dropped (deprecated model or adaptive thinking) + warn-once for others, never a typed kwarg; serialization emits no nulls; real-signature guard.
• pdm run qa:fix clean; full unit suite green on anthropic 1.2.0.
• Verified end-to-end via a manual playground: act() across providers (anthropic/askui/openrouter) × models (Sonnet 4.5/4.6/5) × temperatures — 36/36, with per-cell confirmation of whether temperature was actually sent or safely dropped.

🤖 Generated with Claude Code