Support Claude Opus 5.5 in the Anthropic backend

[MrJoy]

Opus 5.5 rejects forced tool use (tool_choice: type "tool" and "any" are not supported for this model.), and AnthropicBackend forced it on every call. Every role on Opus 5.5 failed on its first request.

The backend now starts forced. On that specific 400 it flips the instance to tool_choice: auto (with disable_parallel_tool_use), adds one system-prompt line naming the tool, and retries. set_model() resets it. Opus 4.8 and older never see a difference. Under auto the model can answer in text, so a response without a tool call now gets a correction and a retry through the existing 3-attempt loop instead of an immediate RuntimeError.

Reading guide: _stream_tool_call in src/pact/backends/anthropic.py is the whole mechanism. Both stream paths now go through it.

Why not strict: true or structured outputs: both require additionalProperties: false on every object. ComponentContract, ContractTestSuite, InterviewResult and ShapingPitch all have dict[str, X] fields, which that can't express. Going strict means reworking those schemas first. Pydantic validation plus the correction loop already covers schema validity.

Also in here:

• claude-opus-5-5 priced at $4/$20. The fuzzy prefix match in pricing_for_model was resolving it to claude-opus-4 at $15/$75. claude-opus-4-8 still hits the same wrong match; I left that alone.
• claude-opus-5-5 max-tokens cap raised to 128K. Thinking can't be turned off on 5.5 and it counts against max_tokens.
• stop_reason: "refusal" raises right away instead of burning two retries.

Verified live against the API with the new code, one run per model:

INFO pact.backends.anthropic: claude-opus-5-5 rejected forced tool_choice; using tool_choice auto
claude-opus-5-5 register ok: _RegisterResponse forced= False 2s
claude-opus-5-5 interview ok: risks 9 questions 7 21s
claude-opus-5-5 contract ok: duration_parser 2 functions 34s
claude-opus-4-8 register ok: _RegisterResponse forced= True 2s
claude-opus-4-8 interview ok: risks 6 questions 7 37s
claude-opus-4-8 contract ok: duration_parser 2 functions 23s

No missed-tool-call retries fired on 5.5 in that run. One run each, so that's a sample and not a reliability number.

Test suite: 27 failures, and main fails the same 27 in the same venv (missing optional deps: openai, tree-sitter, Slack/Linear). The new code adds no failures.

🤖 Generated with Claude Code

https://claude.ai/code/session_011FeZUmAZbyQWDJNAPkUA3S

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[MrJoy]

Does this maintain backwards compatibility with Opus 5 / 4.8?

[MrJoy]

Yes. The request those models get is byte-for-byte the same as before: forced tool_choice, same system prompt. The switch to auto only happens on the specific 400 that 5.5 returns, and neither Opus 5 nor 4.8 ever returns it.

Checked live against both with this branch:

claude-opus-5 register ok: _RegisterResponse forced= True 3s
claude-opus-5 interview ok: risks 10 questions 10 48s
claude-opus-5 contract ok: duration_parser 2 functions 49s
claude-opus-4-8 register ok: _RegisterResponse forced= True 2s
claude-opus-4-8 interview ok: risks 6 questions 7 37s
claude-opus-4-8 contract ok: duration_parser 2 functions 23s

forced= True after the first call means the switch never happened.

Three retry-loop changes do apply to every model, but none of them can fire on a normal forced-mode response:

1. No tool call in the response: retry with a correction instead of an immediate RuntimeError. Under forced tool_choice that only happens when output got cut off.
2. max_tokens with no tool call: this used to raise, and now it bumps max_tokens and retries. Before, the bump only happened when a truncated tool call was present.
3. stop_reason: "refusal" raises right away instead of retrying.

N.B. Opus 5 and 4.8 are both still mispriced at the Opus 4 rate ($15/$75) through the fuzzy match in pricing_for_model. That was already broken before this PR and isn't fixed here. Can add the entries if you want them in this one.

[MrJoy]

Added in 1d2b7fa: claude-opus-5 and claude-opus-4-8 at $5/$25, per the current Anthropic pricing table. claude-opus-4-7 goes through the same fuzzy match and still lands at $15/$75. Left it out since nobody asked, but it's a one-liner if you want it.

[MrJoy]

9b28e00 goes through the whole DEFAULT_MODEL_PRICING table. Every entry was checked against the vendor's own pricing page today.

Wrong before, fixed now:

Model	Was	Now
o3	$10 / $40	$2 / $8
gemini-2.5-flash	$0.15 / $0.60	$0.30 / $2.50
gemini-2.5-flash-lite	$0.075 / $0.30	$0.10 / $0.40
gemini-3-flash-preview	$0.15 / $0.60	$0.50 / $3.00

Missing before, and the fuzzy match was resolving them to the wrong rate:

Model	Resolved to	Now
claude-opus-4-7	$15 / $75 (Opus 4)	$5 / $25
claude-sonnet-5, claude-sonnet-5-5	$3 / $15 (Sonnet 4)	$2 / $10
claude-fable-5, claude-fable-5-1, claude-mythos-5, claude-mythos-5-1	$1 / $5 (Haiku fallback)	$10 / $50

gemini-3-pro-preview is gone. Google shut it down on 2026-03-09 and gemini-3.1-pro-preview ($2 / $12) replaces it. The old entry was wrong anyway, at $1.25 / $10.

claude-haiku-3 is retired and off Anthropic's pricing page. $0.25 / $1.25 was its last published price, so I left it in.

test_gemini_backend.py asserted the old 2.5 Flash price. Nobody noticed because it skips when google-genai isn't installed. Fixed it, and ran it with .[all-backends] installed.

[adaptcom]

Confidence Score: 2/5

Summary

Adds automatic tool-choice fallback and refreshes model pricing. A reproduced concurrency bug can abort parallel adoption, so the fallback needs correction before merge.

Important Files Changed

File	Overview
src/pact/backends/anthropic.py	Adds automatic tool-choice fallback, correction retries, refusal handling, and a higher token cap.
src/pact/budget.py	Updates model pricing and adds newer model entries.
tests/test_anthropic_tool_choice.py	Exercises sequential fallback, model reset, caching, and missing-tool retries.
tests/test_budget.py	Adds exact-match pricing coverage.
tests/test_gemini_backend.py	Updates Flash pricing expectations.
tests/test_validation_retry.py	Expects three attempts when tool calls are absent.

↻ Re-run review · View in Adapt