fix(proxy): add provider-aware Anthropic thinking policy

[DeliciousBuding]

Summary

This is a smaller provider-aware alternative to #4210.

Instead of replacing proactive placeholder injection with a reactive retry path, this PR keeps the current default behavior for compatibility and adds an explicit provider metadata policy for Anthropic-format tool-use thinking history:

• auto / unset: keep historical behavior for reasoning-vendor hints.
• preserve_only: preserve real thinking blocks and safe structural cleanup only; do not synthesize {"type":"thinking","thinking":"tool call"}.
• placeholder_always: explicitly opt a provider into placeholder synthesis, even without a vendor hint.

The intent is to stop treating synthetic thinking placeholders as a hardcoded behavior for every reasoning-vendor Anthropic path. They become a provider capability / compatibility choice.

Context collected

• fix(proxy): normalize DeepSeek Anthropic tool thinking history #3203 introduced DeepSeek/MiMo Anthropic tool-thinking history normalization so assistant tool_use turns always include a plain thinking block.
• Synthetic "tool call" thinking placeholders may break long DeepSeek-compatible Anthropic sessions #4208 reports long DeepSeek-compatible Anthropic sessions where synthetic "tool call" thinking placeholders may correlate with visible text being emitted inside thinking blocks and session freezes.
• [BUG] Anthropic protocol: thinking-only responses without text blocks — end_turn with no visible output (regression in v3.16.x) #3645 reports thinking-only responses with no visible text blocks in DeepSeek Anthropic protocol sessions.
• [Bug] Anthropic System Message Normalization breaks DeepSeek prefix cache (90% → 75%) #3934 reports request transformations affecting DeepSeek prefix-cache behavior; fix(proxy): 规范化 Anthropic system 消息 #3775 was later reverted in main, reinforcing that Anthropic-format proxy transforms should be minimal and provider-aware.
• Claude Code resume can replay invalid signed thinking blocks through Claude provider #3930 covers a nearby but different replay problem: invalid signed historical thinking blocks.
• feat(proxy): strip effort params when thinking:disabled for DeepSeek endpoint #4239 recently added a very targeted DeepSeek official endpoint transform for thinking: disabled + effort conflicts. That is a separate issue, but it follows the same direction: precise provider-scoped behavior instead of broad request rewriting.

Why this shape

A pure reactive retry path is useful as a compatibility fallback, but it has drawbacks:

• it depends on upstream error text staying stable;
• it adds a failed round trip for vendors that still require backpass;
• some relays wrap or normalize errors before cc-switch can match them;
• it still treats placeholder synthesis as an error-path repair rather than an explicit provider capability.

This PR does not claim that every DeepSeek-compatible Anthropic endpoint has relaxed thinking backpass requirements. It only makes the existing synthetic placeholder behavior configurable and safe to disable for providers where it is harmful.

Private A/B validation

I also ran a sanitized private A/B against a DeepSeek-compatible Anthropic relay. No private endpoint names, machine names, or credentials are included here.

• Auth/model probe: bearer auth worked for the test model; x-api-key was rejected by that relay.
• Anthropic /v1/messages, small tool-use history:
  • no synthetic placeholder: HTTP 200, response contained thinking + text blocks.
  • with synthetic "tool call" placeholder: HTTP 200, response contained thinking + text blocks.
• Anthropic /v1/messages, non-streaming long-history fixture (~63 KB request, 12 tool turns, ~21.9k input tokens):
  • no synthetic placeholder: HTTP 200, thinking + text, stop_reason=end_turn.
  • with synthetic placeholder: HTTP 200, but one run emitted only thinking before max_tokens.
• Anthropic /v1/messages, streaming stress fixture (~304 KB request, 18 tool turns, ~107k input tokens):
  • no synthetic placeholder: HTTP 200, thinking + text, stop_reason=end_turn.
  • with synthetic placeholder: HTTP 200, thinking + text, stop_reason=end_turn.

Interpretation: this is not proof that disabling placeholders fixes every long-context freeze. It does confirm that this tested DeepSeek-compatible Anthropic path accepts tool-use history without synthetic placeholders, and that synthetic placeholders can still affect the output block mix / budget under longer histories. That is why this PR exposes a provider-scoped policy instead of changing the global default.

Changes

File	Change
src-tauri/src/provider.rs	Add AnthropicToolThinkingPolicy and ProviderMeta.anthropicToolThinkingPolicy.
src-tauri/src/proxy/providers/claude.rs	Parameterize placeholder synthesis while keeping signature stripping and redacted_thinking conversion; soften outdated comments that implied all DeepSeek-compatible endpoints still require placeholder replay.
src/types.ts	Add frontend TypeScript metadata type for the new policy.

Behavior

• Existing providers are unchanged because unset policy maps to historical auto behavior.
• preserve_only prevents missing / empty thinking blocks from being filled with synthetic "tool call", but still strips invalid signature fields from real thinking blocks.
• placeholder_always allows explicit opt-in for a provider that requires synthetic placeholders even when it does not match the built-in vendor hints.

Verification

Passed:

cargo test --manifest-path src-tauri/Cargo.toml --lib proxy::providers::claude -- --nocapture
# 66 passed

cargo fmt --manifest-path src-tauri/Cargo.toml --check
cargo check --manifest-path src-tauri/Cargo.toml --lib
pnpm typecheck
pnpm format:check

Also ran:

cargo test --manifest-path src-tauri/Cargo.toml --lib

Result: 1654 passed, 8 failed, 2 ignored. The failures are outside this PR's touched area (codex_history_migration, commands::misc::anchored_upgrade_windows, database::dao::usage_rollup) and appear unrelated to Anthropic provider normalization.

Draft notes

This is intentionally draft because the maintainer may prefer one of two follow-ups:

1. expose anthropicToolThinkingPolicy in the provider UI, or
2. seed preserve_only / placeholder_always on specific presets after more provider verification.

I kept this PR backend-focused to separate the protocol behavior from UI/preset policy decisions.

[farion1231]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. 🎉

Reviewed commit: c97604e669

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".