#162 fix(claude-code): show model thinking and cache reads, price cache reads in agent configs

Closes #99.

Claude Code now shows the model's thinking and reports cache reads in its usage. OpenCode, OpenClaw and Prime Agent now get the cache-read price, so a cached prompt costs what it should in their cost display.

Tested manually on prod: Claude Code over two turns (thinking shown, cache reads reported), and OpenCode's cost display.

---

Summary by cubic

Fixes Claude Code translation so model reasoning surfaces as thinking blocks and cache reads are reported, priced, and not double-counted.

• Streams reasoning as an unsigned thinking block; text closes it and vice versa.
• Cache reads move from input_tokens to cache_read_input_tokens, keeping the total unchanged.
• OpenCode, OpenClaw, and Prime Agent configs now use the catalog's cached-input price.

Summary by CodeRabbit

• New Features
  • Reasoning content now appears as thinking blocks in Anthropic responses, including streamed responses.
  • OpenCode, OpenClaw, and Prime Agent cost displays now include catalog-provided cache-read pricing when available.
• Bug Fixes
  • Anthropic usage reporting now accounts for cached input tokens without double-counting them as regular input.
  • Claude Code documentation clarifies unsigned thinking blocks and cache-token reporting.