第二轮真实报告暴露的残留问题:
- agent: grep 要搜"符号族"而非单个字面(修 Apm.start 搜到、Apm.preStart 漏搜)
- prompt: 文件"本次新增"判定必须以 diff 为准(修把既有 AAR 误报为本次引入)
- prompt: 修复建议须标注已知陷阱(修建议 recycle 位图却不提渲染线程竞争)
- prompt: 极端/异常输入才触发的缺陷定级下探到 🟢(收敛过高定级)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

- AI_REVIEW_MODEL: gpt-5.5 -> gpt-5.6-sol(深度审查用最强推理旗舰)
- AI_REVIEW_TRIAGE_MODEL: gpt-5.4-mini -> gpt-5.6-luna(实测延迟不高于 mini 且更稳)
- 均可被 GitLab CI Variables 覆盖,模型 ID 依据本地 CLIProxyAPI /v1/models 实时列表核实
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

- agent: 时序/竞态、代码搬家类结论强制先 grep 取证
- prompt: 定级与证据强度挂钩,未验证假设不进必须立即修复
- prompt: 敏感数据外传做基线对比,同级字段不单独升级
- prompt: 不建议撤销有意设计;建议动作按修复/确认/优化分组
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Lets the deep-review LLM actively call read_file / grep / list_files to
verify cross-file impact instead of returning "请确认调用方" TODOs that
push the work back to humans. standard / triage paths unchanged; on any
agent-loop error the deep path automatically falls back to the previous
one-shot thinking call, so CI never breaks. Toggle via
AI_REVIEW_AGENT_ENABLED.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

- Updates the default `API_URL` and port from `8088` to `8317` in `ai_review/config.py` and documentation to align with CLIProxyAPI defaults.
- Replaces the custom `proxy/codex_proxy.py` and launchd plist with a new configuration example for `CLIProxyAPI` (`proxy/cliproxyapi.config.example.yaml`).
- Updates `README.md` with instructions for installing and configuring `CLIProxyAPI` via Homebrew, including OAuth login and API key setup.
- Adjusts environment variable documentation to reflect that `AI_REVIEW_API_KEY` is now required by the new proxy configuration.

Removes the Feishu notification step when there are no code changes to review. The process now simply prints a skip message and exits to reduce unnecessary notifications.

Introduces `find_symbol_references` in `ai_review/git.py` to detect deleted symbols within a diff and search for remaining references across the repository using `git grep`. This provides the LLM with factual context to determine if callers have been properly updated, reducing "hallucinated" risks regarding unsynchronized changes.

Key changes:
- Implements heuristic-based symbol extraction for deleted lines (supporting Kotlin, Java, Swift, JS/TS, etc.).
- Filters out references within the files already modified in the current diff.
- Appends a formatted "Symbol Reference Context" block to both standard and deep review prompts.
- Limits analysis to the first 8 symbols and 20 references per symbol to manage prompt size.

Improved the accuracy of risk level assessment from LLM responses by implementing a prioritized detection strategy:
1. Matches leading "✅" for skipped or passed reviews.
2. Specifically searches for an emoji (✅, ⚠️, or 🚨) within a "Risk Level" line.
3. Implements a fallback that checks for actual list items within the "High Risk" section to avoid false positives triggered by section headers.

Updates the default configuration and documentation to prioritize the local codex proxy over public cloud endpoints.

- **Config**: Changes default `AI_REVIEW_API_URL` to `http://localhost:8088/v1/chat/completions`.
- **Models**: Updates default `MODEL` and `TRIAGE_MODEL` to `gpt-5.5` and `gpt-5.4-mini` respectively to align with codex naming conventions.
- **LLM**: Refines thinking mode logic to send hint parameters (`enable_thinking`, `thinking_budget`) to OpenAI-compatible gateways while maintaining compatibility with proxies that ignore these fields.
- **Documentation**: Updates `README.md` with detailed instructions for codex proxy deployment, including model resolution rules and GitLab CI variable configuration.
- **Timeouts**: Increases default `AI_REVIEW_DEEP_TIMEOUT` to 600 seconds in the documentation.

Updates `call_llm` to return both the content and the actual model name returned by the API. This ensures that the review summary and notifications display the real model used (e.g., "codex" when using a local proxy) rather than static configuration values. Updated `main.py` to propagate and display these dynamic model names in the final report.

Updates the default `DEEP_TIMEOUT` value from 300 to 600 seconds in `ai_review/config.py` to allow more time for deep review processing.

Introduces `_strip_thinking` to remove reasoning segments from the LLM output. This handles `<think>` tags, residual `</think>` tags, and plain-text thinking prefixes to ensure only the final review content is returned. It also prioritizes `reasoning_content` when provided by OpenAI-compatible gateways.

The CU Cloud gateway base URL is /v1 but the actual endpoint needs
/chat/completions appended. Auto-detect and append when missing.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

The gateway returns HTTP 404 because qwen3.5-plus and qwen3-coder-plus
are not valid model IDs. Updated defaults to Qwen3.5-397B-A17B and
Qwen3-235B-A22B which are available on the CU Cloud gateway.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>