v1.13.56

Upgrade Notes

• Trajectory evaluation now requires Python 3.11+ and uses ATIF v1.7 models with Detector/Verifier APIs. Consumers of the former Evaluator API must migrate to Verifier and TrajectoryAnalysisRun.
• Rebuild existing trajectory Dataset/Run artifacts from the original sessions and labels. Run weekly_report.py --week <YYYY-Www> to rebuild the selected week and its predecessor; use ccr_trajectory_report.py --no-history when rebuilding a single week. Diagnostic caches use a new schema.
• Routed model calls now share a bounded fallback budget, configured through routing.call_timeout_sec (default: 180 seconds).

New Features

• Build labeled trajectory datasets from local review sessions and GitHub/GitLab annotations, with weekly HTML/Markdown reports for review completion, finding quality, cost, and behavior patterns.
• Diagnose selected trajectories incrementally with resumable caches, explicit call budgets, bounded evidence context, and token-usage reporting. Trace known missing findings to the review stage where delivery stopped.
• Enrich code-search results with bounded symbol context and expose additional import and supertype facts to support evidence gathering.

Bug Fixes

• Constrain investigation tools during wrap-up and preserve explicit incomplete outcomes when an execution cannot finish. Routed fallback no longer stacks hidden SDK retry budgets.
• Keep language-analysis caches separate by backend and reuse parsed facts across queries.
• Correct weekly-report denominators and label-availability handling, surface dataset build failures, and reuse canonical analysis results for reports and diagnostics.
• Preserve label links and resolvable step evidence when adapting CCR exports to ATIF; derive measurements before detection and verification.

Full changelog

v1.13.46

Changes

• PR #194 feat: track malformed tool arguments
• PR #193 fix: normalize recoverable tool argument drift
• PR #192 docs: define which findings deserve attention
• PR #191 feat: stream findings as JSONL
• PR #190 fix: align code-search scope and regexp execution
• PR #189 fix: distinguish empty code-search scopes
• PR #188 feat: gate review on external evidence
• PR #187 fix: make review submissions incremental
• PR #186 fix: verify Review 2 premises before consequences
• PR #185 fix: harden review loop completion
• PR #184 feat: expose review session identity
• PR #183 feat: trace review pipeline timing in sessions
• PR #182 feat: make review context trajectory-aware
• PR #181 feat: stream review findings through the pipeline
• PR #180 feat: expose context compaction in sessions
• PR #179 perf: bound ignored review wrap-up turns
• PR #178 docs: describe code graph evolution
• PR #177 docs: define project knowledge and Review 3
• PR #176 perf: bound review unit formation cost
• PR #175 feat: add language-aware file outlines
• PR #174 refactor: make Unit the review aggregate
• PR #173 feat: converge unit review investigation
• PR #172 fix: preserve function ownership across diff hunks
• PR #171 feat: score trajectory health bottlenecks
• PR #170 feat: batch code searches by default
• PR #169 feat: make viewer execution lifecycle explicit
• PR #168 feat: batch file reads by default
• PR #167 feat: diagnose CCR trajectory inefficiency
• PR #166 feat: split trajectory evaluation by review stage
• PR #165 feat: evaluate CCR trajectories with trajectory harness

v1.9.1

Changes

• PR #128 feat(eval): add private review data collection workflow
• PR #127 chore: set ccr version to v1.9.1
• PR #126 fix(language): harden gotreesitter integration
• PR #125 feat(language): add in-process gotreesitter backend
• PR #124 refactor(language): centralize source analysis
• PR #123 feat(unit): add TypeScript function-level review
• PR #122 chore(labels): 收 compforge/devloop#81 的 ccr 标注
• PR #121 eval: label doctor MR 72 findings

v1.9.0

Changes

• PR #119 fix(file_read): return recoverable error on reversed line range
• PR #118 eval: harvest ccr:label replies from GitLab MR discussions
• PR #117 feat(review): findings 结构化 category/severity 字段
• PR #116 fix(tool): code_search 拒绝含 .. 的 file_patterns（仓外逃逸）
• PR #115 docs(eval): §11 review_team 三臂首轮记录（hostel 38 PR，reviewbench 首次全量）
• PR #114 chore(eval): #113 post-review finding 入册（debatable #stale）
• PR #113 docs(eval): labels.py 输出示例补 tags 字段 + 该 finding 入册（minor）
• PR #112 feat(eval): 打标词表补齐——病因 #tag + 漏报 ccr:missed + judge 逐条留痕
• PR #111 fix(reviewbench): PR #110 review 反馈——修 6 项，驳 2 项误报
• PR #110 feat(eval): reviewbench——corpus × arms 的 worksheet 评测编排
• PR #109 feat(eval): finding 人工标注统一约定——ccr:fp footer + ccr:label 回复 + 回收器
• PR #108 feat(board): post_bulletin 怀疑通道——Review Team v0.5（gate 默认关）
• PR #107 fix(board): flag facts 从未上板——按各 tool 的实际参数名解析路径
• PR #106 fix: 运行期诊断统一走 stdout.Err()，防止污染 --format json 输出

v1.8.0

Changes since v1.7.2

Features

• Review Team v0 — 共享案情板（Board）：多 reviewer 间共享案情上下文，feature gate review_team 默认关闭 (#104)

Fixes

• file_read rename awareness + fatal-run JSON contract — file_read 对 rename 文件的感知修复；fatal 运行时 stdout 仍保证输出且仅输出一个 JSON 对象（status: failed），下游解析器不再误报自身 parse 错误 (#105)

Docs

• reviewer 信任边界：只读不变量约定 + Review Team 输入隔离背书 (#103)
• cross-unit 定稿为 Review Team 设计（Board/Bulletin/cross_check）(#91)

v1.7.2

Briefing / Debrief（unit 的交底与复盘）

• Briefing（#86）：briefer 协议按 scope 定预载材料——own source、changed symbols 的 usage-sites 预 grep、大文件 ranged 降级、callchain 邻居函数体；token guard 分级降级
• Debrief（#87 #90 #100）：每 unit 一条终局记录（outcome/formed/降级/成本聚合）+ run manifest（gates/版本/预算/CCR_EVAL_TAG）；unit 生命周期显式化——scope Close 自产 debrief，异步计数归零自动 finalize，panic 可统计
• Finding 锚点（#88）：指纹（path+content，刻意不含行号）+ symbol_id + git_head——后验扫描的钥匙

Message model（review 领域消息，参考 pi）

• 会话货币 []msg.Msg，wire 只在 lowering 边界（#92）；File 消息带 path+range 身份（#93 #99：身份=内容+配对，形态归 Lower()）
• file_dedup（#93）：后读覆盖先读原地 stub；file_evict（#94）：token 压力下先驱逐可再生的文件内容再 LLM 压缩
• typed_briefing（#95，#102 转默认开）：预载 File 化，回归集 A/B 实测 timeout 9→3、成本持平

Eval 基建

• posterior.py（#89）：finding 后验扫描（line_touched 实锤候选），自验精确指回修复 commit
• replay.py + ccr-self corpus（#96）：固定回归集重放，corpus×arms 矩阵——首次全量应用即发放 typed_briefing 翻转许可证，并让 ccr 自审抓到一条真实快照竞态（#101 修复）

其它

• pkg/stdx 孵化毕业为 go-stdx 独立仓，ccr 切外部依赖（#97 #98）
• Review Team 设计定稿（Board/Bulletin/动态 cross_check，v0 待实现）：docs/cross-unit.md（#91）
• eval/README 新增 §7-§9：trajectory 统计驱动、采集面、翻转许可证方法论

v1.7.1

v1.7.1 — empty-search treatment, typed call graph, truncation wrap-up

Eval-driven review-quality release:
- code_search suggests real identifiers on zero hits (#73)
- ranked repo symbol map injected per run, Go+Python backends (#74-#76, #80)
- budget wrap-up turn + unit_incomplete marking (#77)
- code_comment keeps scope-member paths (#79)
- typed Go call graph for caller/callee clues and chain merge (#81, #82)
- eval/ methodology + codegraph design docs (#72, #83)

Measured: empty searches 60%→0 on worst-baseline PRs; caller
neighborhoods exact vs 5/8-wrong under grep.

v1.7.0

v1.7.0 — upstream sync baseline

Engine aligned with upstream open-code-review v1.7.0-era fixes:
panic isolation in fan-out goroutines (#69, upstream #182),
rule.json file references with symlink-safe loading (#70 + #71,
upstream #226).