[josh87523]

---

title: RTK pytest 结果真实性修复合并过程
status: source_ready
owner: Camus
template_id: report.merge_process.v1
template_source: team-ai-pack/templates/merge-process-template.md
related_prd: not_applicable
related_technical_plan: not_applicable
related_test_plan: not_applicable
related_test_set: not_applicable
last_updated: 2026-08-10

合并过程：RTK pytest 结果真实性修复

本改动只修复 pytest 压缩输出的结果判定，并复用捕获层的截断事实；不增加通用超时、不恢复 Safe Runner child deadline、不触碰业务 runtime。

产品汇报（与 Multica 正文对齐）

需求对象与用户价值

• 问题：pytest 已经收集或执行测试，但摘要未被识别时，RTK 会把“无法判断”误报成“没有测试”，让用户误以为测试没有运行。
• 用户价值：用户继续得到真实退出码、可识别的通过/失败/跳过/收集数量、子测试数量和失败位置；摘要不完整时会明确看到“无法解析”，不会得到编造结论。
• 如果不改：长时间运行、静默输出、摘要格式变化或捕获截断都会继续制造错误 no-tests 结论，掩盖真实失败。

目标边界

• In scope：pytest 单一解析主路、直接 pytest 执行的退出码与 stdout 截断事实传入、摘要识别、计数、失败位置和原始日志路径 fixtures。
• Out of scope：Safe Runner、通用超时、10 MiB 捕获上限数值、其他 test/lint wrapper 的行为、业务 runtime、生产发布。
• 历史相关卡 / 依赖：PR Add moon filter — route underlying-tool output through existing rtk filters #1877 的本地异常证据；当前代码基于 origin/develop 的 RTK pytest wrapper。

当前进度判断

• 已完成：红灯复现、根因确认、最小解析修复、24 条 pytest 夹具、捕获截断观测 fixture、直接二进制退出码/失败位置/无法解析日志验证、两轮独立复核问题修正。
• 进行中：本地提交、远程 PR、PR CI 和主线读回。
• 未完成 / 缺口 / 阻塞：本地全量测试仍有 9 个未触达的 curl/SQLite 基线环境失败；远程 CI 尚未消费本提交。

RCA / 解决判断

• 现象：原解析器五类计数全为零时直接输出 No tests collected；摘要缺失、格式未识别和捕获截断都会进入这条分支。
• 根因或判断：摘要缺失不等于无测试；原逻辑没有保留“是否见到摘要”、collected 数量或真实退出码的判定上下文。
• 处理或方案：直接 pytest 改走带真实退出码的既有 runner 入口；解析器单一路径记录 collected、摘要和显式 no-tests 信号，只有显式 no-tests 或 collected 0 + pytest exit 5 才输出 no-tests，其余降级为无法解析并保留退出码与失败位置。
• 验证或待决策点：24 条解析夹具、捕获层 10 MiB 边界 fixture、49 条 Python 命令测试、Clippy 和真实二进制 pass/fail/no-tests/exit 4 已通过；最终独立复核、PR exact-head CI 和目标分支读回仍待完成。

完整改动逻辑

• 改动前：pytest 进入捕获 runner，过滤器只从摘要计数推断结果；所有零计数都被压成 no-tests。
• 触发条件：直接执行 rtk pytest ... 时，既有 runner 把真实子进程退出码传给 pytest 解析器；管道输入继续复用兼容入口。
• 唯一主路径：捕获层提供 stdout 与“是否截断”事实 → 清理 ANSI → 读取 collected → 识别唯一终态摘要 → 校验摘要与退出码 → 解析 passed/failed/skipped/xfail/xpass/error/subtests → 输出压缩结果。
• 阻塞 / 失败路径：pytest exit 5 且 collected 不为正才输出 no-tests；无摘要、多摘要、截断、exit 0/1 与摘要冲突、或 exit 2/3/4 异常退出时统一输出“无法解析”、真实退出码、已知 collected、失败位置和原始日志路径。
• 生产者与消费者：pytest 子进程产生 stdout/stderr 与退出码；既有 stream 捕获层额外暴露 stdout 是否超过既有 10 MiB 上限；runner 仅为 pytest 传入该事实并用 stdout+stderr 计算输出预算；main 继续消费真实返回码，pipe_cmd 继续消费无退出码兼容入口。没有新增旁路 writer 或业务状态文件。
• 兼容与迁移：保留原有 xfail/xpass、失败摘要和管道入口；ANSI、quiet-mode、单等号和子测试摘要进入同一解析主路。
• 回滚：回退本 PR 的 squash commit，恢复原 pytest 过滤器；不需要迁移数据或切换运行时配置。

交付物、现阶段结论与下一步

• 交付物 / 证据索引：src/cmds/python/pytest_cmd.rs；outputs/rtk-pytest-truth/pytest-regression-final3.log；python-consumers-final2.log；clippy-final3.log；all-tests-final.log。
• 现阶段结论：source 层和本地直接消费者层证明 pytest 不再把摘要缺失误报为 no-tests；尚未证明 PR CI、main 或线上消费。
• 下一步：提交并推送分支，创建 PR，等待 required CI，按 exact-head 合并并回读 main；runtime/online regression 不适用。

Discovery Gate

• task classification: C - 已有 wrapper 的局部 bug 修复
• research_status: not_required
• research_depth: baseline
• formal_research_reason: not_applicable
• internal_sources: src/cmds/python/pytest_cmd.rs; src/core/runner.rs; src/core/stream.rs; src/cmds/system/pipe_cmd.rs; CLAUDE.md
• external_sources: not_applicable
• source_intake_artifact: not_applicable
• research_conclusion: 复用现有 pytest runner 和 tee 路径，只收紧解析判定
• research_gap: none
• 调研结论: 复用现有主路；不新增通用 timeout、旁路或状态文件
• authority_plan: not_applicable
• test_plan: not_applicable
• test_set: not_applicable
• acceptance_registry: not_applicable
• mechanism_key: not_applicable
• metadata_note: 运行时退出码由既有 runner 产生；本 PR 只改变 pytest 输出层的可信判定和摘要保留。
• related surfaces: source / pytest direct caller / pipe caller / unit fixtures / local binary smoke / PR CI / main
• contracts / fixtures / smoke evidence: pytest parser fixtures and direct binary smoke recorded under outputs/rtk-pytest-truth/
• surface-freeze-exempt: not_applicable；只修改一个已有解析器文件和一个合并过程记录
• uncertainty / decision boundary: 10 MiB 捕获上限仍是既有合同；本 PR 只把捕获层已有的 stdout 截断事实暴露给 pytest 单一消费者，不改变上限、截断策略或其他 wrapper 行为。
• 预估命中率依据: not_applicable
• 误报代价: not_applicable
• 退出条件: not_applicable

Change Summary

• User / project value: 让 RTK pytest 输出只报告已被底层摘要或退出码证明的结论。
• Actual changed behavior: 摘要缺失或不可识别时不再输出 No tests collected；直接 pytest 解析获得真实退出码，并保留 collected、子测试和失败位置。
• In scope: src/cmds/python/pytest_cmd.rs、src/core/stream.rs 的 stdout 截断观测字段、src/core/runner.rs 的 pytest 专用 capture-aware 入口与本合并过程记录。
• Out of scope: Safe Runner、超时、捕获上限数值、其他 wrapper 行为、hooks、业务 runtime、生产部署。
• Touched source / callers / tests / config / docs: pytest source and fixtures；stream 只暴露已有截断事实，runner 新入口仅由 pytest 消费，既有调用方默认行为不变。
• Runtime or external impact: none
• external_write: not_applicable
• external_endstate: pending_post_merge

Validation Evidence

• targeted local / unit: pytest parser 24 passed; stream 10 MiB 截断边界 fixture passed; Python command consumer suite 49 passed; cargo fmt --all -- --check; cargo clippy --all-targets passed。
• selected stable test cases: normal pass, failure, explicit no-tests, collected-but-summary-missing, quiet/ANSI/non-standard summary, silent interval then pass, 10 MiB 真实捕获截断、伪终态摘要、exit 2/3/4、stderr-only 参数错误、xfail/xpass 和 mixed subtest counts。
• change-specific tests: outputs/rtk-pytest-truth/pytest-regression-final3.log
• smoke / E2E when required: real binary pass exit 0; real failure exit 1 with /private/tmp/...py:2; empty directory real pytest exit 5。
• PR / CI: pending push and exact-head CI
• skipped checks and reason: full local suite is not clean because 9 unrelated curl tee and tracking SQLite tests fail in this environment; see all-tests-serial-isolated.log。
• 旧调用方/旧数据兼容：pipe caller retains the one-argument compatibility function; no data migration。
• legacy fixture 或真实 oracle: parser fixture and real pytest subprocess smoke
• 行为是否变化: describe changed behavior
• online regression: not applicable before merge; no runtime or external consumption surface
• external owner write regression: not_applicable; no external owner write target

Review Evidence

• required by current risk: yes
• fresh-context review（仅新高风险或当前合同要求，否则 not_applicable）: read-only explorer and code-reviewer review requested; no write access
• Codex review artifact / decision: pending final readback
• Fable review artifact / decision: not_applicable
• accepted findings and applied changes: 首轮发现精确伪摘要、伪 no-tests 横幅和 mixed subtests；第二轮发现 exit 2/3/4、真实捕获截断、stderr-only 空输出和冲突分支丢失败位置。全部收口到单一可信度判断和单一无法解析输出，均有 fixture 或真实二进制 smoke。
• rejected / deferred findings and reason: none；stdout truncation provenance 已按复核意见以只读字段实施，runner 保持唯一日志 writer
• reviewer boundary: review evidence is not CI, runtime, online acceptance, or external readback proof

Merge Readiness

• source revision: local HEAD (PR head/CI/main pending; read back with git rev-parse HEAD)
• target branch: develop
• PR URL: pending
• review decision: pending
• required CI: pending
• mergeability: pending
• merge strategy: exact-head squash
• merge owner: repository maintainer
• rollback / revert path: revert the exact squash commit
• ready to merge: no

Merge Result And Handoff

• merge commit: pending
• post-merge main readback: pending
• promoted runtime required: no
• next required document: not_applicable; no runtime/external acceptance surface
• Promote / runtime owner: not_applicable
• Online acceptance owner: not_applicable
• Final closeout remains blocked until: PR exact-head CI and main readback complete

RCA 与风险

• 现象：已收集或已执行的 pytest 被压缩器误报为 no-tests。
• 根因或判断：缺少摘要存在性、collected 计数和退出码参与的可信判定；这是展示层误判，不是 pytest 子进程没有执行。
• 处理或方案：单一 exit-aware parser；未知结果保守降级，真实退出码由既有 runner 返回。
• 验证：红灯夹具先证明旧行为；24 条 pytest fixtures、捕获边界 fixture、49 条 Python consumer、真实二进制 smoke 和 Clippy 通过；全量测试的 9 个非相关失败已单独记录。
• 剩余风险：远程 PR CI/目标分支尚未验证；完整 suite 的 9 项环境基线失败仍由各自 owner 处理。
• 待决策点：无新增产品决策；只待外发授权、CI 与合并读回。

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

Obayo Tian seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[KuSh]

Thank you for the careful writeup of the pytest truth-preservation problem, the failure modes you catalogued are real. Most of what this PR builds on has since been rebuilt on develop: commit 86b34df from 2026-07-22 moved pytest onto the exit-aware runner and already falls back to raw output when a failing run filters down to "Pytest: No tests collected", #2188 had added RunMode::FilteredWithExit as the way a filter receives the exit code, and #3772 on 2026-09-12 rewrote the same guard and tracking block your runner changes patch so that stderr is forwarded and counted. Against that code the new run_filtered_with_capture, guard_combined_output and force_tee_when_output_contains are a third parallel API on a call site that no longer exists, so this cannot be rebased as written. Three gaps you found are still open on develop and would each make a good narrow issue or PR on the current code: a zero-exit run with an unrecognized or duplicated summary still reports no tests collected, the parser has no collection-count or last-line anchor so a summary-shaped line from a test's own output can be mistaken for the real one, and StreamResult carries no truncation flag so a capture cut at the 10 MiB cap looks like a run with no summary. If part of this was missed, please comment and it can be reopened.