v0.1.2

变更 / Changes

本次为三 agent 代码审查（核心逻辑 / 宿主集成与安全 / 测试质量，共 20 项发现）的修复版本。无 P0；含 1 个 P1 产品 bug、全部值得修的 P2 与关键 P3。

Fixes

• watcher：基线误报（P1） — 基线轮询失败不再把历史红色运行误报为 NEW CI FAILURE；基线只由成功轮询建立，建立前抑制检测
• watcher：超时与取消时机 — sleep 以剩余存活时长为上限、醒后复查 deadline；poll 在途时的 cancel 优先于检测结果
• runtime：jobs 服务懒解析（与 shell/storageDomain 同一挂载时序理由）；单个 job 日志拉取失败不再中止整个诊断（以 logError 呈现）；maxLines 硬上限 2000
• 安全：日志摘录围栏加固 — 四反引号围栏 + 折叠内容中 4+ 反引号连写，关闭 prompt-injection 破栏通道；ghBin 统一 shell 引号
• config：resolveConfig 拒绝 NaN/Infinity/越界值（schemastery number 校验放行它们，NaN 间隔会把 watch 退化成 1ms 热轮询）
• signatures：保留并解包 ##[error] 规范失败行（此前被噪声过滤丢弃）；sha256 长度 hex 纳入掩码，跨运行签名 id 稳定
• ledger：record() 串行化 — 并发诊断不再丢 seenCount 增量

Tests

• 85 → 102：createShellRunner 直接单测（宿主形状适配层，此前零覆盖）；经真实 apply() 接线的 ci_diagnose 端到端测试（钉住 4MB 捕获预算与 ghBin 引号）；watcher 三条回归（基线失败/deadline cap/poll 在途取消）；账本并发测试；ci_watch 测试清理空转轮询循环
• verify-web-profile.mjs 账本持久化检查新增 mtime 断言，杜绝陈旧文件虚空通过

验证结论 / Verification

• pnpm format:check / typecheck / build ✅；单元测试 102/102 ✅
• 真实环境：本机 web profile 进程内 boot() 19/19 ✅（含新增 mtime 防空洞检查；cli/cli run #31782742089 实时诊断、真实 watch job 基线与 kill、账本本运行重写落盘）

Full Changelog: v0.1.1...v0.1.2

v0.1.1

Documentation

• README 重写（中英双语）：新增 Usage/使用方法 章节 —— 以自然语言提问示例（"watch CI on this repo"、"nightly 为什么挂了"）说明每个工具何时被触发，并展示对话中实际呈现的 markdown 诊断卡；移除本机真实环境验证细节与过程性说明，账本存储位置改为通用 DSH 存储目录表述 (#2)

Verification

• pnpm format:check / typecheck / build ✅，单元测试 85/85 ✅（docs-only，无运行时代码变更）

Full Changelog: v0.1.0...v0.1.1

v0.1.0 — CI failure auto-diagnosis for DeepSeek Harness

First release of dsh-ci-doctor: watch GitHub Actions for new failures, diagnose raw job logs into structured findings, and remember every failure signature across diagnoses — all through two agent tools and your authenticated gh CLI.

Features

• ci_watch — background watch job on the host jobs registry: baseline-aware polling for new failed runs (historical red never fires the alarm), exponential backoff on transient errors, kill any time via job_kill, detection summary with a ready-made ci_diagnose next step.
• ci_diagnose — fetches failed-run jobs/logs through the host shell + gh CLI, extracts normalized error signatures, classifies failures (test/build/lint/typecheck/dependency/network/permission/timeout/infra), mines suspect files, renders a markdown diagnosis card with honestly trimmed log excerpts.
• Failure-signature ledger — durable via the host storage domain (ci_doctor unit, lazy open with memory fallback); repeat failures surface as seen N×.
• Read-only contract — every canonical result carries repositoryWrites: false; optional invariant companion (dsh-ci-doctor/invariant) fails loudly if a result ever loses the marker.
• Robust gh transport — --jq projections keep API payloads under the host shell's 64 KiB capture cap; raw job logs get a raised stdoutMaxBytes budget; failures classified (auth/not-found/rate-limit/timeout/api).

Tests

• 85/85 unit tests green — signature extraction/normalization, log trimming, gh client, watcher state machine, ledger (memory + domain), plugin wiring, invariant companion.

Docs

• Bilingual README: English · 中文.

Verification conclusions

• Local gate: pnpm typecheck ✅ · pnpm test (85/85) ✅ · pnpm build ✅ · pnpm format:check ✅
• Real environment: in-process boot() of the actual local web profile composition (dsh-app-boot exports, webserver overlaid to port 0) — 18/18 checks passed: tools registered on the real ToolRuntime, both waterfalls accepted the real host exec shape { name, arguments }, live read-only diagnosis of cli/cli run #31782742089 extracted 8 signatures, a real watch job streamed its baseline and settled on kill, and the ledger persisted to ~/.dsh/storages/ci_doctor.json across process exit.
• The real-environment check caught and fixed two genuine bugs before release: a service mount-order race (shell/storageDomain now resolved lazily) and shell capture truncation of large gh api payloads (now --jq-projected).

Full Changelog: https://github.com/jkrandom-sudo/dsh-ci-doctor/commits/v0.1.0