IssueLens Agent v0.2.0 — Verifiable Agent MVP

中文

v0.2.0 将 IssueLens Agent 收敛为一个可运行、可评测、可演示、可审计的 Agent 岗位作品集 MVP。

主要更新

• Docker Shadow Workspace：固定镜像摘要、断网、非 root、只读挂载、资源限制和超时清理。
• 四阶段证据链：失败复现、补丁复验、完整回归、隐藏测试。
• 可恢复 Agent：检查点、仓库/Issue 漂移检测、工具/证据/时间预算、有限重试。
• 人工闸门：本机 Web 审阅界面、追加式决策日志、Draft PR Markdown 导出。
• 8 个评测案例与双语架构、安全、基线、演示和作品集文档。

验证结果

• GitHub Actions：6/6 检查通过，包括 Python 3.11、3.12 与独立 shadow-validation。
• make test：130 个测试。
• make eval：8/8 符合预期；6/6 定位案例 Hit@1、Recall@3、MRR 均为 1.0；2/2 正确拒答。
• Wheel 已从最终合并提交构建并完成临时安装验证。

该版本仍为 Alpha 预发布。未使用 Docker 的本地执行器仅用于可信夹具，不提供隔离保证。

English

v0.2.0 turns IssueLens Agent into a runnable, evaluable, demonstrable, and auditable Agent-role portfolio MVP.

Highlights

• Docker Shadow Workspace with a digest-pinned image, disabled network, non-root execution, read-only mounts, resource limits, and timeout cleanup.
• A four-stage proof chain: failing reproduction, patched reproduction, full regression, and hidden tests.
• Resumable Agent checkpoints with repository/Issue drift detection, tool/evidence/time budgets, and bounded retries.
• Human gates through a loopback-only Web review UI, append-only decisions, and Draft PR Markdown export.
• An 8-case evaluation suite plus bilingual architecture, security, baseline, demo, and portfolio documentation.

Verification

• GitHub Actions: 6/6 checks passed, including Python 3.11, Python 3.12, and independent shadow-validation.
• make test: 130 tests.
• make eval: 8/8 expected outcomes; all 6 localization cases reached 1.0 Hit@1, Recall@3, and MRR; 2/2 abstentions were correct.
• The wheel was built from the final merged commit and verified through a clean temporary installation.

This is still an alpha pre-release. The non-Docker local runner is for trusted fixtures only and provides no isolation guarantee.

Artifact integrity

issuelens_agent-0.2.0-py3-none-any.whl

SHA-256: bb0f9df59faa3d74cf0eb26906e6a9afa8f4e74419fdce02cd17418f2780652d

IssueLens v0.1.0 — Initial Public Baseline / 首个公开基线

IssueLens v0.1.0 是首个公开 Alpha 基线，展示一个本地优先、证据可追溯、默认只读的 Python Coding Agent。

IssueLens v0.1.0 is the first public alpha baseline of a local-first, evidence-grounded, read-only-by-default Coding Agent for Python repositories.

核心能力 / Highlights

• 解析 Issue，并结合文本检索与 Python AST 索引定位候选代码。
  Parse Issues and locate candidate code with text retrieval and a Python AST symbol index.
• 为诊断和修改计划保留文件、行号、代码片段及内容哈希证据。
  Preserve file, line, snippet, and content-hash evidence for diagnoses and change plans.
• 通过路径、符号链接和前后快照校验证明分析过程没有修改目标仓库。
  Use path, symlink, and before/after snapshot checks to prove that analysis did not modify the target repository.
• 提供有步数预算、显式工具契约和完整 Trace 的离线与 LLM Agent Policy。
  Provide offline and LLM Agent policies with step budgets, explicit tool contracts, and complete traces.
• 支持 Unified Diff 策略检查、快照绑定提案、显式哈希审批、临时副本预检和内容寻址审计回执。
  Support Unified Diff policy checks, snapshot-bound proposals, explicit hash approval, temporary-copy preflight, and content-addressed audit receipts.
• 运行时只依赖 Python 标准库；离线演示不需要 API Key。
  Use only the Python standard library at runtime; the offline demo requires no API key.

可复现验证 / Reproducible Verification

• 90 项自动化测试通过，GitHub CI 覆盖 Python 3.11 与 3.12。
  All 90 automated tests pass, with GitHub CI covering Python 3.11 and 3.12.
• 在 3 个合成冒烟案例的 deterministic baseline 上，两个代码定位案例的 Hit@1、Gold File Recall@3 与 MRR 均为 1.0。
  On a three-case synthetic smoke evaluation of the deterministic baseline, both localization cases achieve Hit@1, Gold File Recall@3, and MRR of 1.0.
• 唯一的信息不足案例被安全拒绝，全部 3/3 次运行保持目标仓库快照不变。
  The sole underspecified case is safely rejected, and all 3/3 runs preserve the target-repository snapshot.

这些数字只描述固定合成夹具，不能代表真实公开 Issue 或陌生仓库上的通用准确率。

These numbers describe only fixed synthetic fixtures and do not establish general accuracy on real public Issues or unfamiliar repositories.

快速开始 / Quick Start

make demo
make patch-demo
make eval
make test

当前边界 / Current Scope

• 本版本是可复现的 deterministic baseline，不声称能够自动修复任意 Issue。
  This release is a reproducible deterministic baseline and does not claim to repair arbitrary Issues automatically.
• 补丁预检只在临时副本中进行，不会写入真实目标仓库。
  Patch preflight runs only in a temporary copy and never writes to the real target repository.
• 当前评测不执行目标仓库测试，也不计算 Resolved@1。
  The current evaluation does not execute target-repository tests or calculate Resolved@1.
• Docker Shadow Workspace、真实公开 Issue 时间切分评测和 GitHub Draft PR 集成仍在路线图中。
  Docker shadow workspaces, time-split evaluation on real public Issues, and GitHub Draft PR integration remain on the roadmap.