feat(agent): add bounded implementation retry loop

[dangzitou]

Summary

Related to #76
这个 PR 为 agent implementation 阶段增加了一个有边界的重试 loop，让 OpenMeta 在第一次 implementation draft 返回 needs_review 时，不会立刻收口为 artifact-only，而是先保守地补充少量相邻上下文，再重试生成 concrete patch。

之前的行为更像：

1. patch draft 生成完成
2. 第一次 implementation 尝试返回 needs_review
3. 直接结束为 review/artifact 模式
4. 最终 changed files: none

这在一些“其实已经很接近可改，只是还差几个邻近文件”的场景里会过早放弃。

这次改动的目标不是放宽安全门，也不是引入无限 agent loop，而是给 implementation 一个小范围、可控、保守的二次推进空间。

What changed

• 在 agent orchestration 中加入 bounded implementation retry loop
• 当第一次 implementation draft 需要更多上下文时，按小批量补充相邻上下文后再重试
• 限制重试次数和上下文扩张规模，避免无界增长和 token 浪费
• 在 machine-readable 结果中暴露 retry 诊断信息：
  • implementationAttempts
  • implementationStopReason
  • implementationContextFilesAdded
• 补充测试，覆盖：
  • 上下文扩展后重试成功
  • 达到边界后正确停止
  • machine 输出中的 retry diagnostics
  • workspace context expansion 行为

Design goals

• 保持 agent 默认保守
• 不让上下文扩展变成全仓库无方向蔓延
• 提高“第一次实现只差一点上下文”的场景下生成真实 changed files 的概率
• 如果模型在补充上下文后仍然不能安全改动，继续保留现有 artifact-only / review fallback

Scope

这个 PR 有意保持最小改动：

• 不引入开放式 autonomous loop
• 不做大范围重构
• 不放宽现有安全门

目标是以尽量低的 upstream conflict 风险，提高 implementation 阶段的跟进能力。

Validation

• bun run typecheck
• bun run check
• bun run build
• bun test

Notes

这个 PR 不保证每个 issue 最终都会产出 patch。

它主要解决的是这样一类情况：第一次 implementation draft 已经接近可用，但还缺少少量局部上下文，导致 agent 过早停止。

---

Summary

This PR adds a bounded implementation retry loop to the agent flow so OpenMeta can conservatively expand local context before falling back to artifact-only / review mode.

Previously, the flow often behaved like this:

1. patch draft is generated
2. the first implementation attempt returns needs_review
3. the run exits into artifact/review mode immediately
4. final result is changed files: none

That can be too eager to give up in cases where the model is already close and only needs a few adjacent files to produce a safe patch.

The goal here is not to relax safety gates or introduce an open-ended autonomous loop, but to give implementation a small, controlled, conservative retry window.

What changed

• add a bounded implementation retry loop in agent orchestration
• expand implementation context in small adjacent batches before retrying patch generation
• cap retry attempts and expansion size to avoid unbounded growth and token waste
• expose retry diagnostics in machine-readable results:
  • implementationAttempts
  • implementationStopReason
  • implementationContextFilesAdded
• add tests covering:
  • successful retry after context expansion
  • bounded stop behavior
  • machine-output retry diagnostics
  • workspace context expansion behavior

Design goals

• keep the agent conservative by default
• avoid turning context expansion into repository-wide drift
• improve the chance of producing real changed files when the first implementation attempt is only slightly under-contextualized
• preserve the existing artifact-only / review fallback when the model still cannot produce a safe patch

Scope

This PR intentionally keeps the change small:

• no open-ended autonomous loop
• no broad refactor
• no relaxation of existing safety gates

The goal is to improve implementation follow-through with minimal upstream conflict risk.

Validation

• bun run typecheck
• bun run check
• bun run build
• bun test

Notes

This PR does not guarantee that every issue will end with a patch.

It specifically improves the case where the first implementation draft is close, but still needs a small amount of additional local context before a safe patch can be generated.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.