Coder/deepseek-harness
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(preset): align minimal agent with RL composition

Browse Source
This commit is contained in:
Yichen Jiang
2026-08-10 18:02:11 +08:00
parent d2321d210a
commit 86d5dd4384
48 changed files with 484 additions and 406 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
.agents/notes/implemented/bug-fix/2026-08-10-minimal-preset-owns-rl-composition.i18n.yaml Normal file
View File

@@ -0,0 +1,6 @@
# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
# side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write .agents/notes/implemented/bug-fix/2026-08-10-minimal-preset-owns-rl-composition.md
2026-08-10-minimal-preset-owns-rl-composition.md: 043f2e45e3fe4fbb92aa6652ce099ebfde09de55
2026-08-10-minimal-preset-owns-rl-composition.zh.md: 83f243b56b25237f19fa288f87e15eee6a264c94

37
.agents/notes/implemented/bug-fix/2026-08-10-minimal-preset-owns-rl-composition.md Normal file
View File

@@ -0,0 +1,37 @@
# Agent Note: The minimal preset owns the complete RL agent composition
Status: implemented
English | [中文](2026-08-10-minimal-preset-owns-rl-composition.zh.md)
## Problem
The Web surface offered two owners for the Claude SWE-compatible RL agent: a process-wide `core-web.cordis.yml` patch and the per-session `minimal` preset. Once [agent presets](../architecture/2026-08-03-per-session-agent-presets.md) became the agent-composition boundary, the preset's scoped `deployment:persona` shadowed the overlay's corrected global persona with stale coding-agent text. The overlay test mounted no preset, while the preset test booted without the overlay, so neither exercised the composition users selected.
The split also hid other drift. The preset mounted one-shot Bash rather than the [persistent Bash](../feature/2026-07-29-persistent-bash-str-replace-editor.md) used by the RL harness and omitted the RL compaction policy. Keeping both owners makes every future prompt, tool, and policy change a cross-product.
## Decision
The shipped `minimal` preset is the sole RL agent composition. It declares an entry-local PTY registry and local backend, persistent `bash` with the RL environment description and 300-second timeout, `str_replace_editor`, and an entry-local compaction backend. Tool presentation remains a deployment choice. The compaction policy keeps the RL threshold, absolute retention, generation cap, and retry count; model capacity comes from routed adapter metadata because `contextWindow` is no longer a compact-basic config field. The editor accepts no `requireAbsolutePath` setting because absolute paths are its unconditional contract.
The preset persona is exactly `You are a helpful software engineer assistant.` and sets `complete: true`. A complete `PromptSection` participates in ordinary assembly so tools, contexts, variables, and cooperative listeners still resolve; after the `system-prompt/assemble` waterfall, the prompt registry restores a detached copy of that section as the sole system-prompt section. Multiple effective complete sections reject assembly. This final registry constraint prevents harness identity, Web orientation, tool guidance, or an assembly listener from appending prompt text.
The process-wide `core-web.cordis.yml` patch is absent. Browser UI, workspace attachment, persistence, filesystem, subprocess, sandbox, permission, model routing, and other cross-session services remain host-owned. Selecting `minimal` changes one agent's model-facing composition without changing other sessions in the Web process.
## Verification
System-prompt and persona package tests prove final complete-section enforcement, including waterfall mutation and duplicate rejection. The shipped-preset composition test asserts the exact prompt, Bash description, absolute editor schema, and two-tool catalog under the default native presentation. The keyless Web replay sends a real request through a `minimal` agent while global identity, Web surface text, and a test section are registered, then executes two persistent Bash calls to prove environment and cwd state survive and executes the editor through an absolute path.
## Alternatives considered
**Keep `core-web.cordis.yml` as a compatibility patch.** Rejected because a process patch and a session preset are two independent owners for one agent contract; precedence makes either one capable of silently undoing the other.
**Disable every known prompt contributor in the preset.** Rejected because host rows are process-wide and new contributors would reopen the prompt. A final complete-section constraint expresses the negative guarantee at the registry that assembles the prompt.
**Filter sections only with a prepended waterfall listener.** Rejected because another prepended wrapper can run outside it and append after the filter. Enforcement after the complete waterfall has stable final authority.
**Mount PTY services on the Web host.** Rejected because only the minimal agent consumes them. An entry-local `pty` realm gives the services the same lifetime and scope as their sole consumer without publishing a process-global service from a preset.
## Consequences
The RL prompt is fixed rather than environment-overridable, and `minimal` is the only shipped place that states it. The model sees only persistent `bash` and `str_replace_editor`; shell state is per agent and disappears with that agent. The preset pays for its own PTY and compaction service instances, while other presets pay nothing for them. The local persistent-shell backend requires the supported POSIX terminal substrate, so this preset is not a Windows agent surface.

37
.agents/notes/implemented/bug-fix/2026-08-10-minimal-preset-owns-rl-composition.zh.md Normal file
View File

@@ -0,0 +1,37 @@
# Agent Note: minimal preset 拥有完整的 RL agent 组合
Status: implemented
[English](2026-08-10-minimal-preset-owns-rl-composition.md) | 中文
## 问题
Web surface 同时由两个位置定义与 Claude SWE 兼容的 RL agent（智能体）：进程级 `core-web.cordis.yml` patch，以及逐会话的 `minimal` preset。[agent preset](../architecture/2026-08-03-per-session-agent-presets.md) 成为 agent 组合边界后，preset 中带作用域的 `deployment:persona` 会用陈旧的 coding-agent 文本遮蔽 overlay 修正过的全局 persona。overlay 测试没有挂载 preset，而 preset 测试启动时没有 overlay，因此两者都没有覆盖用户实际选择的组合。
这种拆分还掩盖了其他偏差。preset 挂载了一次性 Bash，而不是 RL harness 使用的[持久 Bash](../feature/2026-07-29-persistent-bash-str-replace-editor.md)，并且遗漏了 RL 压缩（compaction）策略。保留两个所有者，会使今后每次修改提示词、工具或策略时都必须验证二者的交叉组合。
## 决策
随附的 `minimal` preset 是 RL agent 组合的唯一所有者。它声明 entry 本地的 PTY 注册表与本地后端、带 RL 环境描述且超时为 300 秒的持久 `bash`、`str_replace_editor`，以及 entry 本地的压缩后端。工具呈现仍由部署选择。压缩策略保留 RL 的阈值、绝对保留量、生成上限和重试次数；模型容量来自经路由选定的适配器元数据，因为 `contextWindow` 已不再是 compact-basic 的配置字段。编辑器不接受 `requireAbsolutePath` 设置，因为要求绝对路径是它的无条件约定。
preset persona 恰好是 `You are a helpful software engineer assistant.`，并设置 `complete: true`。complete `PromptSection` 参与常规组装，因此工具、上下文、变量和协作式监听器仍会解析；`system-prompt/assemble` waterfall（瀑布式事件）结束后，提示词注册表会将该段落的独立副本恢复为唯一的系统提示词段落。存在多个有效 complete 段时，组装会被拒绝。这项最终注册表约束可防止 harness 身份、Web 定位、工具引导或组装监听器追加提示词文本。
进程级 `core-web.cordis.yml` patch 不再存在。浏览器 UI、workspace 附加、持久化、文件系统、子进程、沙箱、权限、模型路由及其他跨会话服务仍由宿主持有。选择 `minimal` 只会改变一个 agent 面向模型的组合，不会改变 Web 进程中的其他会话。
## 验证
系统提示词与 persona 包测试证明了 complete 段的最终约束，包括 waterfall 修改与重复项拒绝。交付 preset 组合测试在默认原生呈现下断言精确的提示词、Bash 描述、要求绝对路径的编辑器 schema 和双工具目录。无密钥 Web 回放通过 `minimal` agent 发送一个真实请求，同时注册全局身份、Web surface 文本和一个测试段落；随后执行两次持久 Bash 调用，证明环境与 cwd 状态能够保留，并通过绝对路径执行编辑器。
## 考虑过的替代方案
**将 `core-web.cordis.yml` 保留为兼容 patch。** 被拒绝，因为进程 patch 与会话 preset 是同一 agent 约定的两个独立所有者；优先级会使任意一方都能静默撤销另一方的配置。
**在 preset 中禁用每个已知的提示词贡献方。** 被拒绝，因为宿主行属于整个进程，新的贡献方也会重新开放提示词。由组装提示词的注册表实施最终 complete 段约束，才能表达这项否定保证。
**仅使用前置 waterfall 监听器筛选段落。** 被拒绝，因为另一个前置包装层可以在该监听器外执行，并在筛选后追加内容。在整个 waterfall 结束后实施约束，才能稳定拥有最终决定权。
**在 Web 宿主上挂载 PTY 服务。** 被拒绝，因为只有 minimal agent 消费这些服务。entry 本地的 `pty` realm 与唯一消费方具有相同的生命周期和作用域，无需由 preset 发布进程级全局服务。
## 后果
RL 提示词固定不变，不能通过环境覆盖，且 `minimal` 是交付内容中唯一声明该提示词的位置。模型只看到持久 `bash` 与 `str_replace_editor`；shell 状态按 agent 隔离，并随该 agent 一并消失。preset 为自身的 PTY 与压缩服务实例承担开销，其他 preset 无需承担。持久 shell 的本地后端需要受支持的 POSIX 终端基础环境，因此该 preset 不适用于 Windows agent surface。