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

fix(workspace-context): guard baseline restoration

Browse Source
This commit is contained in:
fz
2026-08-03 22:05:17 +08:00
parent eff7b758b1
commit 19be29cd0b
31 changed files with 249 additions and 105 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
.agents/notes/implemented/feature/2026-06-24-workspace-context.i18n.yaml
View File

@@ -2,5 +2,5 @@
# 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/feature/2026-06-24-workspace-context.md
2026-06-24-workspace-context.md: b224eedb03cd1e48842c883637b2097ee83fd4b4
2026-06-24-workspace-context.zh.md: f6c410b9467091b85a3be43c69dd3aecae5ea51a
2026-06-24-workspace-context.md: 19b227c56c6c50f802f1d9cf32ca6f21b3a24180
2026-06-24-workspace-context.zh.md: 86137db7296cce99eeb06e5cc023f20ded0f77c4

4
.agents/notes/implemented/feature/2026-06-24-workspace-context.md
View File

@@ -34,7 +34,7 @@ The injection becomes a durable `user/message` with a typed `workspace-instructi
A resumed agent creates a new loop instance and injects a baseline composed from current files before its first request. This permits current baseline content on resume without mutating an earlier history event. A resume and a hot plugin remount both face a log that may already hold a baseline; they are told apart by `agent/session-start`, which a startup or resume emits before the first step while a remount attaches to an already-live session and never sees it. A remount retains the existing baseline only when its typed event remains in the current visible surface, and still rebuilds scope and provider-version tracking from current files. If compaction has shadowed that event, the remount injects a current baseline. A resume always re-composes.
Compaction can shadow the baseline after this plugin's guarded `agent/step` listener has already run for the session. The `system-prompt/assemble` waterfall therefore delegates first, then checks the final visible surface. When a prior typed baseline exists but none remains visible, it recomposes and injects the current chain before the loop drains its outbox and snapshots derived request history. A per-session settled marker prevents repeated preparation when the current generation produced no baseline; a separate queued marker prevents duplicate assembly before outbox drain and clears when a step or turn closes without a durable baseline, so a cancelled delivery remains eligible for the next request.
Compaction can shadow the baseline after this plugin's guarded `agent/step` listener has already run for the session. The `system-prompt/assemble` waterfall therefore delegates first, but restores only for an assembly explicitly marked for the loop's next model request; diagnostic assemblies such as TUI `/status` remain read-only. When a prior typed baseline exists but none remains visible, the listener recomposes the current chain, rechecks cancellation and the current surface generation after every asynchronous probe, and injects before the loop drains its outbox and snapshots derived request history. A per-session settled marker prevents repeated preparation when the current generation produced no baseline; a separate queued marker plus the synchronous commit-time recheck lets concurrent preparations scan without queuing duplicate baselines.
The baseline is a user-role `<system-reminder>` with `Instructions from: <path>` sections and explicit authority and precedence language. This familiar model-facing frame avoids a harness-specific XML vocabulary. Project paths are root-relative and the user-global path is `~/.dsh/AGENTS.md` for the default home or `$DSH_HOME/AGENTS.md` for a configured home. The final rendering boundary escapes a literal `</system-reminder>` anywhere in instruction content or model-visible path, scope, and budget metadata before byte accounting completes. The package README owns the exact current [prompt shape](../../../../packages/context/workspace-context/README.md#prompt-shape).
@@ -56,7 +56,7 @@ At reconciliation time the plugin scans workspace-sourced `user/message` events
An unchanged path and digest is suppressed. A logged removal is a tombstone, so a reappearing candidate becomes a new `set`. Resume works from persisted metadata. If compaction removes a dynamic instruction event from the visible surface, that state no longer suppresses a later tool-triggered load; if it removes the baseline, prompt assembly restores the complete current chain before the next request. Only changes actually included under the byte budget enter metadata or pending state, so an omitted file remains eligible on a later touch.
The initial baseline's typed changes are comparison state only while its event remains in the visible session surface. Prompt assembly recomposes a shadowed baseline for the current replacement generation and appends it before the first post-replacement request; a queued baseline discarded with its step can be prepared again. Later successful filesystem touches can append edits or removals as dynamic messages. It never rewrites the original event. The in-memory scope marker and provider-version cache only select and accelerate probes, so neither can suppress context the model no longer sees. During resumed or post-replacement baseline preparation the plugin also reconciles visible dynamic scopes, so nested changes made while the agent was offline can append an update before the next request.
The initial baseline's typed changes are comparison state only while its event remains in the visible session surface. Model-request prompt assembly recomposes a shadowed baseline for the current replacement generation and appends it before the first post-replacement request. It rechecks the caller's signal before injection, so an aborted preparation publishes no pending baseline; a queued marker remains until the corresponding durable event confirms delivery. Later successful filesystem touches can append edits or removals as dynamic messages. The plugin never rewrites the original event. The in-memory scope marker and provider-version cache only select and accelerate probes, so neither can suppress context the model no longer sees. During resumed or post-replacement baseline preparation the plugin also reconciles visible dynamic scopes, so nested changes made while the agent was offline can append an update before the next request.
There is intentionally no watcher. Detection occurs at the next successful structured filesystem touch, post-replacement prompt assembly, or resumed baseline preparation. A provider failure produces no removal; absence is only accepted when all configured candidates in that scope were probed successfully.

4
.agents/notes/implemented/feature/2026-06-24-workspace-context.zh.md
View File

@@ -34,7 +34,7 @@ Status: implemented
恢复 agent 会创建新的循环实例，并在其第一次请求前注入由当前文件组合的基线。这样，恢复时可以使用当前基线内容，而无需修改先前的历史事件。恢复与插件热重挂都会面对日志中可能已存在基线的情况；二者通过 `agent/session-start` 区分：启动或恢复会在第一步前发出该事件，而热重挂附着到一个已存活的会话、永远不会看到它。只有当基线的类型化事件仍在当前可见表层中时，热重挂才保留既有基线，同时仍会根据当前文件重建 scope 与提供方版本跟踪。如果压缩（compaction）已遮蔽该事件，热重挂会注入当前基线。恢复则始终重新组合。
在本插件带防护的 `agent/step` 监听器已经为该会话运行后，压缩仍可能遮蔽基线。因此，`system-prompt/assemble` waterfall（瀑布式事件）会先委托，再检查最终可见表层。如果此前存在带类型的基线、但已无基线可见，它会在 loop 排空 outbox 并对派生请求历史创建快照之前，重新组合并注入当前文件链。逐会话的已结算标记会在当前代次没有产生基线时避免重复准备；单独的排队标记会在 outbox 排空前避免重复组装，并在步骤或轮次关闭且未产生持久基线时清除，因此已取消的投递仍可在下一个请求中重试。
在本插件带防护的 `agent/step` 监听器已经为该会话运行后，压缩仍可能遮蔽基线。因此，`system-prompt/assemble` waterfall（瀑布式事件）会先委托，但只有当组装被明确标记为供 loop 的下一个模型请求使用时才恢复；TUI `/status` 等诊断组装保持只读。如果此前存在带类型的基线、但已无基线可见，该监听器会重新组合当前文件链，在每次异步探测后重新检查取消状态和当前表层代次，并在 loop 排空 outbox 和对派生请求历史创建快照之前注入。逐会话的已结算标记会在当前代次没有产生基线时避免重复准备；单独的排队标记加上提交时同步复查，使并发准备可以扫描而不会排入重复基线。
基线是一条 user 角色的 `<system-reminder>`，包含 `Instructions from: <path>` 章节，以及明确的权威性与优先级说明。这种熟悉的模型可见框架避免引入 harness 专用的 XML 词汇。项目路径相对于根目录；使用默认 home 时，用户全局路径为 `~/.dsh/AGENTS.md`，使用已配置 home 时则为 `$DSH_HOME/AGENTS.md`。最终渲染边界会在完成字节核算前，转义指令内容或模型可见的路径、scope 与预算元数据中出现的字面量 `</system-reminder>`。包 README 负责规定当前准确的[提示词形态](../../../../packages/context/workspace-context/README.md#prompt-shape)。
@@ -56,7 +56,7 @@ shell 命令不会触发发现。本地 bash 调用会启动全新的 shell，
路径和 digest 均未变化时会被抑制。日志中的移除操作是一条墓碑记录，因此重新出现的候选项会成为新的 `set`。恢复操作从持久化元数据继续工作。如果压缩从可见表面移除动态指令事件，该状态不再抑制之后由工具触发的加载；如果移除的是基线，提示词组装会在下一个请求前恢复完整的当前指令链。只有真正纳入字节预算的变更才会进入元数据或待处理状态，因此被省略的文件在之后的触碰中仍有资格加载。
只有当初始基线事件仍在可见会话表层中时，其类型化变更才用作比较状态。提示词组装会为当前替换代次重新组合被遮蔽的基线，并在替换后的第一个请求前追加它；随其步骤一起被丢弃的已排队基线可以再次准备。之后成功的文件系统触碰仍可把编辑或移除作为动态消息追加。它绝不重写原始事件。内存中的 scope 标记和提供方版本 cache 只用于选择探测对象并加速探测，因此二者都不能抑制模型已无法看见的上下文。在恢复或替换后准备基线的过程中，插件还会协调可见的动态作用域，因此 agent 离线期间发生的嵌套变更可以在下一个请求前追加更新。
只有当初始基线事件仍在可见会话表层中时，其类型化变更才用作比较状态。面向模型请求的提示词组装会为当前替换代次重新组合被遮蔽的基线，并在替换后的第一个请求前追加它。它会在注入前重新检查调用方的 signal，因此已中止的准备不会发布待处理基线；排队标记会保留，直到相应的持久事件确认投递。之后成功的文件系统触碰仍可把编辑或移除作为动态消息追加。插件绝不重写原始事件。内存中的 scope 标记和提供方版本 cache 只用于选择探测对象并加速探测，因此二者都不能抑制模型已无法看见的上下文。在恢复或替换后准备基线的过程中，插件还会协调可见的动态作用域，因此 agent 离线期间发生的嵌套变更可以在下一个请求前追加更新。
系统刻意不使用文件监视器。检测发生在下一次成功的结构化文件系统触碰、替换后的提示词组装或恢复时的基线准备。提供方失败不会产生移除；只有该作用域中的全部已配置候选项都成功完成探测后，系统才接受「不存在」这一结论。