DeepSeek Harness Plugin Hub

发布与管理完整 Harness Profiles,发现适合你的插件。

探索

插件目录环境预设文档中心动态

社区

发布插件联系我们报告问题

相关链接

Plugin Hub GitHubDeepSeek Harness 官方项目系统状态隐私说明
© 2026 DeepSeek Harness Plugin HubPowered byPaxTech

独立、非官方社区项目,与 DeepSeek 官方无隶属、授权或背书关系。

DeepSeek Harness Plugin Hub
ProfilesPlugins分类动态文档登录管理 Profiles
ProfilesPlugins分类动态文档登录
← Plugins

@playform/hook-dsh-replay-suggester

Hook Dsh Replay Suggester

replay-suggester hook:一个 DeepSeek Harness 插件,可根据会话历史回答“上次执行此操作时,模型实际接触了什么,以及它是如何验证的?”当模型以此前曾追求过的意图进行搜索时,

插件会安装到这里;不确定时保持 web。

npx -y @deepseek-ai/dsh plugin --profile web add @playform/hook-dsh-replay-suggester@0.2.1
README兼容性版本

兼容性与来源证明

Hook Dsh Replay Suggester 以 @playform/hook-dsh-replay-suggester 发布,当前版本为 0.2.1。Plugin Hub 会校验它的 manifest,并保存精确安装来源,便于复现安装结果。

DSH 兼容范围
*
运行环境
any
发布来源
npm
Registry 更新时间
2026/11/4

版本

0.2.1stable
2026/10/28
0.2.0stable
2026/10/25
0.1.0stable
2026/10/25

相关插件

正在加载相关插件…

最新版
0.2.1
DSH
*
HMR
重启进程
Tree shaking
未声明可安全裁剪
解包体积
56.7 kB
文件数
52
Surface
any
许可证
CC0-1.0
发布源
npm
GitHub
★ 2
周下载
288
安全扫描
✓ v0.2.1 扫描通过
最近提交
2026/11/4
查看源码 ↗项目主页 ↗
README Badge

点击下方 Badge 复制 Markdown,粘贴到 README 即可。

这是你的 Plugin?认领权益 · 优先安全扫描

验证 package.json 声明的 GitHub 仓库,即可管理这个公开页面。认领后,Hub 会优先安排当前版本的安全扫描,并在通过后公开展示结果。

认领这个 Plugin →
报告问题

相关插件

继续浏览 memory-context 分类下经过校验的插件。

Experimental Session Search@deepseek-ai/dsh-experimental-session-search为 Agents 提供搜索和读取当前工作区早期会话的工具。Hindsight Coding Agents@vectorize-io/hindsight-coding-agents面向编程代理的仅反思式 Hindsight 长期记忆(可插拔支持多种运行框架:opencode、……),自动在后台导入(无需设置 CLI)。Ab Memorydsh-ab-memory--- description: "dsh-ab-memory 包:一种跨会话、基于文件的智能体记忆,提供五个工具(remember / recall / forget / load_skill / distill)以及始终启用的轻量注入。" kind: "package-reference" ---Fish Memorydsh-fish-memory给 DSH 里的会话一份会淡忘的长期记忆,每条都带时间。

README

@playform/hook-dsh-replay-suggester

v0.2.1 - UNPUBLISHED - plan: SEARCH-SUGGESTER-PLUGIN

The role

The replay-suggester hook is a DeepSeek Harness plugin that answers the question "the last time this was done, what did the model actually touch and how did it verify?" (SEARCH-SUGGESTER-PLUGIN.md §9.3).

When the model searches with an intent it has pursued before, the hook recalls the matching record from its history store and prepends ONE labeled additionalContexts notice: the files touched last time and the verification command.

  • The history store is ~/.dsh/replay-suggester/history.json (the seed of the search-suggester's journal per §9.3, while the offline learner and the journal adapter are the end-phase wiring).

    The store loads once at apply, and a missing or corrupt store degrades to silence with a ledger line, never to a guess.

  • The recall store has a producer (the wave 2c re-spec, 03-SUGGESTERS.md): every judged call is recorded, and a later call recalls the paths that accompanied a similar prior call.

    The producer's store is its own (the replay_history storage domain, or the JSON fallback beside the seeded store), and the seeded store still wins when it matches.

  • Without a recalled record the hook abstains (replay no-memory), because a replay answer without a record is a guess and the plan forbids the guess (§2.4).

  • Privacy posture (plan §5.4): repo-relative paths only, no file contents, no user text, local only.

The surfaces

  • The tools/post-execute waterfall listener (after the ccr's {prepend: true} listener) adds one additionalContexts notice and never a content replacement.

  • The agent/pre-step listener resets the note counter for each turn.

  • The recall store is the replay_history storage-domain KV domain, or the JSON fallback beside the seeded store (~/.dsh/replay-suggester/).

  • Injected services: pluginFactory.

    There is no tool and no provided service.

How it works

  1. The post-execute listener.

    The tools/post-execute listener (waterfall, plain registration, after the ccr's {prepend: true} listener) calls next() first.

    Then the hook runs record (the producer), recall (the seeded store, then the live history), detect, gate, and prepend ONE notice.

    Content is never replaced.

  2. The pre-step listener.

    The agent/pre-step listener resets the note counter each turn when a user message enters.

The recall store (the producer)

On every judged call that names a path, the hook appends one entry { seq, path, intent, outcome, at, probe } to the recall history: the per-agent ring plus the durable tier.

The durable tier is the storage-domain KV domain replay_history (version 1, per-record, tables history and meta) when the harness provides a storage facility, and the JSON fallback at recordPath (~/.dsh/replay-suggester/replay_history.json) otherwise; both are bounded by historyDepth and both are awaited through State.Flush().

The recall matches the most recent prior entry with the same tool and either a contained probe literal or the same path directory, then names the paths recorded within learnWindow calls of it (the episode), which means "the last time this area was touched, these files were touched".

The recorded line, the recovery lines (store recovered <n> entries (replay_history|file), store empty (starting fresh)), and the failure lines are byte-exact in SCHEME.md.

The pack

The Repeat intent is pack-driven: the built-in pack (this monorepo's default) lives in @playform/lib-dsh-pack and is projected here as the replay/ domain.

A repository that ships .dsh/pack.yml (or a user-layer pack at ~/.dsh/playform/workspaces/<slug>.yml) gets its OWN replay/ intent and caveat, and a repository without a pack stays silent.

The recalled companion still comes from the history store, because the pack supplies the gate and the caveat, never the companion.

The ledger

The lines are prefixed replay-suggester: and go to logFile:

activated (...)
suggest <id> <k> <n> candidates call <tool>
suppressed <verdict> <id>
abstained ambiguous <id> <k>
abstained no-companions <id> <k>
abstained no-memory <id> <k>
abstained foreign-workspace
recorded <intent> <path> (<n> in store)
history loaded <n> records
history missing (starting empty)
history corrupt (starting empty) <error>
store recovered <n> entries (<source>)
store empty (starting fresh)
store open failed: <reason>
store write failed: <reason>
replay-suggester: <error>; the note is skipped

The configuration

See SCHEME.md for the full schema.

Every field is a volatile cell:

FieldDefaultMeaning
enabledtruethe plugin ships enabled
logtruethe ledger switch (shared cell)
logFile~/.dsh/playform/hook-dsh-replay-suggester.logthe ledger
threshold0.75the gate threshold k
maxCandidates7the bound on named candidates
maxNoteChars1200the note length bound
cooldownCalls6the cooldown between notes
maxNotesPerTurn3the per-turn cap
intents[]the intent filter
judged tools["bash", "edit", "glob", "grep", "read", "write"]the judged calls
exclude["node_modules", "Target", ".git", "dist"]the excluded paths
historyPath~/.dsh/replay-suggester/history.jsonthe seeded store
learnWindow2the episode width in calls
recordPath~/.dsh/replay-suggester/replay_history.jsonthe producer's JSON fallback
historyDepth32the entry bound

The build and the tests

The build is pnpm prepublishOnly (Source/prepublishOnly.sh): the plugin build (ESM Target/, Test tree excluded), the Test build, the harness selfcheck, and the Test type-check.

The smoke folders are History-Producer, Loader-And-History, Pack-Driven, Pipeline-And-Recall, and Workspace-Gate.

The plan

  • SEARCH-SUGGESTER-PLUGIN - section 9.3, the recall answer.

  • Completion-Spec/03-SUGGESTERS - the wave 2c re-spec (the producer).

  • The package, the Source, the smoke suites, and the build pass are complete, and the recall store's producer landed.

  • The live activation (the profile row flip and the restart) is the remaining end-phase step (POC-BRIEF.md §8).

    It ships enabled: false in the bundle patch.

License

CC0-1.0.