DSH | Tacit | Learns what you leave unsaid in your prompts — automatically — and tells the agent for you

[hackernotfound]

Note

Unofficial community project, independently developed and maintained. Not affiliated with or endorsed by DeepSeek. Read the source and permissions before installing third-party software.

Repo: https://github.com/hackernotfound/dsh-tacit · npm: dsh-tacit · MIT

You keep correcting your agent the same way. Tacit watches how each turn actually went — retries, tool errors, and above all the message you send when the agent got it wrong — and turns that into a few directives the agent follows in every new conversation. Automatically. You do nothing.

you:    "make the login page better"
agent:  …stumbles…
you:    "no, I meant the Next.js app under apps/web"

Tacit:  learns → "The user often omits which app they mean — check apps/web first."
        injects it into every later conversation's system prompt.
        You never have to say it again.

What makes it different from a memory plugin

• Passive. No clicks, no /remember. Messy turns and your corrections are analyzed in the background — $0.001–0.003 per lesson on deepseek-v4-flash, 30/day cap by default. It even learns from the clean turn that follows a messy one: what you included the second time becomes the lesson.
• Visible and editable. What it learned is a short system-prompt section you can read, edit, toggle or delete in Settings. Nothing hidden. Directives can be global or scoped to one workspace.
• Gets better with time, and proves it. A new directive starts as a candidate and is retired if your messy-turn rate gets worse. Settings shows the real trend — messy-turn rate and tokens per turn, first 20 turns vs. latest 20 — so you can see whether it is earning its keep.

There is also a ✨ Improve button that rewrites the draft in your composer using what Tacit has learned — one pass, before/after preview, and it tells you when your draft is already complete.

Install (30 s)

npx @deepseek-ai/dsh plugin --profile web add dsh-tacit

Restart dsh web, refresh. Requires DSH ≥ 0.1.1-rc.1, Node ≥ 22, and a DeepSeek key already configured in the harness — Tacit never reads it. Everything it keeps stays under ~/.dsh/storages/tacit/.

We want 10 testers. Use it in two conversations, then tell us one directive that helped or was wrong, any surprising cost, and your versions → hackernotfound/dsh-tacit#18

Docs: how it works · privacy & cost · 中文

---

中文

Tacit 不是记忆插件。它盯着每一轮对话实际的走向——重试、工具报错、尤其是 Agent 做错之后你发出的那句纠正——把你反复省略的东西提炼成几条指令，注入之后每个新会话的系统提示词。全自动，你什么都不用做。

• 零操作：后台自动分析，每条 $0.001–0.003（deepseek-v4-flash），默认每天 30 次上限。连"出错之后的那轮干净对话"也会学：你第二次补上的信息就是教训。
• 看得见、改得了：学到的指令就在 Settings 里，可编辑、可开关、可删除；可以全局，也可以只对某个工作区生效。
• 越用越好，而且有数据：新指令先是"候选"，如果你的问题轮次比例变差就自动退役。Settings 里有真实趋势（前 20 轮 vs 最近 20 轮），能看到它到底有没有用。

安装：npx @deepseek-ai/dsh plugin --profile web add dsh-tacit，重启 dsh web，刷新。数据全部在 ~/.dsh/storages/tacit/，只有模型调用经由你自己配置的 key 出去。

招 10 位测试者：用两个会话，然后到 Discussion #18 说一条有用或错误的指令、意外的花费、以及版本。中文可以。

中文文档 · 入门

[liyangbing]

这个项目把“可见、可编辑、可退役”的学习状态与隐式记忆区分开，是很好的信任边界。建议测试说明再补一份 sanitized learning receipt，让用户能审计每条 directive 的来源和作用域，但不暴露原始 prompt/工具输出。

receipt 可包含：directive id/version、scope（global/workspace/Session）、candidate/active/retired 状态、触发原因类别（如 correction/tool-error/retry）、创建与最近评估时间、评估窗口大小、token/cost 聚合和删除状态；原文只保存在用户明确选择的本地存储中，默认不上传到第三方分析服务。

重点负向测试包括：纠正文本含 API key/路径/个人数据、同一 directive 在两个 workspace 之间串用、workspace rename/move 后的作用域、删除后冷启动仍复活、压缩/重试重复学习，以及改善按钮把 directive 当作用户意图直接执行。候选 directive 变 active 前应有清晰的 preview/disable/delete，且 directive 不应改变工具权限、审批或 sandbox policy。

成本趋势也建议把 provider/model、输入输出 token、失败/重试与每日 cap 分开统计；“messy-turn rate 下降”不能单独证明因果。手册里的 Session/记忆边界与插件审计原则可参考：sessions vs memory 和 community plugin audit。

[hackernotfound]

Thank you again for this review — it was concrete enough that it became issues #40–#44, all shipped in v0.5.0, and the receipt you proposed got a proper face in v0.6.0 this week. Point by point:

You suggested	What shipped
A sanitized learning receipt per directive	Every directive records id, version, scope, status, trigger categories, created / updated / evaluated dates, its trial window with baselines, evidence counts, and the metered cost of its distillation. Shown per directive in Settings with Copy as JSON, same record via /api/tacit/directive-receipt. Never any prompt text. In v0.6.0 it reads as tiles → lifecycle timeline → trigger chips instead of a flat list.
Raw text stays local, never uploaded	Everything lives in local files under the harness home; the only network traffic is model calls through the harness's own model service. The privacy docs now say this explicitly.
Corrections carrying keys / paths / personal data	Credential-shaped strings (API keys, tokens, JWTs, private keys) are masked at capture, before storage or any model call.
One directive bleeding across workspaces	A workspace is one normalised path; same-name folders read a/web / b/web to the distiller; a directive whose evidence never carried a resolvable scope is dropped, not applied everywhere.
Scope after rename / move	The old scope shows not seen since …, its candidate pauses and frees the trial slot, and every directive has a Move to workspace picker.
Deleted directive reviving on cold start	Removal is a remembered tombstone: the distiller is told not to re-propose it, and a returning copy or near-rewording stays removed.
Double-learning on compaction / retry	Compactions and retries are counters on a single turn digest and reports are keyed per turn, so nothing is learned twice; re-analysing an analysed turn is also free unless you explicitly ask for a fresh one.
✨ Improve executing a directive as user intent	Directives are not part of the Improve prompt at all, and every coach prompt treats its input as evidence, never as instructions.
Clear preview / disable / delete before a candidate goes live	New opt-in reviewCandidates: a distilled directive waits as queued until you press Start trial. Disable and delete already existed; the receipt shows the whole history.
Directives must not touch tool permissions / approval / sandbox	The steering section states it changes none of those, and a directive about them is refused — distilled or typed.
Split cost by provider, tokens, failures, cap	Spend by provider route and by trigger, a tile for failed/repair calls, input and output tokens shown apart, the daily cap beside today's spend.
"Messy-rate down" isn't causation	The trend chips now say they are a before/after over your own turns, not a controlled comparison, and retired reads as what happened during the trial, not a verdict.

The one item I didn't take: per-session directive scope. Steering is frozen per conversation, so that scope would govern exactly one conversation — which the opt-in pre-send note already covers per turn. If you have a concrete case where it matters, I'd genuinely like to revisit it.

Two more things:

• I'd value a second pass. If you have half an hour for v0.6.0 (npx @deepseek-ai/dsh plugin --profile web add dsh-tacit), running your own negative tests against it would be the most useful review it could get. Anything you find — issue, discussion reply, or PR — is welcome, and so are your ideas beyond this list; the receipt exists because you described it. If you know others who'd kick the tires, I'd love their feedback too.
• The handbook chapters you pointed to (sessions vs memory, community plugin audit) were genuinely useful grounding while building this — thank you for those as well.

This review made the plugin noticeably better than I would have made it alone. Thank you.

---

感谢这份评审 — 它被整理成 issue #40–#44，全部在 v0.5.0 发布，你提出的“学习回执”也在本周的 v0.6.0 有了正式的界面。逐条对应：每条指令记录 id/版本/作用域/状态/触发类别/创建·更新·评估时间/试用窗口与基线/证据数量/提炼成本，在设置中按指令展示并可复制为 JSON，同一记录经 /api/tacit/directive-receipt 返回，绝不含 prompt 原文（v0.6.0 起以指标块 → 生命周期时间线 → 触发标签的形式呈现）；所有数据只存本地文件，唯一网络流量是经宿主模型服务的模型调用，隐私文档已明确说明；凭据形状的字符串在采集时即被遮蔽；工作区为规范化路径、同名可区分、无法归属的指令直接丢弃；重命名/移动后显示“自某日未见”、候选暂停并释放试用名额、每条指令可移动作用域；删除有墓碑记忆，改写复活也保持删除；压缩/重试只计数不重复学习，已分析轮次重复分析免费；Improve 的提示词完全不含指令，且所有教练提示把输入当证据而非指令；新增 reviewCandidates，候选需你按“开始试用”才生效；steering 明确不改变工具权限/审批/沙箱，相关指令一律拒绝；成本按线路、触发、失败/修复、输入/输出拆分，每日上限与当日花费并列；趋势明确标注为前后对比而非对照实验。唯一未做的是按会话划分作用域（steering 每会话冻结，只会作用于一个会话，按发送的补充说明已覆盖），如你有具体场景，欢迎继续讨论。

如果你有时间对 v0.6.0 再审一遍 — 尤其是用你自己的负向测试 — 那会是它能得到的最有价值的反馈。issue、PR、新想法都非常欢迎，也欢迎推荐给可能感兴趣的人。你指的手册两章（会话与记忆边界、社区插件审计）在实现过程中帮了很大忙。这份评审让插件比我一个人能做出来的好得多，再次感谢。