Goal tools vanish from the session tool catalog mid-session, while the goal-round prompt instructs an action they would be needed for

[Dewario]

Upstream report — @deepseek-ai/dsh goal tools

Channel: deepseek-ai/deepseek-harness has GitHub Issues disabled (API returns 410); the
README routes bug reports to GitHub Discussions, so this one was posted there, under Q&A.
Component: @deepseek-ai/dsh-tool-goal, @deepseek-ai/dsh-goal-round-driver, session tool catalog
Observed: 2026-09-29, dsh 0.1.7-rc.2 (npx cache _npx/1e7f6d9597241db0), Windows, web profile
Severity: an agent cannot close its own goal; the loss is silent and costs a full-context
round each turn

Symptom

1. Early in one session, create_goal worked and created
   goal-af5f3f0e-ca14-4e50-86d0-dcbe208db085 (maxGoalRounds: 12, phase: active).
2. The goal-round driver then ran 12 rounds normally — each delivering
   <goal_round>…</goal_round> (text matches packages/goal/goal-round-driver/src/prompt.ts:12-26).
3. In later turns of that same session, create_goal, get_goal and update_goal all
   return unknown tool (13 attempts; the goal could never be closed from the agent side).
4. The same session, same window, also lost its native tools: pwsh, edit, write
   became unknown tool. The surviving catalog was MCP servers + subagent only. The change
   coincided with a mid-session model switch (deepseek-v4-pro → deepseek-flash).

Why this looks like a defect and not a policy

• The round prompt instructs the action that is then impossible: "Before claiming
  completion, gather evidence that the whole objective is achieved, read the current goal, and
  mark it complete." (goal-round-driver/src/prompt.ts)
• Authority explicitly permits goal tools on goal rounds:
  tool-goal/src/authority.ts accepts hasDirectHumanInput() or
  isMatchingGoalRound() — the second exists precisely so a goal's own round may mutate it.
• Both shipped and user presets declare the goal rows (command-goal, tool-goal), so a
  missing preset row is not the explanation.

Hypotheses to test (in likely order)

1. Tool-catalog re-composition on model switch drops rows. A model change mid-session
   recomposes the exposed catalog; non-MCP tools (goal, fs, shell) are lost while MCP servers
   survive. Reproduce: create a goal, switch the seat's model, call get_goal.
2. Live profile-patch hot reload drops rows. patchReload: live re-composes on patch
   edits; check whether a reload rebuilds the catalog and omits rows.
3. An external tool-shadowing layer filters the catalog. A local MCP-side integration
   running "shadow mode" (native Read/Grep/Glob denied, its own equivalents mandatory) may be
   rewriting the exposed tool set beyond those three tools.

Impact

• No agent-side closure path: every round re-sends the whole conversation (the dominant cost
  line in this deployment's own audit) with no way to reach completion.
• 12 post-completion rounds were spent on a finished objective before the operator closed the
  goal by hand.
• An honest agent reports it as a blocker each round — noise that hides real blockers.

What to capture for a fix

The per-turn tool list the Host sends (tools in the request header) across a model switch and
across a patch reload; whether the rows are absent from the catalog or present-but-refused
(a refusal would read differently from unknown tool).

[PerryLink]

"工具中途消失、而提示词还在叫它用"——这类不一致最值得按装配一致性提

1. 现象的性质

你写：goal 工具在会话中途从工具目录里消失，而 goal-round 提示词仍在指示模型执行"需要这些工具"的动作。⇒ 这不是"某个工具坏了"，而是工具目录与提示词不一致：

提示词告诉模型"你可以做 X"，而工具目录里已经没有做 X 的工具。

⇒ 后果是模型必然失败（它会去调用一个不存在的工具），而且失败原因看起来像模型的问题。⇒ 建议把这个角度写成摘要第一句。

2. 请补三样（决定它落在哪一层）

1. 消失的时点：是在某一轮之后、某个事件之后，还是随机？（你已有"mid-session"，把触发点缩小到"第几轮/什么操作之后"）
2. 那一轮的工具目录原文（前后各一份，列出 goal 工具在不在）⇒ 两份目录的差分就是证据；
3. 提示词原文（goal-round 那一段）⇒ 证明它确实在指示使用已消失的工具。

"目录差分 + 提示词原文"这两样齐了，这条就无可辩驳。

3. 一个可自查的判据

请确认：同时刻的 list_* / 工具列表输出是否也看不到这些工具（即确实从目录移除），还是只是模型没调用到（目录里还在）。⇒ 这两种完全不同：前者是装配问题，后者是模型行为。

4. 请补版本

你未写版本（当前最新 0.2.1-alpha.1）。你说的组件是 dsh-tool-goal 与 dsh-goal-round-driver ⇒ 请说明这两个是否来自同一个 bundle：若是，**"一个 bundle 内的行被单独移除"**本身就是可疑点（我此前核实过同类情形：bundle 整体切换时行会一起变）。

5. 一条边界

我没有核对 goal 工具目录的装配逻辑（你给的是现象）。⇒ 本条的价值落在**"把它变成可判定的不一致"**（第 2、3 节）；补齐后我可以对着源码核"工具目录在什么条件下会移除已挂载的行"。