@local/dsh-queue-resume-on-stop
用户停止后，让排队中的消息继续流转，而不是滞留在收件箱中
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:evanw0211/dsh-queue-resume-on-stop#7080f25bdeb58b4d1fd0de37d916045f485839e0Queue Resume On Stop 以 @local/dsh-queue-resume-on-stop 发布，当前版本为 0.1.0。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。
A small Host plugin for DSH that keeps queued messages flowing after you press Stop.
While a turn is running, anything you send is queued in the agent's inbox
(agent.inbox.nextTurn) and is claimed at the next turn boundary.
The Stop button calls agent.cancel({ kind: 'user' }, { keepInbox: true })
(dsh-api-session-controller → session.cancel). That aborts the in-flight turn
and deliberately keeps the queue — but nothing re-arms the driver afterwards:
turn() returns true while inbox.hasPending,
so the next queued prompt is claimed immediately.turn(), so kick() unwinds and the driver stops.
The queued prompts stay in the inbox and are never sent to the model.The loop already handles the mirror case: a message that arrives after the abort
is classified as a wake (send() → wakingAfterAbort), which latches
phase.wakeRequested; kick()'s exit path replays that latch. A message queued
before the Stop never sets the latch — that is the gap this plugin closes.
Symptom the user sees: the queued message still sits in the queue (and is never answered) until some new message happens to wake the driver.
On every durable turn/end whose reason is aborted with cause user (the Stop
button) or parent (a parent interrupting a child), if the same agent still has
pending nextTurn prompts, the plugin asks the loop to wake once more — exactly
as a message typed a moment after the Stop would. The next turn then claims the
queued prompts and sends them to the model.
disposed (teardown) and hook (credential/session invalidation) aborts are
ignored: restarting work there would fight the teardown. next-step input
(agent.inject()) is deliberately non-waking and is not resumed.
It is a plain Host bundle: no dependencies, no build step. Point DSH's plugin manager at this directory (absolute path) and it installs the package, adds the bundle to the profile, and activates over HMR:
plugin_manager install_bundle <path-to-this-directory>
Equivalently, from the CLI: dsh plugin --profile <name> add <path-to-this-directory>.
Requires DSH 0.2.0-rc.2. The bundle resolves every Harness package from the
running installation, so it declares no dependencies of its own.
cordis.patch.yml inserts one row; override it in your profile patch:
- id: queue-resume-on-stop
config:
enabled: false # turn the behavior off without uninstalling
| File | Role |
|---|---|
index.js | Host plugin: the session/event trigger and the wake |
cordis.patch.yml | Inserts the queue-resume-on-stop row |
package.json | Bundle manifest (dsh.bundle.patch) |
locale/*.json | Plugin Manager display text |
dev/scan-stranded-queue.mjs | Scans stored session logs for "stop stranded the queue" occurrences |
dev/dump-session.mjs | Dumps a slice of one session.v4.jsonl.zstd log |
The public Agent surface (followup/steer/inject/cancel/inbox) has no
wake primitive. Re-queuing a pending prompt would need inbox.remove(id) +
followup(message), which writes a durable "canceled" inbox splice and re-insert —
and Session.append() refuses to reenter while a session/event observer is
running, which is exactly where this listener executes.
So the plugin calls wakeDriver(true) when it exists: it mutates only the
in-memory phase latch (the identical call send() makes on its post-abort path)
and appends nothing to the session log. If a future loop renames it, the plugin
falls back to the public remove + followup() path, deferred past the append.