dsh-interject

English | 中文

Delivers a queued chat message at the next step boundary instead of waiting for the whole turn to finish — without aborting the model call that is already running.

A DeepSeek Harness host plugin. No client half, no WebUI changes, no app.asar edits.

---

The problem

While an agent is working through a long turn (a single reasoning step can take minutes), a message you send mid-turn goes into the next-turn inbox queue. The loop only claims that queue after the entire turn ends — which may be many steps later.

What this plugin does

DSH's inbox has two queues:

Queue	Written by	Claimed at
next-turn	Enter while busy (queue and send)	after the whole turn ends
next-step	Cmd/Ctrl+Enter (interject and send)	the next step boundary (one step = one model call plus its tools)

The plugin listens for agent/inbox/inserted. When the policy matches, it moves the message from next-turn into next-step. At the start of the next step the model sees your message in its history, and the reasoning currently in flight is not discarded at all.

Why it cannot be read mid-generation: llm.stream(request) freezes the messages array when the request is built and the HTTP body is already on the wire — there is no channel to append input during generation. So there are only two options: wait for the call to finish (the default here), or abort and resend (interrupt).

---

Install

From the plugin market

Search for dsh-interject in the DSH plugin market and install it. The market resolves the package from this repository.

From source

git clone https://github.com/sdlhs/dsh-interject.git
cd dsh-interject
dsh plugin --profile desktop add .

The package declares dsh.bundle.patch, so the profile picks up cordis.patch.yml automatically and the row mounts on the next start. Restart DSH fully after installing or editing lib/index.js — DSH loads plugins as static ESM, so file edits do not affect a running process.

Manual mounting (without the market)

Add to your profile's cordis.patch.yml:

- insert:
    - id: dsh-interject
      name: dsh-interject

name must be the package name, not an absolute path — the market matches loader entries by name, and an absolute path leaves the plugin card reading "installed, not loaded".

---

Configuration

~/.dsh/dsh-interject.json — hot-reloaded, save the file and it takes effect; no restart.

{
  "enabled": true,
  "onQueue": "steer",
  "onSteer": "pass",
  "interruptWhile": "thinking",
  "graceMs": 250,
  "cooldownMs": 2000,
  "includeSubagents": false,
  "stallMs": 120000,
  "wakeFallbackMs": 3000,
  "bootMarker": true,
  "log": true
}

Field	Values	Meaning
onQueue	steer / interrupt / pass	What to do with a message that entered next-turn. steer = promote it to next-step (default, zero waste)
onSteer	pass / interrupt	What to do with a message that entered next-step. pass = hands off (default, it already waits for the next boundary)
interruptWhile	thinking / always	When a hard interrupt is allowed; with thinking, the tool-running phase degrades to a soft interjection instead of killing a long command
graceMs	ms	Delay after a hit, so messages sent in one batch are merged
cooldownMs	ms	Minimum gap between two hard interrupts on one agent
includeSubagents	bool	Also handle sub-agents (default: only root sessions you can see)
stallMs	ms, 0 disables	Write a "still running" line to the log when a single step exceeds this
wakeFallbackMs	ms, 0 disables	How long to wait for the driver to report idle after a hard interrupt; only used on the degraded path
bootMarker	bool	Write ~/.dsh/dsh-interject-boot.json (version / pid / effective config)
log	bool	Diagnostic log switch

The default never interrupts. To make Cmd/Ctrl+Enter a preemption gesture: {"onSteer": "interrupt"}. To switch the plugin off: {"enabled": false}.

Environment overrides: DSH_INTERJECT_CONFIG (config path), DSH_INTERJECT_LOG (log path; empty string disables logging), DSH_INTERJECT_BOOT (boot marker path).

---

Does it cost more tokens?

Not with the default configuration. The plugin sends no extra requests; it only moves a pending message to a different queue. The only increment is that the message enters history earlier, so the remaining requests in this turn carry it once more — a prompt-cache prefix hit, which is cheap.

Only interrupt really costs: the aborted request's produced output is billed as usual, and then a full history is resent.

---

Verifying it works

Get-Content "$env:TEMP\dsh-interject.log" -Tail 20

A working promotion logs next-turn -> next-step <id>. If a single step exceeds stallMs you see step N of turn M has been running Xs (still going) — use it to tell "one request is genuinely stuck" from "a long turn made of many normal steps" (the UI's "thinking, X elapsed" counts the whole turn).

Which build is the live process running? (editing lib/index.js needs a restart, and after a restart it is easy to doubt whether it took):

Get-Content "$env:USERPROFILE\.dsh\dsh-interject-boot.json"

version there is the build the running process actually loaded. The log is append-only, so seeing an old config line proves nothing; this file is rewritten wholesale on every apply().

The test scripts redirect their own boot marker to a temp dir via DSH_INTERJECT_BOOT; otherwise running the self-check would overwrite this evidence. If the pid in this file is not your DSH process, something else wrote it.

Offline self-check, no DSH needed:

node test/test-interject.mjs        # 46 checks: policy, capability probing, degraded paths
node test/test-interject-loop.mjs   # 18 checks: mini driver transcribed from agent-loop

Both scripts import ../lib/index.js by default; set DSH_INTERJECT_ENTRY to test another build.

---

Compatibility

Declared: >=0.2.0-rc.1 with no upper bound. The source of three host versions was read line by line to confirm the public surface this plugin uses:

Version	Status
0.2.0-rc.2	verified (development host)
0.2.1-rc.5	verified (npm latest / next)
0.2.2-alpha.2	verified (npm alpha)

Every capability is probed, not assumed — the plugin imports no @deepseek-ai/* package and uses only the ctx / agent objects the host passes in. That is not for the sake of older versions: it is so a future rename of an internal field degrades the plugin instead of breaking it. Anything undetectable falls back to a public path (for example, no phase means "wait for agent/status to report idle, then steer" instead of the wake latch).

The three versions agree on: ctx.agents.get/roots, agent.session, agent.status, agent.inbox.{nextTurn,nextStep,hasPending,claim,append,prepend,replace,remove,splice}, agent.followup/steer/inject/cancel, the session events step/start, step/end, turn/end, assistant/message, assistant/attempt, tool/call, tool/result, agent/inbox/inserted (payload {agent, message}), agent/disposed (payload {agent}), message.source.kind === 'user', and the internal field agent.phase.wakeRequested.

---

Known limits

• steer is not zero-latency. It can only act at the boundary of the current model call. If that call runs for ten minutes, you still wait ten minutes.
• Moving a message leaves one agent/inbox/spliced with outcome: 'canceled'. That is the shape of the public API (inbox.remove() goes through splice()). Verified harmless: the goal driver's sameQueued() requires source.kind === 'goal', and the UI's retireFailedSubmission() raises nothing for a plain-text submission.
• interrupt mode prefers the internal agent.phase.wakeRequested — latching it keeps the driver alive in the same turn, which is fastest. When it cannot be latched the plugin falls back to a public path: wait for the agent/status event to report idle, then steer() the message, with a wakeFallbackMs timeout. Both paths are covered by offline tests.
• Editing lib/index.js requires a full DSH restart (tray exit, not just closing the window). Only the JSON config is hot-reloaded.

---

License

MIT