Until Done (dsh-until-done)

A DSH plugin. When a turn ends normally but the task is only half done, it sends one more "continue" and pushes the model back to work. A reply counts as truly finished only when it closes with the completion keyword.

Forked from HsiangNianian/dsh-auto-continue; recovery, offline pause and loop guards are kept, while completion detection and failure classification were reworked.

What changed in this fork

• Completion keyword (default 「完成」). With it on, turn/end = completed no longer always means "stop": a reply that does not end with the keyword is continued, and only a reply that ends with it stops. The keyword is configurable; only the tail counts, so 完成之前还要跑测试 does not finish a task.
• Rule injected into the system prompt (order 900, interpolate: false) so the model knows the rule from the first turn instead of learning it after the first auto-continue. The text is editable and {keyword} is replaced with the keyword.
• Finer failure classes: tool failure, 429 rate limit and 5xx upstream errors each get their own continuation text. A tool failure names the tool and carries its error excerpt, and asks the model to diagnose before retrying instead of treating the failure as completion.
• Optional "real work only" mode: a plain chat turn (no tool call, no active goal) is left alone, so "hi" is not answered ten times.
• Independent namespace: contract namespace, HTTP routes, log prefix and UI copy all belong to dsh-until-done, so it can be installed beside the upstream plugin without clashing.

Install

# from the release tarball (recommended)
dsh plugin --profile web add https://github.com/KunKun24/dsh-until-done/releases/download/v1.0.0/dsh-until-done-1.0.0.tgz

# or from a local checkout while developing
dsh plugin --profile web add link:/path/to/dsh-until-done

Restart DSH, then open More settings → Plugins → Until Done in the bottom-left corner.

Key settings

Setting	Default	Meaning
Require completion keyword	on	A reply without the keyword counts as unfinished and is continued
Completion keyword	完成	Prefer a word that rarely appears in normal prose
Put the rule in the system prompt	on	Tell the model from the first turn
Completion rule text	built-in	{keyword} is replaced with the keyword
Only continue turns with real work	off	Plain chat then stops auto-continuing
Continue text (tool failure)	built-in	Supports {tool} / {result}
Continue text (rate limit)	built-in	Supports {status} / {code}
Continue text (upstream error)	built-in	Supports {code} / {status}
Autonomous loop	off	Every normally finished turn is continued until the keyword appears

Retry policy (delay, consecutive cap, backoff, offline pause, classification), startup recovery, loop guards, the status bridge and browser notifications are unchanged.

Decision flow

turn/end(completed)
  └─ autonomous loop on?
       ├─ no  → legacy behaviour (budget cleared only on visible output)
       └─ yes → does the last assistant text end with the keyword?
                ├─ yes → task complete: reset loop counters, stop
                └─ no  → (real-work-only && no tool this turn && no active goal)
                          ├─ yes → plain chat, stop
                          └─ no  → send the continuation

"Last assistant text" skips messages that carry a tool call (those are not a final answer) and tolerates trailing punctuation, quotes, asterisks and invisible format characters.

Development

npm install
npm run build     # esbuild + tsc → lib/
npm test          # includes tests/smart-completion.mjs
npm run typecheck

Layout: src/host/engine.ts (engine), src/shared/core.ts (settings, decisions, templates), src/client/* (settings card, composer toggle, status bridge).

License

MIT. Original work Copyright (c) 2025 HsiangNianian; modifications Copyright (c) 2026 KunKun. See LICENSE.