Auto-continue for DeepSeek Harness

English | 中文

A pure-client composer plugin for DeepSeek Harness (DSH): when a reply is cut off at the model's output-token cap, it sends continue as a normal user message so the model can pick the work back up.

It is deliberately lighter than a host-side retry service:

• no HTTP routes, no polling, no host-side settings file;
• one event subscription for the active conversation;
• bounded in-memory state and persisted counters in localStorage.

Status: local developer preview, tested against DSH 0.2.1-rc.3. The public npm name is chosen at release time; see PUBLISHING.md.

There is an unrelated community plugin also named dsh-auto-continue (spix18) that focuses on rate-limit/quota retries. This repository focuses on output-token truncation, context compaction recovery, loop guards, and statistics.

Features

Truncation continuation	turn/end with reason.kind === 'max-tokens' sends "继续" / "continue" after a short delay.
Chunk-aware fallback	If the truncated turn's latest assistant message has no completed tool call, the follow-up asks for a skeleton and chunked writes instead of a bare “continue”.
Optional error-code retries	RATE_LIMIT, QUOTA / ACCOUNT_QUOTA, EMPTY_RESPONSE, plus user-defined codes. Built-in codes are matched exactly; custom codes are matched against the whole error payload.
Context recovery	CONTEXT_WINDOW_EXCEEDED runs /compact once (configurable), then continues; turning this off stops instead of retrying.
Loop guards	Consecutive limit, “no new output” limit, and auto-compaction budget. A normally completed turn resets the chain. A stop stays visible in the tooltip.
Composer switch	Lucide toggle-left / toggle-right pair next to the model selector, plus a per-conversation count badge.
Tooltip	On/off meaning, current state (waiting, compacting, stopped, error), stop reason, and per-conversation/all-conversation counts.
Settings page	Settings → Auto-continue: enable, hide the composer switch, trigger toggles, custom codes, context strategy, limits, continuation text, statistics, reset.
Statistics	Cumulative and per-conversation counts, counts by reason, successful compactions, conversations tracked.
Bounded memory	128 session states, 200 stat entries, 7-day claim cleanup; no polling and no retained event objects.

Install (local checkout)

dsh plugin --profile <profile> add /absolute/path/to/auto-continue

For the Desktop profile, install through Plugins → Add plugin or use the Desktop-managed command. After installation, open Settings → Auto-continue or click the compact switch beside the model selector.

A published release will use:

dsh plugin --profile <profile> add <published-package>

Install from GitHub

# latest main — review the current commit first
dsh plugin --profile <profile> add github:tkctly/dsh-auto-continue

# or pin a commit (recommended)
dsh plugin --profile <profile> add github:tkctly/dsh-auto-continue#<40-hex-commit>

The repository is a developer preview and currently keeps the local package name @local/dsh-auto-continue; the public npm name is chosen at release. The plugin has no build step, so Git installs do not need allowBuilds.

Usage

1. Open a conversation and click the compact switch to the left of the model selector (gray = off, blue = on).
2. When a reply hits the output-token cap, the plugin sends "继续" (or "continue" in an English UI) as a normal user message.
3. Hover the switch to see the state, stop reason, and counts.
4. Tune behavior under Settings → Auto-continue.

The switch and settings are shared across conversations. Settings and counters live in localStorage; they are not written to the session log or the Host.

Safety and loop guards

The plugin only acts on a completed turn boundary:

• the turn must be the latest turn;
• the session must be idle, open, not blank, not removed, with no queued submission and no prompt error;
• the truncation must have happened after the switch was enabled;
• each turn is claimed once, so reloads and multiple tabs do not send twice.

It stops instead of retrying when:

• the consecutive continuation limit is reached (1–100, default 20);
• N consecutive continuations produce no assistant text, reasoning, or tool call (1–10, default 3);
• the context-window compaction budget is exhausted (default 1);
• /compact is unavailable or fails;
• credentials are invalid;
• the user stops the task.

A normally completed turn resets the chain. Turn the switch off and on, or send a normal message that completes, to resume after a stop.

Permissions and data

• Reads the session's committed event stream through the public Client sessions service.
• Sends a normal user message (prompt(..., 'queue')) and, for context recovery, the official /compact command.
• Stores settings/statistics in localStorage.
• No network requests, telemetry, child processes, filesystem writes, or credential access.
• No DOM scraping or clicks; the switch and settings page are React slot registrations.

Troubleshooting

The model keeps saying it will write a file but never calls a tool

This is usually the model's output budget, not the plugin. Inspect the session log's assistant/message events: if usage.outputTokens equals the model cap and source.replayState.response.stopReason is length, the model ran out of output tokens while reasoning or while filling a large tool-call argument. An incomplete tool call is dropped, so the settled message can look like reasoning-only.

Fixes, in order:

1. Raise the model's output cap (maxTokens / defaultMaxTokens) in the local provider config if the server supports it.
2. Lower the model's reasoning effort, or choose a less reasoning-heavy model.
3. Ask for chunked writes: create a skeleton file first, then append or edit sections; keep one tool call under the output limit.
4. Start a fresh session with a direct, chunked instruction. Repeating “继续” in a long, cluttered session often reinforces the loop.

The plugin detects a max-tokens turn whose latest assistant message has no completed tool call and switches to a chunking-aware follow-up automatically. It cannot enlarge the model's output cap by itself.

It stopped after a few continues

A normally completed turn ends the chain by design. If the model ends the turn while only claiming it will write a file, the plugin cannot know the file is missing. Hover the switch to see the stop reason; the session log's turn/end reason distinguishes completed, max-tokens, and guard stops.

Compatibility

• DSH >=0.2.1-rc.3 <0.3.0 (engines.dsh).
• Node >=20 for development/tests.
• The plugin uses the public Client services sessions, locale, and slots. DSH is in developer preview; slot names and event shapes may change between minor versions.

Development

npm test          # Node tests for classification, backoff, guards, stores, and the engine
npm run check     # syntax + tests
npm pack --dry-run

The tests install a small window.__ModuleLoader__ capture shim, import the real client.js, and exercise its exported __test functions with fake event sources and timers. No React renderer, DOM, or Harness runtime is required.

License

MIT for this plugin. The composer switch uses Lucide toggle-left / toggle-right under the ISC license; see NOTICE.md for the required attribution. The bundle icon (icon.svg) is original artwork.