dsh-auto-continue

DSH Web UI plugin — when a request is interrupted by a network error or any other non-human cause, it automatically types 「继续」 and sends it for you.

English · 中文

What It Does

For DeepSeek Harness (dsh web): whenever a request in the web GUI gets interrupted by a non-human cause, the plugin simulates the user typing 「继续」 and sends it, so the agent keeps working without manual intervention. The message enters the session log exactly like a manual prompt — the model sees it, and the interrupted work resumes. Since 0.8.0 the engine runs inside the host process (single instance), so it keeps watching even with every browser tab closed, and multiple open tabs can never double-send.

Smart recovery (all configurable):

• Error classification — transient failures (network / timeout / 5xx / 429…) are auto-resumed; permanent ones are skipped and notified, because retrying them never helps. A failure counts as permanent when its HTTP status is 401/403 or its code/message matches auth, credential/API-key, balance/quota, unknown-model, or context-length/overflow keywords. Turn classification off to resume everything
• Adaptive backoff — consecutive failures wait longer each time (cooldown × factor: 20s → 40s → 80s…), capped at the max backoff, instead of hammering a broken upstream
• Templated continue text — continueText supports {code} {message} {status} {tool} {turn} {errorCount} {sessionTitle} {elapsed} placeholders, so the resume message can carry the failure context ("继续 (git push failed: UPSTREAM)"); a separate template fires on max-tokens (e.g. "继续输出, 不要重复已生成的内容")
• Idempotency guard — before resuming, the plugin inspects the last tool call: if its result is unconfirmed (the turn died mid-tool, e.g. a git push that may have gone through), the resume message tells the model to check state first and not to rerun; if the tool is confirmed done, it says so and asks not to repeat it; a failed tool gets no guard (retrying it is the point). Both guard texts are configurable ({tool} / {result} placeholders)
• Pause — a global Pause auto-continue toggle in the settings card stops everything (live + scan) instantly; per-session pauses (e.g. via a notification button) suspend only one session until they expire. The Resume now notification button is the one explicit exception: pressing it is the user asking for exactly one send, pause or not
• Notification buttons — notifications carry Resume now (send immediately, ignoring cooldown, the consecutive cap and any pause) and Pause this session 1h actions
• Loop guard — watches running turns too. Three signals trip the guard, which cancels the turn and restarts it with a configurable loop text ("stop repeating, try another way"): the model repeating the exact same message several times (any length — e.g. "Let me test variants of the regex…" ×7), many short messages inside a short time window with no tool call in between (the "Let me read…" spin), or the same tool called repeatedly with the same arguments and the same results (a changed argument or result counts as progress). The cancel carries an internal marker so it is never confused with a user stop — the restart only happens for guard-initiated cancels. Thresholds, the time window and the loop text are configurable
• Stats panel — the settings card shows today's auto-continue count, recoveries, failures, permanent skips, give-ups and loop breaks, broken down by error code, with a one-click reset
• Browser notifications — optional alerts when auto-continue fires, gives up, or hits a permanent error; the browser asks for permission on first use, and nothing is shown again after a denial

It watches the live event streams and reacts to:

Never auto-continues: user-aborted turns (aborted) or policy rejections (blocked); live interrupted turn-ends too — that marker is only written by crash repair when the host reloads, so orphaned turns are recovered by the startup scan, not the live path; sessions the host already resumed itself; running sessions or sessions with queued messages; subagent sessions; anything inside the cooldown / consecutive-cap windows (configurable in the settings card, below).

How It Works

The host-side engine subscribes to the session event firehose inside the dsh host process — exactly one engine, regardless of how many tabs are open (the duplicate-send class of bugs cannot exist by construction). On an interruption it waits a grace period (default 3 s) — if the host starts a new turn by itself (turn/start), the auto-continue is cancelled — then sends the configured text through the agent registry (agent.followup, the same queue the Send button uses).

On host boot it also scans the live sessions: a session whose last turn ended with a non-human reason within the scan window (default 15 minutes), with no later turn/start or user message, gets resumed automatically too (e.g. the host crashed while the browser was closed — the agent-loop resumes the session and the engine picks it up).

The browser half is a thin shell: the settings card, plus a status bridge that shows notifications (with Resume now / Pause this session 1h buttons, routed back to the host engine) and feeds the card's stats / paused-sessions panels.

Quick Start

DSH plugins install into a profile (dsh web → web profile). Install, restart dsh web, done.

Requires DSH ≥ 0.1.0-rc.7 — newer versions register the settings card through the keyed settings.plugin.item slot; on rc.6 and earlier the plugin fails to load (list slot ... requires options.id). Upgrade dsh first.

From npm (recommended)

Published as dsh-client-auto-continue:

dsh plugin --profile web add dsh-client-auto-continue
dsh web

Directly from GitHub (no clone needed)

Installs straight from the repository's default branch — built artifacts are committed, so no local clone or build step:

dsh plugin --profile web add github:HsiangNianian/dsh-auto-continue
dsh web

This tracks the main branch rather than released tags — great for trying the latest changes, while the npm method above is the stable choice. Switching between install sources is just re-running dsh plugin --profile web add <other-spec>; the profile dependency is replaced in place.

From this repository

Requires Node.js ≥ 18.

git clone https://github.com/HsiangNianian/dsh-auto-continue.git
cd dsh-auto-continue
npm install
npm run build

# the package carries its own cordis.patch.yml (dsh.bundle.patch),
# so the plugin row registers itself
dsh plugin --profile web add link:$(pwd)

dsh web

Manual (no pnpm / dsh plugin needed)

ln -sfn "$(pwd)" ~/.dsh/profiles/node_modules/dsh-client-auto-continue
# then append to ~/.dsh/profiles/web/cordis.patch.yml:
#   - insert:
#       - id: auto-continue
#         name: 'dsh-client-auto-continue'
dsh web

Switching from a manual install to dsh plugin add? Remove the manual insert entry first — the bundle patch registers the row and a duplicate would conflict.

Settings exposure: since DSH 0.1.0-rc.7 the web settings surface is registry-driven — every namespace a plugin registers is served, so the settings card works out of the box, no vendor patch needed (the plugin requires rc.7+, see Quick Start).

Verify & uninstall

dsh --profile web --dump-config | grep auto-continue   # config layer mounted

In the browser console (Ctrl/Cmd+Shift+I): [auto-continue] 已启动(文本="继续", …) — every detection and auto-send is logged.

dsh plugin --profile web remove dsh-client-auto-continue   # npm / repo install
# or remove the symlink + the insert entry                  # manual install
dsh web

Configuration

Everything is configurable from the GUI — no file or console edits needed. Open Settings → Plugins and find the dsh-client-auto-continue configuration card, right where every other plugin's config lives. Besides the fields below, the card shows a live stats panel (today's activity with a reset button) and the list of paused sessions (each with a per-session resume button).

Or skip the GUI and edit the config file directly — the engine reads the plugin's section from ~/.dsh/settings.yaml (one shared file for every plugin's sections), so this works in any install, patched or not. The file is watched and re-read automatically, so changes apply live; restart dsh web if a page that was already open doesn't pick them up. Fields you leave out fall back to the defaults in the table below:

auto-continue:
  paused: false
  continueText: '继续'
  continueTextMaxTokens: '继续'
  guardTools: true
  guardPendingText: '(上一步工具「{tool}」可能未完成, 先确认状态再继续, 不要重复执行)'
  guardDoneText: '(上一步工具「{tool}」已完成, 结果: {result}; 不要重复执行, 直接继续)'
  graceMs: 3000
  cooldownMs: 20000
  maxConsecutive: 3
  scanOnBoot: true
  scanLimit: 8
  freshMs: 900000
  verbose: true
  classify: true
  backoffFactor: 2
  backoffMaxMs: 300000
  notify: false
  loopGuard: true
  loopShortChars: 40
  loopWindowMs: 30000
  loopShortCount: 12
  loopRepeatText: 4
  loopToolRepeat: 5
  loopText: '(检测到你可能陷入循环, 请停止重复刚才的动作, 换一种方式继续)'

How the card works:

• Edits are staged — nothing reaches the disk until you hit Save; an unsaved badge marks the card while drafts are pending, and Discard drops them
• A field you changed shows an Overridden badge with a per-field Reset to default button that restores the built-in value
• Boolean fields are tri-state: Inherit (use the default) / On / Off
• Invalid drafts (non-numbers, values below the minimum) block the save with a hint
• In a read-only deployment the card shows the stored values but disables every control
• Changes apply immediately after Save and persist in ~/.dsh/settings.yaml (uninstalling the plugin leaves the section behind — harmless, delete it by hand if you like)

continueText (and continueTextMaxTokens) accept the placeholders {code}, {message}, {status}, {tool} (last tool call before the failure), {turn}, {errorCount} (consecutive failures including this one), {sessionTitle} (from the session list) and {elapsed} (time since the failure, e.g. 1m5s) — e.g. 继续 ({tool}: {code}) becomes 继续 (git push: UPSTREAM). The guard texts accept {tool} and {result} (a truncated excerpt of the last tool output).

Privacy & permissions

The plugin is browser-only and touches no files, credentials, or network beyond the dsh host:

• It opens the same two read-only event streams the web UI already uses (no extra server, no third-party endpoints)
• The engine's only automatic write is sessions.prompt — the same call the Send button makes — with the text you configured (saving the settings card writes the auto-continue section of ~/.dsh/settings.yaml through the normal settings API, exactly like any other setting)
• No browser storage at all: the single host-side engine keeps its cooldowns, send caps, pauses and stats in process memory
• Browser notifications are opt-in (notify setting) and permission is requested on first use only

Development

npm run typecheck   # tsc --noEmit
npm run build       # lib/client.js + lib/index.js + lib/types
npm run watch       # rebuild on change; host HMR hot-reloads without a page refresh
npm run test        # node tests/simulate-host.mjs — 15 host-side behavioral scenarios

While npm run watch runs, the profile's client-hmr row polls lib/client.js every 500 ms and hot-reloads the plugin in the browser — no server restart needed for code changes.

CI runs dsh-plugin-check on every push to main and every pull request, and gates releases — the tag workflow refuses to publish while the check fails.

Activity

Links

• Repository: github.com/HsiangNianian/dsh-auto-continue
• LINUX DO: linux.do
• DeepSeek Harness: github.com/deepseek-ai/deepseek-harness
• dsh-plugin-check: github.com/omdsh-dev/dsh-plugin-check — health-check your own DSH plugin repos

License

MIT © Hsiang Nianian