dsh-status-turbo
Turns "what is it stuck on right now?" into one line of ticking text above the composer.
dsh-status-turbo is a plugin for DeepSeek Harness. It keeps a single thin status bar above the message composer:
● Running tool(bash) 2:14 3.1K tok
For those moments three minutes into a long command — page frozen, no log output, is it dead or not — this line is the answer.
What it shows
It sits inside DSH's own composer footer stats row (the 6 turns · 283 steps / 27.9M tok · cache hit 96% / 10% line),
on the right of the shipped entries — so it never adds an extra row or floats above the composer.
It only fills in what that row does not already show:
| Field | Meaning | Example |
|---|
| Phase | assembling / requesting / streaming / tool(<name>) / approval / compacting | ● Running tool bash · 2:14 |
| Phase elapsed | Refreshed every second. Needs no push — the client computes it from its own clock, so it keeps ticking while the log is silent | ditto |
| Step tokens | Input + output accumulated for the current step (since 0.19.2 only the number) | 3.1K tok |
| Gear | Not on the status bar since 0.19.2 — the app's own effort selector already shows the same value, so it stays in the detail panel. The value is the reasoning effort this step actually sent (read from the logged request/header, not guessed locally) | — |
| No progress | No new output for more than noProgressMs (default 90s) → that cell turns amber and shows how long it has been stuck (since 0.19.2 four words only; the full sentence moves to title) | 2:49 no new output |
The host projects this as the turbo wire view, version v: 13 with 37 keys — v1's 19 keys, the 9 keys added in v2
(mountStatus, prefixEpoch, prefixCause, prefixAt, lastTurnMs, topTools, hookFaults, idleLine, lastSpill),
then one key per later version: v3 activity, v4 waitTools / waitKind / subagents, v5 slowMs, v6 enhancement,
v8 handoff, v9 rotation, v13 newSession (v7 added enhancement.mode inside an existing key, and v10 added the
rotation.childId / seedOk / message fields inside rotation — neither raised the top-level key count). newSession
carries the New session page's copy-ready opening text ({ text, oldId, cwd }). The client pins EXPECTED_V = 13: on a
version mismatch it renders red text 投影版本错配（宿主 vN ≠ 客户端 vM） in place of the bar rather than disappearing.
Deliberately not repeated from the shipped row: session turns/steps, tok/s, total tokens, session-level cache hit rate,
context usage. The per-step cache hit rate stays in the projection data (hitRate) but no longer occupies a cell —
two different "cache hit" numbers on one line only confuse.
While idle the cell does not disappear — it keeps a one-line summary of the last turn: the phase name plus how
long that turn took (● requesting · last turn 42.3s). The host sends that as idleLine / lastTurnMs, so the
summary survives even with no new events.
Clicking the cell opens a detail panel below the composer (slot conversation.input.dock, id dsh-status-turbo.detail,
order 0); clicking again collapses it. While collapsed — or before the first snapshot arrives — the panel renders
nothing at all. The panel has 11 always-present sections, top to bottom: Now (what this step is doing — phase name ·
elapsed · tool name), Activity (the last few 'what the model thought / which tool it called / how it turned out',
newest first), this turn / last turn / cache and usage / prefix / slowest tools (top 5 by maxMs) / waits / rotation /
enhancement / handoff. Two more are conditional and simply do not appear when empty: the last oversized-result spill
with its absolute path (lastSpill) and recent faults (hookFaults).
The activity stream is memory-only (a ring buffer — never written to disk, no new dependencies, no change to the
requests sent): assistant/message records a reasoning excerpt (truncated to reasoningChars characters),
tool/call records the tool name and its redacted arguments, tool/result records the first resultHeadLines
lines of the result / error, and turn/end records the stop reason. Argument values whose key matches redactKeys
(case-insensitive substring match) become '***' before they ever reach the UI, circular references degrade to
[Circular], and a dirty argument never throws.
Install
Manual (profile assembly): drop the package into the profile's node_modules, then in the profile's cordis.patch.yml:
- insert:
- id: dsh-status-turbo
name: 'dsh-status-turbo'
config:
statusBar: true
noProgressMs: 90000
downgrade: true
downgradeFloor: low
minQuietSteps: 2
spillThresholdBytes: 32768
bulkInjectionBytes: 32768
evidenceWhitelist:
- '/exit code:?\s*0\b/mi'
- '/(^|\n)\s*(#\s*(pass|tests|suites)\s+\d+|\d+\s+passing|Tests:\s+\d+\s+passed|Test Suites:\s+\d+\s+passed|ok\s+\d+\s+-)/m'
- '/(^|\n)\s*(Successfully compiled|Compiled successfully|Found 0 errors|\d+\s+files?\s+checked,?\s*0\s+errors|Build succeeded|Done in \d)/mi'
stuckAction: warn
stuckRounds: 3
stuckRepeats: 3
mojibakeMinRuns: 2
Every key is optional; the block above is a short example (12 of the 46 default keys). The single source of truth for all
46 defaults is lib/state.js · DEFAULT_CONFIG, which the package's own cordis.patch.yml mirrors verbatim; every key with
its default and meaning is tabulated under Configuration below.
⚠️ Do not list the same package in both dsh.profile.bundles and the profile's cordis.patch.yml — it would be loaded twice.
Saving compute: per-step reasoning gear (B1)
The plugin watches the tool trajectory and drops the reasoning effort for that step when the step plainly
does not need deep thinking:
- ≥2 consecutive finished steps that ran only read-only tools (
read/glob/grep/web_search …)
with zero errors → drop one gear;
- a write-capable tool (
bash/edit/write/sql_exec …) or any error → back to the original gear
immediately;
- the original gear is whatever you selected for the session. The plugin only lowers, never raises, and
never touches your default model or provider.
Every downgrade leaves two independent traces: a B1 降档: … line in the DSH log, and an extra gear row
in the detail panel (0.19.2 moved it off the status bar).
| Key | Default | Description |
|---|
downgrade | true | Master switch. Only an explicit false disables it (then llm is not even injected) |
downgradeFloor | low | Lowest gear allowed. Compared by index in the model's own ordered effort list; if the model does not publish it, drops one gear at a time |
minQuietSteps | 2 | How many clean read-only steps are required. 1 is more aggressive, and easier to get wrong |
No gear name is hard-coded — everything is asked from ctx.llm.resolveModelInfo(). Whenever it is unsure
(no effort list, current gear not published, service missing, judgment threw), it runs the step at the
original gear and never lets the plugin fail a request.
Saving context: oversized tool results (B3)
When a single tool result is huge, the host half shrinks it before it reaches the model:
- it first asks the host's own
toolResultPruner.pruneContent to trim the body;
- if the result is still at or above
spillThresholdBytes (default 32768), it is handed to the host's existing
spill backend (spillStore.saveText), and the shrunken body carries a notice with the absolute path of the
saved file — a path you can read with offset/limit or grep directly.
The notice is rendered verbatim by lib/spill.js:
（已省略 {bytes} bytes；完整结果落盘于：{locator}。用 read 配 offset/limit，或 grep 这个路径继续。）
If the backend hands back a locator that is not an absolute path, the plugin degrades instead of guessing (no
fake path is ever printed). See the "Writes only its own data" note under Design constraints: for B3 the write is the
host's, and the plugin itself never opens a file for writing in that path.
Seeing a dead cache prefix (B4)
Cache-prefix invalidation is visible as a 前缀已失效 pill on the status bar (only while the invalidation is fresh,
i.e. within the last 5 minutes). The host counts three triggers:
| Trigger | Condition |
|---|
compaction-start / compaction-end | a compaction/start or compaction/end event |
bulk-injection | one tool/result payload ≥ bulkInjectionBytes (default 32768) |
The wire view exposes it as prefixEpoch / prefixCause / prefixAt. On a real recorded session this counter read
17 invalidations (bulk-injection ×9, compaction-start ×4, compaction-end ×4) — see tools/replay-log.mjs.
Notes appended inside the same tool return (C1 / C2 / C3)
Three cheap corrections are appended to the current tool result — they never add a round trip, and the session's
step count does not change:
- C1 — no verifiable evidence: a result that shows no evidence line at all (nothing matching
evidenceWhitelist,
exactly 3 regex sources, e.g. exit code: 0, N passing, Found 0 errors) gets
「上一条没有可核对的证据……别据此下结论。」
- C2 — truncated read: a
read that used offset/limit while the file has more lines gets
「你只看了 N / M 行（绝对路径）；后面还有内容没读，别基于这个片段下结论。」
- C3 — mojibake / noise: broken text (≥
mojibakeMinRuns runs of U+FFFD) or output swamped by noise gets
「这段输出里的中文已经损坏……」 / 「这次输出被噪声淹没（x / y 行不是目标输出）……」
Stuck detection (D) — warn level only
stuckAction defaults to warn, which is zero side effect (it records a fault, nothing more). nudge and
stop only happen if you configure them explicitly — and stop additionally requires the host's
workspace/session-activity to report a live session, because it would otherwise kill long-running work such as an
overnight build. Known gap: stuckRounds currently has no consumer — the host half only sees one half of the
signal (stuckRepeats over consecutive identical aggregate keys), so D is not a complete implementation.
The New session page (0.18.0)
When a session has grown enough to be handed over (see the rotation* keys under Configuration below), the rotation creates a blank
child session: since 0.21.1 nothing is written into it (before that the opening script was appended as its first user/message, so every new
session opened with the same paragraph). The New session page shows that opening text so you can copy it with one click and paste it into
the new session — no more making the new session guess what it was supposed to do.
- The text is built from the template:
{old} becomes the old session id and {cwd} the working directory. The
default template is 继续旧会话 {old}（工作目录 {cwd}）。需要它的上下文时用 @{old} 查阅，不要假设我记得；先跟我确认要接着做的那件事。
- What the page shows: a rotation section, the opening text inside a selectable
<pre>, a Copy button (it reports
'已复制' once the clipboard write succeeds, and '复制失败，请手动选中' when the clipboard is unavailable or refuses),
one line with the working directory, and a hint to paste the text into the new session.
- Where it lives: the bottom of this plugin's settings page (
dsh-status-turbo.settings, right under the config
form). 0.18.2 removes the 0.18.0 left-rail entry — one page instead of two, and no sidebar button to keep in sync.
- Data comes through the shipped hook: the settings page is root-scoped with no session binding, so the card reads
useSessions(state => state.current) and then state.byId[id]?.projectionValues?.turbo — the host already publishes
every session's projection values inside the session list. No new subscription, HTTP route, or host RPC.
- The child session is created through the agent lifecycle (0.19.0): the host half now calls
ctx.agents.create({ sessionId, meta }) instead of ctx.sessions.create(...). A session published outside the agent
lifecycle never gets a session-log writer, so it persists nothing — it lives in memory only and never appears in the
session list. That call is asynchronous: its outcome (and the title pass) is appended to rotation-trace.jsonl once it
settles. Without an agents service the host falls back to the old in-memory path.
- …and attached to its workspace (0.19.1): persisting is not enough. The left-hand session list is built from
each workspace's
sessionIds, not from disk, so once the create promise settles the host half resolves the
workspace by normalized path (workspaceRegistry.resolveByPath(cwd)) and calls attachSession(childId) — the same
step the host's own new-session and fork paths take. The outcome (attached / no-registry / no-workspace /
resolve-failed / attach-failed) lands in rotation-trace.jsonl; anything but attached also logs a warning.
- …carrying a model and a preset, and switching by itself (0.19.3): the child agent is created with the parent's
model (
agentOptions, falling back to ) and the parent's (mounted
through ), because the deployment persona section renders and a child without one fails its first
run.
Not implemented this round: B2
Backgrounding long commands (returning a job id within 10s instead of blocking) is not implemented. The
integration target does not exist on this machine (no dsh-plugin-background-tasks, and PreToolDecision has no
result-replacing branch), and there is no supported hook for it. Evidence and the three follow-up leads are in
docs/plan-B2.md. Please do not expect a B2 usage that is not there.
Configuration
| Key | Default | Description |
|---|
statusBar | true | Mount switch for the status bar. The host projects it into the wire field mountStatus ('enabled' / 'disabled'); the client has three states: enabled → normal bar; disabled → renders nothing; field absent (an older host that does not send it) → a red cell 状态条开关缺失：配置未送达 instead of a silent no-op. Off means "no bar", not "silently dead switch" |
noProgressMs | 90000 | No-progress threshold in milliseconds |
downgrade | true | See the section above |
downgradeFloor | low | See the section above |
minQuietSteps | 2 | See the section above |
readOnlyTools | 10 tool names | 0.9.2 (P1-4): the B1 read-only allowlist — only explicitly read-only tools count as "pure read / list / search"; anything not listed (bash / edit / write / sql_exec / univer_execute…) counts as "may write", and one such call breaks the consecutive run (deliberately conservative). Elements must be non-empty strings and invalid ones are filtered; an empty result falls back to this table ([] means "not configured", not "no read-only tools"). Defaults: read / read_image / glob / grep / list_dir / search_context / web_search / web_fetch / cordis_inspect_list / cordis_inspect_query |
spillThresholdBytes | 32768 | B3: a result body this many UTF-8 bytes or larger may be spilled (AND-ed with the harness tool-result pruner) |
localTrimHeadChars | 4096 | 0.6.5: characters of the head kept by the local trimmer used when the host pruner is unavailable (spillThresholdBytes decides to spill; these three keys decide how much the model sees) |
Restart DSH after changing configuration.
Note: the per-key behaviour of the handoffExport* and rotation* families (0.7.0 handoff export, 0.8.0+ rotation) is
described in Chinese in README.zh.md §「交接文档自动导出」/§「到量轮换」; the defaults above are the
same values.
Manual acceptance
| # | Steps | Expected |
|---|
| 1 | Have the model run ping -n 120 127.0.0.1 (~2 min) | Within 1s the bar appears, phase = Running tool(bash), elapsed ticks from 0 |
| 2 | Do nothing for 2 minutes | Elapsed keeps advancing; a silent log never freezes it |
| 3 | Wait for the command to finish | Phase returns to requesting, then the cell flips to the idle summary last turn … — it does not vanish |
| 4 | Set statusBar: false and restart | No bar at all (the host still folds state, very cheap). If the field mountStatus is missing — i.e. an older host half — the cell instead shows the red 状态条开关缺失：配置未送达; that red cell is a failure signal, not a working "off" |
| 5 | Produce more than spillThresholdBytes (32 KiB) of tool output in one result | The body is pruned and carries a notice with a readable absolute path (the spill locator) |
| 6 | Have the model claim "tests passed" without running them; or make it read a file with offset/limit | A correction note appears inside the same tool return — no extra round trip, session step count unchanged |
| 7 | Run a command that emits GBK mojibake, or one buried in noise | An "encoding broken" / "noise" note appears in the same tool return |
| 8 | Leave one step silent for 2 minutes | At 90s (noProgressMs) the no-progress cell turns amber and counts up |
| 9 | Inject memory / trigger compaction mid-session | A 前缀已失效 pill appears within 5 minutes, cause = 压缩开始 / 压缩结束 / 大块结果注入 |
| 10 | Remove the conversation.composer.dock declaration (or use a non-existent slot name) and restart | UI does not crash; console shows [dsh-status-turbo] 状态条 3 秒内未能挂载到 conversation.composer.dock… |
| 11 | Tamper with the projection version v | Red text "投影版本错配" appears in place of the bar instead of silently vanishing |
The original §6 row 2 ("long commands stop blocking; a background job id within 10s") does not apply this round —
B2 is not implemented, see docs/plan-B2.md. It is deliberately dropped and explained rather than
left as a gap, so the rows above are renumbered into a continuous 1–11.
Development
node --test "test/*.test.mjs" # pure fold + host/client contracts + UI contract + hooks
-
No build step: lib/ is source and artifact at once.
-
No zod: projection schemas use a passthrough stub (see docs/design.md §2.3).
-
Verify without opening the GUI — replay a real committed session log:
node tools/replay-log.mjs <session.v4.jsonl.zstd> [--peek] [--max N] [--stall-ms N]
It runs the exact same fold and labels over the log, printing which event changed the strip to what, and finally the longest tool-silence window — i.e. what the strip was showing during it. Measured on this session:
# jsonl lines 1182 # parseable events 1182 (bad lines skipped 0)
# longest tool window: pwsh silent 50.4s (at session +52m34s,
# strip should read "执行工具(pwsh) 50.4s")
(Session logs live in D:\deepseek-harness\sessions\<workspace-slug>\session-<id>\session.v4.jsonl.zstd and are multi-frame zstd; the tool splits on the 28 B5 2F FD magic and inflates frame by frame.)
-
Layout: lib/state.js (pure fold: no services, no Date.now(), no third-party dependency; one Node builtin import — node:crypto for the A2 tool-argument fingerprint), lib/index.js (host half: the turbo wire projection plus B1 gear switching and the B3/C/D post-execute hooks), lib/config.js (the config schema built from DEFAULT_CONFIG, the live-config reader, and volatile-value unwrapping), lib/gears.js (B1 judgment, zero imports, tool-trajectory only), lib/spill.js (B3 — hands oversized results to the host's spill backend and renders the locator notice), lib/notes.js (C1/C2/C3 note texts and judgments, plus tool aggregation for the detail panel), lib/handoff.js (0.7.0 handoff export: threshold/throttle/freshness judgments, bili command probing, export argv), lib/rotation.js (pure rotation judgment shouldRotate, fed by the same handoffExport* thresholds), lib/rotation-host.js (0.9.0 plan B host action: creates the blank child session and writes the seed message — never throws), lib/client.js (client half: the summary cell at conversation.composer.dock, the detail panel at conversation.input.dock, and the New session card at the bottom of the plugin's settings page), tools/replay-log.mjs (offline replay), test/hooks.test.mjs + (host-hook and UI-contract guards), (0.18.0 New session page guard: text card, copy button and its two messages, three-state fallback, icon and registration), + (config-mirror / wire-version and client-audit guards), plus the per-pillar test files for the fold, gears, notes, spill, handoff, rotation and the two halves.
Design constraints (non-negotiable)
- Reads only committed session-log events (via a projection). No host memory, no polling, no DOM injection, no React-fiber walking.
- No network requests, no routes. Writes only its own data: dsh-status-turbo itself does import write APIs (
existsSync from node:fs and mkdir / readFile / readdir / rename / writeFile from node:fs/promises, lib/index.js:58-59), and it touches writes in exactly two places, both of them its own data: ① the oversized result body for B3 is delegated to the host's existing spill backend (spillStore.saveText) and the plugin only prints back the absolute path it was handed — it opens no file for that path itself; ② since 0.7.0 the handoff export creates its directory and drops a .gitignore into it (lib/index.js:538, content in lib/handoff.js:246). Both of those calls goes through the plugin's own hostIo seam (lib/index.js:436) — it writes to disk no other way, and it never writes user settings.
- Never touches
agent-default-model, compaction config, or any user setting.
- Never rewrites tool arguments or intercepts tools.
- Every failure must surface (in the UI, or at minimum a
console.error naming the cause). Silent return is not allowed.
License
MIT — see LICENSE. Third-party notes in THIRD_PARTY_NOTICES.md.
