dsh-response-meta
English | 中文
A DSH Web plugin that adds one live runtime summary to every model reply:
model, reasoning extent, tokens per second, timestamp, total runtime, and
time to first token (TTFT). The summary appears while the model is responding
and updates as chunks arrive; it does not wait for the reply to finish.
The plugin hides the duplicate native assistant timestamp while preserving user
message timestamps and the native usage and runtime detail buttons. Manually interrupted replies, including replies stopped during
reasoning, receive an inline summary as well.
Example in a Chinese UI (11 px muted text):
# While streaming, inline in the message stream
deepseek-v4-pro · 思考 1.5k 字 · ~42 tok/s
# Completed reply, below the message actions
deepseek-v4-pro · 思考 15 tok · 46 tok/s · 14:16 · 用时 2分24秒 · 首 token 2.1秒
# Interrupted during reasoning, inline in the message stream
deepseek-v4-pro · 思考 1.5k 字
- Model comes from the turn's
request/header session event. Later turns
without a new header inherit the most recent known model.
- Live updates appear as soon as model or streamed-token data is available.
Reasoning extent and estimated throughput refresh at animation-frame cadence;
final provider usage replaces the estimate when the reply completes.
- Reasoning extent prefers provider-reported reasoning tokens and falls
back to the reasoning text length.
- Tokens/s is output tokens divided by decode time, from first token to
message completion, matching DSH's native timing semantics.
- Timestamp shows
HH:mm for today's replies and includes the date for
older replies.
- Total runtime spans the whole turn; TTFT spans the current step from
start to its first token.
- Interrupted replies usually have no usage record, so the plugin shows the
model and streamed reasoning extent. Real usage is retained when a
finalization race provides it.
- Missing fields are omitted individually; nothing is rendered when all fields
are unavailable.
Requirements
DSH 0.1.5-rc.1 with a Web profile. Version 0.4.0 handles transient assistant/live-chunk events and the packed streams embedded in durable messages and interrupted attempts. It reconstructs reasoning counts and token timing after a cold load, and preserves completed throughput and TTFT when the host retires transient chunks. Settlement replaces transient counters to avoid double counting. Keep plugin 0.3.0 with DSH 0.1.2-rc.1, and 0.2.1 with older DSH installations.
Install
dsh plugin --profile web add github:Unintendedz/dsh-response-meta#v0.4.0
Restart the running DSH Web service after installation. Plugins are loaded
when the service starts.
Local development
The local installer requires an explicit DSH_HOME and --profile (or
DSH_PROFILE). It never defaults to an existing home or the web profile.
It removes and re-adds the local file: snapshot, reusing the selected profile's
pnpm storeDir, including paths containing spaces. Restart only that profile's
Web instance afterward.
Use a fresh temporary home and synthetic workspace for every test:
export DSH_HOME="$(mktemp -d "${TMPDIR:-/tmp}/dsh-response-meta.XXXXXX")"
export DSH_PROFILE=audit-response-meta
mkdir -p "$DSH_HOME/workspace"
./scripts/install.sh --profile "$DSH_PROFILE"
Use a dedicated free loopback port and synthetic provider/session data. Never
copy an existing profile's settings, credentials, sessions, caches, or browser
state. Record the home, profile, port, workspace and process ID before any Web
interaction; stop the verified test process and remove only its temporary
directory afterward.
Uninstall
dsh plugin --profile web remove dsh-response-meta
How it works
- Host (
lib/index.js): registers a dependency-free projection unit on
ctx.sessionProjections. It folds session events, tracks the active
turn/step, records the last in-turn request/header model in byTurn, and
inherits the previous model for header-less turns. Headers outside turns,
such as title or compaction requests, are ignored. Projection updates reach
the browser through DSH's existing session-projection channel. The plugin
implements the projection contract (stateSchema + wire.viewSchema).
- Browser (
lib/client.js):
- Running replies publish one incremental
dsh-response-meta-aborted
conversation node from step/start; streamed chunks refresh it at most once
per animation frame. A visible finalized message hides that stable live node
and hands off to the exact completed summary. If the step is interrupted,
failed, or aborted, the same node remains visible.
- Completed replies register in
conversation.chat.assistant-actions. The
component reads usage, step timing, turn timing, reasoning content, and the
per-turn model projection for the exact messageId. A scoped structural
selector hides only the duplicate native assistant timestamp.
- The incremental node renders through a keyed
conversation.chat.node
entry. This also covers reasoning-only steps that never produced
turn/end.
- Actions remain on the first row; the complete summary owns a responsive
second row. Narrow screens wrap without truncation or horizontal overflow.
Verify
npm test
DSH_TEST_IDENTITY=/path/to/test-instance.json \
node tests/live-e2e.mjs http://127.0.0.1:33880
The live command also requires DSH_HOME and DSH_PROFILE from the isolated
launcher. Its identity JSON must contain dshHome, profile, host (exactly
127.0.0.1), numeric port, the running server's numeric pid, workspace,
and synthetic: true, all matching that instance. Both home and workspace must
be beneath the operating system's temporary directory, and the profile must
start with audit-, test-, or fix-. The live harness rejects port 33080.
Configure a local synthetic model provider in this new profile before running
it; no credentials or settings from an existing profile are needed.
The check creates a synthetic session and passes only when the same turn ends
with reason completed, the final visible answer is exactly OK, and that
turn has a nonempty model projection. An early projection, interrupted or
failed partial reply, wrong answer, or timeout fails the check. The browser
scripts in tests/ cover live/interruption behavior and completed-row layout;
run them only in the same isolated instance and a fresh browser context.
