Install / Update
Install the dsh-context plugin on DeepSeek Harness with released NPM package:
dsh plugin --profile web add dsh-context@latestUpdate the dsh-context plugin:
dsh plugin --profile web update dsh-context@latestWhat's New
A timing-insight and density release: the 耗时统计 card now splits every model call into its wait and generation phases using the request's own stream timing from the durable log, and the Context tab packs tighter with halved card padding and gaps.
- The Timing card now splits the model call into TTFT (模型等待) and LLM Gen (模型生成) — the old single 模型调用 slice priced the whole step-start → assistant-message window; the card now reads the request's own stream timing off the durable log: the model wait is step start → first token, and the generation is first token → finalized message — the same rule as the harness's own session-stats fold, and the same TTFT / Generation figures the trajectory view shows per request. The donut grows to four slices (TTFT, LLM Gen, tool runs, overhead), with the legend rows keeping the true sums and call counts.
- Works across the whole support matrix: the assistant-chunk stream the split reads is present on every supported dsh baseline (0.1.1-rc2+ / 0.1.2-alpha2+).
- Calls whose stream recorded no token (legacy logs, aborted steps) stay unattributed and land in the overhead residue instead of being guessed.
- The projection's persisted state bumps to version 12 and cached rows refold from the log, so existing sessions gain the split automatically — no manual action.
- A denser Context tab — every card's inner padding halves (14×16 px → 7×8 px) and every card-to-card gap halves (14 px → 7 px), so more of the pane goes to content. Scoped to the Context tab; the /context dialog keeps the shared spacing.
Commits
- feat(host,client): split the timing model call into TTFT and LLM generation (ffc89ff)
- style(client): halve the Context tab card padding and gaps (c590456)
v0.40.1...v0.41.0 · 3 commits