bowenliang123/dsh-context v0.78.1

on GitHub

Install / Update

Install the dsh-context plugin on DeepSeek Harness with released NPM package:

dsh plugin --profile web add dsh-context@latest

Update the dsh-context plugin:

dsh plugin --profile web update dsh-context@latest

What's New

v0.78.1 is a patch over 0.78.0. The Command Executions rail now runs on one exclusive lens that always reads through the filter box, and its cells match the Context Stats surface; the stats lanes count program invocations instead of records, drop the "+0 more" chip, and let their folded tails jump to the full lists; the Timing card's tool slice becomes a true partition of active wall time; every card title carries an icon.

Features

• Card titles carry icons

  • Token Stats, Timing, Context Stats, Context Trend, Context Events, Context Browser, Command Executions, File Activity, and the Plugin card lead their titles with a matching icon, resolved through the harness's own primitives.

• Folded "+N" chips jump to their full lists

  • The tool lane's tail opens every tool result in the Context Browser; the command lane's tail reveals the Command Executions card without seeding a program lens.
  • Both tails carry a hover hint and the same keyboard-and-click wiring as the named chips.

• Program counts separate invocations from runs

  • The program distribution counts every recognizable invocation, so grep a | grep b && grep c books grep ×3 instead of once per record.
  • Each row also reports the number of distinct run records containing the program, matching what the program filter shows; the summary and section labels spell the two units out.
  • The Context Stats command lane ranks by the same invocation counts, so both surfaces agree.

Fixes

• The Timing card's tool slice partitions active time — parallel tool runs were summed per call, so the tool row could print more wall time than the donut total and read a 100% share beside the other categories. The fold now records the per-step tool busy union (toolBusyMs), every row share derives from the same attributed windows, and the tool row prints its busy window while keeping the per-call sum in its tip when the two differ. Model rows keep their measured durations. The overview KPI and team aggregation follow the same figure with a clamped fallback for pre-field rows, and cached projections refold (projection state version 35).
• One exclusive lens in the Command Executions rail — the attention facts and the program ranking stay session-wide. Every rail click clears the other facets before applying its own, writes its keyword into the filter box (program picks and the Context Stats jump included), and a second click on the active lens clears it; editing the box releases a stale program pick. The active attention row carries aria-pressed and a tonal highlight.
• The skill tool chip lands on its own bucket — a skill call's result folds under Skill Injections, so its chip now opens that bucket instead of an empty "No rows match the current filter".
• No "+0 more" chip — a tool lane whose whole tally fits renders it whole, and the tail chip appears only when it hides something; an empty tool lane keeps "+0" as its figure. Both lanes now share one chip lane: measured fit, fold chip, gauge, and empty state are the same code.
• Attention rows read Failed and Repeated — the labels name the state each ×N counts instead of a streak (失败 / 重复).
• Rail cells match the Context Stats cells — the summary tiles and attention rows take the shared control-track fill, each attention row reads as its own cell, and hover steps to the interactive-active tint.
• Status glyphs — running is now ⏳ and killed/stopped 🛑; completed ✓ and failed ✕ keep their status colors.
• Overview cards tighten — card padding drops to 10×12 on the overview page.

Changes

• Auto overview scan defaults to 100 steps (was 200), in the Config form and the client preferences.
• Production builds stop emitting source maps.
• docs/compatibility.md records the timing card's busy-union limitation and refreshes the verification stamp.
• Coding standards move fully into AGENTS.md; CODING_STANDARD.md is removed.

Commits

• chore(release): 0.78.1 (4e60c910)
• fix(timing): make the timing breakdown a partition of active time (e3c42029)
• style(agents): adjust formatting for clarity and consistency in guidelines (36212c29)
• style(overview): adjust padding for card elements in overview page (b6ce2d92)
• fix(config): disable sourcemap generation for production builds (2e62ed4c)
• feat(stats): let the folded "+N" chips jump to their full lists (dc2928e9)
• fix(stats): rank the command lane by program invocations (0bd16bd2)
• feat(cmds): count program invocations and runs in the distribution (5f2e0c6c)
• Set auto-overview default to 100 steps (b0c095d2)
• feat(context): add icons to card titles (9a839400)
• fix(i18n): label the attention rows Failed and Repeated (1c1c4f05)
• style(commands): fill the rail cells like the Context Stats cells (084757c4)
• refactor(coding): consolidate and enhance coding standards in AGENTS.md (46b4a35e)
• fix(stats): land the skill tool chip on its Skill Injections bucket (3f184ccd)
• fix(stats): drop the zero chip from a fully visible tool lane (5fe1394e)
• style(commands): swap flat status glyphs for emoji (5f9a803d)
• fix(commands): unify the Command Executions rail lens (2e4bd7e0)

v0.78.0...v0.78.1 · 17 commits

Contributors

• @bowenliang123