v0.1.0

Added

• Grouped bar and line charts. chart takes seriesNames plus one values array per item: {seriesNames: ["修前","修后"], data: [{label:"图边", values:[13,41]}]} draws one group of two bars per metric, with a legend. A before/after comparison had no way to be expressed, so it was flattened into one bar per metric-and-time with the time tag concatenated into the label — which stops being a comparison.
• Closing-turn channel (the figure no longer depends on the model remembering). agent/turn-stopping now runs the same structure check the nudge uses: when a drawing session tries to close on an answer that carries structure (a table, five-plus list items, or comparison/plan/flow prose) and no figure, the plugin steers the agent, which re-opens the inbox so the model owes another step and draws the figure. Capped at two steers per session with a two-turn gap, so it cannot loop. Measured cause: an AnGIneer session drew 12 figures over 29 turns while individual answers shipped tables, and the old rate-only rule stayed silent for all 29 turns — the model's own reasoning said "I already drew 9 figures, I'll use a table here".
• Guidance runs as a procedure, not a preference: plan the structures first, make the draw_figure calls in the same batch, write the answer around the figures, then self-check the prepared answer for a table or long list that should have been a figure.
• raster-diagnostic.txt in the session workspace records every raster degradation where the user can read it, plus a session figure-coverage line in the tool result (figures embedded in replies versus figures drawn).

Fixed

• An asymmetric before/after chart is refused with the fix in the message. A chart whose label pairs a metric with 修后 but not 修前 (or the reverse) is not a comparison, and the missing half silently drops a number the answer's own table carries. The validator names the group, the missing side, and the grouped spec to send instead. Measured case: a chart shipped 进图轮次 修前/修后, 图节点 修前/修后, 图边 修后 while the same answer's table read "图规模 5 节点/13 边 → 24 节点/41 边" — the 图边 修前 value 13 was lost.
• The closing channel is gated on the session's own shape. It arms only in a session that has drawn at least once; the cheap reminder channel only in a session that has already produced a structured answer. Without that gate a working session (code, commits, verification logs) would be steered into drawing figures it does not need, at one extra model step each time.
• A raster failure no longer produces a broken image. The GUI serves figure files with Content-Security-Policy: sandbox; default-src 'none', and Chromium refuses to rasterize an SVG loaded that way — so falling back to a relative .svg link (option A's documented fallback) rendered nothing at all. Measured live in the DSH Web GUI on 2026-10-05: the .svg link is a broken image while the PNG twin renders. The fallback order is now PNG → inline SVG data URI (32 KB budget) → relative path with a loud warning.
• A failed sharp load is cached per attempt instead of forever, and the diagnostic moved out of the plugin directory: under the shell sandbox an install below ~/.dsh cannot be written, so the old one-time raster-diagnostic.txt write failed silently and the degradation stayed invisible for two days.
• The nudge's per-session counters are keyed by session and dropped on agent/disposed. The session/event observer cannot iterate a WeakMap, and reading the agent map threw usage.values is not a function — caught by scripts/_repro-nudge-channels.mjs.
• Neither channel could read the answer, so both were silently dead. Two separate causes, both found by probing the live event path rather than reading the code: (1) the usage ledger is created lazily at the first agent/pre-step, so a session's FIRST assistant message arrived before the ledger existed and was dropped; (2) the ledger clears its own text copy at every turn boundary, so a tools/post-execute read always saw an empty string. The reply text now lives in a session-keyed latestText map that only agent/disposed clears. A frame that was measured as "12 draws, 0 nudges" is what this looked like from outside.
• The session key accepts session.id, session.header.id, agent.sessionId and agent.id. Reading only session.id works on the real Session (a getter over its header) but degraded every equivalent object to 'unknown', which is what the repro exposed.
• decideNudge reads the reply text from usage, never from opts: text passed in the wrong argument silently degraded every call to the counter fallback. The readable-text path also had no reminder exit at all — only the steer — so a session that stopped drawing went silent instead of reminding.