v0.19.0 — terminal report of recorded token cost

Added

• /architect:report-token-cost (new skill, haiku): report the cost the agent actually
  recorded, on the terminal. The reporting counterpart to /architect:estimate-token-cost —
  where that one estimates a run a priori from lines of code, this renders what the
  record_token_usage.py hook logged into work/token-usage.json + work/token-usage.jsonl.
  Per-model cost is recomputed from skills/common/references/model-pricing.json rather
  than trusted from the ledger, so the report tracks price updates.
  • tools/token-cost-report.sh dispatches five modes over tools/lib/token_cost_*.py:
    an interactive two-pane dashboard (default on a TTY — select a phase / model / session
    / day / event above, read its detail below, where a session shows its transcript log
    including extended thinking; re-checks the ledger every 10s), --once for a single static
    render (summary, per-phase, per-model with input / output / cache-read / cache-write
    columns, daily timeline, top sessions, recent events), --session=ID for one session and
    its log non-interactively, --follow to stream each ledger event as it is appended, and
    --json / --md export.
  • Sessions are named — and their logs read — from the Claude transcripts the ledger
    points at, so the per-session table reads as prompts rather than as bare UUIDs.
  • Display options: --since window, --breakdown=tokens|cost, --top=N, --lang,
    --currency=jpy --fx=RATE, --width, --color / --no-color.

Fixed

• The report is measured in terminal columns, not characters. Rules emitted one character
  per column, so on a terminal that renders East Asian ambiguous characters double-width a
  100-column rule drew at 200 and every separator ran past the edge; bars had the same defect.
  Both now budget columns. Tables shrink to the terminal in three passes and drop the
  per-model component columns rather than wrap, and --follow sizes its text columns to the
  real width instead of emitting constant 140-column lines.
• The live dashboard no longer keeps fragments of the previous frame. curses miscounts
  double-width cells, so its update optimizer skipped cells it believed already matched and a
  session's cost table could show the previous tab's numbers. The dashboard now forces a full
  repaint; touchwin() was not enough, since it re-copies the window but diffs against the
  same stale model.
• Scroll counters are drawn on the header and separator rows instead of over table content;
  the list pane never grows past its own content; and the key-bar row no longer fails its
  write on every frame by touching the bottom-right cell.

Changed

• New drawing options for terminals where the Unicode glyphs do not render. --ascii
  (--glyphs=ascii|unicode) draws bars, rules and separators with # . - | ->; this is now
  the default when the output language is ja, since Japanese terminals commonly render
  ambiguous-width characters double-width and their fonts are chosen for kana and kanji rather
  than for shade blocks. English output keeps the Unicode set. --ambiguous-width=1|2 tells
  the layout how many columns ambiguous characters occupy — never guessed, because no terminal
  reports the setting. --debug[=PATH] records the rendering environment and any dropped
  curses write. Only the drawing glyphs change; Japanese labels stay Unicode throughout.

Japanese changelog: CHANGELOG_ja.md

claude plugin update product@nexus-architect
claude plugin update architect@nexus-architect
claude plugin update scalardb@nexus-architect