Dashboard and report

July 29, 2026 · View on GitHub

/tt watch demo opens the synthetic dashboard immediately. /tt watch follows real local history: reads pulse, parent folders ripple, and rows fade to their heat color. The visual always pairs color with five-cell bars, h values, and counts.

Controls: f recent focus, h hottest, c coldest, n name order, / prompt history, a live overview, s prompt privacy, r refresh, and q quit. The live view limits proven activity to ten folders and collapses quiet paths; /tt insights retains the complete inventory.

The report keeps heat, lifetime reads, search evidence, routing coverage, trend, task clusters, protected context, retired paths, and review candidates separate. When a current directive manifest exists, it adds per-directive opportunities, followed counts, rate/confidence, unobservable ratio, capture-disabled and awaiting-capture probes, and always-loaded cost. unobserved is always labeled as missing evidence rather than a violation. A stale manifest replaces metrics with a refresh instruction instead of silently evaluating old probes. Its grade is a summary, not a verdict.

Both surfaces also report agent personas when agent capture is on: the report lists invocations, sessions, last use, and never-invoked definitions; the dashboard shows one line with the most used persona and the never-invoked count. Neither says anything when capture is off, because silence would then measure nothing rather than disuse.

The live dashboard includes a compact instruction-adherence panel in demo and measured modes. The complete evidence (up to five recent session IDs per directive), trend, cost method, and never-triggered review prompts remain in /tt insights and /tt instructions.

Its optional SVG visuals preserve the tables beneath them: single-series KPI sparklines, separate single-axis count and search-ratio trends, and a TUI-shaped indented heat tree. Dashed trend segments mark small samples; neutral note ticks carry tooltip text without implying causation.

Locally rendered trigger-tree HTML report with a flat grade header, KPI sparklines, and evidence table

This capture comes from the generated, self-contained HTML report. It is separate from the intentionally synthetic dashboard demo on the project homepage.

First run with no evidence explains the loop: work normally and reads light up, or run /tt watch demo now.