Skip to content

reach view

Read back an evaluation artifact (default: .reach/eval.json) and render human-readable scorecards or interactive standalone HTML reports.

Synopsis

reach view [ARTIFACT] [OPTIONS]

Key Scenarios

Print summary scores, precision, recall, and top confusion pairs for the latest evaluation:

reach view

Output:

all: 58 skills, 35 probes

 skill                             reached   recall    95% CI   absorbed   precision
 ───────────────────────────────────────────────────────────────────────────────────
 google-agents-cli-adk-code            5/5     100%   57-100%                   100%
 google-agents-cli-deploy              5/5     100%   57-100%                   100%
 google-agents-cli-eval                5/5     100%   57-100%                   100%
 google-agents-cli-observability       5/5     100%   57-100%                   100%
 google-agents-cli-publish             5/5     100%   57-100%                   100%
 google-agents-cli-scaffold            5/5     100%   57-100%                   100%
 google-agents-cli-workflow            5/5     100%   57-100%                   100%

              51 resident skills had no query and took no traffic
consistency 100% [65-100%]   top-1 100.0% ± 4.7pp   abstention 0% [0-10%] (0% false)
macro-F1 100.0%  over 7 observed labels  [5 replicates, 7 repeated queries]

Generate a self-contained, standalone HTML diagnostic workbench report with an interactive confusion matrix, skill-level scores, search filtering, and expandable queries:

reach view --format html > report.html
open report.html

The interactive workbench includes:

  • Confusion matrix cross-filtering: Click any hit or miss cell in the confusion matrix to instantly filter queries down to the matching expected/invoked skill pair, with an active filter banner and one-click reset.
  • Dynamic table and query search: Filter skills and query cards in real time by ID, skill name, or query prompt text.
  • Expand/collapse controls: Expand all query cards at once to inspect full prompt text and difficulty ranks, or collapse them for a high-level overview.
  • Air-gapped and self-contained: Zero external script or CDN dependencies; safe to open offline or in air-gapped environments.

List individual queries, difficulty ranks, and probe outcomes in the terminal:

reach view --show-queries

Output:

● google-agents-cli-adk-code  5/5 reached (100%)

 query              reached   rank   leak           selected   text
 ───────────────────────────────────────────────────────────────────────────────────────────────
 q-adk-code   5/5 [57-100%]   1/58   names target   ✓ match    Write custom agent code with ADK

● google-agents-cli-deploy  5/5 reached (100%)

 query              reached   rank   leak           selected   text
 ───────────────────────────────────────────────────────────────────────────────────────────────
 q-deploy-1   5/5 [57-100%]   1/58   names target   ✓ match    Deploy agent service to Cloud Run

● google-agents-cli-eval  5/5 reached (100%)

 query              reached   rank   leak           selected   text
 ───────────────────────────────────────────────────────────────────────────────────────────────
 q-eval-1     5/5 [57-100%]   1/58   names target   ✓ match    Run an evaluation on skill suite

Options

Option Type Default Description
ARTIFACT Path .reach/eval.json Path to the evaluation artifact JSON file. If omitted, defaults to .reach/eval.json.
--out, -o Path - Destination path to write the rendered report (defaults to stdout).
--open, -O Flag false Open the rendered HTML report directly in the default web browser.
--show-queries Flag false Display individual scored query records below the summary scorecard.
--queries, -q Path - Subset query set file used to slice the recorded evaluation run.
--filter-skill String () Glob pattern(s) matching target skill names to slice the recorded run.
--filter-id String () Glob pattern(s) matching query IDs to slice the recorded run.
--format Choice text Output format: text (terminal table), html (standalone interactive HTML), json (raw artifact JSON), jsonl (scored query records as JSON lines).
--verbose Flag false Display full hexadecimal hash digests alongside badges.