[TonsOfFun]

What

A complete evaluation run can now leave the dashboard in three formats, and any piece of its report can be copied as Markdown for an issue or a pull request.

JSON export that round-trips through the existing import

• GET <mount>/api/evaluation_reports/:run_id: authenticated with the ingest key (IngestAuthentication) and scoped to the tenant exactly like POST. It returns the version-1 envelope the POST takes. 404 for another tenant's run, 422 for a run that isn't complete or an unknown format.
  • ?format=html returns the self-contained page run_report renders (text/html). It takes no theme, so it follows the viewer's preference, and its fix actions link into the dashboard by absolute path (request.script_name). The CLI draft (Add the activeagent CLI: evals, catalogs, agents and traces, with JSON import/export between installs #620, evals report --format html) calls this.
  • ?format=md returns the Markdown (text/markdown).
  • ?download=1 sends any format as an attachment with the same file names as the dashboard's downloads.
• How :run_id resolves:
  • Imported run: the run_id it was published under.
  • Run the dashboard executed: the stable dashboard-run-<id>-<created_at YYYYMMDDhhmmss UTC>. The timestamp keeps two installs' run Active Agent generate with stream #12 from colliding on a third install.
• ActiveAgent::Evals::Publisher.envelope builds the envelope for both the publisher and the export (ActionAgent::EvaluationReportExport), so the hash layout lives in one place.
• The report is EvaluationRun#to_report(estimate: false) with the recorded run-level judge usage, so the same run always exports the same envelope. The DB run id is left out of the metadata.
• EvaluationReportImport treats a run's own export as an identical retry (200, duplicate: true). It covers both cases:
  • an imported run, where the export's digest differs from the stored digest of the original payload;
  • a dashboard run, found by its exported run_id, so posting it back creates nothing.
    Different content under the run_id is still a 409. The receipt's run_id is the exported one.
• POST /api/evaluation_reports itself is unchanged.

Dashboard run report

• …/runs/:run_id/report.json serves the same envelope to a dashboard session.
• report.md serves the Markdown. ?part=verdict|fix_items[&model=]|scenario&scenario= serves one piece of it.
• ?download=1 on any format sends Content-Disposition: attachment with a file name like <evaluation-slug>-run-<n>.html|json|md.

UI

• The run report view has Download HTML / JSON / Markdown and Copy as Markdown next to Open standalone. The iframe gets allow="clipboard-write".
• The suite's run header gets the same downloads plus a copy button.
• There are Copy as Markdown buttons on:
  • the Models panel (verdict and comparison);
  • What to fix (follows the model filter);
  • each fix card (its Markdown comes with show_run's fix_items[].markdown);
  • each scenario drill-down.
• The bundle is rebuilt.

Markdown (ActiveAgent::Evals::ReportMarkdown)

• Report#to_markdown now has the HTML page's content in GitHub-flavored Markdown:
  • the verdict's rationale;
  • What to fix from fix_items, which replaces the old Recommendations list;
  • each scenario in a <details><summary>… block with a per-model table (status, score, tools, fault, cost) and the answers quoted.
• Table cells escape | and fold line breaks. Text in summaries and answers escapes </>.
• The pieces are public methods: verdict_markdown, fix_items_markdown(model:), fix_item_markdown(item) and scenario_markdown(key). The CLI, the endpoints, the dashboard buttons and the standalone HTML page all use these methods.
• The standalone HTML page keeps its Markdown in hidden textareas. One small inline script copies it, falling back to execCommand("copy") when the Clipboard API is refused. Nothing is loaded from outside the page.

Rake

• action_agent:evaluations:export RUN=<run_id> | EVALUATION=<id> RUN=<id> FORMAT=json|html|md [FILE=path] writes to standard output when no FILE is given.
• action_agent:evaluations:import FILE=path [ACCOUNT_ID=…] goes through EvaluationReportImport and prints Stored: … or Already stored (identical retry): ….

Tests

• New actionagent/test/evaluation_report_export_test.rb:
  • export → re-POST is an identical retry, for imported and dashboard runs;
  • moving a run to a fresh install and exporting it there gives the same envelope;
  • a modified export is a 409;
  • run_id values with dots work;
  • the API-key format=html page is the same page run_report renders, with absolute fix links, text/html and no theme;
  • download=1 on json/html/md, and an unknown format is a 422;
  • HTML and Markdown exports are covered by auth (401) and tenant isolation (404);
  • 404/422 cases;
  • ingest-key auth;
  • tenant isolation, for imported and dashboard runs;
  • .json/.md/download=1 headers and file names, and the Markdown pieces;
  • fix_items[].markdown;
  • the rake export (json/md/html) and import (stored, then identical retry).
• test/evals/report_test.rb:
  • the <details> blocks, pipe/newline escaping and per-model What to fix;
  • each piece matching the whole report;
  • the HTML copy stores matching the Ruby Markdown;
  • one inline script and no external assets.
• test/evals/publisher_test.rb: Publisher.envelope is what #call sends.
• Frontend: reportExport.test.mjs (URLs, ClipboardItem path and fallback) and reportExportViews.test.mjs (the rendered buttons and links).

Results:

• Follow-up commit (format=html, download=1): engine suite 1794 runs, 0 failures, 0 errors; test/evals 245 runs, 0 failures.
• bin/test (gem + engine): 4056 runs, 0 failures, 5 errors. All 5 are RubyLLMIntegrationTest, which calls OpenAI live whenever OPENAI_API_KEY is set, and the local placeholder key gets a 401. The same 5 fail on a clean origin/main checkout with the same .env.test. CI names that key OPEN_AI_API_KEY, so the file skips there.
• bin/lint: no offenses.
• npm test: 698 pass. npm run build was rerun and its output committed.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KhpgqyUrmSDK8cLDSheH3W