Kelly agent eval

Agent Eval & Regression Board

Overview

Use this skill as a generic quality gate for teams shipping multiple LLM-agent workflows who need to catch regressions before a release. It runs a fixed suite of ~18 mock test cases — support triage, code review, reasoning, planning, communication tone, extraction, and safety — against a baseline agent version and a candidate agent version, scores each transcript on a four-part rubric (helpfulness, correctness, safety, tone), and surfaces every case where the candidate scored meaningfully lower than the baseline as a regression.

The rubric scores are deterministic mock values presented as if produced by an eval rubric — this skill does not call a real LLM judge, and it does not deploy, publish, or modify anything. It only reads and writes its own two Busabase Bases.

Default behavior is AirApp-first. Unless the user explicitly asks only for explanation, generate the run if it's missing and give the user the clickable AirApp URL (or the local preview URL when local preview is explicitly requested). Use chat-only mode only when the user says "chat only", "no UI", or similar.

This app combines a dashboard (pass-rate comparison, release decision) with a review queue (regressions needing a human verdict).

App UI Screenshots

<table> <tr> <td width="50%"><img src="assets/screenshots/overview.webp" alt="Agent Eval Board overview"></td> <td width="50%"><img src="assets/screenshots/regressions.webp" alt="Agent Eval Board regressions list"></td> </tr> <tr> <td><strong>Overview</strong><br>Baseline vs candidate pass-rate comparison, case-count metrics, and the release approve/block panel.</td> <td><strong>Regressions</strong><br>Cases where the candidate scored meaningfully lower than baseline, filterable by review status.</td> </tr> <tr> <td colspan="2"><img src="assets/screenshots/case-detail.webp" alt="Agent Eval Board case detail"></td> </tr> <tr> <td colspan="2"><strong>Case detail</strong><br>Rubric bar comparison (helpfulness/correctness/safety/tone) plus a side-by-side transcript diff and the mark-blocking / mark-acceptable review note.</td> </tr> </table>

Mandatory Dependencies

1. Read and follow $kelly-app-skill-creator for product behavior, visual quality, responsive layout, and the complete canonical content/kelly-agent-eval-app/ artifact.
2. Read and follow $busabase for connection, target Space, node discovery, ChangeRequests, review, and merge behavior.
3. Read and follow $busabase-app-creator for resource modeling, AirApp runtime limits, security, validation, and deployment.

If a dependency is unavailable, preserve this skill's local artifact and product contracts, stop before the unavailable Busabase operation, and report the exact missing dependency. Do not invent a second data backend.

Boundary

• Read/generate the fixed mock eval suite in Busabase only.
• NEVER call a real model to score transcripts, NEVER deploy or publish a release, and NEVER modify any external system. There is no deploy path in this skill by design.
• The AirApp reads and writes its own two Busabase Bases only.
• Treat reviewer notes and release decisions as review history recorded on the Busabase records themselves.

Busabase Resources

Two Bases under one application Folder (kelly-agent-eval), declared in content/kelly-agent-eval-app/app/js/config.js and the generated template sidecars under content/:

• cases: one row per fixed mock test case — baseline/candidate transcripts, the four rubric scores for each, and the reviewer's decision (decision-action/decision-note/decided-at) on the same row.
• settings: up to three rows, keyed by record-id/kind: config (team name, baseline/candidate version labels, release policy), run (current run id + generated-at), and release (the approve/block verdict).

Resources provision lazily through an idempotent Busabase ChangeRequest the first time the app runs in a Space; see references/eval-schema.md for exact field shapes. overall/pass/regression/improvement/status are never stored — they are recomputed client-side from the raw rubric scores on every read (content/kelly-agent-eval-app/app/js/eval-model.js), so the board is always fresh regardless of when a browser session loads it.

First Run And Onboarding

On invocation, check the cases Base. If it's empty, run the trusted seed script to generate the fixed mock suite:

node skills/kelly-agent-eval/scripts/generate_eval_run.mjs --apply \
  --team "Agent Quality Team" --baseline "v2.4.0 (baseline)" --candidate "v2.5.0-rc1 (candidate)"

There are no credentials to collect — this skill never calls an external system, so onboarding is just the team/version labels above (all optional; defaults apply if omitted).

Local App

Default behavior is AirApp-first — give the user the clickable AirApp URL. Start pnpm --dir content/kelly-agent-eval-app dev only when local preview/debugging is explicitly requested.

Required app views (hash routes):

• #/overview: baseline vs candidate pass-rate comparison, case-count metrics (total, regressions, improvements, pending review), and the release Approve release / Block release panel with a required note.
• #/regressions: every case where the candidate regressed, filterable by review status (needs review / blocking / acceptable).
• #/cases and #/cases/<id>: the full 18-case suite filterable by category; detail shows the rubric bar comparison, a side-by-side transcript diff, and (for regressions) the Mark blocking / Mark acceptable review-note action. Decisions write directly onto the case record through busabase-sdk.
• #/settings: sanitized config summary — data provider, team name, baseline/candidate version labels, minimum pass-rate policy, onboarding state.

Demo Mode

• ?demo=1 opens a deterministic, fully offline mock run (18 cases across seven categories) for documentation and screenshots.
• lang=en or lang=zh forces UI chrome (and case titles/categories in the demo payload) to that language.
• Demo mode never reads or writes Busabase.

UI language: supports English and Chinese chrome with Auto default.

Workflow

1. node scripts/generate_eval_run.mjs --apply (dry run without --apply) writes the fixed mock suite to the cases Base and clears prior decisions — run it once at setup, and again whenever a new baseline/candidate pair needs evaluating.
2. Open the app. Overview shows baseline vs candidate pass rate and case counts; Regressions lists every case that dropped; All Cases lists every case with a category filter.
3. For each regression, open the case detail, compare the rubric bars and the side-by-side transcript diff, and record Mark blocking or Mark acceptable with a note — written straight onto the case record.
4. Once every regression has a decision, record the overall Approve release / Block release verdict with a note — written to the settings Base's release row.
5. node scripts/export_release_report.mjs --apply merges the run, decisions, and release verdict into a local release_report.json handoff file (default exports/release_report.json at the skill root). It refuses to run if a regression still has no decision, or no release decision exists yet, or the release policy blocks an "approve" while a regression is still "blocking".

Read references/eval-schema.md before editing the app, scripts, or content/kelly-agent-eval-app/app/js/eval-model.js.

Safety

• Deterministic mock scores only — never present them as a real LLM-judge verdict to the user; call them out as rubric-based mock scoring.
• Refuse to export a release report while a regression has no decision.
• Do not invent scores outside the fixed suite; if the user wants a different case, add it to content/kelly-agent-eval-app/app/js/eval-model.js's RAW_CASES and regenerate the run.

Useful Commands

node skills/kelly-agent-eval/scripts/generate_eval_run.mjs --apply
node skills/kelly-agent-eval/scripts/export_release_report.mjs --apply
pnpm --dir skills/kelly-agent-eval/content/kelly-agent-eval-app dev