benchmark-agents

Benchmark Agents — Advanced AI Systems

Launch real Claude Code sessions with the plugin installed, verify skill injection, monitor PostToolUse validation catches, and produce a coverage report. This skill covers the full eval loop: setup → launch → monitor → verify → fix → release → repeat.

How Evals Work (The Only Correct Method)

Evals are run by you, in this conversation, not by scripts. The process is:

1. You create directories and install the plugin via Bash tool calls
2. You spawn WezTerm panes with wezterm cli spawn — each pane runs an independent Claude Code interactive session
3. You wait, then check debug logs and claim dirs to see what the plugin injected
4. You inspect the generated source code for correctness
5. You read conversation logs to find what the user had to correct
6. You update skills/hooks, run /release, and spawn more evals

Never use claude --print, eval scripts, or Bun.spawn(["claude", ...]). These do not work because:

• Plugin hooks (PreToolUse, PostToolUse, UserPromptSubmit) only fire during interactive tool-calling sessions
• --print mode generates text without executing tools — no files are created, no deps installed, no dev servers started
• No session_id means dedup, profiler, and claim files don't work

The WezTerm interactive approach is the only method that exercises the plugin correctly. Every eval in our history (60+ sessions) used this approach.

DO NOT (Hard Rules)

These are absolute prohibitions. Violating any of them wastes the entire eval run:

• DO NOT use claude --print or -p flag — hooks don't fire, no files created
• DO NOT use --dangerously-skip-permissions — changes agent behavior
• DO NOT create projects in /tmp/ — always use ~/dev/vercel-plugin-testing/
• DO NOT manually create settings.local.json or wire hooks by hand — use npx add-plugin
• DO NOT set CLAUDE_PLUGIN_ROOT manually — the plugin manages this
• DO NOT use bash -c or bash -lc in WezTerm — always use /bin/zsh -ic
• DO NOT use the full path to claude — use the x alias (it's configured in zsh)
• DO NOT create custom debug.log files with stderr redirects — debug logs go to ~/.claude/debug/
• DO NOT write eval runner scripts in TypeScript/JavaScript — do everything as Bash tool calls in the conversation
• DO NOT try to git init or create package.json manually — npx add-plugin + the WezTerm session handle all scaffolding
• DO NOT use uppercase letters in directory names — npm rejects them (e.g. T in timestamps breaks create-next-app)

Copy the exact commands below. Do not improvise.

Setup & Launch (Exact Commands)

Naming convention

Always append a timestamp to directory names so reruns don't overwrite old projects:

<slug>-<yyyymmdd>-<hhmm>

Example: tarot-card-deck-20260309-1227, interior-designer-20260309-1227

Generate the timestamp with: date +%Y%m%d-%H%M

1. Create test directory and install plugin

TS=$(date +%Y%m%d-%H%M)
SLUG="my-app-$TS"
mkdir -p ~/dev/vercel-plugin-testing/$SLUG
cd ~/dev/vercel-plugin-testing/$SLUG
npx add-plugin https://github.com/vercel/vercel-plugin -s project -y

2. Launch session via WezTerm

wezterm cli spawn --cwd /Users/johnlindquist/dev/vercel-plugin-testing/$SLUG -- /bin/zsh -ic \
  "unset CLAUDECODE; VERCEL_PLUGIN_LOG_LEVEL=debug x '<PROMPT>' --settings .claude/settings.json; exec zsh"

Key flags:

• unset CLAUDECODE — prevents nested session detection error
• VERCEL_PLUGIN_LOG_LEVEL=debug — enables hook debug output in ~/.claude/debug/
• x — alias for claude CLI
• --settings .claude/settings.json — loads project-level plugin settings

3. Find the debug log (wait ~25s for SessionStart hooks)

find ~/.claude/debug -name "*.txt" -mmin -2 -exec grep -l "$SLUG" {} +

4. Launch multiple sessions in parallel

Create dirs and install plugin in a loop, then spawn each WezTerm pane:

TS=$(date +%Y%m%d-%H%M)
cd ~/dev/vercel-plugin-testing
for name in tarot-deck interior-designer superhero-origin; do
  d="${name}-${TS}"
  mkdir -p "$d" && (cd "$d" && npx add-plugin https://github.com/vercel/vercel-plugin -s project -y)
done

# Then spawn each (these run in separate terminal panes)
wezterm cli spawn --cwd .../tarot-deck-$TS -- /bin/zsh -ic "unset CLAUDECODE; VERCEL_PLUGIN_LOG_LEVEL=debug x '...' --settings .claude/settings.json; exec zsh"
wezterm cli spawn --cwd .../interior-designer-$TS -- /bin/zsh -ic "unset CLAUDECODE; VERCEL_PLUGIN_LOG_LEVEL=debug x '...' --settings .claude/settings.json; exec zsh"
wezterm cli spawn --cwd .../superhero-origin-$TS -- /bin/zsh -ic "unset CLAUDECODE; VERCEL_PLUGIN_LOG_LEVEL=debug x '...' --settings .claude/settings.json; exec zsh"

Monitoring

Skill injection claims (the key metric)

TMPDIR=$(node -e "import {tmpdir} from 'os'; console.log(tmpdir())" --input-type=module)
CLAIMDIR="$TMPDIR/vercel-plugin-<session-id>-seen-skills.d"

# List all injected skills
ls "$CLAIMDIR"

# Count
ls "$CLAIMDIR" | wc -l

# Check specific skill
ls "$CLAIMDIR/workflow" && echo "YES" || echo "NO"

Hook firing

LOG=~/.claude/debug/<session-id>.txt

# SessionStart hooks
grep -c 'SessionStart.*success' "$LOG"

# PreToolUse calls and injections
grep -c 'executePreToolHooks' "$LOG"        # total calls
grep -c 'provided additionalContext' "$LOG"  # actual injections

# PostToolUse validation catches
grep 'VALIDATION' "$LOG" | head -10

# UserPromptSubmit
grep -c 'UserPromptSubmit.*success' "$LOG"

Quick status check for multiple sessions

TMPDIR=$(node -e "import {tmpdir} from 'os'; console.log(tmpdir())" --input-type=module 2>/dev/null)

for label_id in "slug1:SESSION_ID_1" "slug2:SESSION_ID_2" "slug3:SESSION_ID_3"; do
  label="${label_id%%:*}"
  id="${label_id##*:}"
  claimdir="$TMPDIR/vercel-plugin-$id-seen-skills.d"
  echo "=== $label ==="
  count=$(ls "$claimdir" 2>/dev/null | wc -l | tr -d ' ')
  claims=$(ls "$claimdir" 2>/dev/null | sort | tr '\n' ', ')
  echo "Skills ($count): $claims"
done

Verification — What to Check in Generated Code

After sessions build, verify these patterns in the generated projects:

Project structure

echo -n "src/: "; test -d "$base/src" && echo YES || echo NO          # Should be NO for Workflow SDK projects
echo -n "workflows/: "; test -d "$base/workflows" && echo YES || echo NO
echo -n "withWorkflow: "; grep -q "withWorkflow" "$base"/next.config.* && echo YES || echo NO
echo -n "components.json: "; test -f "$base/components.json" && echo YES || echo NO

Image generation model

# Should use gemini-3.1-flash-image-preview, NOT dall-e-3 or older gemini models
grep -rn "gemini.*image\|dall-e\|experimental_generateImage\|result\.files" "$base/workflows/" "$base/app/" 2>/dev/null | grep "\.ts"

Gateway vs direct provider

# Should use gateway() or plain "provider/model" strings, NOT openai("gpt-4o") directly
grep -rn "from.*@ai-sdk/openai\|openai(" "$base" 2>/dev/null | grep "\.ts" | grep -v node_modules
grep -rn "gateway(\|model:.*\"openai/" "$base" 2>/dev/null | grep "\.ts" | grep -v node_modules

AI Elements installed

find "$base" -path "*/ai-elements/*.tsx" 2>/dev/null | grep -v node_modules | wc -l

Workflow API usage

wf=$(find "$base" -name "*.ts" -path "*/workflow*" 2>/dev/null | grep -v node_modules | head -1)
head -5 "$wf"   # Should show: import { getWritable } from "workflow"

Prompt Design Rules

Describe products, not technologies. Let the plugin infer which skills to inject. This tests whether the plugin's pattern matching and prompt signals work from natural language.

DO:

• "runs a multi-step creation pipeline that streams each phase"
• "generates a portrait image"
• "users can chat with an AI advisor"
• "store all designs in a gallery"

DON'T:

• "use Vercel Workflow SDK with getWritable"
• "use gateway('google/gemini-3.1-flash-image-preview')"
• "install npx ai-elements"
• "add withWorkflow to next.config.ts"

Always end prompts with:

"Link the project to my vercel-labs team so we can deploy it later. Skip any planning and just build it. Get the dev server running."

Phrases that trigger key skills (via promptSignals):

• workflow: "multi-step pipeline", "streams progress", "streams each phase", "durable pipeline", "creation pipeline"
• ai-sdk: Triggered by imports/install patterns (very broad)
• shadcn: Triggered by create-next-app bash pattern
• ai-elements: Triggered when ai-sdk is active + chat UI patterns

Common Issues Found in Evals (and Fixes Applied)

Agent-Browser Verification

After dev server starts, verify with agent-browser. Note: agents currently DO NOT self-verify despite the skill being injected. You must launch verification manually:

agent-browser open http://localhost:<port>
agent-browser wait --load networkidle
agent-browser screenshot
agent-browser snapshot -i

Coverage Report

Write results to .notes/COVERAGE.md with:

1. Session index — slug, session ID, unique skills, dedup status
2. Hook coverage matrix — which hooks fired in which sessions
3. Skill injection table — which of the 43 skills triggered
4. Code quality checks — gateway vs direct, image model, withWorkflow, AI Elements
5. PostToolUse validation catches — outdated models, deprecated APIs
6. Issues found — bugs, pattern gaps, new findings to feed back into skills

Release → Eval Loop

The standard improvement cycle:

1. Run evals — launch 3 sessions with natural language prompts
2. Check results — skill claims, project structure, code quality
3. Identify gaps — what skills didn't trigger, what patterns are wrong
4. Read conversation logs — find user follow-up corrections
5. Fix skills — update SKILL.md content, patterns, validate rules
6. Run gates — bun run typecheck && bun test && bun run validate
7. Release — bump version, bun run build, commit, push
8. Repeat — launch 3 more evals to verify fixes

Scenario Table

Complexity Tiers

Tier 1 — Core AI (30-45 min, --quick)

Scenarios 01, 04, 09 — AI SDK, Gateway, Sandbox, AI Elements without durable workflows.

Tier 2 — Durable Agents (45-60 min)

Scenarios 02, 03, 06, 10 — Workflow SDK, multi-step durability, agent orchestration.

Tier 3 — Platform Integration (45-60 min)

Scenarios 05, 07, 08, 11, 12 — Chat SDK, Queues, Flags, Firewall, cross-platform messaging.

Full Suite

All 12 scenarios, ~3-4 hours.

Cleanup

rm -rf ~/dev/vercel-plugin-testing