rosetta-pr-screenshots

name	rosetta-pr-screenshots
description	Capture AX trace UI, Phoenix trace UI, and Wonder Toys app UI screenshots for a framework's PR, then upload as GitHub release assets and embed in the PR body. Called automatically by rosetta-add-framework-docs as part of new-framework PRs; can also be invoked standalone to retrofit existing PRs. Cross-platform — uses Playwright end-to-end.

Rosetta PR Screenshots

Captures the full screenshot set for a framework's PR — AX session/traces, Phoenix session/traces, app UI — and attaches them via GitHub release assets referenced in the PR body.

Everything runs through Playwright. No AppleScript, no screencapture, no Safari focus games — just page.screenshot() against the actual SPA DOM.

Why release assets and not in-repo or drag-drop

GitHub has three ways to attach images to a PR:

1. Drag-drop into PR body (user-attachments) — produces nice github.com/user-attachments/assets/<uuid> URLs but the upload mechanism uses session cookies, not API tokens. Not cleanly automatable from gh CLI.
2. Commit to the PR's branch — works, but adds ~5MB of PNG bloat to the repo permanently after merge.
3. GitHub release assets — this is what this skill uses. gh release create + gh release upload are 100% reliable from CLI. The asset URLs (github.com/<org>/<repo>/releases/download/<tag>/<file>) work in markdown image tags. Repo stays clean.

We use option 3.

Prerequisites

• Playwright project at .claude/skills/rosetta-add-framework-playwright/ installed:

  cd .claude/skills/rosetta-add-framework-playwright
  npm install && npx playwright install chromium
  

• AX auth captured once via the bootstrap below (saved to $HOME/.rosetta-stone/ax-auth.json)
• Local Phoenix running on :6006 (the skill starts it if missing) — no auth needed
• Framework's agent.py tags spans with session.id (e.g. via OpenInference's using_session(user_id) wrap). Without this the session URLs render empty. This is a per-framework responsibility — fix the framework, not this skill.

One-time setup: AX auth bootstrap

The AX UI requires session cookies. Capture them once:

cd .claude/skills/rosetta-add-framework-playwright
node auth-bootstrap.mjs

This opens Chromium, navigates to https://app.arize.com, watches the URL for the post-login pattern, waits for the authenticated sidebar to render (otherwise the SPA hasn't yet written localStorage tokens and the saved state is incomplete), then saves:

• $HOME/.rosetta-stone/ax-auth.json — Playwright storage state (cookies + localStorage)
• $HOME/.rosetta-stone/ax-auth-meta.json — { ax_org_id, ax_space_id } parsed from the post-login URL, so the orchestrator doesn't need to round-trip the Arize SDK

Re-run only when the session expires (observed: ~5–7 days). Override paths via AX_STORAGE_STATE=/some/path node auth-bootstrap.mjs.

Token rotation

AX rotates the refresh token on every use, so the saved state would be stale after the first Playwright run. The AX screenshot spec writes the context's storageState back to disk at the end of every successful run (page.context().storageState({ path: STATE_PATH })). The next run reads the rotated token. As long as you re-run before AX's underlying session expiry (~5–7 days), you don't need to re-bootstrap.

Inputs

• FRAMEWORK — slug (e.g. agno, crewai)
• FRAMEWORK_DIR — directory name (e.g. agno-py, crewai-py)
• PR_NUMBER — existing PR number to update
• AX_PORT (default 8011) — backend port for AX tier
• PHOENIX_PORT (default 8021) — backend port for Phoenix tier
• NEXT_PORT (default 3010) — Next.js port for UI capture (and dev-server target)
• BRANCH — the PR's branch (for the release tag)

Steps

1. Set up output dir + state

WT=$(git rev-parse --show-toplevel)
OUT_DIR="$WT/.screenshots-staging/$FRAMEWORK"
mkdir -p "$OUT_DIR"

# Distinct session IDs for AX and Phoenix so projects don't collide
AX_SESSION="demo-ax-$(date -u +%Y%m%d-%H%M%S)-$(openssl rand -hex 2)"
PHX_SESSION="demo-phx-$(date -u +%Y%m%d-%H%M%S)-$(openssl rand -hex 2)"

# Time window for AX URL — pad ±15 min around now
START_MS=$(($(date +%s) * 1000 - 900000))
END_MS=$(($(date +%s) * 1000 + 60000))

PW_DIR="$WT/.claude/skills/rosetta-add-framework-playwright"

2. Boot AX tier and run the canned 3-turn conversation

cd "$WT/ax/$FRAMEWORK_DIR"
(set -a && source .env.local && set +a && \
  python -m uvicorn backend.main:app --host 127.0.0.1 --port $AX_PORT) > /tmp/ax-cap.log 2>&1 &
until curl -sf http://127.0.0.1:$AX_PORT/products/featured > /dev/null 2>&1; do sleep 1; done

Send the 3-turn conversation (search dragons → buy plushie → ship) with x-user-id: $AX_SESSION. Use the helper pattern from rosetta-demo-capture step 4 (send_turn shell function over SSE). Reuse that verbatim — it's well-tested.

After the conversation:

echo "waiting 45s for AX ingestion…"
sleep 45

3. Resolve AX project + trace IDs and build URLs

Pull ax_org_id from the meta file the bootstrap wrote:

ARIZE_ORG_ID=$(python3 -c "import json; print(json.load(open('$HOME/.rosetta-stone/ax-auth-meta.json'))['ax_org_id'])")

Then:

set -a; source "$WT/ax/$FRAMEWORK_DIR/.env.local"; set +a

ARIZE_PROJECT_ID=$(ax projects list --space "$ARIZE_SPACE_ID" --limit 100 --output csv 2>/dev/null \
  | awk -F',' -v name="$ARIZE_PROJECT_NAME" '$2 == name {print $1; exit}')

# Trace IDs in this session (filtered via ax CLI by time window, then by session.id)
mkdir -p .arize-tmp-traces
ax traces list "$ARIZE_PROJECT_ID" \
  --space "$ARIZE_SPACE_ID" \
  --start-time "$(python3 -c "import datetime; print(datetime.datetime.fromtimestamp($START_MS/1000, datetime.timezone.utc).strftime('%Y-%m-%dT%H:%M:%SZ'))")" \
  --end-time   "$(python3 -c "import datetime; print(datetime.datetime.fromtimestamp($END_MS/1000, datetime.timezone.utc).strftime('%Y-%m-%dT%H:%M:%SZ'))")" \
  --limit 50 --output .arize-tmp-traces/ax-session.json > /dev/null 2>&1

AX_TRACE_IDS=$(python3 -c "
import json
spans = json.load(open('.arize-tmp-traces/ax-session.json'))
spans = spans.get('spans', spans) if isinstance(spans, dict) else spans
seen=[]
for s in spans:
    if s.get('attributes',{}).get('session.id') != '$AX_SESSION': continue
    tid = s.get('context',{}).get('trace_id')
    if tid and tid not in seen: seen.append(tid)
print(' '.join(seen))
")

Build the session URL and trace URLs using the templates from arize-link / rosetta-demo-capture step 7. Collect trace URLs into a JSON array for the spec env:

AX_SESSION_URL="https://app.arize.com/organizations/$ARIZE_ORG_ID/spaces/$ARIZE_SPACE_ID/projects/$ARIZE_PROJECT_ID?selectedSessionId=$AX_SESSION&queryFilterA=&selectedTab=llmTracing&timeZoneA=America%2FLos_Angeles&startA=$START_MS&endA=$END_MS&envA=tracing&modelType=generative_llm"

AX_TRACE_URLS_JSON=$(python3 -c "
import json, os
ids = '''$AX_TRACE_IDS'''.split()
base = f\"https://app.arize.com/organizations/{os.environ['ARIZE_ORG_ID']}/spaces/{os.environ['ARIZE_SPACE_ID']}/projects/{os.environ['ARIZE_PROJECT_ID']}\"
urls = [f\"{base}?selectedTraceId={tid}&queryFilterA=&selectedTab=llmTracing&timeZoneA=America%2FLos_Angeles&startA=$START_MS&endA=$END_MS&envA=tracing&modelType=generative_llm\" for tid in ids]
print(json.dumps(urls))
" ARIZE_ORG_ID="$ARIZE_ORG_ID" ARIZE_SPACE_ID="$ARIZE_SPACE_ID" ARIZE_PROJECT_ID="$ARIZE_PROJECT_ID")

4. Run AX Playwright spec

cd "$PW_DIR"
OUT_DIR="$OUT_DIR" \
  AX_SESSION_URL="$AX_SESSION_URL" \
  AX_TRACE_URLS="$AX_TRACE_URLS_JSON" \
  npx playwright test ax-screenshots.spec.ts --reporter=list

Outputs ax-01-session.png, ax-02-trace-<tid>.png, …

Then stop the AX backend:

lsof -ti :$AX_PORT | xargs -r kill 2>/dev/null

5. Boot Phoenix tier and run conversation

Same pattern as AX. Boot the Phoenix tier on $PHOENIX_PORT, ensure local Phoenix is on :6006, send the 3-turn conversation with x-user-id: $PHX_SESSION, wait for ingestion (Phoenix is local so ~5s suffices, but use 15s to be safe).

# Start local Phoenix if not running
curl -sf http://localhost:6006/healthz > /dev/null 2>&1 \
  || phoenix serve > /tmp/phoenix-shared.log 2>&1 &

cd "$WT/phoenix/$FRAMEWORK_DIR"
(set -a && source .env.local && set +a && \
  python -m uvicorn backend.main:app --host 127.0.0.1 --port $PHOENIX_PORT) > /tmp/phx-cap.log 2>&1 &
until curl -sf http://127.0.0.1:$PHOENIX_PORT/products/featured > /dev/null 2>&1; do sleep 1; done

# Re-use send_turn from step 2 with BACKEND="http://127.0.0.1:$PHOENIX_PORT"
# and x-user-id: $PHX_SESSION

6. Resolve Phoenix project + trace IDs and build URLs

Phoenix exposes a REST API at :6006. Important: the UI's session URL uses Phoenix's internal session ID (a base64 token like UHJvamVjdFNlc3Npb246MzE=), not the user-supplied session.id attribute. Look it up via the sessions endpoint:

PHX_PROJECT_NAME=$(grep '^PHOENIX_PROJECT_NAME=' "$WT/phoenix/$FRAMEWORK_DIR/.env.local" | cut -d= -f2)
PHX_PROJECT_ID=$(curl -s http://localhost:6006/v1/projects \
  | jq -r --arg n "$PHX_PROJECT_NAME" '.data[] | select(.name == $n) | .id')

# Phoenix trace IDs: query the spans API for ones tagged with the session.id.
# Match by prefix — some frameworks (google-adk) append a per-turn suffix
# (`:0`, `:1`, `:2`) to the session.id so each turn lands in its own session.
PHX_TRACE_IDS=$(curl -s "http://localhost:6006/v1/projects/$PHX_PROJECT_ID/spans?limit=200" \
  | jq -r --arg s "$PHX_SESSION" \
      '.data | map(select(.attributes."session.id" | startswith($s))) | map(.context.trace_id) | unique | .[]')

# Resolve the internal session ID for the UI URL (use the ":0" variant — the
# first turn's session — for frameworks that suffix per turn).
PHX_INTERNAL_SESSION_ID=$(curl -s "http://localhost:6006/v1/projects/$PHX_PROJECT_ID/sessions?limit=100" \
  | jq -r --arg s "${PHX_SESSION}:0" \
      '.data[] | select(.session_id == $s) | .id' \
  | head -1)
# Fallback if the framework doesn't suffix: try the bare session.id
if [ -z "$PHX_INTERNAL_SESSION_ID" ]; then
  PHX_INTERNAL_SESSION_ID=$(curl -s "http://localhost:6006/v1/projects/$PHX_PROJECT_ID/sessions?limit=100" \
    | jq -r --arg s "$PHX_SESSION" '.data[] | select(.session_id == $s) | .id' | head -1)
fi

PHX_SESSION_URL="http://localhost:6006/projects/$PHX_PROJECT_ID/sessions/$PHX_INTERNAL_SESSION_ID"
PHX_TRACE_URLS_JSON=$(python3 -c "
import json
ids = '''$PHX_TRACE_IDS'''.split()
print(json.dumps([f'http://localhost:6006/projects/$PHX_PROJECT_ID/traces/{t}' for t in ids]))
")

7. Run Phoenix Playwright spec

cd "$PW_DIR"
OUT_DIR="$OUT_DIR" \
  PHX_SESSION_URL="$PHX_SESSION_URL" \
  PHX_TRACE_URLS="$PHX_TRACE_URLS_JSON" \
  npx playwright test phoenix-screenshots.spec.ts --reporter=list

Outputs phoenix-01-session.png, phoenix-02-trace-<tid>.png, …

8. Capture Wonder Toys app UI

Boot a Next.js dev server. Any tier's frontend works (same code). Set BACKEND_URL so the SSR product fetch hits the running backend, not the default :8001:

cd "$WT/phoenix/$FRAMEWORK_DIR"
BACKEND_URL="http://127.0.0.1:$PHOENIX_PORT" \
  npx next dev -p $NEXT_PORT > /tmp/next-cap.log 2>&1 &
until curl -sf "http://localhost:$NEXT_PORT" > /dev/null 2>&1; do sleep 1; done

cd "$PW_DIR"
OUT_DIR="$OUT_DIR" \
  UI_BASE_URL="http://localhost:$NEXT_PORT" \
  npx playwright test ui-screenshots.spec.ts --reporter=list

Outputs ui-01-landing.png, ui-02-product.png. Without BACKEND_URL, the product detail page hits a 500 (ECONNREFUSED) because Next.js's SSR tries the default backend port.

9. Upload as GitHub release assets

RELEASE_TAG="screenshots-pr-$PR_NUMBER"

gh release create "$RELEASE_TAG" \
  --title "Screenshots for PR #$PR_NUMBER ($FRAMEWORK)" \
  --notes "Auto-generated screenshots from rosetta-pr-screenshots for PR #$PR_NUMBER." \
  --target main \
  || true   # ignore "already exists" — `gh release create` is not idempotent

gh release upload "$RELEASE_TAG" "$OUT_DIR"/*.png --clobber

REPO_FULL=$(gh repo view --json nameWithOwner -q .nameWithOwner)
ASSET_BASE="https://github.com/${REPO_FULL}/releases/download/${RELEASE_TAG}"

10. Update PR body

CURRENT_BODY=$(gh pr view "$PR_NUMBER" --json body -q .body)

SCREENSHOTS_MD=$(cat <<EOF

## Screenshots

### Wonder Toys app UI

![Landing page]($ASSET_BASE/ui-01-landing.png)
![Product detail]($ASSET_BASE/ui-02-product.png)

### Arize AX

Session view:

![AX session]($ASSET_BASE/ax-01-session.png)

Traces:

$(for f in "$OUT_DIR"/ax-*-trace-*.png; do
  name=$(basename "$f")
  echo "![$name]($ASSET_BASE/$name)"
done)

### Phoenix

Session view:

![Phoenix session]($ASSET_BASE/phoenix-01-session.png)

Traces:

$(for f in "$OUT_DIR"/phoenix-*-trace-*.png; do
  name=$(basename "$f")
  echo "![$name]($ASSET_BASE/$name)"
done)
EOF
)

# Strip any prior Screenshots section (idempotent re-runs)
NEW_BODY=$(echo "$CURRENT_BODY" | awk '/^## Screenshots/{exit} {print}')
NEW_BODY="${NEW_BODY}${SCREENSHOTS_MD}"

gh pr edit "$PR_NUMBER" --body "$NEW_BODY"

11. Cleanup

lsof -ti :$AX_PORT :$PHOENIX_PORT :$NEXT_PORT 2>/dev/null | xargs -r kill 2>/dev/null
rm -rf "$OUT_DIR"

Output

Screenshots attached to PR #<N>
  Release: <RELEASE_TAG>
  Files: <count> AX, <count> Phoenix, <count> UI
  URLs: https://github.com/<repo>/releases/tag/<tag>
  PR body updated.

Failure modes

Symptom	Cause	Fix
Playwright errors with "storage state not found"	Auth bootstrap hasn't run	cd .claude/skills/rosetta-add-framework-playwright && node auth-bootstrap.mjs
AX session screenshot shows login page	Saved session expired	Re-run auth-bootstrap.mjs
AX session URL shows "no recent data"	startA/endA window misses the trace	Widen START_MS/END_MS (default ±15 min should cover any sane ingestion lag)
Phoenix session view empty	Framework doesn't tag spans with session.id	Fix the framework — using_session(user_id) wrap (see CrewAI's agent.py). Skill won't paper over this.
Trace tree not expanded in screenshot	DOM selector drifted in AX/Phoenix UI	Update the EXPAND_JS selectors in tests/{ax,phoenix}-screenshots.spec.ts. Probe via npx playwright test --debug
Phoenix session view shows "Something went wrong"	Using user-supplied session.id in URL instead of Phoenix's internal session ID	The orchestrator must resolve the internal id via GET /v1/projects/<id>/sessions and use .data[].id in the URL (see step 6)
UI product detail page 500s with ECONNREFUSED	Next.js SSR is hitting default backend port :8001, not the one this skill booted	Pass BACKEND_URL=http://127.0.0.1:$PHOENIX_PORT when starting next dev (see step 8)
gh release create fails with "tag already exists"	Re-running against the same PR	Safe to ignore — the `

Re-running

Idempotent for a given PR:

• gh release upload --clobber overwrites existing assets
• gh pr edit --body replaces the existing Screenshots section (awk-based prefix strip)

Safe to re-run if any step failed mid-way.

Integration with rosetta-add-framework-docs

When a new framework PR is created via rosetta-add-framework-docs, this skill is automatically invoked as the final step before the PR opens. See rosetta-add-framework-docs/SKILL.md for the integration point.