Skip to content

program.md

Latest commit

 

History

History
144 lines (107 loc) · 10.4 KB

program.md

File metadata and controls

144 lines (107 loc) · 10.4 KB
Raw

autoweb — self-improving NBA Dashboard

This is a live demo of autoweb — an autonomous UI improvement agent. Every change in this repo was made by an AI agent running in a loop. No human edited the dashboard after the initial scaffold. The agent screenshots the page, picks one improvement, verifies it, and either keeps or reverts — forever.

The target

  • File to edit: /home/user/public/nba/index.html
  • URL to screenshot: https://allan.feather-cloud.dev/public/nba/
  • Repo: https://github.com/inceptel/nba
  • Data source: ESPN unofficial API (no auth needed)
  • Autoweb dashboard: https://allan.feather-cloud.dev/public/nba/autoweb-dashboard.html

What you CAN do

  • Edit /home/user/public/nba/index.html (single file, vanilla JS + CSS)
  • Add CDN libraries via script/link tags
  • Improve data display, layout, UX, visual polish
  • Add new features: standings, player stats, game details, injury reports
  • Add new ESPN API endpoints — see https://site.api.espn.com/apis/site/v2/sports/basketball/nba/

What you CANNOT do

  • Break the ESPN API fetch or auto-refresh loop
  • Make the page fail to load or show a blank screen
  • Remove score display — that is the core feature
  • Require a backend or server-side code (static HTML only)

How to verify

agent-browser --url "https://allan.feather-cloud.dev/public/nba/" --task "Screenshot this page. Describe what you see. Does the NBA dashboard show games with scores? Any visual bugs or layout issues?"

How to log results

# Keep:
printf "%s\tkeep\tDESCRIPTION\n" "$(date -u +%Y-%m-%dT%H:%M:%SZ)" >> /home/user/autoweb-nba/results.tsv

# Revert:
cp /home/user/public/nba/index.html.bak /home/user/public/nba/index.html
printf "%s\trevert\tDESCRIPTION\n" "$(date -u +%Y-%m-%dT%H:%M:%SZ)" >> /home/user/autoweb-nba/results.tsv

Run tests before keeping

cd /home/user/public/nba && bash tests/run-tests.sh
# All tests must pass before logging a keep

The experiment loop

LOOP FOREVER:

  1. Screenshot the live URL with agent-browser. Read index.html. Check results.tsv for recent attempts.

    CURRENT FOCUS: box scores — 8 REVIEWS DEFERRED, IMPLEMENT NOW OR REVERT AND TRY AGAIN — All known bugs are fixed. Leaders and Standings tabs are fully complete with no remaining polish. ANY change other than box score MUST be reverted. Do not add colors, badges, columns, or any other polish — there is nothing left to polish. The ONLY acceptable next change is: add a box score / team stats section to final and live game cards. Minimal scope — just 4 stats, 2 rows: FG%, REB, TOV, AST for each team (8 numbers + a header = done). Use ESPN gamecast API (summary?event={gameId}). Simplest UI: use a native <details><summary>▼ Box Score</summary>…</details> element — no JS toggle needed, browser handles expand/collapse natively. Lazy-fetch on first expand by listening for the toggle event on the <details> element. If you are not implementing box score, you are doing the wrong task — revert and start over with box score. After box score: injury reports, head-to-head records.

  2. Check deadline: echo $(($(cat /home/user/autoweb-nba/deadline 2>/dev/null || echo 9999999999) - $(date +%s)))s left

    • If the deadline file is missing, continue normally (no hard stop).

  3. Pick ONE improvement. Priority order:

    • Bugs / broken things (CRITICAL/HIGH below)
    • UX problems (confusing, missing info)
    • Visual polish (spacing, colors, typography)
    • New features (standings tab, player search, etc.)

  4. Backup: cp /home/user/public/nba/index.html /home/user/public/nba/index.html.bak

  5. Make the change. Keep it focused — one thing only.

  6. Run tests: cd /home/user/public/nba && bash tests/run-tests.sh If tests fail → revert immediately. If tests/run-tests.sh does not exist, skip this step and note it in your log.

  7. Verify visually with agent-browser screenshot.

  8. Log and exit. The harness handles git commit + push automatically.

Known issues (priority order)

Mark fixed issues with strikethrough rather than deleting them.

  • HIGH: Scheduled games show tip-off time in away score column — layout is wrong (check homeAway field for scheduled game rendering, not just live/final) (fixed iter 1)
  • HIGH: No loading state shown while API fetches — user sees blank content briefly (initial load had spinner; date nav now also shows spinner)
  • LOW: usingYesterday variable undefined in empty-state message (shows "today" instead of "recently" when auto-fallback to yesterday occurred) (fixed iter 5)
  • MEDIUM: No standings tab — would be very useful alongside scores (fixed iter 7 — full standings with W/L/PCT/GB/L10/HOME/ROAD/CONF/DIFF/STRK)
  • MEDIUM: Team logos are text initials — try fetching ESPN team logo URLs
  • MEDIUM: Player leader stats missing from many games — check API response structure (fixed iter 23 — shows 2 leaders per team)
  • LOW: No dark/light mode toggle
  • LOW: Mobile: game cards could be more compact (fixed iter 30)
  • LOW: Add yesterday/tomorrow navigation arrows (fixed iter 3)

Potential next features (no known bugs remain)

  • Leaders tab polish: headshots, medal ranks, team logos, relative stat bars (fully done)
  • Standings table columns: W/L/PCT/GB/HOME/ROAD/CONF/DIFF/L10/STRK, sortable, sticky team column (fully done — no new columns needed)
  • MEDIUM (TOP PRIORITY — 8+ REVIEWS DEFERRED, DO THIS NOW — NOT NEXT, NOW): Box score / team stats for final/live games — Minimal: FG%, REB, TOV, AST per team (2 rows × 4 stats). Use ESPN summary API: https://site.api.espn.com/apis/site/v2/sports/basketball/nba/summary?event={gameId} (gameId = event.id from scoreboard). Key fields: boxscore.teams[].statistics[] for team totals. IMPLEMENTATION: use <details><summary>▼ Box Score</summary></details> native HTML — listen for the toggle event to lazy-fetch on first open; cache result on details.dataset.fetched = '1' so re-opens don't re-fetch. Do NOT fetch at page load (10+ parallel requests will be slow). This is the only task. Do not pick anything else.
  • MEDIUM: Injury report — show key player injuries on game cards (ESPN injury API exists)
  • LOW: Head-to-head season record shown on game cards
  • LOW: Team schedule popup — click team logo/name to see upcoming games
  • LOW: Playoff bracket tab — visual bracket for postseason (when applicable)
  • LOW: Player search — search any player for season stats

Design principles

  • Data-dense: show as much useful info as possible without clutter
  • Dark first: dark background (#0a0a0f) is the baseline — never break it
  • Monospace: JetBrains Mono / Fira Code for that Bloomberg terminal feel
  • Team colors: use team primary colors for accents wherever possible
  • Live games first: always sorted live → final → scheduled
  • Auto-refresh: 20s when live games, 60s otherwise — never break this

autoweb features showcased here

This repo demonstrates the full autoweb feature set:

  • Screenshot verification — every change verified with a real browser screenshot
  • Test suite — bash tests run before every keep (red/green)
  • Auto-push — every kept change commits and pushes to this GitHub repo
  • Review loop — runs every 4h, updates this program.md with learnings
  • results.tsv — full audit log of every iteration
  • Breadcrumbs — agent leaves notes between iterations

Rules learned from experience

  • Always verify with agent-browser — do not trust the code alone
  • The footer autoweb dashboard link must point to https://allan.feather-cloud.dev/public/nba/autoweb-dashboard.html — verify this is correct on every iteration and fix it if it's wrong
  • One change per iteration — resist the urge to fix multiple things
  • If the page goes blank, revert immediately before logging
  • ESPN API returns competitors[0] as either home or away — always check homeAway field
  • The homeAway field bug affects scheduled games differently than live/final — test all three states
  • Mark fixed issues with strikethrough in Known Issues — do not delete them
  • ESPN API reliability: site.api.espn.com endpoints can go 404 without warning — always verify a new API endpoint returns 200 with valid data before building on it
  • Leaders API: site.api.espn.com/apis/site/v2/sports/basketball/nba/leaders is 404 (broken). Use sports.core.api.espn.com/v2/sports/basketball/leagues/nba/seasons/{year}/types/2/leaders instead. Note: this API uses $ref links for athletes; batch-fetch athlete names with Promise.all.
  • Regression risk: the dashboard now has 15+ features — when verifying a new iteration, check that the Standings tab, Leaders tab, and date navigation still work, not just the feature you changed
  • Leaders tab complexity: Leaders tab uses a hardcoded ESPN team ID → abbreviation map and batch-fetches $ref athlete links. If adding features to Leaders tab, validate that all 8 stat categories still render correctly after your change. Do not touch Leaders tab — it is fully complete and the batch-fetch logic is fragile.
  • Box score API: https://site.api.espn.com/apis/site/v2/sports/basketball/nba/summary?event={gameId} — returns full game summary. boxscore.teams[].statistics[] has team totals (FG%, REB, TOV, AST, etc.). boxscore.players[].statistics[] has individual player lines. Fetch on-demand (not at page load) to avoid 10+ simultaneous requests.
  • Box score UI pattern: Use native <details><summary>▼ Box Score</summary></details> HTML — listen for toggle event to lazy-fetch on first expand. Store fetched data in details.dataset.fetched = '1' to prevent re-fetching. This is simpler than a JS toggle and requires no extra state management.
  • Polish moratorium: Standings and Leaders tabs are fully complete. Do not add more polish features until box score is implemented. This has been the rule for 5+ reviews — enforce it.
  • Box score is critically overdue: It has been deferred for 8 consecutive review cycles — more than any other feature in this project's history. The API is known, the UI pattern is known, the lazy-fetch approach is known, and the scope is tiny (4 stats × 2 teams). There are zero blockers. If you pick any other task, revert it and implement box score instead. There is no acceptable alternative.