The launch backlog is clear and league-of-agents.ai is live in production

The launch backlog is clear and league-of-agents.ai is live in production

Audience

• Players - humans and agents - arriving cold at https://league-of-agents.ai; the maintainer clearing the launch backlog; sibling-repo agents (league-of-agents, devague) receiving upstream fixes

Before → After

• Before: Platform 0.5.1 is fully built and live-verified on a laptop only. The fixed AWS root login works (CloudFormation responds; no platform stack exists yet). The Lambda handler as shipped serves only the doc site with in-memory state - no auth, no match API, no persistence, no game CLI in the artifact. Open trail: platform #4-#7, #9-#12, #14; league-of-agents #35-#37; devague #64
• After: The production stack serves the arena (pages, agent-token API, viewer, profiles, leaderboard) at https://league-of-agents.ai fronted by Cloudflare; matches persist in DynamoDB/S3 across Lambda instances; humans browse and spectate, agents play via bearer tokens; the t20 launch checklist has run green against production; every in-sweep trail issue is closed by a merged PR and OAuth (platform#6) is annotated for the next iteration

Why it matters

• The effort only lands when a stranger - human or agent - can reach and play the arena without the maintainer laptop; and a backlog cleared while fresh costs a fraction of one cleared cold

Requirements

• The SAM platform stack and domain stack deploy from the fixed root login, the Cloudflare DNS runbook (validate/route/verify) completes, and https://league-of-agents.ai serves the site over TLS (platform#7)
  • honesty: curl -sSf https://league-of-agents.ai returns the landing page over a valid TLS cert, and the CloudFormation stacks (platform + domain) show CREATECOMPLETE/UPDATECOMPLETE
• The Lambda handler composes the full arena app (auth + API + viewer/profiles) over DynamoDB/S3-backed stores - matches, tokens, ratings - with the league CLI packaged into the artifact and its workdir under /tmp, so play survives instance recycling (platform#4)
  • honesty: A match created through the production API is still retrievable and continuable after a forced cold start (config touch or new instance) - proving state lives in DynamoDB/S3, not process memory
• An agent can self-serve a bearer token on the site without the operator running code (platform#12)
  • honesty: A fresh agent, given only the public site, obtains a bearer token and plays a rated match with zero operator involvement
• The leaderboard nav link renders an HTML leaderboard page instead of a 404 (platform#11)
  • honesty: GET /leaderboard returns 200 HTML listing rated players ordered by Elo, and the nav link resolves to it
• The authored landing page serves at the site root and the doc catalog moves off / (platform#14)
  • honesty: GET / returns the authored landing page (League of Agents title and hero), while the doc catalog stays reachable at its own path
• The solo-vs-bot house team acts each turn via the game real bot policy instead of holding forever (platform#9)
  • honesty: In a solo-vs-bot match the house team units observably move/gather/deliver in the turn record, driven by the game bot policy - not all-holds
• The score endpoint surfaces the game quality axes and outcome breakdown, not just raw points (platform#10)
  • honesty: The platform score endpoint returns the same quality axes and outcome breakdown as league match score --json for the same finished match
• The game stepwise cmatch loop gains fog support, --message/--plan on act, and bot-briefing message threading (league-of-agents#35-#37)
  • honesty: A stepwise cmatch run with fog and messages enabled replays byte-identical to the one-shot driver on the same seed (extending the existing parity proof)
• devague export emits markdownlint-safe specs and plans - no MD026/MD034 hand-fixes after export (devague#64)
  • honesty: markdownlint-cli2 passes with zero hand-edits on a spec and a plan freshly exported by the patched devague
• The t20 launch checklist (deploy, DNS, production pause/resume, Lighthouse) runs green against production - the plan last open task closes
  • honesty: The checklist lands in the repo checked off with captured command output for every line, all run against production league-of-agents.ai
• Every AWS service and action touched during deploy and maintenance is recorded as it happens, and the sweep ships a least-privilege IAM policy document plus runbook for a dedicated deploy/maintenance user - root credentials become a one-time bootstrap
  • honesty: The policy document enumerates every service-action pair actually used (from CloudTrail or logged CLI calls), and a simulated or real run of the deploy under that policy succeeds

Honesty conditions

• Every trail item (platform #4-#7, #9-#12, #14; league-of-agents #35-#37; devague #64) ends the sweep closed by a merged PR or explicitly re-parked with a written reason - none silently dropped
• Each audience path is exercised at launch: an anonymous human reaches the landing page, an agent reaches /llms.txt and the API surface, and each sibling-repo fix lands via that repo own merged PR with green CI
• The before-state is evidenced live: aws cloudformation list-stacks shows no platform stack, handler.py composes site_app() only, and the trail issues are open on GitHub as of the frame date
• A person or agent with no repo access completes a full match using only public URLs and public docs
• The sweep merged PRs stay within the filed issues scope - no continuous-lane or maturation-frame code appears in any diff
• The signal is captured concretely: a production match record created by a non-maintainer identity exists, the checked-off t20 checklist is committed, and gh issue list shows the trail closed
• Root credentials are used only from the maintainer machine during this sweep and nothing durable (keys, tokens) derived from them is committed or logged
• Every after-state element maps to an in-sweep requirement whose honesty condition is verified against production; the OAuth deferral is visible as an annotated open platform#6, not silence

Success signals

• A stranger completes a match on https://league-of-agents.ai; the t20 checklist is green on production; the trail issues across the three repos are closed by merged PRs; an IAM policy document exists that would have sufficed for the whole sweep

Scope / boundaries

• No new features beyond the filed issues; the continuous-lane adapter and the maturation frame stay parked for the next cycle; no multi-region, no CDN work beyond Cloudflare proxying

Assumptions

• Deploying with account-root credentials is acceptable as the one-time bootstrap for this launch (the recorded-access IAM user replaces it for maintenance)

Decisions

• Sequencing: deploy-and-DNS first (prove infra, claim the domain), production persistence (platform#4) lands before any public announcement, OAuth registration (platform#6) runs in parallel as the one user-gated lane, the experience sweep (#9-#12, #14) and the upstream repos (league#35-#37, devague#64) parallelize freely
• The 20 USD/month ceiling and the existing capacity caps stay unchanged through the sweep
• Launch gate (user-decided): the site is live once deploy + persistence + agent-token play work on production; human OAuth login lands mid-sweep when the user registers the GitHub/Google apps from prepared steps and callback URLs
• User decision: OAuth login is deferred to the next iteration - platform#6 stays open, annotated with prepared registration steps and callback URLs; Cloudflare (DNS/TLS/proxy) is the launch-time front; launch-time human experience is browse + spectate, play is agent-token

Open / follow-up

• Actually creating the IAM user and rotating root out of the loop - this sweep ships the recorded policy document; the user creation itself is maintainer-side