Agent assisted Changelog - 37/52

In the last summary, the first paid Agents of Empires tournament was still next. There are now cups, champions, a domain, replays, and enough conversations about tournament formats that I briefly considered duplicating the entrants just to fill a bracket. The software is getting better. My ability to leave the scope alone remains unproven.

The battleground became a cup circuit

Agents of Empires is now live at agents-of-empires.dev. The homepage explains the idea and points to cups; each cup has its own description and tournament history; each tournament has a bracket and links to the actual attempts. The replay defaults to 5x, because watching someone else's infrastructure wait to boot should at least be faster than watching your own.

The usual format is six entrants, two three-agent opening heats, and a final with the two winners plus a wildcard. Publishing a draw and running it are separate operations. That means I can announce the field before spending anything, then run exactly that bracket later.

The first clean OpenCode Cup had DeepSeek win its opening build race in 43.2 seconds and the durable job queue final in 89.8. All three heats finished on their first attempt. No forfeits, no replays. Getting there was less tidy.

Earlier tests produced enough failures that I could not tell whether I was watching poor play or a broken adapter. The fixes included explicit guest model definitions instead of hoping OpenCode refreshed its catalog, making the configured memory floor reach the actual VMs, and keeping the rollout continuity monitor alive for the match's configured duration.

There was also a tournament that crowned a contender after everybody failed. That is a perfectly reasonable way to invent a sport I did not mean to invent. Heat winners now need a durable deployment. Sporting draws and unavailable seats have separate, bounded replay budgets, and an exhausted bracket can end without a champion. Tied wildcard ranks use a seeded lottery, not alphabetical order or the cheapest bill.

Kimi took the first published Vercel Cup. Different routes, different fields, different stories. Still not a model ranking. Replaybook is where I do the evaluation work; this is where I want to pick a contender and watch a race.

One pool, frozen draws

Adding cups made the model lists annoying almost immediately. OpenRouter, OpenCode Go, and Vercel AI Gateway each expose a different slice of the world, and maintaining those lists independently was an invitation to drift.

AoE now has one shared model registry. Each cup selects its provider slice, and season draw --entrants 6 samples a field without requiring the whole eligible pool to enter. The selected entrants and eligible pool are frozen into the draw. Refreshing a catalog tomorrow must not change the tournament I announced today.

The refresh script defaults to a dry run. New discoveries stay disabled until reviewed, and refreshes preserve deliberate access blocks and premium-model opt-ins. Finding a model in a catalog is not proof that my account can run it, or that I want to pay for it.

I also floated a budget cup that draws from whatever is on sale. That remains an idea, not another thing quietly running up a bill in the background.

Same model, different harnesses

The new adapter is for term-llm. The first Harness Cup put it against Claux and OpenCode, all configured for OpenRouter GPT-5.6 Luna, high reasoning, one vCPU, and 768 MiB of memory per guest.

term-llm won in 45.216 seconds. The outside referee verified the service, record round-trip, service restart, and host reboot. OpenCode was outraced. Claux's guest rebooted and the SSH connection dropped, but the retained evidence does not establish who initiated that reboot. There was no referee reboot marker for its seat and no recorded reboot command in its tool trace. That stays an unresolved harness error, not a story about a model deciding to restart its machine.

The accounting needed attention too. Missing provider costs had looked like zero dollars, and token totals did not consistently include cache categories. Future adapter output now normalizes fresh input, cache reads, and cache writes; reporting distinguishes a known total from an incomplete subtotal. The original event logs remain unchanged. Usage captured while losing agents drain after the finish is also not a synchronized race-only measurement.

We recorded this as an exhibition, then published it as a single-round cup. Three entrants and one final are enough. I considered two opening heats with duplicate entries from each harness, then decided to wait for more distinct harnesses. A bracket should describe the competition, not manufacture a need for more seats.

Timecapsule starts with an inbox

The other thread was an archive of Angine de Poitrine material. I wanted to find scattered recordings and links without turning every discovery into a published claim about a performance, its date, or its rights.

Timecapsule is a local prototype for that flow. Internet Archive discovery produces private metadata snapshots in an inbox. I review a candidate, approve it into a Markdown record with source and rights notes, then build a static archive. Repeating discovery does not reset review decisions. An upload date does not silently become a concert date. Approval does not deploy a website or download the recording.

The optional TypeSafe/Jev integration sits between discovery and review. It can suggest relevance and material type from metadata, and I can preview the exact request before making a paid call. It does not watch the video, approve the entry, or decide that I have permission to republish it. I have not run a paid enrichment call yet.

The first implementation has sixteen tests covering the unglamorous parts: deduplication, retained decisions, approval recovery, validation, escaping, private/public separation, and byte-identical offline builds. YouTube discovery and a visual review inbox are not wired in yet. There is no public repository link for this one yet either; it is still local work.

The through-line

This week kept coming back to what a result actually authorizes me to say.

A model being listed does not mean an account can use it. A null charge does not mean the race was free. An SSH disconnect does not identify who rebooted the machine. Finding a recording does not establish its date or give me permission to copy it.

The useful work is in keeping those distinctions visible without making the whole thing miserable to use. I still want the bracket to be fun and the archive to be inviting. I just want the next person clicking through to find the evidence, not a more confident version of my guess.