Prosper Nwo

Gate

Decides whether a code diff from any agent can be merged: approve, review or reject. Runs as a CLI, a GitHub Action, or an MCP server the agent calls on its own work. 472 tests

Ledger

Append-only decision log. Records are SHA-256 hash-chained in Postgres, chain roots are Ed25519-signed, and the auditor export maps to EU AI Act Articles 12 and 14. 244 tests

Warrant

When one agent delegates to another, it issues a short-lived token scoped to that job. Every tool call passes through a proxy the model can't reach. 157 tests

Harbour

Runtime for long-running agents. Goals survive crashes, you can attach to one from any machine, and two-phase commit means a retried action never runs twice. 158 tests

Bench

Turns a repo's own bug fixes into benchmark tasks, then ranks coding agents on them. It answers "which agent should we use?" with your code instead of a public leaderboard. 107 tests

Proof

For crawlers: records what robots.txt, content signals and the page licence said at fetch time, in signed manifests, and reports when those rules change later. 191 tests

How I know Gate works

I built a labelled set of 71 diffs across Go, Python, JavaScript and TypeScript: 44 with a planted bug and 27 clean. This is Gate with no LLM, using only its static and security checks:

Precision 1.0, recall 0.21. The deterministic checks never raise a false alarm, but they miss every bug that needs you to understand the intent. Catching those is the LLM review stage's job, and I haven't benchmarked it yet. I published both numbers because the low one tells you more.

By the numbers

From Prosper's own build database: 246 games and 2,285 units in total. The second period adds the genre starter games, a stricter play test that fails any runtime error, and Jules. Because the play test is stricter, only 22% pass it first time. Regeneration and repair lift that to 72%. The first period didn't record first-test results separately.

Jules, the repair agent

When a game fails the play test, Prosper copies it to a repair folder and hands it to Jules, an OpenCode agent running on free models. Jules gets the spec, the exact failure evidence, and a check command that runs the real merge and play test. Two sessions race on separate API keys, and the first to pass wins.

Prosper doesn't take Jules's word for anything. After every round it re-runs the play test itself, restores any pipeline-generated file Jules touched, and copies a fix back only if it passes. Every round is logged with its diff, and when the same kind of repair keeps showing up, I fix that bug in the pipeline instead.

In its first two days, Jules worked on 7 games that failed the play test and made 5 of them playable, over 33 rounds and about 22 hours of agent time. One fix took a single round: a ReferenceError from a variable used before it was initialized. Another took three rounds and four files, after a sprite call that didn't exist crashed the game loop. The two it couldn't fix both came from add-ons calling methods the game didn't have, so I fixed that at the source: each add-on now declares what it needs from the game, and installation fails immediately if the game doesn't provide it.

• speed

  Single steps ran for over 10 minutes because step time grows with reasoning tokens. I switched the default to low reasoning effort.

• stalls

  29 of the 33 rounds ended on a "no progress" timeout. A 7-minute stall limit killed sessions in the middle of a step. I raised it to 13 minutes, and a hung request now retries in the same session after 4 minutes instead of ending the round.

• scope

  The repair folder sits inside the repo, so Jules wandered into pipeline code. It's now blocked from reading or searching outside the game.

• restarts

  A new repair session now starts from the previous session's edits, so a restart keeps the progress already made.

Genre coverage

Nine hand-built starter games across eight genres, like Star Lancer (space shooter), Gem Cascade (match-3) and Bastion Road (tower defense), come with asset kits and 16 add-ons like bosses, extra waves and relics that scale a game up. Seven genres have a play test that actually plays the game. The shooter test, for example, fails a run where one enemy shot destroys the ship.

To check coverage, I classify 101 realistic game requests. 97 land in a genre (96%) with 0 misclassified, and 81 (80%) have a starter game today. The four it can't place (3D shooter, MMORPG, rhythm game, fighting game) are out of scope for now. Management and story games have no starter yet.

What broke first

In August I audited the pipeline against four real builds. The verification looked strict but was mostly letting everything through:

• gate

  The browser play check set ok = False on failure, but the function never returned it. No build could fail.

• motion

  The "is anything moving" check compared two array objects instead of their contents, so it always saw motion, even in a frozen game.

• boot

  The Phaser boot check skipped itself because the platform string was phaser+js, not phaser.

• cache

  Empty units all hashed to the SHA-256 of an empty string, so one unit's failed verdict was reused for every later empty unit.

All four are fixed. Every JavaScript and TypeScript file is now parsed and bundled with esbuild before approval, and the play test fails a game on blank rendering, placeholder art, game-code errors, or a run that ends without taking any input.

What broke

Literal queries worked. "How does octopus camouflage work" put the octopus pages first. Paraphrased ones didn't:

"what makes something intelligent or smart" missed the octopus article, which says octopuses are "considered the most intelligent of all invertebrates." It returned thesaurus pages instead, with the best score at 0.16.

There were two causes. Each page was embedded as a single vector, so one sentence in a 10,774-word article got drowned out. And search always returned the top results, however weak they were. The next day I split pages into overlapping chunks, each with its own vector and a link back to its page, and added a minimum score so a weak match returns nothing instead of noise.