A more agile city.
A finish worth reaching.

This is a revision of the Astra edition of Nothing@lose, following the creator’s playtesting feedback. It is not a new agent generation. The original Opus 4.8 edition remains preserved and playable.

The brief

Improve jittery movement, make completing the word at the Capitole satisfying, add rare homing missiles, replace short-limbed robots, improve trees and give Toulouse more varied, recognizable districts. Reframe the homepage around the game and its growing AI development archive.

What changed

• Movement commands are processed in fixed steps, with fractional rendering and smoothed server corrections. The camera follows the same rendered character position. Slow rendering no longer reduces the intended command rate.
• Taller articulated robots have hip, knee, ankle, shoulder and elbow joints, two-handed weapon support, airborne poses, landing compression and a victory gesture.
• Pigeon Guidé missiles acquire targets over 0.85 seconds. Their steering is limited, buildings break guidance, and targets receive warnings. A pickup supplies two missiles and takes 75 seconds to return.
• The playable bounds grow from 246 × 302 to 390 × 430 game units. The Saint-Étienne quarter and Grand Rond garden provide the final letter types, with alternate routes. Trees use branching trunks and layered leaf textures with wind.
• Completing the run now preserves the result, protects the winner, hides gameplay controls and presents a celebration, elapsed time, eliminations and the leaderboard. Another run is an explicit choice. Other players continue.
• The expanded course has a separate leaderboard. Earlier course scores remain preserved on disk.
• Nothing@lose leads the homepage. Agent names identify editions within a playable archive, alongside development evidence and gaps in the record.

Time, effort and attribution

Revision began

6 September 2026, 18:57:35 UTC — observed work-start timestamp.

Resumed

7 September 2026, 06:55:16 UTC — after an overnight pause. The interval is not a measure of coding time.

Active coding time

Not measured. No reliable total is claimed.

Earlier rebuild effort

Not recorded. The initial Astra rebuild and original Opus development have no reliable time or cost baseline.

Cost / token usage

Not recorded.

Human contribution

Concept, direction, approval and repeated playtesting feedback on movement, combat, graphics, landmarks, completion and the homepage. No intervention count was tracked.

Agent labels

Astra is the designation for this project. Opus 4.8 is the creator’s attribution for the original; an independent historical model log is unavailable.

Verification

The revised implementation passed 19 simulation tests covering traversal, letter reachability, bot completion, collision, combat, input queues, finishing/restarting and homing acquisition/cover behavior. A two-client check verified movement synchronization, stopping, disconnects and separation from the original room.

Browser checks covered desktop and emulated phone joining/rendering, simultaneous touch movement and high jumps, double jumps, air dashes, drag-to-aim firing and input release. An isolated fixture verified the full desktop and phone finale, saved leaderboard and explicit replay. Its test scores are deleted and never enter the public leaderboard.

The movement fixture injected 60–80 ms input-delivery delays and checked for backwards snaps and excessive speed variation during steady travel. This is a regression check, not a frame-rate benchmark. Tests run in headless Chromium using software graphics on this machine.

The 63 original compiled client files were compared with the pre-rebuild backup and remain byte-identical. Browser checks also load the original edition.

What this does not establish

Real-phone frame rates, sustained 16-player load and difficult mobile networks still need hands-on evaluation. The city and characters are stylized procedural art; this is not a claim of console-production fidelity. Completion fixtures test the finish reliably but do not replace human playtesting of the whole run.

The editions use different implementations and have received different amounts of feedback. Their difference cannot be attributed solely to the agent. This archive makes that comparison playable and documents its limitations.

Assets and city references

The original soundtrack is retained. Robot, building, foliage, sky and water visuals are generated procedurally in the game. The countryside panorama was generated for the Astra edition; its prompt is recorded in the project’s asset README.

Landmark inspiration includes the official tourism descriptions of Grand Rond and Saint-Étienne Cathedral. The playable layout simplifies their geography and architecture.

7 September follow-up: weapon balance and time displays

After further playtesting, all run-time displays now use minutes and seconds. Ordinary rockets travel twice as fast (96 game units per second). The scattergun reaches 60 units instead of 32, with a tighter spread and a 0.55-second firing interval instead of 0.7 seconds. Two additional simulation checks cover effective mid-range shotgun damage and fast-rocket impacts. These changes respond to human feedback; they do not constitute a new agent edition.

7 September follow-up: animation, letters and melee

Further feedback led to an SVG robot favicon, somersaults on jumps and double jumps, smoother body turning with independent weapon aiming, subtle idle movement, and an always-available long sword. Press E or use the SWORD touch button. The sweep has a 0.14-second wind-up, a 0.7-second cooldown, 4.8-unit reach, 60 damage and knockback; walls and vertical separation block hits.

Letter spawn count increases from 16 to 24. Uncollected copies remain visible even when that letter type is already complete in the player’s inventory. The required word remains TOULOUSE. These availability and combat changes can affect completion times; scores are retained, so the leaderboard is not a fixed-rules benchmark across revisions.

Keeping the next record

For future agent generations, record the initial brief, exact available model identification, start/stop intervals, human interventions, asset provenance, test conditions and known issues from the outset. Preserve each released game and its ruleset with its own scores.

Read the structured version record ↗ · Play the latest edition ↗