picopilot@0.1.4

Patch Changes

• d98efab: Make the agent JUDGE the rendered frame, not just confirm it drew, so readability and layout-fairness defects get caught. A large-model run verified every state transition thoroughly but shipped a game where a hazard was the same colour as the background and obstacle gaps were too tight to pass, because it only ever confirmed screenshots "rendered" and never critiqued them.

  The scaffolded AGENTS.md definition of done now adds a "judge the frame" check: when you look at a gameplay screenshot, actively hunt for what is WRONG (can you tell every entity apart from the background and each other; are hazards spaced so a human can react in ~250ms rather than frame-perfect; does it look intentional), because "it renders" is not "it is good". game-design-reference gains the matching self-check that applies its readability + fairness lenses to the actual rendered image, not just the logic. Kept concrete (distinguishable? clearable? intentional?) rather than a vague "make it pretty".

picopilot@0.1.3

Patch Changes

• f884a6f: Ship a PICO-8 API reference and stop the agent guessing, and make "done" catch mislabeled outcomes and too-easy games. These target the exact ways a weak model wasted runs and shipped bugs.

  • New reference/pico8-api.md resource under picopilot-code: the exact function names + signatures for the common API, plus a table of the wrong names an LLM reaches for (rand/ranf, rectcol, math.floor, table.insert, poke4 for saving, spr with extra args) and their real PICO-8 equivalents. picopilot-code and the always-loaded AGENTS.md now point at it with a "do NOT guess an API name, look it up" rule, so a nil-call crash does not cost a whole playtest round. The resource ships to installed skills via the existing resource-copy seam.
  • Stronger definition of done in AGENTS.md: beyond "state behaves, not just renders", it now adds two generic checks: drive EVERY distinct kind of interaction and confirm its effect has the intended SIGN (the classic bug is a hazard that adds score instead of costing a life), and run the design self-checks against what you actually saw (can the player genuinely lose, or is it too easy / does it escalate / is it fair). Kept generic, no fixed mechanic checklist.

picopilot@0.1.2

Patch Changes

• 00b4f00: Make "done" mean the game's behaviour is verified, not just that a frame renders, so state-transition bugs (a pickup that never despawns, lives that do not reset on restart) get caught before handoff.

  The scaffolded AGENTS.md build loop now tells the agent to SEE the cart PLAY (drive it via picopilot playtest run or the temporary run probe and compare a SEQUENCE of frames for motion and state change), and its definition of done requires driving the core loop and confirming state actually behaves. The check is deliberately GENERIC: it refuses a fixed mechanic checklist and instead tells the agent to enumerate the state transitions ITS OWN game defines and verify each (with a printh state-dump fallback), so it neither assumes mechanics the game lacks nor checks only a canned list and misses the rest.

picopilot@0.1.1

Patch Changes

• 528b9b4: Steer the build toward a visually-verified cart and away from the serve tar pit, so weaker models follow the same disciplined loop the strong ones already do.

  • Scaffolded AGENTS.md now spells out the build loop + a definition of done: every iteration ends by SEEING the cart run (a temporary extcmd("screen") probe screenshot, or picopilot playtest run), with the probe recipe inline. It states plainly that a green verify or a bare run timeout is NOT proof the cart works, and that the finished-cart handoff is pico8 -run <path> (not a browser serve), with no extra summary docs / scripts / banners as deliverables.
  • picopilot run's timeout CTA is now screenshot-aware: a timeout with NO screenshot says a timeout alone does not prove it works and points at the probe / playtest / verify; a timeout WITH a screenshot is framed as the expected interactive-game case.
  • serve is scoped to its real role (sharing a playable build with a human, not self-verification): the command description and the labelless export-failed message now redirect to run/playtest for verifying your own build, and point at the real label remedy (export --label) instead of hand-editing main.p8.

picopilot@0.1.0

Minor Changes

• a989c1a: First feature release of picopilot, the agent-first PICO-8 toolchain (a single TypeScript CLI on the incur framework that is also an MCP server and a set of auto-installable agent skills). Highlights:

  • Core cart loop: init scaffolding (the #include main.lua layout), the tokens budget loop, verify as the single static acceptance gate (tokens + lint + integrity, with the honest gate-incapable outcome when shrinko is absent), and minify.
  • Eyes (art): the gfx set (char-grid) -> gfx render (viewable PNG) -> look-and-fix loop, with the gfx/map overlap smart-refuse.
  • Ears (audio, v2): sfx from-mml and music from-patterns over a documented picopilot-MML subset tuned to PICO-8's exact capabilities, SFX filters (!dampen/!reverb), and record-based audio render.
  • Run + playtest: headless run (boot + screenshot + printh, ends on a done-sentinel), and playtest which drives an arbitrary cart through scripted input and captures real gameplay, one-shot plus a resumable live session (btn/btnp -> serial transform over a fixed-block + ACK transport; ADR-0011/0012).
  • Export: export produces a browser-playable PICO-8 HTML bundle, with an optional --label splash.
  • Skills: auto-installable discipline skills (overview, code, art, audio, debug) plus a composed game-design pair (game-jam + game-design-reference) carrying fairness, human-reaction-budget, visibility, originality-method, and budget-scaled progression guidance.
  • Hygiene: every PICO-8 launch is isolated with -home so it never writes config/data into the working tree.
  • Includes a hands-on platformer tutorial (examples/platformer/) and a curated game showcase.