Spaces:
Sleeping
Sleeping
Commit History
-
Option A: route live codegen path through pipeline (gate + retry + fallback)
app.py codegen branch now uses pipeline.build instead of single-shot codegen_assemble: every generation is metric-gated, retried with the failure fed back, and falls back to the matching analytic primitive on exhaustion — so a buggy model generator (e.g. icosahedron) self-corrects to a correct primitive instead of rendering lopsided. Regeneration stays model-free (_regen_modelfree re-runs the cached generator, re-falling-back to the primitive if needed). pipeline meta now surfaces the accepted spec for caching. Enable codegen on the Modal image (TRIKONA_CODEGEN=1 env). 5 new tests; suite 130. Verified live: octahedron/dodecahedron/vent pass on the model generator; icosahedron falls back to the primitive (20 faces) and passes; vane regen 8->12 in 0.05s with no model call.c42f749 -
Option A · Phase A4: codegen system prompt + Modal/app integration (flag-gated)
- prompts/system_prompt_codegen.txt: model emits a build_spec (structured metadata + a numpy build(p) generator), not coordinates; carries the motion guardrail + guided/simplified behaviors; 3 worked examples. - modal_endpoint.py: TRIKONA_CODEGEN selects the codegen prompt and returns the raw build_spec (no count normalization). Pure LLM; no geometry runs on Modal. - app.py: codegen_assemble() runs the generator + assembles the frozen contract in the backend; regenerate re-runs the CACHED generator with no model call; generator failure falls back to the analytic primitive. Default path (flag off) byte-for-byte unchanged. Offline gate via canned build_spec fixtures: contract + metric gates pass. 11 tests; suite 112. FLAGGED for user: add TRIKONA_CODEGEN to the Modal image env and 'modal deploy' to run codegen live.36b02e9
