Skip to content

AGENTS.md

Latest commit

 

History

History
39 lines (33 loc) · 2.07 KB

AGENTS.md

File metadata and controls

39 lines (33 loc) · 2.07 KB
Raw

AGENTS.md — cppmega.mlx agent contract

RULE #1 (HIGHEST PRIORITY) — NO AUTOMATED FALLBACKS. FAIL FAST, FAIL LOUD.

Every op goes through ONE clear path. When the clear path does not work as expected, errors, or misbehaves, RAISE IMMEDIATELY with WHERE + WHAT failed and crash — surface the bug, fix the root cause. FORBIDDEN: silent/automated fallbacks (try/except→degraded path, except:pass, eager-replaces-fused, MSL-instead-of-tvm-ffi, gated-off-by-default, clamp-instead-of-raise, return zeros/garbage, best-effort, silent precision/shape downgrade). Fail-closed by RAISING a clear error is correct; a silent fallback producing degraded/wrong/zero output is forbidden. See CLAUDE.md. This overrides every rule below.

Tensor memory rule

  • Wrappers/adapters must not silently allocate or copy large tensors.
  • If a design appears to require staging a large tensor, casting a large tensor, or copying a large tensor only to satisfy a wrapper boundary, treat that design as wrong and keep looking.
  • Prefer passing references to existing GPU buffers, fusion, views/broadcasts, or IR-level lowering through the TileLang/TVM/tvm-ffi pipeline.
  • If zero-copy/fused lowering is not currently possible, FAIL EXPLICITLY (raise with where + what). Do NOT silently fall back to another path (Rule #1).
  • Dtype/shape mismatches must be solved at the right level:
    • first look higher in the graph and make the producer create the tensor in the required dtype/shape from the start;
    • if the current dtype/layout is the right system-level format (for example FP8 for the active config), move lower and teach the kernel to consume that format directly;
    • only use casts/reshapes as explicit graph decisions, not hidden adapter staging.

Native subagents

  • Use Codex native subagents for independent, bounded parallel work when it materially improves throughput.
  • Every native subagent must use gpt-5.6-sol with ultra or max reasoning. Use ultra by default.
  • Do not use another model or a lower reasoning level unless the user explicitly overrides this rule.