Imported from mewhhaha/tt (AGENTS.md). Install upstream with npx skills add mewhhaha/tt. Copyright stays with the author.

Working on TT

The owner authorizes direct commits to mewhhaha/tt main and ongoing exploration of a fast refinement-first structural language. Preserve concurrent changes and never force-push. Publishing packages, deploying, editing permissions, and modifying other repositories are not authorized.

Current implementation and local execution

The implementation is dependency-free Node.js ES modules. WebAssembly is the ONLY executable output target. run compiles and executes real Wasm; no TT bytecode VM, JavaScript executable output, native code target, or embedded interpreter is allowed. The host validates/instantiates Wasm and decodes values; it does not evaluate TT operations. No native compiler, Python, package installation, CI artifact, hosted checker, or third-party runtime is a prerequisite.

The current branch must not contain a parallel C++/Python implementation. Historical notes and benchmark evidence may remain as documentation/data, but active or dormant .cpp, .hpp, .py, and CMakeLists.txt implementation files are forbidden by scripts/repository-policy.mjs and local verification.

Read docs/STATUS.md, docs/DESIGN.md, docs/WASM.md, docs/ROADMAP.md, docs/LOCAL_DEVELOPMENT.md, docs/PRODUCTION_READINESS.md and recent commits before choosing work. Inspect live main; do not assume a previous temporary workspace or checkpoint is current.

Run locally:

npm test
npm run verify
npm run bench -- --sizes 500,1000,2000 --samples 11
npm run bench:runtime

The direct equivalent of verify is node scripts/verify.mjs. CI is optional corroboration, never a substitute for a claimed local run. Report missing execution or repository access honestly.

Iteration discipline

Choose a bounded useful improvement and a falsifiable hypothesis. Add acceptance, rejection, runtime and adversarial tests; run relevant work/performance regressions. Record exact commands, counts, failures, source/runtime provenance, and next work in docs/STATUS.md and docs/ITERATIONS.md. Never fabricate tests or reuse stale hashes. Benchmark startup, checking, emitting, Wasm engine compilation/instantiation/execution, and incidental GC as distinct boundaries. RSS snapshots are not peak memory. Do not infer asymptotic guarantees from one median. Reread main before publication and push a coherent state.

Architectural constraints

Structured type predicates are not arbitrary Boolean callbacks run on unknowns. Keep shape, refinement, effects, ownership, phase, identity, and layout judgments separate even when their surface is uniform. Check ordinary functions compositionally. Do not recheck source bodies at every generic call. Do not erase callable input preconditions through polymorphism, records, arrays, joins, aliases or annotations. Function inputs are contravariant. Branch evidence belongs to immutable binder IDs. Unsupported or exhausted proof attempts never justify acceptance. Preserve strict left-to-right demand, checked i64 overflow, and bounds checks. Keep TT names in Maps, not host object prototypes; malformed artifacts must fail deliberately.

Future ECS inference needs real declaration-owned seeds/providers, not storage-type guesses or accidental closure capture graphs. No native or host authority may be forged through structural records or reflection.

Do not call the compiler production-ready until every documented gate has evidence. Do not shrink the gates to finish. Once genuinely complete, report the evidence and disable the recurring development task. Until then, keep implemented semantics distinct from plans.