typhon

Typhon — Language, Compiler, Runtime, and VM Skill

Typhon is a statically-typed, stricter superset of Python that emits clean CPython 3.13+ with zero runtime dependency on the toolchain. The compiler, language server, formatter, debugger wrapper, REPL, and tree-walking interpreter are all a single Rust binary called tyc. Every .ty file emits valid, idiomatic .py. Not all .py is valid Typhon.

Current release: v1.0.0-alpha.9 (2026-08-21) — a maintenance release on top of alpha.8, with no language change. The warn-level tyc::contains_secret_literal keyword table grows from 16 entries to 55: seven overlapping proposals consolidated into one longest-first-ordered table (AUTHORIZATION, CREDENTIALS / CREDENTIAL, WEBHOOK, SIGNING, COOKIE, DSN, the DB_-prefixed password / pass / pwd / secret variants plus a secret-only APP_ / CLIENT_ / JWT_, the ACCESS_ / AUTH_ / BEARER_ / CSRF_ / JWT_-prefixed token variants, and PRIVATE_KEY / PUBLIC_KEY / SSH_KEY / SECRET_KEY, each with its squashed-acronym form), so DB_PASSWORD now reports DB_PASSWORD rather than the bare PASSWORD. One more name-word boundary lands with it — an uppercase keyword directly followed by a lowercase letter (TOKENs, dbPASSWORDstring); MONKEY still does not match KEY. Alongside: three more allocation reductions on compiler AST walks (auto_gather's candidate scan, module_all_names, and tyc build's import scan), the mid-August dependency wave (with the compact_str 0.10 bump reverted — get-size2 0.8 pins 0.9.1, and the duplicate broke the vendored Ruff build), and docs-site keyboard-accessibility / reduced-motion polish. No new syntax and no new error-level diagnostic. v1.0.0-alpha.8 (2026-08-08) was a maintenance release on top of alpha.7. One VM ↔ CPython parity fix: the VM's CPython-compatible MT19937 was seeded from a fixed constant when a program never called random.seed(), so an unseeded program drew the same sequence on every tyc run while tyc build && python drew a different one each time (CPython seeds from OS entropy at import). The default now seeds from entropy to match; explicit random.seed(n) is unaffected and still yields byte-identical sequences under both surfaces. Alongside it: the warn-level tyc::contains_secret_literal name lint now treats digits and UPPERCASE→TitleCase junctions as word boundaries (myPASSWORD123, dbPASSWORDString are flagged; MONKEY still does not match KEY), two allocation reductions on AST walks (parallel_lints, the desugarer's class-marker collection), the early-August dependency wave (five crates, two GitHub Actions majors), and docs-site card / link-card transition polish. No new syntax and no new error-level diagnostic. v1.0.0-alpha.7 (2026-07-29) was the 2026-07-28 full-codebase-review remediation release (docs/codebase-review-2026-07-28.md). It closes all ten of the review's 1.0 blockers plus the Tier-0 gates that make them verifiable: type-checker soundness fixes (instance-attribute assignment is now type-checked, constructor field order follows reverse MRO, model constructors are keyword-only, recursive type aliases terminate, parametric sealed-union match exhaustiveness is checked, and a new nullable-field-dereference check lands at warn level), resolver fixes (let immutability is now enforced inside loop bodies and through global/nonlocal), a shared-lexical-mask rewrite of the preprocessor fixing several literal-corrupting bugs (gather: dependency scanning, with-chain parsing, enum body rewriting), five emitter/preprocessor miscompilation fixes, and VM ExceptionGroup/except* support (PEP 654) plus five other VM↔CPython divergence fixes. Two new CI gates ship alongside: a VM↔CPython differential gate over the full 1342-file example+stress corpus (first run found 126 divergences, now pinned as a baseline), and an opt-in-knob codegen matrix. No new syntax; most changes are narrowings on code that was already crashing or relying on unsound typing, plus one warn-level diagnostic-surface addition. v1.0.0-alpha.6 (2026-07-21) was a maintenance release on top of alpha.5, driven by the 2026-07-20 release-readiness review (docs/release-readiness-2026-07-20.md). The July dependency wave is carried safely across the toml 0.8 → 1.x major — including the fix, before it could ship, for the one regression that bump introduced (the tyc-venv dependency allow-list reader silently returning empty, which would have no-opped third-party arg/type checking). Six squashed-acronym keywords join the secret-name table (API_TOKEN, APITOKEN, API_SECRET, APISECRET, API_PASSWORD, APIPASSWORD), now hoisted into a single shared tyc_analyse::SECRET_NAME_KEYWORDS constant so the tyc::contains_secret_literal lint and the tyc build scan can no longer drift. The release pipeline's artifact download is re-pinned to the proven v4 line (matching the v4 upload), and docs-site accessibility, diagnostics-index completeness (all 87 pages indexed), and repo hygiene (the .Jules/.jules case-collision fix) round it out. No new syntax; no previously-correct program changes behaviour — the only diagnostic-surface change is warn-level (six more secret-shaped names warned). v1.0.0-alpha.5 (2026-07-09) was a performance-focused release on top of alpha.4. VM performance Tier 1 (tyc run): a two-representation VmInt (Small(i64) inline, Big(BigInt) on overflow) keeps common-path integer arithmetic allocation-free while preserving CPython's exact arbitrary-precision semantics; a per-class resolved-method cache (negative results included) and a direct obj.method(args) dispatch path skip redundant find_method walks and BoundMethod allocation; and functions with no closures/global/nonlocal resolve locals to a fixed slot computed once instead of a HashMap lookup per read/write. Cumulative effect: the VM's steady-state slowdown vs tyc build + CPython compresses from ~5–18× to ~3–14× (startup-adjusted); the VM still wins on startup latency. [optimise] config profile + tyc build -O: a single level dial flips the default of auto-memoise / auto-gather / auto-parallel / pgo-memoise to true (an explicit [strictness] entry always wins). Seven new advice-only lints in tyc-analyse/src/perf.rs, gated by [strictness] suggest-perf: perf_membership_in_loop, perf_list_shift_in_loop, perf_str_concat_in_loop, perf_sort_in_loop, perf_sorted_first, perf_keys_membership, and lazy_import_opportunity. Free-threading parallelisation wave ([python] free-threaded = true): auto-parallel widens to comprehension filters, multi-argument calls, and nested pure calls; a new auto-parallel-reductions knob folds a provably-bounded integer accumulator loop (mut total: int; for x in xs: total += EXPR) into a parallel sum(map_pure(...)); two new advice lints (parallel_opportunity, shared_mut_across_tasks); and a parallel-backend = "interpreters" option tries a PEP 734 InterpreterPoolExecutor before falling back to threads. Native PEP 810 lazy imports: [python] target = "3.15" / "3.15t" lowers lazy import ALIAS = MODULE to CPython's native lazy import MODULE as ALIAS statement instead of the typhon_runtime helper call; 3.13/3.14 output is byte-for-byte unchanged. No new syntax; every rewrite here is opt-in or advice-only, so no previously-correct program changes behaviour. v1.0.0-alpha.4 (2026-07-08) was a focused hardening pass on top of alpha.3. Type-checker soundness: it closes H5, the one HIGH finding the release-readiness review deferred — the v0.15.0 qualified↔bare tail-unification let a locally declared class satisfy a same-named class from another module (a user class Response: type-checked into an httpx.Response-typed slot, and vice versa). is_assignable now refuses that unification when the bare side names a class declared in the module being checked and the qualified side's declaration — resolved through its exact module key — has a different shape. The guard is evidence-gated and degrades to the previous permissive unification on any uncertainty (unresolvable module, unknown shape, facade re-export with an equivalent shape, interfaces, bare names of unknown provenance), so the example/stress corpus is byte-identically unchanged. Diagnostics: the secret-name keyword tables in tyc-analyse and the tyc build scan match longest-first again — APIKEY precedes the bare KEY, so KEY_APIKEY reports the more-specific suffix. Supply-chain & packaging: every third-party GitHub Action is SHA-pinned (Dependabot keeps the pins fresh), the installers resolve the newest release including pre-releases (a default install no longer silently fetches older binaries), and a round of dependency / advisory bumps lands (crossbeam-epoch 0.9.20 for RUSTSEC-2026-0204, plus regex, memchr, compact_str). No new syntax; like the alpha.2 diagnostics, the H5 fix is a conservative narrowing that only rejects programs passing a provably-different-shaped class across a module boundary, so no previously-correct program changes behaviour. v1.0.0-alpha.3 (2026-07-03) was a release-readiness remediation pass (RELEASE_READINESS_REVIEW.md), the licensing / packaging / robustness counterpart to alpha.2's soundness sweep. Licensing & packaging: a repository-root MIT LICENSE, the upstream Ruff MIT notice vendored beside the fork, SECURITY.md / CONTRIBUTING.md / Dependabot, and release-workflow hygiene (pre-release tags flagged as such; auto-tag gated on green CI). Compiler & VM: identical errors at distinct source locations are all reported now (the module- and resolver-level dedupe no longer swallows the 2nd+ occurrence), nested-generic assignability is linear again (not O(2^depth)), four more flow-narrowing invalidation holes are closed, and six VM↔CPython parity gaps are fi

Content truncated.