AILang/docs/roadmap.md

# AILang Roadmap

Priority-ordered list of upcoming work — milestones, features, todos,
and ideas. The orchestrator maintains this file. The user can request
additions; the orchestrator chooses what to remove or reprioritise as
work progresses.

## Conventions

- One checkbox per entry. `- [ ]` is open; `- [~]` is in progress
  (work has started — a plan exists, a branch is open, or commits are
  landing); `- [x]` is done. A finished entry may stay briefly for
  context, then is removed (with a one-line mirror in
  `docs/journals/`).
- Each entry is tagged by **kind** and lives under a **priority**
  bucket:
  - **\[milestone\]** — big chunk that will get a `docs/specs/<milestone>.md`.
  - **\[feature\]** — smaller addition inside a milestone, no full
    spec.
  - **\[todo\]** — concrete task that can run without a brainstorm
    (cleanup, doc fix, mechanical refactor, test backfill).
  - **\[idea\]** — not yet decision-ready, no commitment.
- Optional `depends on:` line names another entry that has to land
  first.
- Optional `context:` line points to the journal entry (per-iter
  file under `docs/journals/` or, for pre-2026-05-11 entries, the
  archived `docs/journal-archive.md`) where the rationale lives.
  The roadmap is intentionally terse; rationale stays in the
  journals.
- Priority buckets:
  - **P0** — in flight. Spec or plan already exists.
  - **P1** — next up. Decision made; not yet started.
  - **P2** — medium-term. Decided in principle, scheduled later.
  - **P3** — ideas. No commitment; may be cut.

## P0 — In flight

_(empty — canonical-type-names milestone closed 2026-05-11; Floats
closed 2026-05-10 and was removed once it stopped being load-bearing
context. Pick the next milestone from P1.)_

## P1 — Next

- [x] **\[milestone\]** Post-22 Prelude — Eq/Ord — shipped milestone
  23 (2026-05-12). `Ordering` ADT, `Eq`/`Ord` classes, primitive
  instances on Int / Bool / Str, polymorphic free fns
  `ne`/`lt`/`le`/`gt`/`ge`, runtime `ail_str_eq` + `ail_str_compare`,
  end-to-end coverage including user-ADT-instance integration. Float
  has neither instance by design; polymorphic helpers at Float fire
  Float-aware `NoInstance` per DESIGN.md §"Float semantics".

- [ ] **\[todo\]** Brainstorm Step 7.5 — Grounding-check is the *last*
  step before the spec is committed. Any post-PASS edit to the spec
  (Boss-side polish, user-requested change, anything) must
  re-dispatch `ailang-grounding-check` before the commit. Today the
  skill's Step 7.5 + Step 8 wording allows a "fix issues inline and
  re-run Step 7" loop *without* re-running Step 7.5 — closes the
  loop on linguistic self-review only, not on grounding. Update
  `skills/brainstorm/SKILL.md`: any change to the spec file after
  the grounding-check PASS report invalidates that report and
  requires a fresh dispatch before Step 8 commit.
  - context: surfaced 2026-05-12 during iter rt.1 planning — Boss
    edited the roundtrip-invariant spec to add acceptance #7
    ("tests are pure readers") *after* the PASS report from
    `ailang-grounding-check`, then committed the spec without a
    re-dispatch. User flagged the skill gap that allowed this.

- [ ] **\[milestone\]** Post-22 Prelude — Show + print rewire — ship
  `Show` typeclass with `Show Int/Bool/Str/Float` instances; rewire
  `print` through `Show.show`. Gated on heap-`Str` ABI runtime work
  (`int_to_str` needs to allocate). Originally queued as 22b.4 in
  the typeclass milestone, kept on hold pending demand and the
  heap-`Str` prerequisite.
  - context: brainstorm 2026-05-12 (iter 23.5 wrap). Heap-`Str` ABI
    is the load-bearing prerequisite; until it ships, `Show Int`
    cannot allocate a `Str` to return.

## P2 — Medium-term

- [ ] **\[milestone\]** Cross-model authoring-form test — empirically
  measure whether `.ail.json` or `.ailx` is the form a foreign LLM
  author reaches for and succeeds with. Setup: three configurations
  per task — (i) JSON-mandated, (ii) `.ailx`-mandated, (iii) free
  choice (both corpora available, model picks). Multiple subjects to
  check whether form-preference is model-robust. For each (model ×
  form × task): system prompt includes DESIGN.md verbatim (~28k
  tokens); user prompt is the task plus the form mandate; an error-
  roundtrip loop feeds `ail check` / `ail parse` errors back to the
  model until green or iter-limit. Measure: first-attempt success
  (yes/no), iterations to green, total tokens spent, error classes
  encountered (schema violation, type mismatch, hash mismatch, syntax
  error in the chosen form); for free-choice runs, which form the
  model picked unprompted. The output is a JOURNAL'd dataset +
  interpretation + a concrete DESIGN.md edit on Decision 6 —
  either ratify "ailx is the AI authoring projection" with empirical
  backing across N subjects, or retire the form if JSON wins
  consistently across all subjects.
  - context: brainstorm 2026-05-12 (aborted in favour of two roadmap
    entries). The Decision-6 prose currently asserts `.ailx` is the
    AI authoring projection while §"Feature-acceptance criterion"
    cites JSON-as-canonical-authoring-surface as a behavioural
    counterexample to human-driven feature picks. The tension is
    real; only empirical data over multiple downstream subjects
    resolves it.
  - depends on: Round-trip completeness invariant (P1 above) —
    without it, measured preference is confounded by expressive-
    power gaps. Strongly benefits from (but is not strictly
    blocked by): `ail check`/`build`/`run` accept `.ailx`
    extension (P2 todo below); workspace search beyond entry-
    module's directory (P2 todo below). Both will be encountered
    during the test if not closed first.
  - smoke-test evidence (2026-05-12): Qwen3-Coder-Next reads
    DESIGN.md and produces a structurally correct AILang program
    that passes `ail check` in two turns; first turn missed the
    module-name-matches-file-path rule, second turn fixed it after
    `ail check` error was fed back. ~28k prompt tokens, ~150
    completion tokens, ~2-5 s latency per turn.
  - how to address the IONOS endpoint:
    - URL: `https://openai.inference.de-txl.ionos.com/v1/chat/completions`
    - protocol: OpenAI-compatible chat completions API.
      `/v1/models` returns the catalogue.
    - auth: HTTP header `Authorization: Bearer <token>`. Store the
      token outside the repo (e.g. `~/.ionos_token`, `chmod 600`).
      Tokens are JWTs; the one used in the 2026-05-12 smoke test
      was IONOS-issued and time-bound — refresh when re-engaging.
    - model IDs to try (subset of the IONOS catalogue as of
      2026-05-12, sorted by relevance to this test):
      - `Qwen/Qwen3-Coder-Next` — code-specialised, smoke-tested.
      - `meta-llama/Meta-Llama-3.1-405B-Instruct-FP8` — large
        general instruct, useful as a contrast subject.
      - `meta-llama/Llama-3.3-70B-Instruct` — mid-size instruct.
      - `openai/gpt-oss-120b` — OpenAI open-weight 120B.
      - `mistralai/Mistral-Small-24B-Instruct` — smaller subject,
        stresses spec comprehension.
      - `meta-llama/CodeLlama-13b-Instruct-hf` — older code model,
        useful as a low-capability floor.
    - request shape: standard `messages: [{role: "system", ...},
      {role: "user", ...}, ...]`, with DESIGN.md (107 KB / ~28k
      tokens) in the system message and the task + form mandate in
      the first user message. Subsequent turns alternate
      `assistant` / `user` with the `ail check` error verbatim in
      the user message.
    - response shape: `choices[0].message.content` carries the
      generated program (request "no markdown fences, no
      explanation" in the prompt to skip post-processing).
      `usage.prompt_tokens` + `usage.completion_tokens` give per-
      turn token spend.
    - cost estimate: ~56k tokens per task per (model × form)
      configuration including a few correction roundtrips; a
      full matrix (4 tasks × 3 forms × 4 models × 1 run each) is
      ~2.7M tokens total. Multiply by repetition count if a
      median-of-3 design is chosen.

- [ ] **\[feature\]** Operator routing through `Eq` / `Ord` — `==`,
  `<` etc. resolved via the typeclass instead of the built-in
  primitive comparators. No commitment; gated on bench re-baselining
  to make sure the indirection doesn't tank latency.
  - context: JOURNAL 2026-05-09
  - depends on: Post-22 Prelude — Eq/Ord (shipped 23.5)
- [ ] **\[todo\]** `types` / `ctor_index` overlay shape question —
  decide whether the env's two parallel ctor maps should collapse
  into one overlay, or stay split. Surfaced during the
  env-construction unify audit.
  - context: JOURNAL 2026-05-10 ("Audit close: env-construction unify")
- [ ] **\[todo\]** CLI human-mode diagnostic surface for `WorkspaceLoadError` —
  non-JSON `ail check` routes ct.1 errors via `anyhow`/`thiserror`
  Display (`Error: <message>`) instead of going through
  `workspace_error_to_diagnostic`, so the bracketed `[code]` prefix
  the JSON path carries is missing. Plausibly applies to every CLI
  subcommand that calls `load_workspace(&path)?` directly.
  - context: JOURNAL 2026-05-11 ("Iteration ct.1") — surfaced by ct.1.8 tester.
- [ ] **\[todo\]** Retire dead `KindMismatch` arm — `validate_classdefs`'s
  `walk_kind_mismatch` path is structurally unreachable through
  well-formed schema post-ct.1 (the canonical-form validator catches
  the malformed `Type::Con { name: param }` shape earlier). The
  enum variant + `walk_kind_mismatch` helper stay as dead-but-defensive
  code; a future tidy can delete both.
  - context: JOURNAL 2026-05-11 ("Iteration ct.1").
- [ ] **\[todo\]** Re-key `Registry.type_def_module` to handle
  bare-name-collision-across-modules — `BTreeMap<String, String>` keyed
  by bare type name silently overwrites when two modules each define
  `type Foo`; `normalize_type_for_registry` would then collapse `M.Foo`
  and `N.Foo` to whichever insert won. Acceptable for current corpus
  (distinct bare type names across modules), but the proper fix is to
  key by `(owning_module, bare_name) → defining_module`.
  - context: JOURNAL 2026-05-11 ("Iteration ct.1") — flagged by ct.1.5a quality reviewer.
- [ ] **\[feature\]** 22c — typeclass corpus expansion. User-defined
  classes beyond the prelude four; multi-parameter classes; superclass
  chains; richer instance bodies. Deferred from milestone 22.
  - context: JOURNAL 2026-05-09
- [ ] **\[milestone\]** Module-qualified class names + type-driven
  method dispatch — retire the `MethodNameCollision` workaround
  (`crates/ailang-core/src/workspace.rs:472`) that today keeps bare
  class names viable by enforcing workspace-global method-name
  uniqueness. Replaces name-driven method resolution
  (`ModuleGlobals::class_methods: IndexMap<String, ClassMethodEntry>`,
  `crates/ailang-check/src/lib.rs:983`; consumed by
  `mono::rewrite_class_method_calls`, `crates/ailang-check/src/mono.rs:657`)
  with type-driven dispatch — look up `eq` candidates by argument
  type, pick the unique class instance, fail closed on ambiguity.
  Once shipped, two libraries can each declare `class Eq` with their
  own `eq`. Carries its own DESIGN spec: inference rules, ambiguity
  diagnostics, and the canonical-form extension for class-reference
  fields (`InstanceDef.class`, `SuperclassRef.class`, `Constraint.class`)
  that the canonical-type-names milestone explicitly out-of-scoped.
  - context: `docs/specs/2026-05-10-canonical-type-names.md` "Out of
    scope: Class names" — the workaround is named there so it stays
    visible until this milestone retires it.
- [ ] **\[todo\]** Boehm full retirement — remove the transitional
  Boehm GC path now that RC + uniqueness is the canonical memory
  story.
  - context: JOURNAL pre-22, "Boehm transitional"
- [ ] **\[feature\]** Closure-pair slab / pool — codegen tweak to
  pool the env+code closure pairs instead of one-shot heap allocs.
  Bench-gated.
- [ ] **\[todo\]** `FnDef::synthetic` flag — formalise the
  monomorphiser-emitted FnDefs so downstream passes can tell
  user-authored from synthesised at a glance. Currently inferred from
  symbol naming.
- [ ] **\[todo\]** 21'h iteration — final 21' carry-over (latency
  methodology pass). Numbering kept for continuity with the 21' arc.

- [ ] **\[todo\]** `io/print_float` always-emit-`.0` — surface
  printer always emits `.` or `e/E` so re-lex routes to Float;
  the runtime printer (`printf("%g\n", v)`) doesn't, so `2.0`
  prints as `2` (Int-shaped). Asymmetric. Either switch the
  runtime path to a `.0`-fallback printer (matching surface) or
  document the `%g` contract in DESIGN.md §"Float semantics" so
  the LLM-author knows `io/print_float`'s output is for-humans
  not round-trip.
  - context: `docs/specs/2026-05-10-fieldtest-floats.md` finding F1.
- [ ] **\[todo\]** Rustdoc warning sweep — `cargo doc --no-deps`
  reports 16 pre-existing warnings (15 in `ailang-check`, 1 in
  `ailang-core`: private-item links from public doc, unresolved
  intra-crate links). All predate the design-md-consolidation
  milestone; treat as a one-off sweep.
  - context: JOURNAL 2026-05-10 ("Audit close: design-md-consolidation").
- [ ] **\[todo\]** `design_schema_drift.rs` fidelity widening —
  current test checks anchor *presence* anywhere in DESIGN.md;
  the audit found that `[high]` schema gaps in §"Data model"
  are invisible because anchors live in Decision 11 instead.
  Constrain the test to scan only §"Data model" + ParamMode
  block, or extract JSON-schema blocks into a machine-readable
  file the test consumes.
  - context: JOURNAL 2026-05-10 ("Audit close").
- [ ] **\[todo\]** Split `BadCrossModuleTypeRef` into two diagnostics —
  the current single shape collapses unknown-owner and
  known-owner / unknown-type-in-owner into one message. The two
  cases suggest different fixes (add `(import <owner>)` vs. fix the
  type name). In the known-owner branch, list the owner's available
  type defs as candidates the way `bare-cross-module-type-ref`
  lists candidates from imports.
  - context: fieldtest 2026-05-11 — `examples/ct_3*.ailx` exhibits both branches with identical-shape diagnostics.
- [ ] **\[todo\]** Workspace search beyond entry-module's directory —
  `load_workspace` only finds sibling `.ail.json` files in the same
  directory as the entry module, so any consumer of prelude/std in a
  subdirectory has no way to resolve cross-module imports. Add either
  a `--workspace-root` flag on `ail check` / `ail build` / `ail run`,
  or upward-search from the entry module's directory. Alternatively
  ratify the flat-workspace assumption in DESIGN.md if intentional.
  - context: fieldtest 2026-05-11 — fieldtest fixtures could not be
    placed under `examples/fieldtest/` because of this; predates the
    canonical-type-names milestone but surfaces every time.
- [ ] **\[todo\]** `ail check`/`build`/`run` accept `.ailx` extension —
  today `ail check foo.ailx` produces a misleading JSON-parse error
  (`json: expected value at line 1 column 1`) because the loader only
  accepts `.ail.json`. Either teach the CLI subcommands to auto-parse
  `.ailx` internally, or detect the extension and emit a
  "did you mean `ail parse foo.ailx`?" hint. The LLM-author's natural
  first command should not be the one that produces a misleading
  diagnostic.
  - context: fieldtest 2026-05-11 — orthogonal to canonical-type-names
    but every fieldtest invocation hits this on its first command.
- [ ] **\[todo\]** `check_in_workspace` per-module overlay narrowing —
  `crates/ailang-check/src/lib.rs:1234` still clears+rebuilds
  `env.ctor_index` per-module; ct.3.2 narrowed the analogous mono
  overlay to types-only. The typecheck-side overlay's `env.ctor_index`
  half serves the duplicate-detection diagnostic at workspace-build
  time, not the runtime ctor lookup (which is type-driven post-ct.2.2).
  Narrowing is mechanical but needs a careful read to confirm no
  other consumer survives.
  - context: JOURNAL 2026-05-11 ("Iteration ct.4") — milestone close
    follow-up.

## P3 — Ideas

- [ ] **\[todo\]** `compare_primitives_smoke.ailx` counterpart —
  the canonical happy-path exhibit for canonical-type-names ships
  only as JSON. Per Decision 6, Surface is the LLM-author surface;
  the milestone should have a `.ailx` counterpart. Two paths: author
  `examples/compare_primitives_smoke.ailx` to round-trip into the
  existing JSON, OR retire the JSON fixture in favour of
  `examples/ct_1_ordering_signum.{ailx,ail.json}` as the new canonical
  happy-path exhibit.
  - context: fieldtest 2026-05-11 (spec_gap).
- [ ] **\[todo\]** Codegen `lookup_ctor_in_pattern` type-anchoring —
  `crates/ailang-codegen/src/lib.rs:1790` still walks every module's
  ctor_index by bare ctor name. Plumbing the scrutinee's
  qualified `Type::Con` through pattern lowering would type-anchor
  it symmetric to the ct.2.2 typecheck-side fix. Not load-bearing
  (uniqueness is enforced at typecheck), but cleaner.
  - context: JOURNAL 2026-05-11 ("Iteration ct.3" + ct.4 close).
- [ ] **\[todo\]** `compare_primitives_smoke` IR-shape assertion —
  observe `compare__Int` / `compare__Bool` / `compare__Str` symbols
  in the emitted IR. Blocked on `emit-ir` CLI not running mono;
  resolution paths include extending the CLI to run mono, using
  library APIs directly in e2e.rs, or adding `--dump-ir` to `ail
  build`. The E2E stdout assertion already covers correctness; the
  IR-shape test would catch refactor regressions that rename
  mono symbols.
  - context: JOURNAL 2026-05-11 ("Iteration ct.4") — dropped from
    ct.4.4 when the BLOCKED condition surfaced.

- [ ] **\[idea\]** Latency methodology rework — switch from per-run
  timing to a histogram-based approach so tail-latency regressions
  are visible without re-running. Queued from 21'g.
- [ ] **\[idea\]** Parser-test backfill for 22b.4a-era duplicate-clause
  sites (class × 3, class method × 2, instance × 3, instance method
  × 1). No regression pin today; only worth it if a 22b.4a-era
  diagnostic needs to change.
  - context: JOURNAL 2026-05-10 ("Iteration 22-tidy.7")
- [ ] **\[idea\]** `write_type` `Type::Forall` arm in
  `crates/ailang-prose/src/lib.rs` silently drops constraints in
  inline-type rendering. Dormant — surface forms today only carry
  forall at fn-signature top level. Fix only if a future feature
  carries forall + constraints inline.
  - context: JOURNAL 2026-05-10 ("Iteration 22-tidy.6")
- [ ] **\[idea\]** Richer integration paths between RC and
  uniqueness — deferred from the 21' arc; revisit once the
  uniqueness inference covers more program shapes.