一键在 Manus 中运行任何 Skill

spec-iteration

星标52

分支5

更新时间2026年7月6日 22:10

Continue iterating space_gen against real OpenAPI specs. Each PR closes one gap surfaced by a real spec — could be an `UnimplementedError` stub, ugly generated Dart, a broken edge case, a missing feature, a lint that fires on generated code, anything that prevents the output from reading as cleanly as handwritten Dart. Bar: best-in-class Dart OpenAPI/Swagger generator. Use when asked to "continue space_gen work", "next gap to fix", "what should we improve next", "pick something to work on", or any spec-driven follow-up.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

在 Manus 中运行

来源

eseidel

eseidel/space_gen

打开 GitHub 仓库查看创作者相关仓库

下载

在 Manus 中运行

Current state — the `dart fix` reduction arc (2026-07-06)

This block supersedes the older sections for current status. It documents an in-progress arc; earlier blocks are retained for history.

The arc: stop leaning on `dart fix`

Profiling generation speed (via the Dart VM profiler, and by timing the internal subprocess logs) found that dart fix --apply dominates wall-clock: ~47s of github's ~57s total, vs ~7s for space_gen's own Dart code and ~3s for the two dart format passes. github is the slowest spec; discord ~23s (fix ~14s).

The pivotal finding: dart fix is analysis-bound, not apply-bound. Its cost is repeated whole-package static-analysis passes over 1000+ files, essentially independent of how many fixes it applies. Removing ~4,300 unused_import fixes (#250) shaved only ~2.6s / 5.5% off it. So closing one lint category at the generator gives ~no wall-clock win on its own — the payoff only lands at the very end, when every category is clean enough that the generated package is dart analyze-clean without fix and the _runDart(['fix', ...]) call in lib/src/render/formatting.dart (Formatter.formatAndFix) can be deleted. Even a fully-clean package pays ~7s for dart fix to confirm nothing, so dropping the call reclaims that too. Target: github ~57s → ~10s.

dart fix is load-bearing for the "analyze-clean output" promise: the generated package includes very_good_analysis, so every category fix currently cleans would otherwise fail a consumer's dart analyze (fix-disabled petstore alone has 112 issues; github ~35k). So the categories can't just be ignored — they must be emitted clean.

Each PR in the arc moves one lint category from "laundered by dart fix" to "emitted clean by the generator." Because dart fix was already applying the transform, the final post-fix output stays byte-identical — that's the validation gate (see below). These are output-quality / correct-generation wins, not individually perf wins.

Landed this arc (all merged unless noted)

PR	Lint closed (github raw count)	Where / fix
#249	`unnecessary_this` (6307)	`==` body emitted `this.field == other.field`; drop `this.` except for a field literally named `other` (which the `Object other` param shadows — a real bug caught in A/B: `mapsEqual(other, other.other)`). `equalsExpression` + `_equalsReceiver` guard.
#250	`unused_import` (4337; 97.6% model imports)	`importsForApi`/`importsForModel` imported the whole schema subtree; now gate each import on a `referencedIdentifiers(body)` token scan (mirrors `_referencedModelHelpers`). Closed the pre-existing TODO.
#251	`unnecessary_brace_in_string_interps` (2014) + `unnecessary_string_interpolations` (1617)	`_pathArgLine` wrapped path params in `'${x}'`; String params substitute bare (`replaceAll('{org}', org)`), non-String interpolate braceless (`'$petId'`). One fix, two categories.
#252	`noop_primitive_operations` (2045)	`.toString()` on already-`String` values: `response.body.toString()` → `response.body` (template); query/header param `_stringifyWireValue` helper; enum `toString()` gated on new `valueIsString` (int enums keep `.toString()`).
#253	`prefer_const_constructors_in_immutables` (1185)	`@immutable` model ctors now `const` via `canBeConst` = `!mutableModels && assignmentsLine == null`. Smooshed variants covered (sealed parent already `const`). Cascade: making ctors const raised `prefer_const_constructors` at call sites (40→987, mostly generated tests) — see "what's next."
#254	`always_put_required_named_parameters_first` (946)	Constructor params partitioned required-first (stable); only the ctor reorders (fields/toJson/etc. keep spec order). `entries` (additionalProperties, always required) renders between the two groups.
#255	`unnecessary_parenthesis` (778)	`RenderNumeric.fromJsonExpression` wrapped bare casts: `(json['id'] as int)`. Drop parens only when no `.toDouble()` and no `?? default` (keeping `(x as int?) ?? 0` matches fix).
#256	`avoid_redundant_argument_values` (681)	Three sources: `bodyContentType: BodyContentType.json` (invokeApi default — `defaultBodyContentTypeExpression` constant + `isDefaultBodyContentType` compare; the JSON body still declares json literally); `DateTime.utc(2024, 1, 1)` → `DateTime.utc(2024)` (month/day default to 1); example args equal to a const field default skipped in `RenderObject.exampleValue`.
#257	`prefer_final_locals` (68) — OPEN as of this writing	Shape-dispatch `switch` patterns emitted `Map<String, dynamic> v =>` without `final`; other dispatch sections already had it. One-word template fix.

Also this session: #246 (unintended_html lint on code spans), #247 (regen in-repo gen_tests goldens), #248 (release 1.3.0 — published to pub.dev).

Remaining lint tail (github, fix-disabled, after #256/#257)

~1,143 lints, 91% the const cascade:

count	lint	where
987	`prefer_const_constructors`	985 generated tests, 2 lib/messages
169	`prefer_const_literals_to_create_immutables`	tests
18	`unnecessary_lambdas`
11	`unused_import`	doc-comment `[Type]` refs + meta tail
9	`prefer_int_literals`
8	`prefer_const_declarations`
4	`avoid_escaping_inner_quotes`
3	`unnecessary_this`	(residual, not `==`)
2	`use_raw_strings`

The const-example refactor (the big remaining lever — design settled, not started)

Closing prefer_const_constructors (987) + prefer_const_literals (169) means making the generated round-trip tests' example-instance construction const (Widget(id: 0, name: '...') → const Widget(...)). This is the const-ctor cascade from #253.

It does NOT need DartType (a question that came up — see doc/dart_type.md note added alongside this). Const-example-ness is an expression/structure property, and dart_type.md's North star explicitly keeps that class of thing off DartType (same as fromJson/toJson; #227 prototyped scalar-codec-on-DartType and dropped it — "doesn't generalize"). DartType can't see an object's fields.

Design: a bool exampleValueIsConst(context) sibling to exampleValue on RenderSchema, recursive:

RenderPod: false for uri/uriTemplate (Uri.parse, UriTemplate(...) aren't const), true for bool/dateTime/ email/uuid.
RenderString/RenderNumber/RenderInteger/RenderDate/ RenderEmptyObject: true.
RenderEnum: false today (values.first is a getter call) — or make it true by emitting the first member (Type.member, which is const).
RenderArray/RenderMap: the element/value's.
RenderObject: canBeConst && every property.exampleValueIsConst && entries-const.
RenderOneOf: the chosen variant's; RenderBase64Bytes: false (Uint8List.fromList).

Then prefix const at the outermost const-able point only (a const context makes nested literals const automatically; an inner const would trip unnecessary_const). Watch the round-trip template (schema_round_trip_test.mustache, final instance = {{{ exampleValue }}}) and the object arg builder.

~15 exampleValue impls to touch; intricate but all in the Render* layer. Entirely affects generated tests.

What's next — an open decision the user is weighing

Grind the const refactor — biggest dent (~1,164), most complex, tests-only.
Keep clearing small categories — unnecessary_lambdas (18), unused_import doc-tail (11), prefer_int_literals (9), etc.; clean one-shots (some touch lib), but each closes little and the tail is long (~7 more categories) before dart fix can be dropped.
Bank the wins and stop — ~20k of the original ~35k fix operations already moved to the generator; the rest is a long thin tail that only pays off when the dart fix call is actually deleted.

Dropping dart fix requires all categories clean (the package must be analyze-clean without it). That's the whole tail, including the tiny categories.

Methodology learnings (reuse these)

Measure raw lint counts by disabling fix. Temporarily comment out the _runDart(['fix', '.', '--apply'], ...) line in formatting.dart, regen, dart analyze the output. Restore after. Tally: dart analyze 2>&1 | grep -oE '[a-z_]+$' | sort | uniq -c | sort -rn.
Byte-identical-after-fix is the validation gate. Regen origin/main (old) vs branch (new) into two dirs, normalize the package name (sed 's/package:OLD/package:PKG/' both ways), then diff -rq. Expect 0 — content — differences.
Benign dart_style reflows are expected and OK. When the generator emits a shorter expression than the old dart fix produced, dart_style sometimes lays it out on one line instead of a trailing-comma-wrapped block. A handful of files differ by whitespace + a trailing comma only. Confirm with: strip whitespace and ,)→), compare — must be identical. Seen in #249, #251, #255, #257.
dart fix is analysis-bound. Don't expect a category PR to cut wall-clock. Proven: #250 removed 4,315 fixes, saved ~2.6s.
A/B diffs catch real bugs. #249's blind this. strip broke a field named other (mapsEqual(other, other.other)); only the byte-diff surfaced it (fix had been keeping this.other, which masked the generator bug). Always A/B, don't trust "fix removed it so it's fine."
Commit messages with backticks: use git commit -F <file>, not -m "...". Backticks inside double-quoted -m trigger shell command substitution and silently eat text (`default: false` → empty). Same for gh pr create --body-file.
github regen is ~57s each (fix-dominated) — two regens for an A/B exceed a 2-min Bash timeout; run them as separate commands.
Rotation after each change: github (the bar) + discord + backstage + petstore + spacetraders must stay analyze-clean; petstore (28) + discord (1530) generated round-trip suites are the fast correctness check.

What space_gen is / validation targets

Iterate toward a best-in-class Dart OpenAPI generator: run real specs through it, fix one gap per PR, leave every tracked spec at least as good as before.

Validation targets live in ~/Documents/GitHub/personal/gen_tests/ (parallel to the repo, not checked in; see the external-tracker memory). Regen each after an output-changing PR:

github (api.github.com.json, ~1000 ops) — the primary punching bag and the bar: dart analyze-clean with all generated round-trip tests passing. Exercises oneOf, discriminators, multi-status, big enums, recursion, status-code ranges.
discord (discord.json, ~511 schemas, OpenAPI 3.1) — integer enums, regex-pattern string newtypes; generated suite green (1530 tests), 6 genuine multi-variant-union stubs remain.
backstage (backstage.yaml) — analyze-clean, zero stubs.
petstore / spacetraders / train-travel — smoke + YAML coverage. Plus private_gen_tests/watchcrunch-api.json (a user-reported third-party spec).

A "gap" is any of: an UnimplementedError stub; generated Dart that doesn't compile / analyze-clean / pass its round-trip tests; correct-but-ugly output; a missing OpenAPI feature; a lint that fires on generated code; or a real-spec author's reported snag.

Architecture

The pipeline:

Load → Parse → Resolve → Dispatch → Naming → Render
                            │          │
                            ▼          ▼
                       sidecar lookups (Decisions, Names)

lib/src/dispatch.dart (#160). Pure-structural pass: for each ResolvedSchemaCollection, picks a DispatchDecision (Discriminator / Shape / Hybrid / Predicate / NoDispatch). Variants are referenced as ResolvedSchema, no Dart names. Predicate IR (KeyExists, ArrayElementHasKey, Always) — adding a new predicate is a subtype, not a new mode.
lib/src/naming.dart (#161–#163, #165, #166, #167). The source of truth for every Dart class name and every file basename. Single NameAllocator keyed by snake names; camel is derived via camelFromSnake on read. Each entity submits a preference list (best/shortest first, longest/safest last); the allocator runs a fixpoint that advances colliding entities to their next preference, then numeric-suffix-disambiguates any remaining ties. Order-independent.
- Two-phase resolution: phase 1 claims schemas + synthesized op-response names; phase 2 claims wrapper subclass names (which depend on phase-1 resolved variant names). AssignedNames exposes both snakeFor(pointer) and []/maybeGet (camel, derived).
- Title-derived names (#167): inline oneOf/anyOf variants with a spec title: opt into a multi-tier preference list [<title-snake>, <parent>_one_of_<i>]. When the title is unique, the variant gets the descriptive name (file + class paired). github example: org-ruleset-conditions variants flip to RepositoryNameAndRefName etc.
- Doubling fallback (#165): when a wrapper's <parent><variantTag> would double the parent prefix (because the parser-synthesized inline variant name already starts with the parent), the wrapper falls back to <parent>Variant<i>.
SpecResolver (in render_tree.dart) consults both pre- passes: decideDispatch(source) per oneOf, and a names= field populated from assignNames(spec) at the start of toRenderSpec. Test helpers (renderTestSchema, renderTestOperation, …) populate names with assignNamesForSchema(...) / assignNamesForOperation(...).

Strict-mode invariant (#162): RenderSchema.typeName and RenderOneOf.wrapperTypeName throw StateError if the lookup is missing. Production crashes loudly if any render path forgets to populate names. Tests that build schemas directly without going through SpecResolver must pass assignedName: 'Foo' explicitly.

Resolver stays Dart-blind. It carries snakeName (parser- level) and resolves snake-name collisions with _1 suffixes; it knows nothing about Dart class names.

Smoosh (the structural matrix)

For an inline ResolvedObject variant of a oneOf/anyOf, today's output is smooshed (#168, #169, #170): the variant data class itself extends the sealed parent, inlined in the parent's .dart file, with no separate wrapper subclass and no value: shim.

Pattern matching destructures fields directly:

switch (req) {
  case ProjectsCreateCardRequestOneOf0(:final note):  // direct
    print(note);
}

instead of unwrapping value::

case ProjectsCreateCardRequestVariant0(value: final v):
  print(v.note);  // pre-smoosh

Smoosh applies when:

Structural eligibility (_isStructurallySmooshable): variant is ResolvedObject AND its pointer is a child of the collection's pointer (so it's inline-exclusive by construction). ResolvedAllOf is excluded — it collapses to a synthesized RenderObject at render but isn't a ResolvedObject upstream.
Render support (_dispatchEmitsSmooshed): the parent dispatch's render template emits the smooshed form. Today this is exhaustive — every DispatchDecision kind capable of smoosh does it (predicate-required, discriminator, shape, hybrid Map sub-arms). Array-element predicate is excluded structurally (variants are arrays, not objects); NoDispatch doesn't claim wrappers at all.

Sealed subclasses must live in the same library — Dart's sealed modifier restricts cross-library extension — so smooshed variants are emitted inline in the parent's .dart file, not their own. RenderObject.parentSealedTypeName: String? carries the parent's class name; RenderSchema.isSmooshed is the check used by the dispatch builders and file_renderer.

Render-side dispatch builders

Each dispatch kind has a _buildXMode(...) helper on RenderOneOf that splits variants into dispatch (per-arm caseExpressions), variants (non-smooshed wrapper subclass declarations), and smooshedVariants (full toTemplateContexts rendered inline via the schema_object partial). The four builders follow the same shape:

_buildDiscriminatorMode — switch (json[propertyName]).
_buildShapeMode — switch (json) { T v => ... }.
_buildHybridMode — non-Map shape arms + Map sub-dispatch with optional unguarded fallback.
_buildPredicateMode — if-chain (required-field or array- element).

Each caseExpression is the full Dart expression to return from its switch arm — pre-composed in render code so the template just splices it. Smooshed: <Variant>.fromJson(<arg>). Non-smooshed: <Wrapper>(<Variant>.fromJson(<arg>)).

Open gaps (pick from here)

Discovery: don't pick from this list blind — regen first

Before assuming the gaps below are the right gaps, run a -v regen against the spec you're targeting and tally the warnings:

spec=api.github.com  # or discord, petstore, spacetraders
rm -rf /tmp/${spec//./_}_v && \
  dart run space_gen -v -i ~/Documents/GitHub/personal/gen_tests/${spec}.json \
    -o /tmp/${spec//./_}_v 2>&1 | tee /tmp/${spec//./_}_v.log >/dev/null

grep -E "^(Ignoring|Unused|Skipping)" /tmp/${spec//./_}_v.log \
  | sed -E 's/ in #.*//; s/=[^[:space:]]+ /=… /; s/dynamic/…/; s/\([A-Za-z]+\)/(…)/' \
  | sort | uniq -c | sort -rn

The categories at the top of the tally are dropping spec semantics on the floor. Pick from this or from the list below — don't assume the visible-output gaps are always the highest- leverage ones.

Warnings come from _warnUnused / _warnIgnored in lib/src/parser.dart. To see what fields a parser visit handles vs ignores, search there.

Backlog beyond the `dart fix` arc

Mine a -v regen (Discovery above) for fresh gaps first — the tally is usually more current than any list here. Standing tracks:

Discord's 6 remaining stubs — genuine multi-variant unions (*_actions_inner, *_options_inner, *_components_inner, error_details) with no single-tag discriminator; need real dispatch. The last stub frontier.
Const-modeling arc — #239 (recognize single-legal-value properties by resolved-schema cardinality-1, not the narrow allOf-ref syntax _constTagValue matches; would delete it and unify github's ~299 bare single-value-enum tags with Discord's pinned refs, but changes github output) and #240 (a lone scalar const → a plain constant, not a single-value enum type + file).
#185 — web/dart2js-aware codegen. format: int64 (246 github sites) silently breaks on dart2js for near-2^63 IDs. Needs a real target with such values (Discord doesn't qualify — snowflakes are JSON strings; Stripe is the next candidate to check). Eric's preference: Int64 from package:fixnum, no opt-in flag.
#186 — request/response split for readOnly/writeOnly. Servers can reject requests carrying readOnly fields (81 github sites). Eric's preference: per-class duplication (UserRequest / UserResponse). PR #175 laid parser-side groundwork before being closed; rebuild on it.
Error-status response unions — multi-status dispatch only handles 2xx variation; 4xx/5xx with distinct bodies still fall back to ApiException<Object?>. Related: a 2XX range mixed with explicit 2xx codes synthesizes a source: null RenderOneOf that emits the legacy stub.

Conventions

From the project's CLAUDE.md and saved memory:

Bar: "as good as handwritten" generated Dart. Push back on shortcuts that produce ugly output.
Validation target required: pick the next PR by what real specs need. github is the primary punching bag.
Don't add a new strategy when a new predicate / candidate / knob will do. The dispatch + naming + smoosh series taught this. Adding a 6th dispatch mode or a 5th RenderOneOf codepath is almost never the right move; extend an existing structure with a new variant.
Architectural changes ship as their own PRs first, with byte-identical regen, before the algorithm change that motivated them. Pattern from the dispatch series (#158 → #159), the naming series (#161 → #162 → #163 → #165 → #166), and the smoosh series (the platform piece in each landed dispatch type before the next). Introduce the abstraction without behavior change, lock it in, then swap the algorithm in a focused diff.
Snake-keyed allocator, camel derived. A snake collision IS a camel collision (and vice versa) — one allocator handles both. No parallel allocator. File names and class names share storage.
Sealed subclasses live in the same library (Dart's sealed modifier requires it). Smooshed variants are emitted inline in the parent's .dart file, not their own file. The rendersToSeparateFile predicate gates per-schema file emission; smooshed RenderObjects return false.
Byte-identical regen as verification gate. When a refactor claims no behavior change, prove it: regen ~/.../api.github.com.json before and after, normalize package names + strip the post-format // ignore_for_file: lines_longer_than_80_chars lint header (which is package-name- length sensitive), then diff -r. Zero content diffs = clean lift.
Self-review before pushing. Read the diff cold with two questions: (1) is anything here doing nothing useful (dead gates, speculative branches), (2) are there inline checks duplicated 3+ times that want a helper. Eric will ask if you don't.
Unit tests over gen_tests: each layer (parser, resolver, dispatch, naming, render) has its own unit test file. Add narrow unit tests on the layer changing; only add a gen_tests/ fixture for genuine end-to-end coverage.
Don't use !: bind nullable fields to a local, null-check the local. Same for tests — use is patterns or if (x == null) fail(...) instead of postfix !.
American English. cspell catches British spellings — use the American form. For coined / informal words (smooshing, smooshable, smooshed, dispatchable, regen, unioned, fixpoint, prefs, camelization), just add them to cspell.config.yaml. Rephrase or rename for British spellings.
required on internal types: every constructor parameter on Schema* / Resolved* / Render* classes is required, including fields that default to null/[].
Sealed types over nullable-pair encodings when a field encodes "exactly one of N." Examples: OperationReturn, DispatchDecision, Predicate, _DispatchMode.
Exhaustive switches on sealed types, not if-chains, when the decision is mutually exclusive across known kinds. Adding a new subtype becomes a compile-time prompt to decide. The _dispatchEmitsSmooshed post-self-review change is the recent example.
80-col wrap for code; loose for doc comments.
Don't edit CHANGELOG.md in regular PRs — the release skill generates it from PR bodies.
Separate small PRs ("one gap per PR") over bundled changes — keeps github regen diff crisp per PR.
Every PR branches from origin/main. Never stack a new gap fix on top of an in-progress branch. If you accidentally stack (forgot to branch fresh), the cleanest recovery is to combine related work into one PR and update the title/description rather than try to untangle the git history.
Inline small follow-ups in the same PR — don't file separate issues for them. Human review time is the expensive resource; every separate PR is another full review cycle. When self- review surfaces a small refactor or cleanup, do it now in the same PR. Examples folded inline during the 2026-04-29 review: markUsed API rename + four call-site updates (#182), dropping the dead _collectionSnakeStem('allOf') branch (#182), splitting _pickPropertyArrayElementShape into its own picker (#183), refactoring all four _buildXMode builders to take their dispatch object directly (#183), filing the parser-side required- references-unknown-property warn instead of the render-side invalidJsonExample band-aid (#184). Each was small enough to fit; collectively they kept the PRs at proper checkpoints.
Each PR is a measurable checkpoint. Some changes legitimately span multiple PRs, but each one should land at a meaningful state, not "halfway through a refactor." If review surfaces "we're just a few little bits away from the checkpoint," pull those bits in.
Don't dodge symptoms with band-aids when the real fix is upstream. PR #184 originally filtered invalidJsonExample to silence an asserted-throw test for github's package-version-metadata-docker typo (required: [tags] for a property called tag). Real fix: parser detects the unknown name, drops it from requiredProperties, and warns. Downstream stages then only see real names — no per-stage workarounds needed. Apply this everywhere: a render-time filter compensating for a parser-time miss is a smell.
Verbose-log-mining is the primary discovery channel. The -v regen surfaces gaps that visible-output-only review misses. Run the tally one-shot before assuming the previously-tracked gap list is current — it's typically out of date by 1-2 iterations of the project.

Working rhythm

Branch off origin/main (git checkout -b es/<slug> origin/main). Always — even if the worktree currently sits on the previous PR's branch and that PR hasn't merged yet. Don't stack PRs. Don't git checkout main in this worktree (the primary checkout owns main).
Read the relevant layer (lib/src/parse/, lib/src/resolver.dart, lib/src/dispatch.dart, lib/src/naming.dart, lib/src/render/render_tree.dart, lib/src/render/file_renderer.dart, lib/templates/*.mustache) for context.
Add a unit test exercising the gap.
Make the fix in the right layer.
Render-layer unit tests in test/render/render_schema_test.dart, test/render/render_operation_test.dart, or test/naming_test.dart. Don't snapshot full template output — contains() checks read better and survive whitespace tweaks.
End-to-end smoke test (when warranted): write a small spec at /tmp/<name>_repro/spec.json, run dart run space_gen -i ... -o ..., verify generated code in a Dart test.
Regen against every tracked spec — github first (the bar to maintain), then the others (which surface different patterns). Don't ship a PR that regresses any spec already at "clean"; for specs not yet at "clean" (Discord today), each PR should at least not make things worse:
```
for spec in api.github.com discord petstore spacetraders; do
  out=/tmp/${spec//./_}_out
  rm -rf "$out"
  dart run space_gen \
    -i ~/Documents/GitHub/personal/gen_tests/${spec}.json \
    -o "$out" || echo "FAIL: $spec"
  (cd "$out" 2>/dev/null && dart analyze 2>&1 | tail -1)
done
```
github should be No issues found!. Others surface their own gaps — useful signal for picking what to burn down next.

Add -v (--verbose) to surface detailed warnings that the default run swallows — unsupported feature usage, name collisions resolved by renaming, schema-shape fallbacks, etc.
For "byte-identical" claims: baseline with git stash, regen, restore, regen again, diff. The package-name-in-paths problem only shows up if the two output dirs have different names — keep the path identical (e.g. always /tmp/github_out) to avoid that source of noise.

Count newly-fixed sites if the PR closes a stub-emitting gap:

grep -l "throw UnimplementedError" /tmp/github_out/lib -r | wc -l

Self-review before pushing. Read your diff cold. If something is doing nothing (dead gate, speculative branch), delete it. If a check is duplicated across 3+ files, extract a helper. If a comment claims something that's no longer true ("this is a partial-rollout switch"), update it. Eric will ask "self-review please" and you should already have done it.
Push, open PR. CI will:
- cspell trips on coined words — add them to cspell.config.yaml. Rephrase only for British spellings.
- codecov/patch checks new-line coverage.
- codecov/project may fail with no-drop default — mostly an accounting artifact when adding lines that shift uncovered code's denominator. Address only if patch coverage is genuinely low.
- dart format must pass.

Don't

Don't snapshot-match full template output in tests — fragile.
Don't pivot to gen_tests fixtures for unit-level coverage.
Don't use postfix ! to silence Dart's flow analysis.
Don't bundle multiple gap fixes in one PR — separate PRs.
Don't merge without dart analyze clean on the github regen.
Don't touch the merge-conflict-prone CHANGELOG.md in feature PRs.
Don't conflate operation-level constructs with schema-level ones. (#148 made this lesson explicit: multi-status responses look like a oneOf but aren't.)
Don't add a new dispatch mode / template branch / _DispatchMode subtype when a new predicate kind or new picker that builds the existing mode will do.
Don't compute names in render. Naming is the source of truth; render reads. If a code path needs a name, it goes through AssignedNames (or a strict-mode _requireAssignedName() throws telling you why).
Don't ship infrastructure with no production user. The NameAllocator multi-tier path was almost shipped untested in production — the mistake (a multi-tier list that was worse than single-tier in conflict) was caught in self-review. Today every dispatch builder is the production user of isSmooshed; multi-tier is the production user via title-based renaming. Hold the line.
Don't add a smooshed-variant inline emission case for "future- proofing" without a concrete trigger. Removed one such speculative branch in #170 self-review — it would silently mis- render if ever reached. Better to crash on missing-case than silently produce wrong output.
Don't ship "swap on better arrival" mutation in immutable caches. Tried this for #180's name selection: kept the cache, swapped the entry when a more-canonical pointer arrived. Doesn't work — the FIRST caller already holds the loser instance, so callers end up with two distinct instances and dedup silently fails. Either fix at parse-order time or accept first-arrival semantics (#180 ended up with a narrower scope rule instead).
Don't dispatch fan-out subagents without telling them to stay in their assigned worktree. CLAUDE.md absolute paths can mislead subagents into editing the parent worktree. One agent during the fan-out caught itself before pushing — only because nothing else had uncommitted changes. Every fan-out prompt should explicitly say: "Edit and commit ONLY in the directory you start in (pwd to confirm). Do NOT touch the parent worktree." See memory file feedback_subagent_worktree_isolation.md.

Helpful one-shots

# Burn-down rotation: regen each tracked spec, report analyze
# status for each. Use this to spot regressions or pick the
# next target.
for spec in api.github.com discord petstore spacetraders; do
  out=/tmp/${spec//./_}_out
  rm -rf "$out"
  dart run space_gen \
    -i ~/Documents/GitHub/personal/gen_tests/${spec}.json \
    -o "$out" 2>&1 | tail -1
  echo "$spec: $((cd "$out" 2>/dev/null && dart analyze 2>&1 | tail -1) \
    || echo 'GEN FAILED')"
done

# Discovery: tally what `-v` regen says we're dropping. Run this
# first when picking the next gap — categories at the top of the
# tally are usually higher-leverage than visible-output ugliness.
# Substitute `discord_v.log` etc. for other specs.
grep -E "^(Ignoring|Unused|Skipping)" /tmp/github_v.log \
  | sed -E 's/ in #.*//; s/=[^[:space:]]+ /=… /; s/dynamic/…/; s/\([A-Za-z]+\)/(…)/' \
  | sort | uniq -c | sort -rn

# Drill into one category — see the actual sites
grep "Unused: oneOf" /tmp/github_v.log | head -20

# Are we still emitting stubs? (works for any spec's regen dir)
grep -l "throw UnimplementedError" /tmp/github_out/lib -r | wc -l

# Sealed dispatch sites in models/
grep -l "^sealed class " /tmp/github_out/lib/models/ -r | wc -l

# Multi-status dispatch sites in api/
grep -c "switch (response.statusCode)" /tmp/github_out/lib/api/*.dart 2>/dev/null \
  | awk -F: '{ sum +=  } END { print sum }'

# Smoosh sites — variants that extend a sealed parent directly
grep -E "^final class \w+ extends \w+" /tmp/github_out/lib/messages/*.dart | wc -l

# Wrappers still using the post-#165 `Variant<i>` doubling fallback
grep -rE "class \w+Variant[0-9]+ extends \w+" /tmp/github_out/lib | wc -l

# Generated test suite — sanity check after a render-layer change
(cd /tmp/github_out && dart pub get && dart test 2>&1 | tail -3)

# Coverage for a specific file (after running ./coverage.sh)
awk '/^SF:.*\/parser.dart$/,/^end_of_record$/' coverage/lcov.info \
  | grep "DA:.*,0$"

# Spec-side analysis (Python is fine; the spec is JSON):
python3 -c "
import json
spec = json.load(open('$HOME/Documents/GitHub/personal/gen_tests/api.github.com.json'))
def walk(n, p):
    if isinstance(n, dict):
        if 'oneOf' in n:
            print(p, [v.get('type', 'ref') if isinstance(v, dict) else '?' for v in n['oneOf']])
        for k, v in n.items(): walk(v, p + '/' + str(k))
    elif isinstance(n, list):
        for i, x in enumerate(n): walk(x, p + '/' + str(i))
walk(spec, '')
" | head -20

Where to look

lib/src/parser.dart, lib/src/parse/ — spec → Spec parse tree.
lib/src/resolver.dart — Spec → ResolvedSpec, $ref + snake-name collision resolution. Stays Dart-blind.
lib/src/dispatch.dart — decideDispatch(ResolvedSchemaCollection) per oneOf/anyOf, sealed DispatchDecision family + Predicate IR.
lib/src/naming.dart — assignNames(ResolvedSpec) walks the resolved tree and dispatch decisions to assign every Dart class name. NameAllocator is the multi-tier preference allocator. AssignedNames exposes both snakeFor(pointer) and []/ maybeGet (camel, derived). Smoosh metadata in smooshedParentSnakeByPointer / parentSealedTypeFor.
lib/src/render/render_tree.dart — ResolvedSpec → RenderSpec. RenderOneOf._buildXMode per dispatch kind. RenderObject. parentSealedTypeName carries the smoosh marker; RenderSchema.isSmooshed is the predicate used by builders + file_renderer.
lib/src/render/file_renderer.dart — rendersToSeparateFile gates per-schema file emission (smooshed variants return false). Per-schema and per-API import collection.
lib/templates/ — Mustache templates. schema_object.mustache emits final class X extends Y when parentSealedTypeName is set. schema_one_of.mustache has four dispatch sections; each reads caseExpression per arm and includes the schema_object partial for smooshedVariants.
test/render/render_schema_test.dart, test/render/render_operation_test.dart, test/naming_test.dart — unit tests per layer.
~/Documents/GitHub/personal/gen_tests/api.github.com.json — primary validation target.
ARCHITECTURE.md — pipeline diagram, per-phase notes, planned next steps. Update when phases shift.
Issue #144 on github — running list of oneOf-specific follow-ups. Other gaps surface fresh from regen runs.

name

spec-iteration

description

Current state — the `dart fix` reduction arc (2026-07-06)

This block supersedes the older sections for current status. It documents an in-progress arc; earlier blocks are retained for history.

The arc: stop leaning on `dart fix`

Landed this arc (all merged unless noted)

PR	Lint closed (github raw count)	Where / fix
#249	`unnecessary_this` (6307)	`==` body emitted `this.field == other.field`; drop `this.` except for a field literally named `other` (which the `Object other` param shadows — a real bug caught in A/B: `mapsEqual(other, other.other)`). `equalsExpression` + `_equalsReceiver` guard.
#250	`unused_import` (4337; 97.6% model imports)	`importsForApi`/`importsForModel` imported the whole schema subtree; now gate each import on a `referencedIdentifiers(body)` token scan (mirrors `_referencedModelHelpers`). Closed the pre-existing TODO.
#251	`unnecessary_brace_in_string_interps` (2014) + `unnecessary_string_interpolations` (1617)	`_pathArgLine` wrapped path params in `'${x}'`; String params substitute bare (`replaceAll('{org}', org)`), non-String interpolate braceless (`'$petId'`). One fix, two categories.
#252	`noop_primitive_operations` (2045)	`.toString()` on already-`String` values: `response.body.toString()` → `response.body` (template); query/header param `_stringifyWireValue` helper; enum `toString()` gated on new `valueIsString` (int enums keep `.toString()`).
#253	`prefer_const_constructors_in_immutables` (1185)	`@immutable` model ctors now `const` via `canBeConst` = `!mutableModels && assignmentsLine == null`. Smooshed variants covered (sealed parent already `const`). Cascade: making ctors const raised `prefer_const_constructors` at call sites (40→987, mostly generated tests) — see "what's next."
#254	`always_put_required_named_parameters_first` (946)	Constructor params partitioned required-first (stable); only the ctor reorders (fields/toJson/etc. keep spec order). `entries` (additionalProperties, always required) renders between the two groups.
#255	`unnecessary_parenthesis` (778)	`RenderNumeric.fromJsonExpression` wrapped bare casts: `(json['id'] as int)`. Drop parens only when no `.toDouble()` and no `?? default` (keeping `(x as int?) ?? 0` matches fix).
#256	`avoid_redundant_argument_values` (681)	Three sources: `bodyContentType: BodyContentType.json` (invokeApi default — `defaultBodyContentTypeExpression` constant + `isDefaultBodyContentType` compare; the JSON body still declares json literally); `DateTime.utc(2024, 1, 1)` → `DateTime.utc(2024)` (month/day default to 1); example args equal to a const field default skipped in `RenderObject.exampleValue`.
#257	`prefer_final_locals` (68) — OPEN as of this writing	Shape-dispatch `switch` patterns emitted `Map<String, dynamic> v =>` without `final`; other dispatch sections already had it. One-word template fix.

Also this session: #246 (unintended_html lint on code spans), #247 (regen in-repo gen_tests goldens), #248 (release 1.3.0 — published to pub.dev).

Remaining lint tail (github, fix-disabled, after #256/#257)

~1,143 lints, 91% the const cascade:

count	lint	where
987	`prefer_const_constructors`	985 generated tests, 2 lib/messages
169	`prefer_const_literals_to_create_immutables`	tests
18	`unnecessary_lambdas`
11	`unused_import`	doc-comment `[Type]` refs + meta tail
9	`prefer_int_literals`
8	`prefer_const_declarations`
4	`avoid_escaping_inner_quotes`
3	`unnecessary_this`	(residual, not `==`)
2	`use_raw_strings`

The const-example refactor (the big remaining lever — design settled, not started)

Design: a bool exampleValueIsConst(context) sibling to exampleValue on RenderSchema, recursive:

RenderPod: false for uri/uriTemplate (Uri.parse, UriTemplate(...) aren't const), true for bool/dateTime/ email/uuid.
RenderString/RenderNumber/RenderInteger/RenderDate/ RenderEmptyObject: true.
RenderEnum: false today (values.first is a getter call) — or make it true by emitting the first member (Type.member, which is const).
RenderArray/RenderMap: the element/value's.
RenderObject: canBeConst && every property.exampleValueIsConst && entries-const.
RenderOneOf: the chosen variant's; RenderBase64Bytes: false (Uint8List.fromList).

~15 exampleValue impls to touch; intricate but all in the Render* layer. Entirely affects generated tests.

What's next — an open decision the user is weighing

Grind the const refactor — biggest dent (~1,164), most complex, tests-only.
Keep clearing small categories — unnecessary_lambdas (18), unused_import doc-tail (11), prefer_int_literals (9), etc.; clean one-shots (some touch lib), but each closes little and the tail is long (~7 more categories) before dart fix can be dropped.
Bank the wins and stop — ~20k of the original ~35k fix operations already moved to the generator; the rest is a long thin tail that only pays off when the dart fix call is actually deleted.

Dropping dart fix requires all categories clean (the package must be analyze-clean without it). That's the whole tail, including the tiny categories.

Methodology learnings (reuse these)

Measure raw lint counts by disabling fix. Temporarily comment out the _runDart(['fix', '.', '--apply'], ...) line in formatting.dart, regen, dart analyze the output. Restore after. Tally: dart analyze 2>&1 | grep -oE '[a-z_]+$' | sort | uniq -c | sort -rn.
Byte-identical-after-fix is the validation gate. Regen origin/main (old) vs branch (new) into two dirs, normalize the package name (sed 's/package:OLD/package:PKG/' both ways), then diff -rq. Expect 0 — content — differences.
Benign dart_style reflows are expected and OK. When the generator emits a shorter expression than the old dart fix produced, dart_style sometimes lays it out on one line instead of a trailing-comma-wrapped block. A handful of files differ by whitespace + a trailing comma only. Confirm with: strip whitespace and ,)→), compare — must be identical. Seen in #249, #251, #255, #257.
dart fix is analysis-bound. Don't expect a category PR to cut wall-clock. Proven: #250 removed 4,315 fixes, saved ~2.6s.
A/B diffs catch real bugs. #249's blind this. strip broke a field named other (mapsEqual(other, other.other)); only the byte-diff surfaced it (fix had been keeping this.other, which masked the generator bug). Always A/B, don't trust "fix removed it so it's fine."
Commit messages with backticks: use git commit -F <file>, not -m "...". Backticks inside double-quoted -m trigger shell command substitution and silently eat text (`default: false` → empty). Same for gh pr create --body-file.
github regen is ~57s each (fix-dominated) — two regens for an A/B exceed a 2-min Bash timeout; run them as separate commands.
Rotation after each change: github (the bar) + discord + backstage + petstore + spacetraders must stay analyze-clean; petstore (28) + discord (1530) generated round-trip suites are the fast correctness check.

What space_gen is / validation targets

Iterate toward a best-in-class Dart OpenAPI generator: run real specs through it, fix one gap per PR, leave every tracked spec at least as good as before.

Validation targets live in ~/Documents/GitHub/personal/gen_tests/ (parallel to the repo, not checked in; see the external-tracker memory). Regen each after an output-changing PR:

github (api.github.com.json, ~1000 ops) — the primary punching bag and the bar: dart analyze-clean with all generated round-trip tests passing. Exercises oneOf, discriminators, multi-status, big enums, recursion, status-code ranges.
discord (discord.json, ~511 schemas, OpenAPI 3.1) — integer enums, regex-pattern string newtypes; generated suite green (1530 tests), 6 genuine multi-variant-union stubs remain.
backstage (backstage.yaml) — analyze-clean, zero stubs.
petstore / spacetraders / train-travel — smoke + YAML coverage. Plus private_gen_tests/watchcrunch-api.json (a user-reported third-party spec).

Architecture

The pipeline:

Load → Parse → Resolve → Dispatch → Naming → Render
                            │          │
                            ▼          ▼
                       sidecar lookups (Decisions, Names)

lib/src/dispatch.dart (#160). Pure-structural pass: for each ResolvedSchemaCollection, picks a DispatchDecision (Discriminator / Shape / Hybrid / Predicate / NoDispatch). Variants are referenced as ResolvedSchema, no Dart names. Predicate IR (KeyExists, ArrayElementHasKey, Always) — adding a new predicate is a subtype, not a new mode.
lib/src/naming.dart (#161–#163, #165, #166, #167). The source of truth for every Dart class name and every file basename. Single NameAllocator keyed by snake names; camel is derived via camelFromSnake on read. Each entity submits a preference list (best/shortest first, longest/safest last); the allocator runs a fixpoint that advances colliding entities to their next preference, then numeric-suffix-disambiguates any remaining ties. Order-independent.
- Two-phase resolution: phase 1 claims schemas + synthesized op-response names; phase 2 claims wrapper subclass names (which depend on phase-1 resolved variant names). AssignedNames exposes both snakeFor(pointer) and []/maybeGet (camel, derived).
- Title-derived names (#167): inline oneOf/anyOf variants with a spec title: opt into a multi-tier preference list [<title-snake>, <parent>_one_of_<i>]. When the title is unique, the variant gets the descriptive name (file + class paired). github example: org-ruleset-conditions variants flip to RepositoryNameAndRefName etc.
- Doubling fallback (#165): when a wrapper's <parent><variantTag> would double the parent prefix (because the parser-synthesized inline variant name already starts with the parent), the wrapper falls back to <parent>Variant<i>.
SpecResolver (in render_tree.dart) consults both pre- passes: decideDispatch(source) per oneOf, and a names= field populated from assignNames(spec) at the start of toRenderSpec. Test helpers (renderTestSchema, renderTestOperation, …) populate names with assignNamesForSchema(...) / assignNamesForOperation(...).

Resolver stays Dart-blind. It carries snakeName (parser- level) and resolves snake-name collisions with _1 suffixes; it knows nothing about Dart class names.

Smoosh (the structural matrix)

Pattern matching destructures fields directly:

switch (req) {
  case ProjectsCreateCardRequestOneOf0(:final note):  // direct
    print(note);
}

instead of unwrapping value::

case ProjectsCreateCardRequestVariant0(value: final v):
  print(v.note);  // pre-smoosh

Smoosh applies when:

Structural eligibility (_isStructurallySmooshable): variant is ResolvedObject AND its pointer is a child of the collection's pointer (so it's inline-exclusive by construction). ResolvedAllOf is excluded — it collapses to a synthesized RenderObject at render but isn't a ResolvedObject upstream.
Render support (_dispatchEmitsSmooshed): the parent dispatch's render template emits the smooshed form. Today this is exhaustive — every DispatchDecision kind capable of smoosh does it (predicate-required, discriminator, shape, hybrid Map sub-arms). Array-element predicate is excluded structurally (variants are arrays, not objects); NoDispatch doesn't claim wrappers at all.

Render-side dispatch builders

_buildDiscriminatorMode — switch (json[propertyName]).
_buildShapeMode — switch (json) { T v => ... }.
_buildHybridMode — non-Map shape arms + Map sub-dispatch with optional unguarded fallback.
_buildPredicateMode — if-chain (required-field or array- element).

Open gaps (pick from here)

Discovery: don't pick from this list blind — regen first

Before assuming the gaps below are the right gaps, run a -v regen against the spec you're targeting and tally the warnings:

spec=api.github.com  # or discord, petstore, spacetraders
rm -rf /tmp/${spec//./_}_v && \
  dart run space_gen -v -i ~/Documents/GitHub/personal/gen_tests/${spec}.json \
    -o /tmp/${spec//./_}_v 2>&1 | tee /tmp/${spec//./_}_v.log >/dev/null

grep -E "^(Ignoring|Unused|Skipping)" /tmp/${spec//./_}_v.log \
  | sed -E 's/ in #.*//; s/=[^[:space:]]+ /=… /; s/dynamic/…/; s/\([A-Za-z]+\)/(…)/' \
  | sort | uniq -c | sort -rn

The categories at the top of the tally are dropping spec semantics on the floor. Pick from this or from the list below — don't assume the visible-output gaps are always the highest- leverage ones.

Warnings come from _warnUnused / _warnIgnored in lib/src/parser.dart. To see what fields a parser visit handles vs ignores, search there.

Backlog beyond the `dart fix` arc

Mine a -v regen (Discovery above) for fresh gaps first — the tally is usually more current than any list here. Standing tracks:

Discord's 6 remaining stubs — genuine multi-variant unions (*_actions_inner, *_options_inner, *_components_inner, error_details) with no single-tag discriminator; need real dispatch. The last stub frontier.
Const-modeling arc — #239 (recognize single-legal-value properties by resolved-schema cardinality-1, not the narrow allOf-ref syntax _constTagValue matches; would delete it and unify github's ~299 bare single-value-enum tags with Discord's pinned refs, but changes github output) and #240 (a lone scalar const → a plain constant, not a single-value enum type + file).
#185 — web/dart2js-aware codegen. format: int64 (246 github sites) silently breaks on dart2js for near-2^63 IDs. Needs a real target with such values (Discord doesn't qualify — snowflakes are JSON strings; Stripe is the next candidate to check). Eric's preference: Int64 from package:fixnum, no opt-in flag.
#186 — request/response split for readOnly/writeOnly. Servers can reject requests carrying readOnly fields (81 github sites). Eric's preference: per-class duplication (UserRequest / UserResponse). PR #175 laid parser-side groundwork before being closed; rebuild on it.
Error-status response unions — multi-status dispatch only handles 2xx variation; 4xx/5xx with distinct bodies still fall back to ApiException<Object?>. Related: a 2XX range mixed with explicit 2xx codes synthesizes a source: null RenderOneOf that emits the legacy stub.

Conventions

From the project's CLAUDE.md and saved memory:

Bar: "as good as handwritten" generated Dart. Push back on shortcuts that produce ugly output.
Validation target required: pick the next PR by what real specs need. github is the primary punching bag.
Don't add a new strategy when a new predicate / candidate / knob will do. The dispatch + naming + smoosh series taught this. Adding a 6th dispatch mode or a 5th RenderOneOf codepath is almost never the right move; extend an existing structure with a new variant.
Architectural changes ship as their own PRs first, with byte-identical regen, before the algorithm change that motivated them. Pattern from the dispatch series (#158 → #159), the naming series (#161 → #162 → #163 → #165 → #166), and the smoosh series (the platform piece in each landed dispatch type before the next). Introduce the abstraction without behavior change, lock it in, then swap the algorithm in a focused diff.
Snake-keyed allocator, camel derived. A snake collision IS a camel collision (and vice versa) — one allocator handles both. No parallel allocator. File names and class names share storage.
Sealed subclasses live in the same library (Dart's sealed modifier requires it). Smooshed variants are emitted inline in the parent's .dart file, not their own file. The rendersToSeparateFile predicate gates per-schema file emission; smooshed RenderObjects return false.
Byte-identical regen as verification gate. When a refactor claims no behavior change, prove it: regen ~/.../api.github.com.json before and after, normalize package names + strip the post-format // ignore_for_file: lines_longer_than_80_chars lint header (which is package-name- length sensitive), then diff -r. Zero content diffs = clean lift.
Self-review before pushing. Read the diff cold with two questions: (1) is anything here doing nothing useful (dead gates, speculative branches), (2) are there inline checks duplicated 3+ times that want a helper. Eric will ask if you don't.
Unit tests over gen_tests: each layer (parser, resolver, dispatch, naming, render) has its own unit test file. Add narrow unit tests on the layer changing; only add a gen_tests/ fixture for genuine end-to-end coverage.
Don't use !: bind nullable fields to a local, null-check the local. Same for tests — use is patterns or if (x == null) fail(...) instead of postfix !.
American English. cspell catches British spellings — use the American form. For coined / informal words (smooshing, smooshable, smooshed, dispatchable, regen, unioned, fixpoint, prefs, camelization), just add them to cspell.config.yaml. Rephrase or rename for British spellings.
required on internal types: every constructor parameter on Schema* / Resolved* / Render* classes is required, including fields that default to null/[].
Sealed types over nullable-pair encodings when a field encodes "exactly one of N." Examples: OperationReturn, DispatchDecision, Predicate, _DispatchMode.
Exhaustive switches on sealed types, not if-chains, when the decision is mutually exclusive across known kinds. Adding a new subtype becomes a compile-time prompt to decide. The _dispatchEmitsSmooshed post-self-review change is the recent example.
80-col wrap for code; loose for doc comments.
Don't edit CHANGELOG.md in regular PRs — the release skill generates it from PR bodies.
Separate small PRs ("one gap per PR") over bundled changes — keeps github regen diff crisp per PR.
Every PR branches from origin/main. Never stack a new gap fix on top of an in-progress branch. If you accidentally stack (forgot to branch fresh), the cleanest recovery is to combine related work into one PR and update the title/description rather than try to untangle the git history.
Inline small follow-ups in the same PR — don't file separate issues for them. Human review time is the expensive resource; every separate PR is another full review cycle. When self- review surfaces a small refactor or cleanup, do it now in the same PR. Examples folded inline during the 2026-04-29 review: markUsed API rename + four call-site updates (#182), dropping the dead _collectionSnakeStem('allOf') branch (#182), splitting _pickPropertyArrayElementShape into its own picker (#183), refactoring all four _buildXMode builders to take their dispatch object directly (#183), filing the parser-side required- references-unknown-property warn instead of the render-side invalidJsonExample band-aid (#184). Each was small enough to fit; collectively they kept the PRs at proper checkpoints.
Each PR is a measurable checkpoint. Some changes legitimately span multiple PRs, but each one should land at a meaningful state, not "halfway through a refactor." If review surfaces "we're just a few little bits away from the checkpoint," pull those bits in.
Don't dodge symptoms with band-aids when the real fix is upstream. PR #184 originally filtered invalidJsonExample to silence an asserted-throw test for github's package-version-metadata-docker typo (required: [tags] for a property called tag). Real fix: parser detects the unknown name, drops it from requiredProperties, and warns. Downstream stages then only see real names — no per-stage workarounds needed. Apply this everywhere: a render-time filter compensating for a parser-time miss is a smell.
Verbose-log-mining is the primary discovery channel. The -v regen surfaces gaps that visible-output-only review misses. Run the tally one-shot before assuming the previously-tracked gap list is current — it's typically out of date by 1-2 iterations of the project.

Working rhythm

Branch off origin/main (git checkout -b es/<slug> origin/main). Always — even if the worktree currently sits on the previous PR's branch and that PR hasn't merged yet. Don't stack PRs. Don't git checkout main in this worktree (the primary checkout owns main).
Read the relevant layer (lib/src/parse/, lib/src/resolver.dart, lib/src/dispatch.dart, lib/src/naming.dart, lib/src/render/render_tree.dart, lib/src/render/file_renderer.dart, lib/templates/*.mustache) for context.
Add a unit test exercising the gap.
Make the fix in the right layer.
Render-layer unit tests in test/render/render_schema_test.dart, test/render/render_operation_test.dart, or test/naming_test.dart. Don't snapshot full template output — contains() checks read better and survive whitespace tweaks.
End-to-end smoke test (when warranted): write a small spec at /tmp/<name>_repro/spec.json, run dart run space_gen -i ... -o ..., verify generated code in a Dart test.
Regen against every tracked spec — github first (the bar to maintain), then the others (which surface different patterns). Don't ship a PR that regresses any spec already at "clean"; for specs not yet at "clean" (Discord today), each PR should at least not make things worse:
```
for spec in api.github.com discord petstore spacetraders; do
  out=/tmp/${spec//./_}_out
  rm -rf "$out"
  dart run space_gen \
    -i ~/Documents/GitHub/personal/gen_tests/${spec}.json \
    -o "$out" || echo "FAIL: $spec"
  (cd "$out" 2>/dev/null && dart analyze 2>&1 | tail -1)
done
```
github should be No issues found!. Others surface their own gaps — useful signal for picking what to burn down next.

Add -v (--verbose) to surface detailed warnings that the default run swallows — unsupported feature usage, name collisions resolved by renaming, schema-shape fallbacks, etc.
For "byte-identical" claims: baseline with git stash, regen, restore, regen again, diff. The package-name-in-paths problem only shows up if the two output dirs have different names — keep the path identical (e.g. always /tmp/github_out) to avoid that source of noise.

Count newly-fixed sites if the PR closes a stub-emitting gap:

grep -l "throw UnimplementedError" /tmp/github_out/lib -r | wc -l

Self-review before pushing. Read your diff cold. If something is doing nothing (dead gate, speculative branch), delete it. If a check is duplicated across 3+ files, extract a helper. If a comment claims something that's no longer true ("this is a partial-rollout switch"), update it. Eric will ask "self-review please" and you should already have done it.
Push, open PR. CI will:
- cspell trips on coined words — add them to cspell.config.yaml. Rephrase only for British spellings.
- codecov/patch checks new-line coverage.
- codecov/project may fail with no-drop default — mostly an accounting artifact when adding lines that shift uncovered code's denominator. Address only if patch coverage is genuinely low.
- dart format must pass.

Don't

Don't snapshot-match full template output in tests — fragile.
Don't pivot to gen_tests fixtures for unit-level coverage.
Don't use postfix ! to silence Dart's flow analysis.
Don't bundle multiple gap fixes in one PR — separate PRs.
Don't merge without dart analyze clean on the github regen.
Don't touch the merge-conflict-prone CHANGELOG.md in feature PRs.
Don't conflate operation-level constructs with schema-level ones. (#148 made this lesson explicit: multi-status responses look like a oneOf but aren't.)
Don't add a new dispatch mode / template branch / _DispatchMode subtype when a new predicate kind or new picker that builds the existing mode will do.
Don't compute names in render. Naming is the source of truth; render reads. If a code path needs a name, it goes through AssignedNames (or a strict-mode _requireAssignedName() throws telling you why).
Don't ship infrastructure with no production user. The NameAllocator multi-tier path was almost shipped untested in production — the mistake (a multi-tier list that was worse than single-tier in conflict) was caught in self-review. Today every dispatch builder is the production user of isSmooshed; multi-tier is the production user via title-based renaming. Hold the line.
Don't add a smooshed-variant inline emission case for "future- proofing" without a concrete trigger. Removed one such speculative branch in #170 self-review — it would silently mis- render if ever reached. Better to crash on missing-case than silently produce wrong output.
Don't ship "swap on better arrival" mutation in immutable caches. Tried this for #180's name selection: kept the cache, swapped the entry when a more-canonical pointer arrived. Doesn't work — the FIRST caller already holds the loser instance, so callers end up with two distinct instances and dedup silently fails. Either fix at parse-order time or accept first-arrival semantics (#180 ended up with a narrower scope rule instead).
Don't dispatch fan-out subagents without telling them to stay in their assigned worktree. CLAUDE.md absolute paths can mislead subagents into editing the parent worktree. One agent during the fan-out caught itself before pushing — only because nothing else had uncommitted changes. Every fan-out prompt should explicitly say: "Edit and commit ONLY in the directory you start in (pwd to confirm). Do NOT touch the parent worktree." See memory file feedback_subagent_worktree_isolation.md.

Helpful one-shots

# Burn-down rotation: regen each tracked spec, report analyze
# status for each. Use this to spot regressions or pick the
# next target.
for spec in api.github.com discord petstore spacetraders; do
  out=/tmp/${spec//./_}_out
  rm -rf "$out"
  dart run space_gen \
    -i ~/Documents/GitHub/personal/gen_tests/${spec}.json \
    -o "$out" 2>&1 | tail -1
  echo "$spec: $((cd "$out" 2>/dev/null && dart analyze 2>&1 | tail -1) \
    || echo 'GEN FAILED')"
done

# Discovery: tally what `-v` regen says we're dropping. Run this
# first when picking the next gap — categories at the top of the
# tally are usually higher-leverage than visible-output ugliness.
# Substitute `discord_v.log` etc. for other specs.
grep -E "^(Ignoring|Unused|Skipping)" /tmp/github_v.log \
  | sed -E 's/ in #.*//; s/=[^[:space:]]+ /=… /; s/dynamic/…/; s/\([A-Za-z]+\)/(…)/' \
  | sort | uniq -c | sort -rn

# Drill into one category — see the actual sites
grep "Unused: oneOf" /tmp/github_v.log | head -20

# Are we still emitting stubs? (works for any spec's regen dir)
grep -l "throw UnimplementedError" /tmp/github_out/lib -r | wc -l

# Sealed dispatch sites in models/
grep -l "^sealed class " /tmp/github_out/lib/models/ -r | wc -l

# Multi-status dispatch sites in api/
grep -c "switch (response.statusCode)" /tmp/github_out/lib/api/*.dart 2>/dev/null \
  | awk -F: '{ sum +=  } END { print sum }'

# Smoosh sites — variants that extend a sealed parent directly
grep -E "^final class \w+ extends \w+" /tmp/github_out/lib/messages/*.dart | wc -l

# Wrappers still using the post-#165 `Variant<i>` doubling fallback
grep -rE "class \w+Variant[0-9]+ extends \w+" /tmp/github_out/lib | wc -l

# Generated test suite — sanity check after a render-layer change
(cd /tmp/github_out && dart pub get && dart test 2>&1 | tail -3)

# Coverage for a specific file (after running ./coverage.sh)
awk '/^SF:.*\/parser.dart$/,/^end_of_record$/' coverage/lcov.info \
  | grep "DA:.*,0$"

# Spec-side analysis (Python is fine; the spec is JSON):
python3 -c "
import json
spec = json.load(open('$HOME/Documents/GitHub/personal/gen_tests/api.github.com.json'))
def walk(n, p):
    if isinstance(n, dict):
        if 'oneOf' in n:
            print(p, [v.get('type', 'ref') if isinstance(v, dict) else '?' for v in n['oneOf']])
        for k, v in n.items(): walk(v, p + '/' + str(k))
    elif isinstance(n, list):
        for i, x in enumerate(n): walk(x, p + '/' + str(i))
walk(spec, '')
" | head -20

Where to look

lib/src/parser.dart, lib/src/parse/ — spec → Spec parse tree.
lib/src/resolver.dart — Spec → ResolvedSpec, $ref + snake-name collision resolution. Stays Dart-blind.
lib/src/dispatch.dart — decideDispatch(ResolvedSchemaCollection) per oneOf/anyOf, sealed DispatchDecision family + Predicate IR.
lib/src/naming.dart — assignNames(ResolvedSpec) walks the resolved tree and dispatch decisions to assign every Dart class name. NameAllocator is the multi-tier preference allocator. AssignedNames exposes both snakeFor(pointer) and []/ maybeGet (camel, derived). Smoosh metadata in smooshedParentSnakeByPointer / parentSealedTypeFor.
lib/src/render/render_tree.dart — ResolvedSpec → RenderSpec. RenderOneOf._buildXMode per dispatch kind. RenderObject. parentSealedTypeName carries the smoosh marker; RenderSchema.isSmooshed is the predicate used by builders + file_renderer.
lib/src/render/file_renderer.dart — rendersToSeparateFile gates per-schema file emission (smooshed variants return false). Per-schema and per-API import collection.
lib/templates/ — Mustache templates. schema_object.mustache emits final class X extends Y when parentSealedTypeName is set. schema_one_of.mustache has four dispatch sections; each reads caseExpression per arm and includes the schema_object partial for smooshedVariants.
test/render/render_schema_test.dart, test/render/render_operation_test.dart, test/naming_test.dart — unit tests per layer.
~/Documents/GitHub/personal/gen_tests/api.github.com.json — primary validation target.
ARCHITECTURE.md — pipeline diagram, per-phase notes, planned next steps. Update when phases shift.
Issue #144 on github — running list of oneOf-specific follow-ups. Other gaps surface fresh from regen runs.

spec-iteration

同仓库更多 Skills

同仓库更多 Skills

Current state — the dart fix reduction arc (2026-07-06)

The arc: stop leaning on dart fix

Landed this arc (all merged unless noted)

Remaining lint tail (github, fix-disabled, after #256/#257)

The const-example refactor (the big remaining lever — design settled, not started)

What's next — an open decision the user is weighing

Methodology learnings (reuse these)

What space_gen is / validation targets

Architecture

Smoosh (the structural matrix)

Render-side dispatch builders

Open gaps (pick from here)

Discovery: don't pick from this list blind — regen first

Backlog beyond the dart fix arc

Conventions

Working rhythm

Don't

Helpful one-shots

Where to look

Current state — the dart fix reduction arc (2026-07-06)

The arc: stop leaning on dart fix

Landed this arc (all merged unless noted)

Remaining lint tail (github, fix-disabled, after #256/#257)

The const-example refactor (the big remaining lever — design settled, not started)

What's next — an open decision the user is weighing

Methodology learnings (reuse these)

What space_gen is / validation targets

Architecture

Smoosh (the structural matrix)

Render-side dispatch builders

Open gaps (pick from here)

Discovery: don't pick from this list blind — regen first

Backlog beyond the dart fix arc

Conventions

Working rhythm

Don't

Helpful one-shots

Where to look

Current state — the `dart fix` reduction arc (2026-07-06)

The arc: stop leaning on `dart fix`

Backlog beyond the `dart fix` arc

Current state — the `dart fix` reduction arc (2026-07-06)

The arc: stop leaning on `dart fix`

Backlog beyond the `dart fix` arc