Deferred: first-party quality rules and agent workflow calibration (preserve TSGo and Ultracite)

[BleedingDev]

Deferred first-party quality-policy improvements

Decision and scope

Deferred for later approval, not implementation authorization and not a DAG.
The approved first wave is S1–S4, A9, B1 and B2: separate our blocking/advisory
quality gates, assess bot findings rather than automatically obey them, remove
blanket zero-clone/complexity gating, keep feature scope narrow, repair our
analyzers instead of distorting valid code, offer focused local checks, and
remove implicit formatting from installation.

Do not weaken, override, filter or disable upstream Effect TSGo or Ultracite.
preferSucceedSomeOrNone and preferTypedSchemaDecoder are upstream Effect TSGo
rules, not OntOS rules. The earlier proposal to downgrade them is withdrawn,
not deferred. Sources:

• https://github.com/Effect-TS/tsgo/blob/main/internal/rules/prefer_succeed_some_or_none.go
• https://github.com/Effect-TS/tsgo/blob/main/internal/rules/prefer_typed_schema_decoder.go

Only first-party rules/tooling are eligible below. Existing custom Oxlint
severities remain unchanged pending separate approval. No test should merely
copy configuration or severity maps; validate observable behavior and actual
regressions. Do not rewrite already-correct application code just because a
rule is relaxed.

Deferred A proposals

Proposal	Candidate policy	Medium-term impact	Long-term impact	Risks and required protection
A1: first-party syntax/constructor preferences only	Warning, or off for a proven preference with no useful signal. No upstream TSGo changes.	Fewer behavior-preserving rewrites.	Less coupling to one stylistic idiom.	Verify ownership first; missing input validation is never a syntax preference.
A2: custom no-manual-tag-comparison in test assertions	Warning; consider off only for that narrow scope.	Less assertion-matcher migration churn.	Less bespoke detector maintenance.	Preserve assertions of variant and content; do not conflate with no-raw-effect-adt-tag-check.
A3: custom no-literal-union-type-alias	Warning for internal types; consider narrow off.	Avoid runtime schemas for purely static types.	Simpler internal models.	Keep real runtime validation at API/persistence boundaries.
A4: custom require-context-service-for-service-interface	Warning.	Avoid dependency-injection reshaping solely for local heuristics.	Fewer artificial services/layers.	Keep runtime provisioning and integration tests; absence of local tag evidence is not proof of a missing service.
A5: custom require-concurrency-option	Missing explicit option advisory; demonstrated unbounded fan-out remains blocking.	Avoid mechanical concurrency options.	Capacity and ordering drive policy.	Preserve resource bounds, transaction order and timeout behavior; never mechanically parallelize writes.
A6: custom no-native-json-parse / no-native-json-stringify	Warning for trusted internal use, narrow off only with provenance; boundary contracts remain hard.	Simpler tooling/reporting.	Fewer schemas without a contract purpose.	Scripts can ingest hostile data; protect identity/hash/signature and persisted-document semantics.
A7: custom no-native-error-construction	Warning/narrow off for synchronous validators; real error-channel contract violations remain hard.	Fewer lint-only adapters.	Clearer synchronous/Effect boundaries.	Expected failures must not turn into unhandled defects; test conversion at the boundary.
A8: custom no-unmanaged-mutable-state	Mere module let/Map/WeakMap advisory; proven isolation violations blocking.	Avoid unnecessary cache/registry rewrites.	Explicit state ownership/lifetime.	Verify cross-request/tenant isolation, invalidation and capacity where shared state exists.

Deferred B proposals

Proposal	Implementation direction	Medium-term impact	Long-term impact	Risks and protection
B3: deduplicate repeated review/warnings	Reuse disposition for the same finding until relevant code or assumptions change.	Less repeated agent work and comment noise.	Attention stays on new defects.	Exceptions can become stale; bind them to scope and assumptions, not permanent blanket suppression.
B4: impact-select expensive tooling checks	Run selection in shadow mode beside the full suite before reducing CI work.	Potentially shorter CI.	Lower tooling maintenance/runtime cost.	Indirect dependency misses; rules/generators/lockfile/shared-contract changes require wider checks and safe fallback.
B5: stronger targeted deployment proofs	Fresh DB/migrations, relocated artifact, actual CLI argument forwarding.	Catch deployment errors missed by green PR checks.	Better alignment of tested and deployed behavior.	Avoid another generic test platform; add scenarios demonstrated by historical failures.

Evidence motivating reconsideration

• OntOS PR Migrate all tests to native Effect Rstest and enforce native interfaces #501: Migrate all tests to native Effect Rstest and enforce native interfaces #501 — large test/rule migration and repeated detector repairs; migration counts are not counts of product defects.
• OntOS PR Calibrated unused-code, duplication, and complexity reports #492: Calibrated unused-code, duplication, and complexity reports #492 — custom quality-model and reporting work.
• OntOS PR fix(scaffolding): preserve nested fluent slot calls #503: fix(scaffolding): preserve nested fluent slot calls #503 — valid nested fluent-chain recognition needed a parser repair.
• OntOS PR [codex] implement commerce customer context #618: [codex] implement commerce customer context #618 and OntOS PR Adopt published UltraModern 3.9.0-ultramodern.8 and remove vendored framework tooling #619: Adopt published UltraModern 3.9.0-ultramodern.8 and remove vendored framework tooling #619 — mixed feature/framework/tooling and artifact-proof repair loops.
• OntOS PR Make Auth persistence native Effect and isolate Better Auth adapter #500: Make Auth persistence native Effect and isolate Better Auth adapter #500 and OntOS PR Require Effect-only permission clients and defer SDK requests #497: Require Effect-only permission clients and defer SDK requests #497 — concrete lifecycle/runtime defects justify keeping behavioral protection.

Audit window: 2026-08-17 through 2026-09-16; census of 134 updated PRs and
589 unique main-history commits. It does not establish hours lost or a causal
false-positive percentage. This issue intentionally contains the proposals,
trade-offs and primary evidence without depending on local report files.

Conditions for any later change

• Obtain separate approval; confirm each rule is first-party.
• Keep known real defects detectable and legitimate cases nonblocking.
• Do not add configuration snapshots or tests that just duplicate constants.
• Preserve compiler, authorization, tenant isolation, transaction, secret-handling
  and module-ownership protections. No blanket exclusions for tests/scripts.
• Measure actual repair loops, confirmed false positives, CI wait and escaped
  defects. A short incident-free period is not proof of safety.
• Do not automatically extract shared cross-deployment abstractions to satisfy
  clone metrics. Do not change TSGo or Ultracite enforcement.