app-ai-guardrails v0.0.3: org-preset lint source, greenfield-ideal thresholds, eslint lessons

[a-tokyo]

Summary

Folds the transferable lessons from our org's shared ESLint-config release (and its staged consumer migration) into the skill, plus one production regression fix its confirmation run caught.

Lesson	Disposition
Org preset package (factory + thin overrides)	Adopted as the preferred Phase-0 lint source when one exists; inline canon remains the self-contained default. Scorer resolves via eslint --print-config when text-grep misses — validated on a scratch preset consumer (268 sonarjs rules, cc 15)
Flat-config option retention (severity-only override keeps old options)	Adopted: explicit {} in templates + anti-defang list entry
ESLint ≥10 breaks eslint-config-next	Adopted as the documented escape route: @next/eslint-plugin-next directly
.d.ts no-var override · no-alphabetical-sort/e2e/** relaxations · prefer-read-only-props off	Adopted with rationale comments
tsconfig noFallthroughCasesInSwitch	Adopted; noUncheckedIndexedAccess documented as deliberate omission
Coverage floors	Raised to the greenfield ideal — 90, branches 85 (old 85/71/76/86 were one repo's brownfield ratchet floors). Scorer is behavioral, ledger unaffected; golden regenerated, self-test ALL CHECKS PASSED
antfu-foundation wholesale switch	Rejected: the canon layers onto what the scaffolder generates
sonarjs/deprecation off	Rejected (kept ON): a guardrails scaffold wants the deprecation signal

Regression fix (found by this release's confirmation run): v0.0.2's consent gate could stall non-interactive runs with an unanswerable approval question. Phase 0 now defines non-interactive (no reply channel) and forbids asking in it; the consent gate's non-interactive branch is the TODO(skills-install) block. The benchmark prompt also regained its designed do-not-ask line.

Also: all deferral/roadmap wording removed from shipped text (no-op context); version synced in frontmatter + marketplace.json.

Test plan

• Self-test (fixtures regenerated at new floors): ALL CHECKS PASSED (golden=100 / bare=22)
• Preset-path probe (scratch consumer of a real preset package): maximal-ruleset + complexity resolve correctly
• Behavioral confirmation on the final text: fresh sonnet one-shot run → 94, all gates green, no question-stall
• Compliance: SKILL.md 237/500 · description 949/1024 · zero defer/roadmap greps · all scripts syntax-clean

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Updates the app-ai-guardrails skill and its benchmark/scorer for the v0.0.3 release, incorporating lessons from an org-shared ESLint preset migration, raising “greenfield ideal” coverage floors, and hardening non-interactive behavior to avoid stalling runs.

Changes:

• Bumped app-ai-guardrails to v0.0.3 and synced the version in .claude-plugin/marketplace.json.
• Raised coverage thresholds across canon docs, adapters, and benchmark fixtures (generally to 90%+ with stack-appropriate axes).
• Improved lint/scoring guidance and benchmark behavior: org-preset lint source path, ESLint flat-config option-retention note, “do not ask questions” benchmark prompt, and scorer support for preset-composed ESLint configs via eslint --print-config.

Reviewed changes

Copilot reviewed 19 out of 19 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
skills/app-ai-guardrails/SKILL.md	v0.0.3 bump; Phase 0 parameter/non-interactive wording updates and lint-source parameterization.
skills/app-ai-guardrails/references/canon/gate-interface.md	Adds an ESLint flat-config anti-pattern entry about severity-only overrides retaining prior options.
skills/app-ai-guardrails/references/canon/coverage.md	Updates per-stack coverage threshold table to the raised “greenfield ideal” floors.
skills/app-ai-guardrails/references/canon/agent-surface.md	Clarifies consent gate behavior for non-interactive runs; adjusts TODO block wording.
skills/app-ai-guardrails/references/adapters/springboot.md	Raises JaCoCo verification minimums and updates explanatory text.
skills/app-ai-guardrails/references/adapters/rust.md	Raises cargo-llvm-cov fail-under thresholds.
skills/app-ai-guardrails/references/adapters/next.md	Adds org-preset guidance, ESLint ≥10 escape route notes, test/e2e relaxations, d.ts overrides, and raised coverage thresholds.
skills/app-ai-guardrails/references/adapters/nest.md	Raises coverage thresholds and points to reusing updated Next ESLint/org-preset guidance.
skills/app-ai-guardrails/references/adapters/go.md	Raises Go coverage gate threshold; refines gomod-age note.
skills/app-ai-guardrails/references/adapters/django.md	Raises pytest-cov fail-under threshold in narrative and config block.
skills/app-ai-guardrails/README.md	Updates scope wording to “out of scope” (no roadmap language).
README.md	Updates Claude Code plugin install examples to use the renamed marketplace (@agent-skills).
benchmarks/app-ai-guardrails/task/prompt-template.md	Restores/strengthens “Do not ask questions” instruction to prevent stall regressions.
benchmarks/app-ai-guardrails/results/runs.csv	Adds additional benchmark run records for the release confirmation.
benchmarks/app-ai-guardrails/fixtures/make-golden.sh	Updates vitest coverage thresholds in golden fixture generation.
benchmarks/app-ai-guardrails/check-guardrails.mjs	Enhances Next static-analysis probes to resolve rules via eslint --print-config when using preset-composed configs.
.claude-plugin/marketplace.json	Renames marketplace to agent-skills and syncs app-ai-guardrails plugin version/description.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 19 out of 19 changed files in this pull request and generated 2 comments.

[a-tokyo]

Review dispositions (all rounds — surfaced here because inline replies collapse under outdated threads)

Round 1 (fixed in e361d9c):

1. ✅ Non-interactive proceed-vs-ask contradiction — precedence now explicit: stack stated/derivable → proceed with declared defaults; stack missing in a no-reply session → abort (the one never-defaulted parameter); ask only when the user is reachable. Behavioral note: run 14 (94, all gates green) confirmed the exercised path; the edit sharpens the stack-missing abort edge the benchmark prompt cannot reach.
2. ✅ TODO(skills-install) commands now carry -y — consistent with the install-set rule.
3. ◐ Suppression flags — names kept (verified-real ESLint 10 core, production-used; concrete default > vague gesture) and anchored to the currency rule: verify against eslint --help at use time.

Round 2 (fixed in 2f77ec3):
4. ✅ Watchdog PID-reuse race — KILL escalation verifies the PID still runs a claude/node command before firing.
5. ✅ --print-config probe on non-src layouts — resolves a real source file via the teeth-probe walker; skips resolution (keeping text-grep verdicts) when none exists. Self-test ALL CHECKS PASSED.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Copilot reviewed 19 out of 19 changed files in this pull request and generated no new comments.

Comments suppressed due to low confidence (1)

skills/app-ai-guardrails/references/canon/agent-surface.md:108

• In the TODO(skills-install) block, -y is currently placed after add, which passes it to the skills CLI rather than to npx. That can still allow npx to prompt for installing the skills package (defeating the non-interactive fallback) and it’s inconsistent with the earlier guidance to run installs with -y/--yes on npx.

    npx skills add -y a-tokyo/agent-skills --skill production-grade --skill tribunal --skill database-documentation
    npx skills add -y github/awesome-copilot --skill autoresearch