fix(ai-gateway): remove overbroad model-slug rule (closes #60)

[ramseydallal-lab]

Summary

Removes the \d+-\d+[)'"] model-slug validation rule on the ai-gateway skill. The rule was supposed to catch typo'd hyphenated slugs in gateway('…') calls but in practice produces false positives that compound into the broader noise problem documented in #38 / #59.

Closes #60.

The bug, by example

Authoritative Anthropic docs (platform.claude.com/docs/en/docs/about-claude/models, verified 2026-04-30):

| Claude API ID | claude-sonnet-4-6 |
| Claude API alias | claude-sonnet-4-6 |
| Pricing | $3 / $15 per MTok |

Anthropic's canonical model identifiers use hyphens. They have used hyphens consistently across claude-3-5-sonnet, claude-opus-4-1, claude-sonnet-4-5, claude-opus-4-7 — every model.

A real production app shipped this code:

import { anthropic } from '@ai-sdk/anthropic'
const result = await generateText({
  model: anthropic('claude-sonnet-4.6'),  // typo: dot, written this way because the validator said so
  ...
})

Anthropic API responded model_not_found, every coach/critique call surfaced as ai_error to end users for several days. Fixing the slug to claude-sonnet-4-6 (the form the validator was flagging as wrong) restored service.

Three failure modes of the deleted rule

1. Inverts Anthropic's actual convention. Direct provider SDK calls (anthropic('…'), openai('…')) take the provider's native model IDs, which Anthropic publishes with hyphens. The rule fires on every such call as a false positive.

2. Matches outside model-slug context. The pattern \d+-\d+[)'"] greedily matches any digit-hyphen-digit followed by ), ', or ". Real-world hits include date strings ("01-15-2024"), package versions, file paths, semver ranges in JSDoc — anywhere the substring appears in code.

3. Trains operators to dismiss validator output. Combined with the over-matching documented in posttooluse-validate: regex matches fire inside JSDoc and line comments — false positives on documentation prose #59 (validator scans full file content and re-matches its own SKILL.md rule strings), users learn to ignore all plugin warnings, losing the value of the rules that are correct.

Why the gateway-vs-direct asymmetry exists, and why a rule can't fix it

The Vercel AI Gateway's catalog uses dots in version numbers. Per vercel/docs/ai-gateway/models-and-providers, the Gateway example is:

model: gateway('anthropic/claude-opus-4.6')

So both these calls are correct in their respective contexts:

gateway('anthropic/claude-opus-4.6')   // Gateway catalog form, dots
anthropic('claude-opus-4-7')           // Direct provider SDK, hyphens

A regex on the file content can't tell which surface a string is destined for, so any version of this rule will be wrong half the time. The remaining provider/-prefix rule (gateway\(['"][^'"/]+['"]\)) still enforces the actual hard constraint: gateway calls must use provider/model format. Slug typos within that format will surface from the gateway itself with a clearer error than a regex lint can produce.

Files changed

File	Change
skills/ai-gateway/SKILL.md	Remove the validate: rule (kept as comment with #60 pointer); rewrite Model Slug Rules prose to document gateway-vs-direct asymmetry honestly
generated/skill-manifest.json	Drop the corresponding rule entry
tests/validate-rules.test.ts	Invert the direct assertion (line 426) — both forms now must NOT raise the deleted message; update 3 multi-skill overlap tests to use gateway('opaque') (existing missing-provider rule) as the ai-gateway trigger; extend the negative test to cover both forms
tests/posttooluse-validate.test.ts	Convert "detects gateway from 'ai' with hyphenated model slug" test to assert NO violation, looped over both hyphen and dot forms; ai-gateway skill match still expected
tests/posttooluse-chain.test.ts	Cosmetic: fix anthropic('claude-sonnet-4.6') → anthropic('claude-sonnet-4-6') in test fixture (test logic unchanged)

Test plan

• git diff --stat — 5 files, +73/-54
• Manually validated assertion polarity — for each touched test, confirmed expected violations match remaining ai-gateway rule set
• CI: bun test — local environment doesn't have bun; relying on PR CI

Out of scope

• The broader #59 (validator over-matches inside SKILL.md content) — separate refactor; this PR's scope is just the slug rule.
• Migration from @ai-sdk/anthropic direct calls to AI Gateway — that's a downstream consumer concern, not the plugin's job to enforce.

🤖 Generated with Claude Code