diff --git a/model-overlays/sonnet-5.md b/model-overlays/sonnet-5.md new file mode 100644 index 0000000000..39bf42311f --- /dev/null +++ b/model-overlays/sonnet-5.md @@ -0,0 +1,17 @@ +{{INHERIT:claude}} + +**Instructions are read literally.** Sonnet 5 does not silently generalize an +instruction from one item to the next, and it does not infer requests you didn't +make. When something should apply broadly, say so ("apply this to every section, +not just the first"). Re-baseline holdover style directives — they now land at +face value. + +**Scope work to the request.** At lower effort especially, Sonnet 5 scopes to +exactly what was asked rather than going above and beyond. If reasoning looks +shallow on a genuinely complex task, that is an effort signal: raise effort rather +than adding prose guardrails. + +**Verbosity tracks task complexity.** Responses calibrate length to how complex +the task looks — shorter on lookups, longer on open-ended analysis. If you need a +specific length or format, state it; a positive example of the target beats a +"don't be verbose" instruction. diff --git a/scripts/models.ts b/scripts/models.ts index b6d1d368d5..cf144bb187 100644 --- a/scripts/models.ts +++ b/scripts/models.ts @@ -14,6 +14,7 @@ export const ALL_MODEL_NAMES = [ 'claude', 'opus-4-7', + 'sonnet-5', 'gpt', 'gpt-5.4', 'gemini', @@ -53,6 +54,7 @@ export function resolveModel(input: string): Model | null { if (/^gpt(-|$)/.test(s)) return 'gpt'; if (/^o[0-9]+(-|$)/.test(s)) return 'o-series'; if (/^claude-opus-4-7(-|$)/.test(s)) return 'opus-4-7'; + if (/^claude-sonnet-5(-|$)/.test(s)) return 'sonnet-5'; if (/^claude(-|$)/.test(s)) return 'claude'; if (/^gemini(-|$)/.test(s)) return 'gemini'; diff --git a/test/model-overlay-sonnet-5.test.ts b/test/model-overlay-sonnet-5.test.ts new file mode 100644 index 0000000000..4665efc30a --- /dev/null +++ b/test/model-overlay-sonnet-5.test.ts @@ -0,0 +1,56 @@ +/** + * Sonnet 5 model overlay — gate-tier assertions on the family nudges. + * + * sonnet-5 inherits the claude base and adds Sonnet-5 family nudges: literal + * instruction following (state scope explicitly), scope work to the request + * (raise effort rather than prompting around shallow reasoning), and + * verbosity that tracks task complexity. + */ +import { describe, test, expect } from 'bun:test'; +import * as fs from 'fs'; +import * as path from 'path'; +import type { TemplateContext } from '../scripts/resolvers/types'; +import { HOST_PATHS } from '../scripts/resolvers/types'; +import { generateModelOverlay } from '../scripts/resolvers/model-overlay'; + +function makeCtx(model: string): TemplateContext { + return { + skillName: 'test-skill', + tmplPath: 'test.tmpl', + host: 'claude', + paths: HOST_PATHS.claude, + preambleTier: 2, + model, + }; +} + +const ROOT = path.resolve(__dirname, '..'); + +describe('Sonnet 5 overlay — family nudges', () => { + test('raw sonnet-5.md contains the literal-instructions nudge', () => { + const raw = fs.readFileSync(path.join(ROOT, 'model-overlays/sonnet-5.md'), 'utf-8'); + expect(raw).toContain('Instructions are read literally'); + }); + + test('resolved overlay inherits from claude base (INHERIT:claude)', () => { + const out = generateModelOverlay(makeCtx('sonnet-5')); + expect(out).toContain('Todo-list discipline'); + expect(out).toContain('subordinate'); + }); + + test('resolved overlay carries the Sonnet 5 nudges', () => { + const out = generateModelOverlay(makeCtx('sonnet-5')); + expect(out).toContain('Instructions are read literally'); + expect(out).toContain('Scope work to the request'); + }); + + test('resolved overlay has no unresolved INHERIT directive', () => { + const out = generateModelOverlay(makeCtx('sonnet-5')); + expect(out).not.toContain('{{INHERIT:'); + }); + + test('claude overlay (base) does not carry the Sonnet 5 nudge', () => { + const out = generateModelOverlay(makeCtx('claude')); + expect(out).not.toContain('Instructions are read literally'); + }); +});