chore(scaffold): default examples to gpt-5-mini; pre-approve esbuild build in templates

[blove]

Summary

Post-smoke-test cleanup (2 of 2 follow-up PRs). Finishes the move off gpt-4o-mini for user-facing examples and fixes the scaffold's non-interactive pnpm install.

• Model → gpt-5-mini (the project's canonical default — the research template already uses it) in: the basic scaffold template route, root README.md (the raw-LangGraph + With-Dawn + 30-second snippets), @dawn-ai/sdk README, the landing Hero/FeatureRouting snippets, content/templates/AGENTS.md, content/prompts, and the @dawn-ai/evals llmJudge default.
• Deliberately NOT swept: the known-ids registry, model-validation tests, recorded aimock fixtures (gpt-4o-mini-2024-07-18), api.mdx type-union, frozen superpowers plans, CHANGELOGs, and dated blog posts — those keep gpt-4o intentionally.
• Scaffold pnpm.onlyBuiltDependencies: ["esbuild"] in both template package.json.template files — fixes ERR_PNPM_IGNORED_BUILDS: esbuild that breaks pnpm install non-interactively in CI/Docker (found while smoke-testing the docker blueprint).

Test Plan

• pnpm build / typecheck / lint clean
• @dawn-ai/devkit (9) + create-dawn-ai-app (5) tests pass
• check-docs + pack:check pass; changeset gate passes
• grep gpt-4o over the migrated dirs is empty; only the 10 intended files changed

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
dawnai	Ready	Preview, Comment	Jun 27, 2026 10:42pm

[github-actions]

Automated approval: this PR received an intelligent (AI) code review. See the review comments on this PR.

[github-actions]

Claude finished @blove's task in 1m 31s —— View job

---

PR Review: chore(scaffold): default examples to gpt-5-mini; pre-approve esbuild build in templates

• Read PR diff and changed files
• Verify model ID validity in known-ids registry
• Inspect llmJudge behavioral impact
• Check changeset correctness
• Review pnpm.onlyBuiltDependencies fix
• Check for security/type-safety issues

---

Summary

Clean, well-scoped cleanup PR. No bugs, no security issues, no type-safety problems. Two small observations worth noting:

---

Observations

1. llmJudge default is a silent runtime behavior change

packages/evals/src/llm-judge.ts:23 changes the fallback model from gpt-4o-mini to gpt-5-mini. This is correct and intentional, but it's worth being explicit: any user calling llmJudge({criteria: "..."}) without passing model will silently start using gpt-5-mini after upgrading. The changeset description covers this, but the changelog entry (when generated) won't surface it as prominently as a minor bump would. Given the fixed group in changeset/config.json, all packages bump together regardless — so the bump level doesn't change user behavior here. Just noting it for awareness.

2. Changeset lists only @dawn-ai/evals, but devkit templates also changed

packages/devkit/templates/app-basic/package.json.template and packages/devkit/templates/app-research/package.json.template both changed. This is fine because all packages are in a fixed group in .changeset/config.json — bumping @dawn-ai/evals triggers a version bump for @dawn-ai/devkit and every other package too. No action needed.

---

What Looks Good

• gpt-5-mini is in OPENAI_MODEL_IDS at packages/sdk/src/known-model-ids.ts:9 — the model ID is valid and won't trigger a validation warning.
• pnpm.onlyBuiltDependencies: ["esbuild"] is correctly placed in both templates and is the right fix for ERR_PNPM_IGNORED_BUILDS in non-interactive installs.
• The sweep exclusions (aimock recordings, model-validation test fixtures, api.mdx type union, changelogs, dated blog posts) are all well-reasoned — especially keeping gpt-4o-mini-2024-07-18 in recorded fixtures where replay fidelity matters.
• No secrets, no injection points, no unsafe input handling introduced.

LGTM. The PR does exactly what it says and nothing more.

[github-actions]

Claude finished @blove's task in 1m 50s —— View job

---

PR Review: chore(scaffold): default examples to gpt-5-mini; pre-approve esbuild build in templates

• Read PR diff and changed files
• Verify model ID validity
• Check pnpm.onlyBuiltDependencies fix correctness
• Check llmJudge behavioral impact
• Review test fixture updates
• Check for security/type-safety issues

---

Summary

Clean, well-scoped cleanup PR. No bugs, no security issues, no type-safety problems. LGTM.

---

What Looks Good

Model migration is correct and complete:

• gpt-5-mini is present at packages/sdk/src/known-model-ids.ts:9 — valid model ID, no validation warnings.
• All targeted files (app-basic template route, both READMEs, landing components, content/prompts, AGENTS.md, llmJudge) are migrated. grep gpt-4o-mini over migrated dirs returns empty.
• Intentional exclusions are well-reasoned: api.mdx type union, blog posts, aimock fixtures, and the known-ids registry itself keep gpt-4o-mini correctly.

pnpm.onlyBuiltDependencies: ["esbuild"] fix:

• Correctly placed in both app-basic/package.json.template and app-research/package.json.template.
• Test fixtures (basic.expected.json, custom-app-dir.expected.json) updated to match the new template output.
• run-generated-app.test.ts:481 updated to include onlyBuiltDependencies: ["esbuild"] in the internal fixture builder — the test merges this with pnpm.overrides, so no duplication issue.
• The handwritten runtime fixture (handwritten-runtime-app/package.json) correctly omits pnpm.onlyBuiltDependencies since it's not a scaffolded app.

Changeset and versioning:

• Changeset bumps @dawn-ai/evals as patch. All packages are in the same fixed group in .changeset/config.json, so @dawn-ai/devkit and every other package bump together automatically — no gap here.

---

Minor Observations (non-blocking)

llmJudge default is a silent runtime change: Any caller using llmJudge({ criteria: "..." }) without an explicit model will start using gpt-5-mini after upgrading. This is intentional and correct — just worth being aware of in release notes.

Changeset only names @dawn-ai/evals: Devkit templates also changed, but this is covered by the fixed group bumping all packages together. No action needed.

[github-actions]

Claude finished @blove's task in 1m 20s —— View job

---

PR Review: chore(scaffold): default examples to gpt-5-mini; pre-approve esbuild build in templates

• Read PR diff and changed files
• Verify model ID validity in known-ids registry
• Check pnpm.onlyBuiltDependencies fix correctness
• Check llmJudge behavioral impact
• Review test fixture updates
• Check for security/type-safety issues
• Post findings

---

Summary

Clean, well-scoped cleanup PR. No correctness bugs, no security issues, no type-safety problems. LGTM.

---

What Looks Good

• gpt-5-mini is a valid model ID — present at packages/sdk/src/known-model-ids.ts:9 in OPENAI_MODEL_IDS. No validation warnings will fire.
• pnpm.onlyBuiltDependencies: ["esbuild"] is correctly placed in both app-basic and app-research package.json.template files. This is the right fix for ERR_PNPM_IGNORED_BUILDS in non-interactive pnpm installs (CI/Docker).
• Research template is already consistent — app-research/src/app/research/index.ts and researcher/index.ts already use gpt-5-mini; this PR just catches up the basic template and documentation.
• Test fixtures properly updated — basic.expected.json and custom-app-dir.expected.json both gain the pnpm.onlyBuiltDependencies block, and run-generated-app.test.ts adds it to the fixture builder (createExpectedInternalFixture). Runtime fixtures don't reference these, so no other fixtures needed updating.
• Sweep exclusions are well-reasoned — gpt-4o-mini correctly remains in: the known-ids registry (it's a valid model), agent.test.ts test fixtures, api.mdx type union, dated blog posts, and CHANGELOG.

One Note Worth Mentioning

llmJudge default is a silent runtime behavior change (packages/evals/src/llm-judge.ts:23). Any caller using llmJudge({ criteria: "..." }) without passing model will start using gpt-5-mini after upgrading. This is intentional and the right call — just worth calling out that it affects eval cost/behavior for existing users who omit model. The patch bump level is arguably correct since users can always pass model explicitly, but anyone running automated evals without pinning a model will see a change.

No action required — just awareness.