Add self-improving retries for effect generation with versioned prompt templates

[smcga]

Motivation

• Improve observability and robustness of model-generated effect creation by recording failed attempts and enabling automatic, bounded self-improvement of the generator prompt template.
• Allow prompt template updates at runtime (with version history) so operators can inspect and evolve the system prompt over time.
• Reduce friction for users by automatically retrying generation after a failure and surface concise status/attempt information in the UI.

Description

• Introduced a KV-backed prompt-template store and versioning under effects:generate:prompt-template and a failure log stream effects:generate:failure-logs, with sanitization and seeding logic (new types and helpers in api/effects.ts).
• Reworked server generation flow (POST /api/effects?action=generate) to run a bounded self-improvement loop (configurable constant MAX_SELF_IMPROVEMENT_ATTEMPTS = 5) that: logs failures, requests a prompt-template tweak from the model, appends the new template as a version, and retries generation; includes a leak-guard to prevent copying user idea text into templates.
• Persisted failure diagnostics (timestamp, requestId, exact template text used, full model response, user-visible error, attempt index) to KV via effects:generate:failure-logs and surface self-improvement metadata in API JSON responses under selfImprovement.
• Added requestResponsesApi / requestPromptTemplateImprovement helpers to centralize calls to the Responses API and to request prompt-template improvements.
• Updated client plumbing (src/effectIdeas.ts) and UI (src/main.ts) to propagate and display selfImprovement metadata, remove the manual retry dependence, show clearer busy/error copy, and display attempt counters in the busy/error modal; removed retry-button click wiring.
• Added/updated tests to cover the new behavior and persisted artifacts (api/effects.test.ts, src/effectIdeas.test.ts), and updated README.md to document the prompt-template KV key and failure-log stream.

Testing

• Ran unit tests: npm test -- api/effects.test.ts src/effectIdeas.test.ts src/style.test.ts and all tests passed (42 tests, all green).
• Added tests that assert: self-improvement metadata appears in API responses, prompt-template store is updated (effects:generate:prompt-template), failure logs are appended (effects:generate:failure-logs), and the client surfaces attempt info; these tests passed as part of the run.
• Verified existing generation/monitoring tests still pass after changes.

---

Codex Task

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
codexperimental	Ready	Preview, Comment	Apr 8, 2026 11:27pm