Files
proof-of-work/proof-of-work/lib/ai/sse.ts
T
Keysat 974c3eb07d v1.1.0:2 — model-agnostic AI program generation (5 providers)
Five providers behind one streaming abstraction:
  - claude              (Anthropic)
  - openai              (api.openai.com)
  - openai-compatible   (any base URL — OpenRouter / LiteLLM /
                         vLLM / Together / your own gateway)
  - gemini              (Google)
  - ollama              (self-hosted; no key; LAN URL like
                         http://ollama.embassy:11434)

The "self-hosted Ollama on Start9" angle is the killer use case —
configure Settings → AI integration with the LAN URL of your Ollama
service and no API keys ever leave your network.

Architecture
  - lib/ai/types.ts              LLMProvider streaming interface
  - lib/ai/sse.ts                shared SSE + NDJSON line iterators
  - lib/ai/providers/*.ts        5 implementations + factory
  - lib/ai/programSchema.ts      Zod schema + JSON-schema-for-prompt +
                                  parseAIProgram with markdown-fence
                                  stripping and balanced-brace JSON
                                  extraction
  - lib/ai/apply.ts              materializes parsed AIProgram into
                                  Program tree (validates exerciseIds,
                                  rejects unresolved nulls, atomic
                                  transaction, sets aiGenerated=true)

Schema
  - UserPreferences gets aiProvider/aiModel/aiBaseUrl/aiApiKey
    (plaintext — same threat model as the rest of /data). Dead
    enableClaudeAI/claudeApiKey columns from v1.0.0:1-7 stay as
    no-op fields.
  - AIPromptTemplate (userId nullable; userId=NULL = built-in)
  - AIGeneration (raw response + parsed program + status +
    appliedProgramId + token counts)
  - All compat-ALTER'd in docker_entrypoint.sh on first boot.

API
  - POST   /api/ai/generate              SSE streaming: emits
                                           generation/text/usage/complete
                                           events; persists AIGeneration
                                           row up front so failures show
                                           in history too
  - POST   /api/ai/apply                 takes user-edited AIProgram,
                                           creates Program, marks
                                           generation as applied
  - GET    /api/ai/templates             built-ins + this user's own
  - POST   /api/ai/templates             create user-owned template
  - PATCH  /api/ai/templates/[id]        edit; built-ins admin-only
  - DELETE /api/ai/templates/[id]        delete; built-ins admin-only
  - GET    /api/ai/generations           list (paginated)
  - GET    /api/ai/generations/[id]      full row
  - DELETE /api/ai/generations/[id]      delete one (Program survives)
  - GET    /api/ai/config                returns aiKeyConfigured flag,
                                           never plaintext key
  - POST   /api/ai/config                update provider config
  - DELETE /api/admin/ai/generations     admin-only "clear all" with
                                           optional userId / olderThanDays

UI
  - Settings → AI integration            provider/model/URL/key form;
                                           plaintext key warning visible
  - /main/ai                             hub page with cards
  - /main/ai/generate                    template picker + textarea +
                                           live SSE stream + cancel +
                                           ProgramPreview with inline
                                           unknown-exercise resolver +
                                           apply button + redirect to
                                           the new Program
  - /main/ai/templates                   list + create + edit + delete;
                                           per-row "show prompt" expand;
                                           built-in delete warns about
                                           reconcile re-creation
  - /main/ai/history                     list + delete; status badges;
                                           link to applied Program
  - Nav: "AI" entry between Programs and Exercises (Sparkles icon)

Built-in templates
  - prisma/aiTemplates.seed.json: 5 starter templates (hypertrophy /
    strength / endurance / recovery / custom)
  - prisma/ensurePromptTemplates.cjs: per-boot reconcile,
    INSERT-or-UPDATE keyed on (userId IS NULL AND name=...);
    user-created templates never touched

Tests
  - tests/ai-programSchema.test.ts: extractJson + parseAIProgram
    edge cases (markdown fences, balanced braces, malformed JSON,
    Zod shape rejection, unresolved-exerciseId tolerance)
  - tests/ai-apply.test.ts: materializes valid AIProgram, rejects
    cross-user exerciseIds, rejects unresolved exercises, honors
    isActive flag
  - tests/routes-ai-templates.test.ts: built-in vs user permissions,
    cross-user template isolation, /api/ai/config plaintext-key safety,
    provider enum validation
  - 123 tests across 14 files, all passing.

No data migration. Existing /data is augmented with the new columns
+ tables only.
2026-05-10 15:35:35 -05:00

68 lines
2.1 KiB
TypeScript

/**
* Minimal SSE-line iterator for provider responses.
*
* Reads a fetch response body as a stream and yields each `data:`
* payload exactly once. Handles event boundaries (`\n\n`), CRLF,
* and the "[DONE]" sentinel that OpenAI-style providers emit. Skips
* comments (lines starting with `:`).
*
* Usage:
* for await (const data of sseLines(response)) {
* if (data === '[DONE]') break;
* const evt = JSON.parse(data);
* ...
* }
*/
export async function* sseLines(
response: Response,
): AsyncGenerator<string, void, void> {
if (!response.body) return;
const reader = response.body.getReader();
const decoder = new TextDecoder();
let buffer = '';
while (true) {
const { value, done } = await reader.read();
if (done) break;
buffer += decoder.decode(value, { stream: true });
let idx: number;
while ((idx = buffer.indexOf('\n\n')) >= 0) {
const event = buffer.slice(0, idx);
buffer = buffer.slice(idx + 2);
const dataLines: string[] = [];
for (const raw of event.split('\n')) {
const line = raw.replace(/\r$/, '');
if (!line || line.startsWith(':')) continue;
if (line.startsWith('data:')) {
dataLines.push(line.slice(5).trimStart());
}
}
if (dataLines.length > 0) yield dataLines.join('\n');
}
}
}
/**
* NDJSON line iterator (Ollama). Yields each non-empty line as a
* raw string (not parsed) — caller decides what to do with it.
*/
export async function* ndjsonLines(
response: Response,
): AsyncGenerator<string, void, void> {
if (!response.body) return;
const reader = response.body.getReader();
const decoder = new TextDecoder();
let buffer = '';
while (true) {
const { value, done } = await reader.read();
if (done) break;
buffer += decoder.decode(value, { stream: true });
let idx: number;
while ((idx = buffer.indexOf('\n')) >= 0) {
const line = buffer.slice(0, idx).replace(/\r$/, '');
buffer = buffer.slice(idx + 1);
if (line) yield line;
}
}
if (buffer.trim()) yield buffer.trim();
}