- call_log._default_write: swallow any exception from the default file/ stdout sink and report to stderr, so a log-write failure (disk full, bad permissions, misconfigured CALL_LOG_PATH) never turns a successful narration into a 500 (charter §13). Injected write= sinks (used by tests) are left to surface their own errors. - ollama_client.chat: catch ValueError alongside httpx.HTTPError in the retry loop so a 200 response with a non-JSON body (JSONDecodeError is a ValueError subclass) counts as a failed attempt and falls through to ModelError after the one retry, instead of escaping chat() uncaught (charter §12 — one retry, then ModelError, nothing else escapes). - main.py: update stale docstrings — /dm/narrate is now fully wired (routing + model call + logging); the other four roles remain stubs. Regression tests added for both fixes (TDD: watched RED, then GREEN). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
/api — FastAPI proxy
The guarding proxy from charter §4. Not a generic API — it exists to keep the API key, the prompts, and the model routing server-side, and to log every call from day one.
Stack: Python, FastAPI. Models: Ollama (dev/staging) → Replicate (prod), routed per-role server-side.
Role endpoints (charter §4)
POST /dm/narrate Narrator — scene/outcome prose (good model)
POST /dm/adjudicate Adjudicator — free text → legal action (small/fast, strict JSON)
POST /dm/improvise Improviser — minor sandboxed event (charter §7 limits)
POST /npc/speak NPC — voice one character (bounded moves, §6)
POST /party/banter Banter — companion callbacks (cacheable, §9)
The client does not know which model serves a role or what the prompt is.
Responsibilities
- Auth · metering (retrofit later — middleware + Stripe webhook, §4)
- Prompt ownership — prompts live in
prompts/, never in the client - Model routing — role → model is config, a deploy not a client patch
- Logging — log seed and full prompt with every call (§10) for replay/eval
Contracts
Every AI response is untrusted (§2). Parse, validate, be ready to discard. One retry on parse failure, then authored fallback (§12).
See docs/ for endpoint schemas, model routing, and deploy notes.