class_id is the rulebook's word; the world has callings — that was the whole §17
reconcile. The contract the Narrator reads should say what the world says. There
are no saves (M9) and no deployed clients, so this is as cheap as it will ever be
and strictly more expensive every milestone after.
race_id reaches the log because the AI must be able to describe the player:
api/app/prompts.py rendered 'a sellsword' and now renders 'a beastfolk cutpurse'.
It humanizes the id (the model must never read 'hedge_mage') and picks the right
article ('an elf', not 'a elf').
LogPlayer's hardcoded CLASSES const is deleted — the roster is Callings.IDS. A
parity test reads the JSON Schema from the client and asserts its enum equals the
code, so the two sides of the HTTP boundary cannot drift.
The deserter allowed three callings that no longer exist. He now allows all seven;
a thematic gate is a content decision for the Greywater authoring.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QYa9u7Kdxv5gX4AnwWexy8
After the prompts fix, /dm/narrate reached the model call and failed with
ConnectError: Connection refused — the container's localhost is not the
host, so the default OLLAMA_BASE_URL=http://localhost:11434 hit nothing.
Ollama runs on the host (M1's venv smoke reached it because uvicorn ran on
the host too; the container never had a route).
Add extra_hosts host.docker.internal:host-gateway so the container can
reach the host, and document using http://host.docker.internal:11434 in
.env.example. Verified a container reaches host Ollama's /api/tags (200).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The narrator pipeline (M1) reads role prompts from /app/prompts, but the
Dockerfile copied only api/app and docs/schemas — never api/prompts — so
/dm/narrate 500'd with FileNotFoundError on narrator.md in any Docker run.
The M1 live smoke passed only because it ran uvicorn from the local venv,
where ../prompts resolves on the host; the container path was never
exercised until the client HTTP loop drove a real request through compose.
Dockerfile now COPYs api/prompts; compose mounts it :ro so prompt edits
reload live like the app code (prompts are source code, §16). Verified by
building the image and confirming /app/prompts/narrator.md is present.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- 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>
/dm/narrate now makes a real Ollama call (Task 7). The role-endpoint
loop tests in test_endpoints.py had no mock, so every default pytest
run was silently hitting a live Ollama instance on localhost:11434 —
breaking the hermetic-suite constraint. Add an autouse fixture that
stubs narrate.ollama_client.chat and narrate.call_log.record so all
five endpoints stay hermetic; assertions (200 on valid, 422 on
invalid) are unchanged.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- canon_log.py: stringify jsonschema error-path segments before sorting so a
future schema mixing array indices and object keys can't TypeError (#1)
- main.py: RequestValidationError handler reshapes pydantic body failures into
the same {"detail": {"canon_log_errors": [...]}} envelope as schema failures,
so the client parses ONE 422 shape (#2a) + tests
- pytest.ini: filter the starlette/httpx TestClient deprecation → pristine output (#4)
- docs/canon-log.md: living contract for the client (Plan B) — schemas, the §7
boundary, the unified error envelope, per-role injection (#3)
Full suite 26 passing, 0 warnings.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add a Docker-based FastAPI proxy skeleton (charter §4) for parity across
dev / homelab VPS / fly.io, plus root .gitignore for Godot 4.7 and Python.
- api/Dockerfile: python:3.12-slim, non-root, PORT-aware CMD
- api/app/main.py: /health + five role endpoint stubs (§4)
- api/requirements.txt: fastapi/uvicorn/httpx/pydantic, pinned
- docker-compose.yml: local dev with live reload
- api/fly.toml: prod deploy stub with /health check
- api/.env.example: key lives in the proxy, never the client (§4)
- .gitignore: ignores .env, whitelists .env.example
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
