Per-provider vision blocks; parser repairs local models' almost-JSON

config.toml now carries a [vision.<provider>] block per backend —
model/base_url/key_env — with vision_provider picking the active one,
so the committed file documents every recipe and switching is a
one-line flip. Only the active block applies; typo'd block names and
keys warn like every other config mistake.

First real Ollama run (qwen2.5vl:7b) surfaced what local models emit:
almost-JSON with trailing commas. parse_vision_response now makes one
cheap repair pass before declaring a response unusable.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016jXZFSTZQKzAC8fqpWSz9g

README.md

@@ -70,7 +70,7 @@ Every stage is idempotent and resumable: kill it mid-run, restart, lose nothing.
 ## Requirements
 
 - macOS or Linux, Python 3.12+, [uv](https://docs.astral.sh/uv/)
-- A vision model for extraction — an [Anthropic API key](https://console.anthropic.com/) by default, or any OpenAI-compatible endpoint via `config.toml` (`vision_provider = "openai-compatible"` + `vision_base_url`): OpenAI, OpenRouter, or a free local runtime like [Ollama](https://ollama.com/) with a vision-capable model. Local models read spines noticeably worse than frontier ones — expect a longer proofread pass on the Titles page, not a broken pipeline.
+- A vision model for extraction — an [Anthropic API key](https://console.anthropic.com/) by default, or any OpenAI-compatible endpoint: OpenAI, OpenRouter, or a free local runtime like [Ollama](https://ollama.com/) with a vision-capable model. `config.toml` carries a `[vision.<provider>]` block for each; `vision_provider` picks one. Local models read spines noticeably worse than frontier ones — expect a longer proofread pass on the Titles page, not a broken pipeline.
 - A BoardGameGeek account **and a registered BGG application** — as of BGG's [2025 API policy](https://boardgamegeek.com/using_the_xml_api), the XML API requires a Bearer token from a registered app. Register a free non-commercial application at [boardgamegeek.com/applications](https://boardgamegeek.com/applications) (approval can take a week or more, so **apply on day one**), then create a token. Each user needs their own; tokens must not be shared.
 
 ## Quick start

config.toml

@@ -3,12 +3,19 @@
 
 photos_dir = "photos"
 data_dir = "data"
-model = "claude-sonnet-5"
 rate_limit_seconds = 2.0
 
-# Vision backend. The default ("anthropic") reads ANTHROPIC_API_KEY.
+# Which vision backend reads your shelf photos. Both recipes below stay
-# "openai-compatible" covers OpenAI, OpenRouter, and local runtimes:
+# on file; this line picks one.
-#   vision_provider = "openai-compatible"
+vision_provider = "anthropic"
-#   vision_base_url = "http://localhost:11434/v1"   # e.g. Ollama
+
-#   model = "qwen2.5vl"                             # a vision-capable model
+[vision.anthropic]
-#   vision_key_env = ""                             # local: no key needed
+# reads ANTHROPIC_API_KEY from the environment
+model = "claude-sonnet-5"
+
+[vision."openai-compatible"]
+# OpenAI, OpenRouter, or a local runtime (Ollama, LM Studio, vLLM).
+# key_env names the env var holding the key; "" = endpoint needs none.
+base_url = "http://localhost:11434/v1"
+model = "qwen2.5vl:7b"
+key_env = ""

src/bggpipe/config.py

@@ -140,17 +140,39 @@ def load_config(path: Path | None = None) -> Config:
             "model": str,
             "rate_limit_seconds": float,
             "vision_provider": str,
-            "vision_base_url": str,
-            "vision_key_env": str,
         }
         updates = {key: caster(raw[key]) for key, caster in known.items() if key in raw}
-        if unknown := sorted(raw.keys() - known.keys()):
+        if unknown := sorted(raw.keys() - known.keys() - {"vision"}):
             # a typo'd key silently falling back to defaults is a debugging
             # trap ("No photos found in photos/") — say so up front
             warnings.warn(
                 f"{p}: ignoring unknown key(s): {', '.join(unknown)}",
                 stacklevel=2,
             )
+        # [vision.<provider>] blocks: every provider's recipe can live in
+        # the committed file; only the ACTIVE provider's block applies
+        vision_blocks = raw.get("vision") or {}
+        if stray := sorted(vision_blocks.keys() - set(VISION_PROVIDERS)):
+            warnings.warn(
+                f"{p}: ignoring [vision.*] block(s) for unknown provider(s): "
+                f"{', '.join(stray)}",
+                stacklevel=2,
+            )
+        provider = updates.get("vision_provider", cfg.vision_provider)
+        block = vision_blocks.get(provider) or {}
+        block_known = {
+            "model": "model",
+            "base_url": "vision_base_url",
+            "key_env": "vision_key_env",
+        }
+        if odd := sorted(block.keys() - block_known.keys()):
+            warnings.warn(
+                f"{p}: [vision.{provider}]: ignoring unknown key(s): {', '.join(odd)}",
+                stacklevel=2,
+            )
+        for key, field in block_known.items():
+            if key in block:
+                updates[field] = str(block[key])
         cfg = replace(cfg, **updates)
     if cfg.vision_provider not in VISION_PROVIDERS:
         # a typo here would surface as a confusing extract failure later

src/bggpipe/extract.py

@@ -126,7 +126,12 @@ def parse_vision_response(text: str) -> tuple[list[dict], list[dict], int]:
         if end == -1:
             raise ValueError(f"unterminated JSON in vision response: {text[:200]!r}")
         cleaned = cleaned[start : end + 1]
-    data = json.loads(cleaned)
+    try:
+        data = json.loads(cleaned)
+    except json.JSONDecodeError:
+        # local models emit almost-JSON (trailing commas, mostly): one
+        # cheap repair pass before declaring the response unusable
+        data = json.loads(re.sub(r",\s*([}\]])", r"\1", cleaned))
     if isinstance(data, list):
         titles_raw, unidentified_raw = data, []
     elif isinstance(data, dict):

tests/test_config.py

@@ -52,3 +52,47 @@ def test_unknown_toml_keys_warn(tmp_path, monkeypatch):
 def test_explicit_missing_config_errors_instead_of_silent_defaults(tmp_path):
     with pytest.raises(FileNotFoundError, match="does not exist"):
         load_config(tmp_path / "nope.toml")
+
+
+def test_vision_blocks_apply_only_for_the_active_provider(tmp_path):
+    p = tmp_path / "config.toml"
+    p.write_text(
+        """
+vision_provider = "openai-compatible"
+
+[vision.anthropic]
+model = "claude-sonnet-5"
+
+[vision."openai-compatible"]
+base_url = "http://localhost:11434/v1"
+model = "qwen2.5vl:7b"
+key_env = ""
+"""
+    )
+    cfg = load_config(p)
+    assert cfg.model == "qwen2.5vl:7b"  # the active block's model wins
+    assert cfg.vision_base_url == "http://localhost:11434/v1"
+    assert cfg.vision_key_env == ""
+    # flip the provider: the other block applies, this one is inert
+    p.write_text(p.read_text().replace('"openai-compatible"\n', '"anthropic"\n', 1))
+    cfg = load_config(p)
+    assert cfg.model == "claude-sonnet-5"
+    assert cfg.vision_base_url == ""  # untouched default
+
+
+def test_vision_block_typos_warn(tmp_path):
+    p = tmp_path / "config.toml"
+    p.write_text(
+        """
+[vision.anthropic]
+modle = "oops"
+
+[vision.anthorpic]
+model = "claude-sonnet-5"
+"""
+    )
+    with (
+        pytest.warns(UserWarning, match="modle"),
+        pytest.warns(UserWarning, match="anthorpic"),
+    ):
+        load_config(p)

tests/test_extract.py

@@ -120,6 +120,15 @@ def test_parse_drops_malformed_entries():
     assert len(unidentified) == 1  # empty {} and the bare string are dropped
 
 
+def test_parse_repairs_trailing_commas():
+    # the almost-JSON local vision models emit
+    titles, unidentified, dropped = parse_vision_response(
+        '{"titles": [{"title_raw": "Catan",},], "unidentified": [],}'
+    )
+    assert [t["title_raw"] for t in titles] == ["Catan"]
+    assert unidentified == [] and dropped == 0
+
+
 def test_parse_raises_on_garbage():
     with pytest.raises(ValueError):
         parse_vision_response("I couldn't see any games clearly.")