#!/usr/bin/env python3 """Validated, deterministic source snapshots for bundled model inventories. Every externally maintained bundled inventory uses the same envelope. The catalog generator consumes only the reviewed ``catalog`` records in these snapshots; route-specific source evidence remains alongside each record so refreshes can be audited without making catalog generation depend on a network. """ from __future__ import annotations import math import re from collections.abc import Mapping, Sequence from datetime import date, datetime from decimal import Decimal, InvalidOperation from pathlib import Path from typing import Any from urllib.parse import urlsplit from mmo_util import read_json, stable_hash, valid_absolute_uri, validate_id from mmo_version import APP_SERVER_PROTOCOL_CODEX_VERSION, MMO_SCHEMA_VERSION SNAPSHOT_FIELDS = { "schema_version", "inventory", "adapter", "fingerprint_fields", "as_of", "dynamic", "sources", "discovery", "captures", "models", "models_sha256", } MODEL_RECORD_FIELDS = {"catalog", "evidence"} CAPTURE_FIELDS = {"source", "retrieved_at", "response_sha256"} DISCOVERY_FIELDS = { "endpoint", "endpoint_selections", "ignored_model_prefixes", "known_deprecated_models", } SOURCE_REFERENCE_FIELDS = { "source", "availability_source", "capability_source", "pricing_source", } OPENROUTER_REASONING_ORDER = ("none", "minimal", "low", "medium", "high", "xhigh", "max") OPENROUTER_SOURCE_IDS = { "openrouter-models-api": "https://openrouter.ai/api/v1/models", "openrouter-zdr-endpoints": "https://openrouter.ai/api/v1/endpoints/zdr", "openrouter-openapi": "https://openrouter.ai/openapi.json", "openrouter-reasoning": "https://openrouter.ai/docs/guides/best-practices/reasoning-tokens", "openrouter-tool-calling": "https://openrouter.ai/docs/guides/features/tool-calling", "openrouter-usage": "https://openrouter.ai/docs/cookbook/administration/usage-accounting", } OPENCODE_ZEN_SOURCE_IDS = { "models-dev-opencode-zen": "https://models.dev/api.json", "opencode-zen-docs": "https://opencode.ai/docs/zen", "opencode-zen-docs-source": ( "https://raw.githubusercontent.com/anomalyco/opencode/" "dev/packages/web/src/content/docs/zen.mdx" ), "opencode-zen-models": "https://opencode.ai/zen/v1/models", } OPENCODE_ZEN_ROUTE_BY_NPM = { "@ai-sdk/anthropic": "opencode_zen_anthropic_messages", "@ai-sdk/google": "opencode_zen_google_catalog", "@ai-sdk/openai": "opencode_zen_responses", "@ai-sdk/openai-compatible": "opencode_zen_openai_chat", } OPENCODE_GO_SOURCE_IDS = { "models-dev-opencode-go": "https://models.dev/api.json", "opencode-go-docs": "https://opencode.ai/docs/go/", "opencode-go-docs-source": ( "https://raw.githubusercontent.com/anomalyco/opencode/" "dev/packages/web/src/content/docs/go.mdx" ), "opencode-go-models": "https://opencode.ai/zen/go/v1/models", } OPENCODE_GO_ROUTE_BY_NPM = { "@ai-sdk/anthropic": "opencode_go_anthropic_messages", "@ai-sdk/openai": "opencode_go_responses", "@ai-sdk/openai-compatible": "opencode_go_openai_chat", } CODEX_REASONING_ORDER = ("none", "minimal", "low", "medium", "high", "xhigh", "max", "ultra") FULL_FINGERPRINT_FIELDS = ( "upstream_id", "canonical_slug", "pricing", "limits", "modalities", "tools", "reasoning", "structured_output", "supported_parameters", "deprecation", "endpoint_metadata", ) def infer_model_maker(upstream_id: str) -> str: """Return a stable model-maker identity, distinct from access operator.""" slug = _nonempty_string(upstream_id, "upstream model id").casefold().removeprefix("~") if "/" in slug: return slug.split("/", 1)[0] prefixes = { "claude": "anthropic", "deepseek": "deepseek", "gemini": "google", "glm": "zai", "gpt": "openai", "o1": "openai", "o3": "openai", "o4": "openai", "kimi": "moonshotai", "minimax": "minimax", "nemotron": "nvidia", "qwen": "qwen", } return next((maker for prefix, maker in prefixes.items() if slug.startswith(prefix)), "unknown") def _object(value: Any, label: str) -> dict[str, Any]: if not isinstance(value, Mapping): raise ValueError(f"{label} must be an object") if not all(isinstance(key, str) for key in value): raise ValueError(f"{label} keys must be strings") return dict(value) def _nonempty_string(value: Any, label: str) -> str: if not isinstance(value, str) or not value.strip(): raise ValueError(f"{label} must be a non-empty string") return value def _string_list(value: Any, label: str, *, allow_empty: bool = True) -> list[str]: if not isinstance(value, list) or not all(isinstance(item, str) and item for item in value): raise ValueError(f"{label} must be a list of non-empty strings") if not allow_empty and not value: raise ValueError(f"{label} must not be empty") if len(value) != len(set(value)): raise ValueError(f"{label} must not contain duplicates") return list(value) def _iso_date(value: Any, label: str) -> str: text = _nonempty_string(value, label) try: parsed = date.fromisoformat(text) except ValueError as exc: raise ValueError(f"{label} must use YYYY-MM-DD") from exc if parsed.isoformat() != text: raise ValueError(f"{label} must use canonical YYYY-MM-DD") return text def _sha256(value: Any, label: str) -> str: text = _nonempty_string(value, label) if len(text) != 64 or any(char not in "0123456789abcdef" for char in text): raise ValueError(f"{label} must be a lowercase SHA-256 digest") return text def _utc_timestamp(value: Any, label: str) -> str: text = _nonempty_string(value, label) try: parsed = datetime.strptime(text, "%Y-%m-%dT%H:%M:%SZ") except ValueError as exc: raise ValueError(f"{label} must use YYYY-MM-DDTHH:MM:SSZ") from exc if parsed.strftime("%Y-%m-%dT%H:%M:%SZ") != text: raise ValueError(f"{label} must use canonical UTC timestamp form") return text def _https_url(value: Any, label: str) -> str: text = _nonempty_string(value, label) if not valid_absolute_uri(text): raise ValueError(f"{label} must be an absolute HTTPS URL without userinfo") try: parsed = urlsplit(text) _ = parsed.port except ValueError as exc: raise ValueError(f"{label} must be an absolute HTTPS URL without userinfo") from exc if ( parsed.scheme.lower() != "https" or not parsed.hostname or parsed.username is not None or parsed.password is not None ): raise ValueError(f"{label} must be an absolute HTTPS URL without userinfo") return text def _positive_int(value: Any, label: str, *, allow_zero: bool = False) -> int: minimum = 0 if allow_zero else 1 if not isinstance(value, int) or isinstance(value, bool) or value < minimum: qualifier = "non-negative" if allow_zero else "positive" raise ValueError(f"{label} must be a {qualifier} integer") return value def _validate_catalog_record( inventory: str, key: str, value: Any, sources: Mapping[str, str], ) -> dict[str, Any]: validate_id(key, "inventory snapshot model key") catalog = _object(value, f"inventory {inventory} model {key}.catalog") if catalog.get("inventory") != inventory: raise ValueError(f"inventory {inventory} model {key} has mismatched inventory membership") for field in ("maker", "route", "upstream_id", "display_name", "description", "source"): _nonempty_string(catalog.get(field), f"inventory {inventory} model {key}.{field}") route_key = validate_id(str(catalog["route"]), f"inventory {inventory} route id") if not key.startswith(f"{route_key}__"): raise ValueError( f"inventory {inventory} model {key}: catalog key must start with exact route " f"namespace {route_key!r} followed by '__'" ) _positive_int( catalog.get("context_window"), f"inventory {inventory} model {key}.context_window", allow_zero=True, ) for field in SOURCE_REFERENCE_FIELDS: source = catalog.get(field) if source is not None and source not in sources: raise ValueError( f"inventory {inventory} model {key}.{field} references unknown source {source!r}" ) return catalog def validate_inventory_snapshot( value: Any, *, label: str = "inventory snapshot", expected_inventory: str | None = None, ) -> dict[str, Any]: """Validate and copy one common inventory snapshot envelope.""" snapshot = _object(value, label) unknown = sorted(set(snapshot) - SNAPSHOT_FIELDS) missing = sorted(SNAPSHOT_FIELDS - set(snapshot)) if unknown or missing: raise ValueError(f"{label} fields are invalid: missing={missing}, unknown={unknown}") schema = snapshot["schema_version"] if not isinstance(schema, int) or isinstance(schema, bool) or schema != MMO_SCHEMA_VERSION: raise ValueError(f"unsupported {label} schema_version") inventory = validate_id(_nonempty_string(snapshot["inventory"], f"{label}.inventory")) adapter = validate_id(_nonempty_string(snapshot["adapter"], f"{label}.adapter")) fingerprint_fields = _string_list( snapshot["fingerprint_fields"], f"{label}.fingerprint_fields", allow_empty=False ) if expected_inventory is not None and inventory != expected_inventory: raise ValueError( f"{label} inventory {inventory!r} does not match filename {expected_inventory!r}" ) _iso_date(snapshot["as_of"], f"{label}.as_of") if not isinstance(snapshot["dynamic"], bool): raise ValueError(f"{label}.dynamic must be boolean") sources_raw = _object(snapshot["sources"], f"{label}.sources") if not sources_raw: raise ValueError(f"{label}.sources must not be empty") sources = { validate_id(key, f"{label} source id"): _https_url(url, f"{label}.sources.{key}") for key, url in sources_raw.items() } discovery = _object(snapshot["discovery"], f"{label}.discovery") unknown_discovery = sorted(set(discovery) - DISCOVERY_FIELDS) if unknown_discovery: raise ValueError(f"{label}.discovery has unknown fields: {unknown_discovery}") if "endpoint" in discovery: _https_url(discovery["endpoint"], f"{label}.discovery.endpoint") if "endpoint_selections" in discovery: selections = _object( discovery["endpoint_selections"], f"{label}.discovery.endpoint_selections", ) if not selections: raise ValueError(f"{label}.discovery.endpoint_selections must not be empty") discovery["endpoint_selections"] = dict( sorted( ( _nonempty_string(model_id, f"{label} endpoint-selection model"), _nonempty_string(tag, f"{label} endpoint-selection tag"), ) for model_id, tag in selections.items() ) ) for field in ("ignored_model_prefixes", "known_deprecated_models"): if field in discovery: _string_list(discovery[field], f"{label}.discovery.{field}") captures = snapshot["captures"] if not isinstance(captures, list): raise ValueError(f"{label}.captures must be a list") normalized_captures: list[dict[str, Any]] = [] for index, raw_capture in enumerate(captures): capture = _object(raw_capture, f"{label}.captures[{index}]") unknown_capture = sorted(set(capture) - CAPTURE_FIELDS) missing_capture = sorted(CAPTURE_FIELDS - set(capture)) if unknown_capture or missing_capture: raise ValueError( f"{label}.captures[{index}] fields are invalid: " f"missing={missing_capture}, unknown={unknown_capture}" ) source = _nonempty_string(capture["source"], f"{label}.captures[{index}].source") if source not in sources: raise ValueError(f"{label}.captures[{index}] references unknown source {source!r}") _utc_timestamp(capture["retrieved_at"], f"{label}.captures[{index}].retrieved_at") _sha256(capture["response_sha256"], f"{label}.captures[{index}].response_sha256") normalized_captures.append(capture) records = _object(snapshot["models"], f"{label}.models") if not records: raise ValueError(f"{label}.models must not be empty") expected_hash = _sha256(snapshot["models_sha256"], f"{label}.models_sha256") actual_hash = stable_hash(records) if actual_hash != expected_hash: raise ValueError( f"{label}.models integrity mismatch: expected {expected_hash}, calculated {actual_hash}" ) normalized_records: dict[str, dict[str, Any]] = {} binding_keys: dict[tuple[str, str], str] = {} for key, raw_record in records.items(): record = _object(raw_record, f"{label}.models.{key}") unknown_record = sorted(set(record) - MODEL_RECORD_FIELDS) missing_record = sorted({"catalog"} - set(record)) if unknown_record or missing_record: raise ValueError( f"{label}.models.{key} fields are invalid: " f"missing={missing_record}, unknown={unknown_record}" ) catalog = _validate_catalog_record(inventory, key, record["catalog"], sources) binding = (str(catalog["route"]), str(catalog["upstream_id"])) previous = binding_keys.get(binding) if previous is not None: raise ValueError( f"{label}.models.{previous} and {key} duplicate route/upstream binding {binding!r}" ) binding_keys[binding] = key normalized = {"catalog": catalog} if "evidence" in record: normalized["evidence"] = _object(record["evidence"], f"{label}.models.{key}.evidence") normalized_records[key] = normalized return { "schema_version": MMO_SCHEMA_VERSION, "inventory": inventory, "adapter": adapter, "fingerprint_fields": fingerprint_fields, "as_of": snapshot["as_of"], "dynamic": snapshot["dynamic"], "sources": dict(sorted(sources.items())), "discovery": discovery, "captures": normalized_captures, "models": dict(sorted(normalized_records.items())), "models_sha256": expected_hash, } def load_inventory_snapshots(directory: Path) -> list[dict[str, Any]]: """Load every ``*.json`` snapshot in a directory in deterministic order.""" if not directory.is_dir(): raise FileNotFoundError(f"inventory snapshot directory not found: {directory}") snapshots: list[dict[str, Any]] = [] seen: set[str] = set() paths = sorted(directory.glob("*.json"), key=lambda item: item.name) if not paths: raise ValueError(f"inventory snapshot directory is empty: {directory}") for path in paths: if path.is_symlink() or not path.is_file(): raise ValueError(f"inventory snapshot must be a regular non-symlink file: {path}") snapshot = validate_inventory_snapshot( read_json(path), label=str(path), expected_inventory=path.stem ) inventory = str(snapshot["inventory"]) if inventory in seen: raise ValueError(f"duplicate inventory snapshot: {inventory}") seen.add(inventory) snapshots.append(snapshot) return snapshots def build_inventory_snapshot( *, inventory: str, adapter: str, fingerprint_fields: Sequence[str], as_of: str, dynamic: bool, sources: Mapping[str, str], discovery: Mapping[str, Any], captures: Sequence[Mapping[str, Any]], models: Mapping[str, Mapping[str, Any]], ) -> dict[str, Any]: """Build and validate a common envelope from catalog/evidence records.""" records = {str(key): dict(value) for key, value in sorted(models.items())} snapshot: dict[str, Any] = { "schema_version": MMO_SCHEMA_VERSION, "inventory": inventory, "adapter": adapter, "fingerprint_fields": list(fingerprint_fields), "as_of": as_of, "dynamic": dynamic, "sources": dict(sorted(sources.items())), "discovery": dict(discovery), "captures": [dict(value) for value in captures], "models": records, "models_sha256": stable_hash(records), } return validate_inventory_snapshot(snapshot) def rehash_inventory_snapshot(value: Any) -> dict[str, Any]: """Recalculate the common model-record digest after an intentional review edit.""" snapshot = _object(value, "inventory snapshot") records = _object(snapshot.get("models"), "inventory snapshot.models") updated = dict(snapshot) updated["models_sha256"] = stable_hash(records) return validate_inventory_snapshot(updated) def model_record_fingerprint(value: Mapping[str, Any]) -> dict[str, Any]: """Return the complete, route-aware fingerprint represented by one record. Fingerprints intentionally retain upstream evidence that cannot be flattened into the executable catalog (tiered prices, exact endpoint metadata, and gateway-supported parameters). An ID match therefore cannot masquerade as a capability match. """ record = _object(value, "inventory model record") catalog = _object(record.get("catalog"), "inventory model record.catalog") evidence = _object(record.get("evidence", {}), "inventory model record.evidence") models_dev = _object(evidence.get("models_dev", {}), "models-dev evidence") docs = _object(evidence.get("docs", {}), "provider docs evidence") codex_runtime = _object(evidence.get("codex_runtime", {}), "Codex runtime evidence") selected_endpoint = _object(evidence.get("selected_endpoint", {}), "selected endpoint evidence") raw_pricing = evidence.get("pricing") if raw_pricing is None: raw_pricing = docs.get("pricing") if raw_pricing is None: raw_pricing = models_dev.get("cost") if raw_pricing is None: raw_pricing = { key: catalog[key] for key in ( "input_cost_per_million", "output_cost_per_million", "cached_input_cost_per_million", "cache_write_input_cost_per_million", ) if key in catalog } if codex_runtime: raw_pricing = {"billing_mode": "chatgpt_subscription"} if selected_endpoint: raw_pricing = {"model": raw_pricing, "selected_endpoint": selected_endpoint["pricing"]} limits_evidence: Any = None if "context_length" in evidence or "top_provider" in evidence: limits_evidence = { "context_length": evidence.get("context_length"), "top_provider": evidence.get("top_provider"), } elif models_dev.get("limit") is not None: limits_evidence = models_dev.get("limit") if limits_evidence is None: limits_evidence = { "context_window": catalog.get("context_window"), "max_output_tokens": catalog.get("max_output_tokens"), } if codex_runtime: limits_evidence = {"context_window": codex_runtime.get("context_window")} if selected_endpoint: limits_evidence = { "model": limits_evidence, "selected_endpoint": { "context_length": selected_endpoint["context_length"], "max_completion_tokens": selected_endpoint["max_completion_tokens"], }, } modalities_evidence: Any = evidence.get("architecture") if modalities_evidence is None: modalities_evidence = models_dev.get("modalities") if modalities_evidence is None: modalities_evidence = { "input": catalog.get("modalities", []), "output": catalog.get("output_modalities", []), } if codex_runtime: modalities_evidence = { "input": codex_runtime.get("input_modalities", []), "output": ["text"], } supported_parameters = evidence.get("supported_parameters") if supported_parameters is None: supported_parameters = models_dev.get("supported_parameters") if supported_parameters is None: supported_parameters = sorted( key for key, enabled in ( ("tools", catalog.get("tool_calling")), ("parallel_tool_calls", catalog.get("parallel_tool_calls")), ("structured_outputs", catalog.get("structured_output")), ("reasoning", catalog.get("reasoning_levels") not in (None, ["none"])), ) if enabled ) if selected_endpoint: supported_parameters = selected_endpoint["supported_parameters"] if codex_runtime: codex_tools = _object(codex_runtime.get("tools", {}), "Codex runtime tools") supported_parameters = sorted( key for key, value in codex_tools.items() if value not in (None, False, "none") ) raw_reasoning: Any = evidence.get("reasoning") if raw_reasoning is None and "reasoning" in models_dev: raw_reasoning = { "enabled": models_dev.get("reasoning"), "options": models_dev.get("reasoning_options", []), } if raw_reasoning is None: raw_reasoning = { "levels": catalog.get("reasoning_levels", []), "default": catalog.get("default_reasoning"), "summaries": catalog.get("supports_reasoning_summaries", False), } if codex_runtime: raw_reasoning = codex_runtime.get("reasoning") endpoint_metadata: Any = evidence.get("endpoint_metadata") protocol_resolution = evidence.get("protocol_resolution") if endpoint_metadata is None and docs.get("endpoint") is not None: endpoint_metadata = docs.get("endpoint") if protocol_resolution is not None: endpoint_metadata = { "documented_endpoint": endpoint_metadata, "protocol_resolution": _object( protocol_resolution, "provider protocol-resolution evidence" ), } if endpoint_metadata is None and models_dev: endpoint_metadata = { "provider": models_dev.get("provider"), "interleaved": models_dev.get("interleaved"), } if selected_endpoint: endpoint_metadata = { "model": endpoint_metadata, "selected_endpoint": { key: selected_endpoint[key] for key in ( "model_id", "provider_name", "tag", "quantization", "status", "zdr", ) }, } if endpoint_metadata is None: endpoint_metadata = { "route": catalog.get("route"), "route_policy": catalog.get("route_policy"), } if codex_runtime: endpoint_metadata = { key: codex_runtime.get(key) for key in ( "comp_hash", "multi_agent_version", "service_tiers", "supported_in_api", ) } deprecation: Any = { "availability": catalog.get("availability"), "expiration_date": evidence.get("expiration_date"), "deprecation_date": docs.get("deprecation_date"), } canonical_slug = ( evidence.get("canonical_slug") or models_dev.get("id") or catalog.get("upstream_id") ) return { "upstream_id": catalog.get("upstream_id"), "canonical_slug": canonical_slug, "pricing": raw_pricing, "limits": limits_evidence, "modalities": modalities_evidence, "tools": ( codex_runtime.get("tools") if codex_runtime else { "tool_calling": models_dev.get("tool_call", catalog.get("tool_calling")), "parallel_tool_calls": catalog.get("parallel_tool_calls"), } ), "reasoning": raw_reasoning, "structured_output": ( codex_runtime.get("structured_output") if codex_runtime else models_dev.get("structured_output", catalog.get("structured_output")) ), "supported_parameters": supported_parameters, "deprecation": deprecation, "endpoint_metadata": endpoint_metadata, } def model_record_evidence_coverage(value: Mapping[str, Any]) -> dict[str, bool]: """Report which fingerprint fields were established by observed evidence.""" record = _object(value, "inventory model record") evidence = _object(record.get("evidence", {}), "inventory model record.evidence") if evidence.get("unverified_live_only") is True: return {field: field == "upstream_id" for field in FULL_FINGERPRINT_FIELDS} if not evidence: return {field: False for field in FULL_FINGERPRINT_FIELDS} models_dev = evidence.get("models_dev") docs = evidence.get("docs") is_openrouter = "supported_parameters" in evidence and "top_provider" in evidence if is_openrouter: return {field: True for field in FULL_FINGERPRINT_FIELDS} if isinstance(models_dev, Mapping) and isinstance(evidence.get("live"), Mapping): # A Models.dev model without an override inherits the access product's # validated default OpenAI-compatible package, so its endpoint family # is still known. endpoint_known = bool(isinstance(docs, Mapping) and docs.get("endpoint")) or bool( models_dev ) return { "upstream_id": True, "canonical_slug": True, "pricing": isinstance(models_dev.get("cost"), Mapping) or bool(isinstance(docs, Mapping) and docs.get("pricing")), "limits": isinstance(models_dev.get("limit"), Mapping), "modalities": isinstance(models_dev.get("modalities"), Mapping), "tools": isinstance(models_dev.get("tool_call"), bool), "reasoning": isinstance(models_dev.get("reasoning"), bool), "structured_output": isinstance(models_dev.get("structured_output", False), bool), "supported_parameters": isinstance(models_dev.get("tool_call"), bool), "deprecation": isinstance(docs, Mapping), "endpoint_metadata": endpoint_known, } # Other adapters must opt in with an explicit list. This is what prevents # an authenticated /models response containing only IDs from being called # a complete verification. verified = evidence.get("verified_fingerprint_fields", []) return {field: field in verified for field in FULL_FINGERPRINT_FIELDS} def compare_inventory_fingerprints( expected: Mapping[str, Any], observed: Mapping[str, Any] ) -> dict[str, Any]: """Compare two normalized snapshots, requiring full observed evidence.""" expected_snapshot = validate_inventory_snapshot(expected) observed_snapshot = validate_inventory_snapshot(observed) if expected_snapshot["inventory"] != observed_snapshot["inventory"]: raise ValueError("cannot compare snapshots from different inventories") def bindings(snapshot: Mapping[str, Any]) -> dict[tuple[str, str], Mapping[str, Any]]: return { (str(record["catalog"]["route"]), str(record["catalog"]["upstream_id"])): record for record in snapshot["models"].values() } expected_records = bindings(expected_snapshot) observed_records = bindings(observed_snapshot) missing = sorted(expected_records.keys() - observed_records.keys()) unknown = sorted(observed_records.keys() - expected_records.keys()) mismatches: list[dict[str, Any]] = [] incomplete: list[dict[str, Any]] = [] for binding in sorted(expected_records.keys() & observed_records.keys()): expected_fingerprint = model_record_fingerprint(expected_records[binding]) observed_fingerprint = model_record_fingerprint(observed_records[binding]) changed_fields = [ field for field in FULL_FINGERPRINT_FIELDS if expected_fingerprint[field] != observed_fingerprint[field] ] if changed_fields: mismatches.append( { "route": binding[0], "upstream_id": binding[1], "fields": changed_fields, "expected_sha256": stable_hash(expected_fingerprint), "observed_sha256": stable_hash(observed_fingerprint), } ) coverage = model_record_evidence_coverage(observed_records[binding]) missing_evidence = [field for field in FULL_FINGERPRINT_FIELDS if not coverage[field]] if missing_evidence: incomplete.append( { "route": binding[0], "upstream_id": binding[1], "fields": missing_evidence, } ) exact = not missing and not unknown and not mismatches and not incomplete return { "expected_count": len(expected_records), "observed_count": len(observed_records), "missing_from_observed": [list(item) for item in missing], "unknown_to_catalog": [list(item) for item in unknown], "fingerprint_mismatches": mismatches, "incomplete_evidence": incomplete, "verified_fields": list(FULL_FINGERPRINT_FIELDS) if exact else [], "exact": exact, } def route_catalog_key(route_key: str, upstream_id: str, used: set[str]) -> str: """Derive a route-qualified, bounded key with a deterministic collision suffix.""" validate_id(route_key, "route id") slug = _nonempty_string(upstream_id, "upstream model id") normalized = ( "".join( char if char.isascii() and char.isalnum() else "_" for char in slug.casefold() ).strip("_") or "model" ) prefix = route_key + "__" if len(prefix) >= 64: raise ValueError(f"route id {route_key!r} is too long for a qualified model key") candidate = prefix + normalized if len(candidate) <= 64 and candidate not in used: return candidate digest = stable_hash(slug)[:10] suffix = "_" + digest candidate = prefix + normalized[: 64 - len(prefix) - len(suffix)] + suffix counter = 1 while candidate in used: counter_suffix = f"_{counter}" candidate = candidate[: 64 - len(counter_suffix)] + counter_suffix counter += 1 return candidate def _openrouter_model_rows(document: Any) -> list[dict[str, Any]]: root = _object(document, "OpenRouter models response") raw_rows = root.get("data") if not isinstance(raw_rows, list): raise ValueError("OpenRouter models response.data must be a list") rows = [ _object(row, f"OpenRouter models response.data[{index}]") for index, row in enumerate(raw_rows) ] if not rows: raise ValueError("OpenRouter models response.data must not be empty") if "total_count" not in root or "links" not in root: raise ValueError("OpenRouter models response lacks required pagination metadata") total_count = root["total_count"] _positive_int(total_count, "OpenRouter models response.total_count", allow_zero=True) if total_count != len(rows): raise ValueError( "OpenRouter models response is paginated or incomplete: " f"total_count={total_count}, rows={len(rows)}" ) links_object = _object(root["links"], "OpenRouter models response.links") if "next" not in links_object: raise ValueError("OpenRouter models response.links lacks required next field") if links_object["next"] not in (None, ""): raise ValueError("OpenRouter models response has an unconsumed next page") return rows def _openrouter_evidence(raw: Mapping[str, Any]) -> dict[str, Any]: upstream_id = _nonempty_string(raw.get("id"), "OpenRouter model id") name = _nonempty_string(raw.get("name"), f"OpenRouter model {upstream_id}.name") context = _positive_int( raw.get("context_length"), f"OpenRouter model {upstream_id}.context_length" ) architecture = _object(raw.get("architecture"), f"OpenRouter model {upstream_id}.architecture") input_modalities = _string_list( architecture.get("input_modalities"), f"OpenRouter model {upstream_id}.architecture.input_modalities", allow_empty=False, ) output_modalities = _string_list( architecture.get("output_modalities"), f"OpenRouter model {upstream_id}.architecture.output_modalities", allow_empty=False, ) parameters = sorted( _string_list( raw.get("supported_parameters"), f"OpenRouter model {upstream_id}.supported_parameters", ) ) pricing = _object(raw.get("pricing"), f"OpenRouter model {upstream_id}.pricing") top_provider = _object(raw.get("top_provider"), f"OpenRouter model {upstream_id}.top_provider") top_context = top_provider.get("context_length") if top_context is not None: _positive_int(top_context, f"OpenRouter model {upstream_id}.top_provider.context_length") max_output = top_provider.get("max_completion_tokens") if max_output is not None: _positive_int( max_output, f"OpenRouter model {upstream_id}.top_provider.max_completion_tokens", ) expiration = raw.get("expiration_date") if expiration is not None: _iso_date(expiration, f"OpenRouter model {upstream_id}.expiration_date") evidence: dict[str, Any] = { "id": upstream_id, "canonical_slug": _nonempty_string( raw.get("canonical_slug"), f"OpenRouter model {upstream_id}.canonical_slug" ), "name": name, "context_length": context, "architecture": { "input_modalities": input_modalities, "output_modalities": output_modalities, }, "supported_parameters": parameters, "pricing": dict(sorted(pricing.items())), "top_provider": { "context_length": top_context, "max_completion_tokens": max_output, "is_moderated": top_provider.get("is_moderated"), }, "expiration_date": expiration, "endpoint_metadata": { "links": _object(raw.get("links", {}), f"OpenRouter model {upstream_id}.links"), "per_request_limits": raw.get("per_request_limits"), }, } if raw.get("reasoning") is not None: evidence["reasoning"] = _object( raw["reasoning"], f"OpenRouter model {upstream_id}.reasoning" ) if raw.get("alias_target") is not None: evidence["alias_target"] = _object( raw["alias_target"], f"OpenRouter model {upstream_id}.alias_target" ) return evidence def _openrouter_zdr_endpoint(document: Any, model_id: str, endpoint_tag: str) -> dict[str, Any]: root = _object(document, "OpenRouter ZDR endpoints response") rows = root.get("data") if not isinstance(rows, list) or not rows: raise ValueError("OpenRouter ZDR endpoints response.data must be non-empty") matches = [ _object(row, "OpenRouter ZDR endpoint") for row in rows if isinstance(row, Mapping) and row.get("model_id") == model_id and row.get("tag") == endpoint_tag ] if len(matches) != 1: raise ValueError( f"OpenRouter model {model_id} must have exactly one ZDR endpoint {endpoint_tag!r}" ) row = matches[0] parameters = sorted( _string_list( row.get("supported_parameters"), f"OpenRouter endpoint {model_id}/{endpoint_tag}.supported_parameters", allow_empty=False, ) ) required = {"tools", "tool_choice", "reasoning_effort"} if not required.issubset(parameters): raise ValueError( f"OpenRouter ZDR endpoint {endpoint_tag!r} for {model_id} lacks {sorted(required)}" ) status = row.get("status") if not isinstance(status, int) or isinstance(status, bool) or status != 0: raise ValueError(f"OpenRouter ZDR endpoint {endpoint_tag!r} for {model_id} is not healthy") context = _positive_int( row.get("context_length"), f"OpenRouter endpoint {model_id}/{endpoint_tag}.context_length", ) max_completion = row.get("max_completion_tokens") if max_completion is not None: max_completion = _positive_int( max_completion, f"OpenRouter endpoint {model_id}/{endpoint_tag}.max_completion_tokens", ) return { "model_id": model_id, "provider_name": _nonempty_string( row.get("provider_name"), f"OpenRouter endpoint {model_id}/{endpoint_tag}.provider_name", ), "tag": endpoint_tag, "quantization": _nonempty_string( row.get("quantization"), f"OpenRouter endpoint {model_id}/{endpoint_tag}.quantization", ), "context_length": context, "max_completion_tokens": max_completion, "pricing": _object( row.get("pricing"), f"OpenRouter endpoint {model_id}/{endpoint_tag}.pricing" ), "supported_parameters": parameters, "zdr": True, "status": status, } def _decimal_price(pricing: Mapping[str, Any], field: str, model_id: str) -> Decimal | None: raw = pricing.get(field) if raw is None: return None if isinstance(raw, bool) or not isinstance(raw, (str, int, float)): raise ValueError(f"OpenRouter model {model_id}.pricing.{field} must be numeric") try: value = Decimal(str(raw)) except InvalidOperation as exc: raise ValueError(f"OpenRouter model {model_id}.pricing.{field} must be numeric") from exc if not value.is_finite(): raise ValueError(f"OpenRouter model {model_id}.pricing.{field} must be finite") return value def _per_million(value: Decimal, model_id: str, field: str) -> float: converted = float(value * Decimal(1_000_000)) if not math.isfinite(converted) or converted < 0: raise ValueError(f"OpenRouter model {model_id}.pricing.{field} is out of range") return converted def openrouter_catalog_pricing(evidence: Mapping[str, Any]) -> dict[str, float]: """Translate only scalar text-token prices that the MMO schema can represent.""" model_id = _nonempty_string(evidence.get("id"), "OpenRouter evidence model id") pricing = _object(evidence.get("pricing"), f"OpenRouter model {model_id}.pricing") prompt = _decimal_price(pricing, "prompt", model_id) completion = _decimal_price(pricing, "completion", model_id) if prompt is None or completion is None: raise ValueError(f"OpenRouter model {model_id} lacks prompt/completion pricing") request = _decimal_price(pricing, "request", model_id) or Decimal(0) reasoning = _decimal_price(pricing, "internal_reasoning", model_id) cache_read = _decimal_price(pricing, "input_cache_read", model_id) cache_write = _decimal_price(pricing, "input_cache_write", model_id) for field, value in ( ("request", request), ("internal_reasoning", reasoning), ("input_cache_read", cache_read), ("input_cache_write", cache_write), ): if value is not None and value < 0: raise ValueError(f"OpenRouter model {model_id}.pricing.{field} must not be negative") negative_tokens = prompt < 0 or completion < 0 if negative_tokens: if prompt != Decimal(-1) or completion != Decimal(-1): raise ValueError( f"OpenRouter model {model_id} has inconsistent dynamic token-price sentinels" ) return {} overrides = pricing.get("overrides", []) if not isinstance(overrides, list): raise ValueError(f"OpenRouter model {model_id}.pricing.overrides must be a list") if overrides or request != 0 or (reasoning is not None and reasoning != completion): return {} result = { "input_cost_per_million": _per_million(prompt, model_id, "prompt"), "output_cost_per_million": _per_million(completion, model_id, "completion"), } if cache_read is not None: result["cached_input_cost_per_million"] = _per_million( cache_read, model_id, "input_cache_read" ) if cache_write is not None: result["cache_write_input_cost_per_million"] = _per_million( cache_write, model_id, "input_cache_write" ) return result def openrouter_reasoning(evidence: Mapping[str, Any]) -> tuple[list[str], str]: """Return only reasoning efforts explicitly advertised by OpenRouter.""" model_id = _nonempty_string(evidence.get("id"), "OpenRouter evidence model id") raw = evidence.get("reasoning") if raw is None: return ["none"], "none" reasoning = _object(raw, f"OpenRouter model {model_id}.reasoning") mandatory = reasoning.get("mandatory", False) default_enabled = reasoning.get("default_enabled") if not isinstance(mandatory, bool): raise ValueError(f"OpenRouter model {model_id}.reasoning.mandatory must be boolean") if default_enabled is not None and not isinstance(default_enabled, bool): raise ValueError( f"OpenRouter model {model_id}.reasoning.default_enabled must be boolean or null" ) if "supported_efforts" not in reasoning: return ["none"], "none" supported_raw = reasoning["supported_efforts"] all_gateway_efforts = supported_raw is None supported = ( list(OPENROUTER_REASONING_ORDER) if all_gateway_efforts else _string_list( supported_raw, f"OpenRouter model {model_id}.reasoning.supported_efforts", allow_empty=False, ) ) unknown = sorted(set(supported) - set(OPENROUTER_REASONING_ORDER)) if unknown: raise ValueError(f"OpenRouter model {model_id} has unknown reasoning efforts: {unknown}") if mandatory and "none" in supported: if all_gateway_efforts: supported.remove("none") else: raise ValueError( f"OpenRouter model {model_id} is mandatory-reasoning but advertises none" ) levels = [item for item in OPENROUTER_REASONING_ORDER if item in supported] if not mandatory and "none" not in levels: levels.insert(0, "none") default_effort = reasoning.get("default_effort") if default_effort is not None and default_effort not in supported: raise ValueError(f"OpenRouter model {model_id}.reasoning.default_effort is not supported") if mandatory or default_enabled is True: default = default_effort if isinstance(default_effort, str) else "medium" else: default = "none" if default not in levels: raise ValueError(f"OpenRouter model {model_id} has an unusable default reasoning effort") return levels, default def openrouter_catalog_record(evidence: Mapping[str, Any], *, as_of: str) -> dict[str, Any]: """Derive one conservative text-only executable catalog record.""" audit_date = date.fromisoformat(_iso_date(as_of, "OpenRouter snapshot as_of")) model_id = _nonempty_string(evidence.get("id"), "OpenRouter evidence model id") name = _nonempty_string(evidence.get("name"), f"OpenRouter model {model_id}.name") context = _positive_int( evidence.get("context_length"), f"OpenRouter model {model_id}.context_length" ) architecture = _object( evidence.get("architecture"), f"OpenRouter model {model_id}.architecture" ) inputs = _string_list( architecture.get("input_modalities"), f"OpenRouter model {model_id}.architecture.input_modalities", allow_empty=False, ) outputs = _string_list( architecture.get("output_modalities"), f"OpenRouter model {model_id}.architecture.output_modalities", allow_empty=False, ) if "text" not in inputs or "text" not in outputs: raise ValueError(f"OpenRouter model {model_id} is not a text-input/text-output model") parameters = set( _string_list( evidence.get("supported_parameters"), f"OpenRouter model {model_id}.supported_parameters", ) ) reasoning_levels, default_reasoning = openrouter_reasoning(evidence) expiration = evidence.get("expiration_date") expired = False availability = "current" if expiration is not None: expiration_date = date.fromisoformat( _iso_date(expiration, f"OpenRouter model {model_id}.expiration_date") ) expired = expiration_date < audit_date availability = f"{'expired' if expired else 'expires'}-{expiration}" tool_calling = "tools" in parameters catalog: dict[str, Any] = { "maker": infer_model_maker(model_id), "route": "openrouter_openai_chat", "upstream_id": model_id, "display_name": f"{name} via OpenRouter", "description": f"{name} served through OpenRouter's reviewed text-only Switchyard route", "kind": "chat", "agent_compatible": tool_calling and not expired, "context_window": context, "reasoning_levels": reasoning_levels, "default_reasoning": default_reasoning, "modalities": ["text"], "output_modalities": ["text"], "tool_calling": tool_calling, "parallel_tool_calls": tool_calling and "parallel_tool_calls" in parameters, "supports_reasoning_summaries": False, "structured_output": "structured_outputs" in parameters, "availability": availability, "capability_confidence": "openrouter-models-api-snapshot", "source": "openrouter-models-api", "availability_source": "openrouter-models-api", "capability_source": "openrouter-models-api", "pricing_source": "openrouter-models-api", "inventory": "openrouter", "resource_group": "openrouter", } top_provider = _object( evidence.get("top_provider"), f"OpenRouter model {model_id}.top_provider" ) max_output = top_provider.get("max_completion_tokens") if max_output is not None: catalog["max_output_tokens"] = _positive_int( max_output, f"OpenRouter model {model_id}.top_provider.max_completion_tokens" ) catalog.update(openrouter_catalog_pricing(evidence)) selected_endpoint = evidence.get("selected_endpoint") if selected_endpoint is not None: endpoint = _object(selected_endpoint, f"OpenRouter model {model_id}.selected_endpoint") endpoint_tag = _nonempty_string( endpoint.get("tag"), f"OpenRouter model {model_id}.selected_endpoint.tag" ) catalog["route_policy"] = { "only": [endpoint_tag], "allow_fallbacks": False, "require_parameters": True, "data_collection": "deny", "zdr": True, } quantization = endpoint.get("quantization") if quantization not in (None, "unknown"): catalog["route_policy"]["quantizations"] = [quantization] return catalog def build_openrouter_snapshot( document: Any, zdr_document: Any, *, as_of: str, retrieved_at: str, response_sha256: str, zdr_response_sha256: str, source_url: str, zdr_source_url: str, endpoint_selections: Mapping[str, str], ) -> dict[str, Any]: """Normalize an official OpenRouter models response into the common envelope.""" _sha256(response_sha256, "OpenRouter source response SHA-256") _sha256(zdr_response_sha256, "OpenRouter ZDR source response SHA-256") _utc_timestamp(retrieved_at, "OpenRouter retrieval timestamp") source_url = _https_url(source_url, "OpenRouter models source URL") zdr_source_url = _https_url(zdr_source_url, "OpenRouter ZDR endpoints source URL") sources = dict(OPENROUTER_SOURCE_IDS) sources["openrouter-models-api"] = source_url sources["openrouter-zdr-endpoints"] = zdr_source_url selections = { _nonempty_string(model_id, "OpenRouter endpoint-selection model"): _nonempty_string( tag, "OpenRouter endpoint-selection tag" ) for model_id, tag in endpoint_selections.items() } if not selections: raise ValueError("OpenRouter endpoint_selections must not be empty") used: set[str] = set() records: dict[str, dict[str, Any]] = {} seen_ids: set[str] = set() for raw in sorted(_openrouter_model_rows(document), key=lambda row: str(row.get("id", ""))): evidence = _openrouter_evidence(raw) model_id = str(evidence["id"]) endpoint_tag = selections.get(model_id) if endpoint_tag is not None: evidence["selected_endpoint"] = _openrouter_zdr_endpoint( zdr_document, model_id, endpoint_tag ) if model_id in seen_ids: raise ValueError(f"duplicate OpenRouter model id: {model_id}") seen_ids.add(model_id) architecture = _object( evidence["architecture"], f"OpenRouter model {model_id}.architecture" ) if ( "text" not in architecture["input_modalities"] or "text" not in architecture["output_modalities"] ): continue key = route_catalog_key("openrouter_openai_chat", model_id, used) used.add(key) records[key] = { "catalog": openrouter_catalog_record(evidence, as_of=as_of), "evidence": evidence, } if not records: raise ValueError("OpenRouter response did not contain any text-input/text-output models") missing_selections = sorted(set(selections) - seen_ids) if missing_selections: raise ValueError( "OpenRouter endpoint selections reference unlisted models: " + ", ".join(missing_selections) ) return build_inventory_snapshot( inventory="openrouter", adapter="openrouter_models_api", fingerprint_fields=FULL_FINGERPRINT_FIELDS, as_of=as_of, dynamic=True, sources=sources, discovery={ "endpoint": source_url, "endpoint_selections": dict(sorted(selections.items())), }, captures=[ { "source": "openrouter-models-api", "retrieved_at": retrieved_at, "response_sha256": response_sha256, }, { "source": "openrouter-zdr-endpoints", "retrieved_at": retrieved_at, "response_sha256": zdr_response_sha256, }, ], models=records, ) def opencode_zen_reasoning(metadata: Mapping[str, Any]) -> tuple[list[str], str]: """Translate only Models.dev effort controls representable by Codex.""" model_id = _nonempty_string(metadata.get("id"), "OpenCode Zen metadata model id") reasoning = metadata.get("reasoning") if not isinstance(reasoning, bool): raise ValueError(f"OpenCode Zen model {model_id}.reasoning must be boolean") if not reasoning: return ["none"], "none" raw_options = metadata.get("reasoning_options", []) if not isinstance(raw_options, list): raise ValueError(f"OpenCode Zen model {model_id}.reasoning_options must be a list") effort_values: list[str] | None = None has_toggle = False seen_types: set[str] = set() for index, raw_option in enumerate(raw_options): option = _object(raw_option, f"OpenCode Zen model {model_id}.reasoning_options[{index}]") option_type = _nonempty_string( option.get("type"), f"OpenCode Zen model {model_id}.reasoning_options[{index}].type" ) if option_type in seen_types: raise ValueError( f"OpenCode Zen model {model_id} has duplicate reasoning option {option_type!r}" ) seen_types.add(option_type) if option_type == "effort": effort_values = _string_list( option.get("values"), f"OpenCode Zen model {model_id}.reasoning_options[{index}].values", allow_empty=False, ) unknown = sorted(set(effort_values) - set(CODEX_REASONING_ORDER)) if unknown: raise ValueError( f"OpenCode Zen model {model_id} has unknown reasoning efforts: {unknown}" ) elif option_type == "toggle": has_toggle = True elif option_type != "budget_tokens": raise ValueError( f"OpenCode Zen model {model_id} has unknown reasoning option {option_type!r}" ) if effort_values is None: # Toggle and token-budget controls have no faithful Codex effort value. return ["none"], "none" levels = [item for item in CODEX_REASONING_ORDER if item in effort_values] if has_toggle and "none" not in levels: levels.insert(0, "none") if "none" in levels: default = "none" elif "medium" in levels: default = "medium" elif "high" in levels: default = "high" else: default = levels[0] return levels, default def opencode_zen_catalog_pricing(docs_evidence: Mapping[str, Any]) -> dict[str, float]: """Retain one non-tiered rate row from an OpenCode billing table.""" model_id = _nonempty_string(docs_evidence.get("id"), "OpenCode Zen docs model id") raw_rows = docs_evidence.get("pricing", []) if not isinstance(raw_rows, list): raise ValueError(f"OpenCode Zen docs model {model_id}.pricing must be a list") rows = [ _object(row, f"OpenCode Zen docs model {model_id}.pricing[{index}]") for index, row in enumerate(raw_rows) ] # Multiple distinct rows mean the rate depends on request context, which # the catalog's flat per-million fields cannot represent. if len(rows) != 1: return {} row = rows[0] input_rate = row.get("input") output_rate = row.get("output") if input_rate is None and output_rate is None: # A complete dash-valued provider row means the operator publishes no # scalar token rate. Preserve the raw row as evidence without guessing # that an undocumented price is zero. return {} if (input_rate is None) != (output_rate is None): missing = "input" if input_rate is None else "output" raise ValueError(f"OpenCode Zen docs model {model_id}.{missing} is required") result: dict[str, float] = {} for source, target in ( ("input", "input_cost_per_million"), ("output", "output_cost_per_million"), ("cache_read", "cached_input_cost_per_million"), ("cache_write", "cache_write_input_cost_per_million"), ): value = row.get(source) if value is None: if source in {"input", "output"}: raise ValueError(f"OpenCode Zen docs model {model_id}.{source} is required") continue if ( isinstance(value, bool) or not isinstance(value, (int, float)) or not math.isfinite(float(value)) or value < 0 ): raise ValueError( f"OpenCode Zen docs model {model_id}.{source} must be non-negative and finite" ) result[target] = float(value) return result def _markdown_table( document: str, heading: str, expected_headers: Sequence[str] ) -> list[list[str]]: lines = document.splitlines() try: heading_index = next(index for index, line in enumerate(lines) if line.strip() == heading) except StopIteration as exc: raise ValueError(f"OpenCode Zen docs lack {heading!r}") from exc def cells(line: str) -> list[str]: return [item.strip().strip("`") for item in line.strip("|").split("|")] tables: list[list[str]] = [] current: list[str] = [] for line in lines[heading_index + 1 :]: stripped = line.strip() if stripped.startswith("## "): break if stripped.startswith("|") and stripped.endswith("|"): current.append(stripped) elif current: tables.append(current) current = [] if current: tables.append(current) table_lines = next( ( table for table in tables if len(table) >= 3 and cells(table[0]) == list(expected_headers) ), None, ) if table_lines is None: observed = [cells(table[0]) for table in tables if table] raise ValueError( f"provider docs {heading!r} lack the required table: " f"expected={list(expected_headers)}, observed={observed}" ) headers = cells(table_lines[0]) separator = cells(table_lines[1]) if len(separator) != len(headers) or any( not re.fullmatch(r":?-{3,}:?", item.replace(" ", "")) for item in separator ): raise ValueError(f"OpenCode Zen docs {heading!r} separator is invalid") rows = [cells(line) for line in table_lines[2:]] if any(len(row) != len(headers) for row in rows): raise ValueError(f"OpenCode Zen docs {heading!r} contains a malformed row") return rows def _opencode_zen_name_key(value: str) -> str: return "".join(character.casefold() for character in value if character.isalnum()) _OPENCODE_PRICING_TIER_SUFFIX = re.compile( r"\s+\((?:off-peak|peak|[<>≤≥]=?\s*[0-9]+(?:\.[0-9]+)?\s*[KMGT]?\s+tokens)\)$", re.IGNORECASE, ) def _opencode_pricing_base_name(value: str) -> str: """Remove provider-defined context/time tier labels from a model name.""" base_name = value.strip() while True: stripped = _OPENCODE_PRICING_TIER_SUFFIX.sub("", base_name).strip() if stripped == base_name: return base_name base_name = stripped def _opencode_zen_price(value: str, label: str) -> float | None: if value == "-": return None if value == "Free": return 0.0 match = re.fullmatch(r"\$([0-9]+(?:\.[0-9]+)?)", value) if not match: raise ValueError(f"{label} has an unsupported price {value!r}") return float(match.group(1)) def _opencode_zen_docs( document: Any, metadata_by_id: Mapping[str, Mapping[str, Any]] ) -> dict[str, dict[str, Any]]: if not isinstance(document, str) or not document.strip(): raise ValueError("OpenCode Zen docs source must be non-empty UTF-8 text") endpoint_rows = _markdown_table( document, "## Endpoints", ("Model", "Model ID", "Endpoint", "AI SDK Package"), ) pricing_rows = _markdown_table( document, "## Pricing", ("Model", "Input", "Output", "Cached Read", "Cached Write"), ) deprecation_rows = _markdown_table( document, "### Deprecated models", ("Model", "Deprecation date"), ) docs_by_id: dict[str, dict[str, Any]] = {} name_to_id: dict[str, str] = {} def register_name(name: str, model_id: str) -> None: key = _opencode_zen_name_key(name) if not key: raise ValueError(f"OpenCode Zen docs contain an empty model name for {model_id}") previous = name_to_id.get(key) if previous is not None and previous != model_id: raise ValueError( f"OpenCode Zen docs model name {name!r} is ambiguous: {previous}, {model_id}" ) name_to_id[key] = model_id for model_id, metadata in metadata_by_id.items(): register_name( _nonempty_string(metadata.get("name"), f"OpenCode Zen model {model_id}.name"), model_id, ) for name, model_id, endpoint, npm in endpoint_rows: model_id = _nonempty_string(model_id, "OpenCode Zen docs endpoint model id") endpoint = _https_url(endpoint, f"OpenCode Zen docs model {model_id}.endpoint") npm = _nonempty_string(npm, f"OpenCode Zen docs model {model_id}.npm") if npm not in OPENCODE_ZEN_ROUTE_BY_NPM: raise ValueError(f"OpenCode Zen docs model {model_id} has unsupported package {npm!r}") expected_suffix = { "@ai-sdk/openai": "/responses", "@ai-sdk/openai-compatible": "/chat/completions", "@ai-sdk/anthropic": "/messages", "@ai-sdk/google": f"/models/{model_id}", }[npm] if endpoint != f"https://opencode.ai/zen/v1{expected_suffix}": raise ValueError( f"OpenCode Zen docs model {model_id} has unexpected endpoint {endpoint}" ) evidence = docs_by_id.setdefault(model_id, {"id": model_id}) if "endpoint" in evidence: raise ValueError(f"duplicate OpenCode Zen docs endpoint row: {model_id}") evidence["endpoint"] = {"name": name, "url": endpoint, "npm": npm} register_name(name, model_id) for label, input_price, output_price, cache_read, cache_write in pricing_rows: base_name = _opencode_pricing_base_name(label) pricing_model_id = name_to_id.get(_opencode_zen_name_key(base_name)) if pricing_model_id is None: raise ValueError(f"OpenCode Zen docs pricing model is unresolved: {label!r}") row: dict[str, Any] = { "label": label, "input": _opencode_zen_price(input_price, f"OpenCode Zen docs {label}.input"), "output": _opencode_zen_price(output_price, f"OpenCode Zen docs {label}.output"), } read_rate = _opencode_zen_price(cache_read, f"OpenCode Zen docs {label}.cache_read") write_rate = _opencode_zen_price(cache_write, f"OpenCode Zen docs {label}.cache_write") if read_rate is not None: row["cache_read"] = read_rate if write_rate is not None: row["cache_write"] = write_rate pricing = docs_by_id.setdefault(pricing_model_id, {"id": pricing_model_id}).setdefault( "pricing", [] ) if not isinstance(pricing, list): raise ValueError( f"OpenCode Zen docs pricing evidence is malformed for {pricing_model_id}" ) if row not in pricing: pricing.append(row) for name, raw_date in deprecation_rows: deprecated_model_id = name_to_id.get(_opencode_zen_name_key(name)) if deprecated_model_id is None: # A deprecated, no-longer-inventoried row is irrelevant to the live # join but remains present in the exact source capture. continue try: deprecation_date = datetime.strptime(raw_date, "%B %d, %Y").date().isoformat() except ValueError as exc: raise ValueError( "OpenCode Zen docs model " f"{deprecated_model_id} has invalid deprecation date {raw_date!r}" ) from exc evidence = docs_by_id.setdefault(deprecated_model_id, {"id": deprecated_model_id}) if "deprecation_date" in evidence: raise ValueError(f"duplicate OpenCode Zen docs deprecation row: {deprecated_model_id}") evidence["deprecation_date"] = deprecation_date return docs_by_id def _opencode_zen_listing_rows(document: Any) -> list[dict[str, Any]]: root = _object(document, "OpenCode Zen models response") if root.get("object") != "list": raise ValueError("OpenCode Zen models response.object must be 'list'") raw_rows = root.get("data") if not isinstance(raw_rows, list) or not raw_rows: raise ValueError("OpenCode Zen models response.data must be a non-empty list") return [ _object(row, f"OpenCode Zen models response.data[{index}]") for index, row in enumerate(raw_rows) ] def _opencode_zen_live_evidence(row: Mapping[str, Any]) -> dict[str, Any]: """Validate a live row while excluding its request-time ``created`` value. Zen currently assigns the same current Unix timestamp to every model on each listing request. The raw response digest retains that exact capture; catalog-record evidence keeps only stable availability fields so a refresh does not manufacture 62 model changes when the ID set is unchanged. """ model_id = _nonempty_string(row.get("id"), "OpenCode Zen live model id") if row.get("object") != "model": raise ValueError(f"OpenCode Zen live model {model_id}.object must be 'model'") if row.get("owned_by") != "opencode": raise ValueError(f"OpenCode Zen live model {model_id}.owned_by must be 'opencode'") _positive_int(row.get("created"), f"OpenCode Zen live model {model_id}.created") return {"id": model_id, "object": "model", "owned_by": "opencode"} def _opencode_zen_metadata(document: Any) -> dict[str, dict[str, Any]]: root = _object(document, "Models.dev response") provider = _object(root.get("opencode"), "Models.dev opencode provider") if provider.get("id") != "opencode": raise ValueError("Models.dev opencode provider has an unexpected id") if provider.get("api") != "https://opencode.ai/zen/v1": raise ValueError("Models.dev opencode provider has an unexpected API base") if provider.get("env") != ["OPENCODE_API_KEY"]: raise ValueError("Models.dev opencode provider has an unexpected credential contract") models = _object(provider.get("models"), "Models.dev opencode models") if not models: raise ValueError("Models.dev opencode models must not be empty") return { _nonempty_string(key, "Models.dev OpenCode Zen model key"): _object( value, f"Models.dev OpenCode Zen model {key}" ) for key, value in models.items() } def _opencode_zen_catalog_record( metadata: Mapping[str, Any], docs_evidence: Mapping[str, Any], *, as_of: str ) -> dict[str, Any]: model_id = _nonempty_string(metadata.get("id"), "OpenCode Zen metadata model id") if docs_evidence.get("id") != model_id: raise ValueError(f"OpenCode Zen docs evidence id mismatch for {model_id}") name = _nonempty_string(metadata.get("name"), f"OpenCode Zen model {model_id}.name") description = _nonempty_string( metadata.get("description"), f"OpenCode Zen model {model_id}.description" ) limit = _object(metadata.get("limit"), f"OpenCode Zen model {model_id}.limit") context = _positive_int(limit.get("context"), f"OpenCode Zen model {model_id}.limit.context") output = _positive_int(limit.get("output"), f"OpenCode Zen model {model_id}.limit.output") modalities = _object(metadata.get("modalities"), f"OpenCode Zen model {model_id}.modalities") inputs = _string_list( modalities.get("input"), f"OpenCode Zen model {model_id}.modalities.input", allow_empty=False, ) outputs = _string_list( modalities.get("output"), f"OpenCode Zen model {model_id}.modalities.output", allow_empty=False, ) if "text" not in inputs or outputs != ["text"]: raise ValueError(f"OpenCode Zen model {model_id} lacks the required text contract") tool_call = metadata.get("tool_call") if not isinstance(tool_call, bool): raise ValueError(f"OpenCode Zen model {model_id}.tool_call must be boolean") structured = metadata.get("structured_output", False) if structured is not None and not isinstance(structured, bool): raise ValueError(f"OpenCode Zen model {model_id}.structured_output must be boolean or null") provider_metadata = metadata.get("provider") if provider_metadata is None: npm = "@ai-sdk/openai-compatible" else: npm = _nonempty_string( _object(provider_metadata, f"OpenCode Zen model {model_id}.provider").get("npm"), f"OpenCode Zen model {model_id}.provider.npm", ) route_key = OPENCODE_ZEN_ROUTE_BY_NPM.get(npm) if route_key is None: raise ValueError(f"OpenCode Zen model {model_id} has unsupported provider package {npm!r}") endpoint = docs_evidence.get("endpoint") deprecation_date = docs_evidence.get("deprecation_date") if endpoint is not None: documented_npm = _nonempty_string( _object(endpoint, f"OpenCode Zen docs model {model_id}.endpoint").get("npm"), f"OpenCode Zen docs model {model_id}.endpoint.npm", ) if documented_npm != npm: raise ValueError( f"OpenCode Zen model {model_id} protocol mismatch: " f"Models.dev={npm}, docs={documented_npm}" ) availability = ( "live-undocumented" if endpoint is None and deprecation_date is None else "current" ) if deprecation_date is not None: deprecation_date = _iso_date( deprecation_date, f"OpenCode Zen docs model {model_id}.deprecation_date" ) state = "deprecated" if deprecation_date <= as_of else "deprecation-scheduled" availability = f"{state}-{deprecation_date}-live-listed" executable = route_key != "opencode_zen_google_catalog" upstream_modalities = ["file" if item == "pdf" else item for item in inputs] upstream_modalities = list(dict.fromkeys(upstream_modalities)) reasoning_levels, default_reasoning = opencode_zen_reasoning(metadata) catalog: dict[str, Any] = { "maker": infer_model_maker(model_id), "route": route_key, "upstream_id": model_id, "display_name": f"{name} via OpenCode Zen", "description": description, "kind": "chat" if executable or "image" not in upstream_modalities else "vision_chat", "agent_compatible": executable and tool_call, "context_window": context, "max_output_tokens": output, "reasoning_levels": reasoning_levels, "default_reasoning": default_reasoning, "modalities": ["text"] if executable else upstream_modalities, "output_modalities": ["text"], "tool_calling": tool_call, "parallel_tool_calls": False, "supports_reasoning_summaries": False, "structured_output": structured is True, "availability": availability, "capability_confidence": "opencode-maintained-catalog", "source": "models-dev-opencode-zen", "availability_source": "opencode-zen-models", "capability_source": "models-dev-opencode-zen", "pricing_source": "opencode-zen-docs-source", "inventory": "opencode-zen", "resource_group": "opencode_zen", } catalog.update(opencode_zen_catalog_pricing(docs_evidence)) return catalog def build_opencode_zen_snapshot( listing_document: Any, models_dev_document: Any, docs_document: Any, *, as_of: str, retrieved_at: str, listing_sha256: str, models_dev_sha256: str, docs_sha256: str, listing_url: str, models_dev_url: str, docs_url: str, ) -> dict[str, Any]: """Join live Zen availability with capability and provider-doc evidence.""" as_of = _iso_date(as_of, "OpenCode Zen snapshot as_of") _sha256(listing_sha256, "OpenCode Zen listing SHA-256") _sha256(models_dev_sha256, "Models.dev response SHA-256") _sha256(docs_sha256, "OpenCode Zen docs source SHA-256") _utc_timestamp(retrieved_at, "OpenCode Zen retrieval timestamp") listing_url = _https_url(listing_url, "OpenCode Zen models source URL") models_dev_url = _https_url(models_dev_url, "Models.dev source URL") docs_url = _https_url(docs_url, "OpenCode Zen docs source URL") metadata_by_id = _opencode_zen_metadata(models_dev_document) docs_by_id = _opencode_zen_docs(docs_document, metadata_by_id) rows = _opencode_zen_listing_rows(listing_document) sources = dict(OPENCODE_ZEN_SOURCE_IDS) sources["opencode-zen-models"] = listing_url sources["models-dev-opencode-zen"] = models_dev_url sources["opencode-zen-docs-source"] = docs_url seen_ids: set[str] = set() used: set[str] = set() records: dict[str, dict[str, Any]] = {} for row in sorted(rows, key=lambda value: str(value.get("id", ""))): live_evidence = _opencode_zen_live_evidence(row) model_id = str(live_evidence["id"]) if model_id in seen_ids: raise ValueError(f"duplicate OpenCode Zen live model id: {model_id}") seen_ids.add(model_id) metadata = metadata_by_id.get(model_id) if metadata is None: raise ValueError(f"OpenCode Zen live model lacks Models.dev metadata: {model_id}") if metadata.get("id") != model_id: raise ValueError(f"OpenCode Zen Models.dev id mismatch for {model_id}") docs_evidence = docs_by_id.get( model_id, {"id": model_id, "documented_endpoint": False}, ) catalog_record = _opencode_zen_catalog_record(metadata, docs_evidence, as_of=as_of) key = route_catalog_key(str(catalog_record["route"]), model_id, used) used.add(key) records[key] = { "catalog": catalog_record, "evidence": { "docs": docs_evidence, "live": live_evidence, "models_dev": dict(sorted(metadata.items())), }, } return build_inventory_snapshot( inventory="opencode-zen", adapter="opencode_zen_join", fingerprint_fields=FULL_FINGERPRINT_FIELDS, as_of=as_of, dynamic=True, sources=sources, discovery={"endpoint": listing_url}, captures=[ { "source": "opencode-zen-models", "retrieved_at": retrieved_at, "response_sha256": listing_sha256, }, { "source": "models-dev-opencode-zen", "retrieved_at": retrieved_at, "response_sha256": models_dev_sha256, }, { "source": "opencode-zen-docs-source", "retrieved_at": retrieved_at, "response_sha256": docs_sha256, }, ], models=records, ) def _opencode_go_metadata(document: Any) -> dict[str, dict[str, Any]]: root = _object(document, "Models.dev response") access = _object(root.get("opencode-go"), "Models.dev opencode-go access product") if access.get("id") != "opencode-go": raise ValueError("Models.dev opencode-go access product has an unexpected id") if access.get("api") != "https://opencode.ai/zen/go/v1": raise ValueError("Models.dev opencode-go access product has an unexpected API base") if access.get("env") != ["OPENCODE_API_KEY"]: raise ValueError("Models.dev opencode-go has an unexpected credential contract") models = _object(access.get("models"), "Models.dev opencode-go models") if not models: raise ValueError("Models.dev opencode-go models must not be empty") return { _nonempty_string(key, "Models.dev OpenCode Go model key"): _object( value, f"Models.dev OpenCode Go model {key}" ) for key, value in models.items() } def _opencode_go_docs(document: Any) -> dict[str, dict[str, Any]]: if not isinstance(document, str) or not document.strip(): raise ValueError("OpenCode Go docs source must be non-empty UTF-8 text") endpoint_rows = _markdown_table( document, "## Endpoints", ("Model", "Model ID", "Endpoint", "AI SDK Package"), ) pricing_rows = _markdown_table( document, "## Usage limits", ("Model", "Input", "Output", "Cached Read", "Cached Write", "Usage"), ) docs_by_id: dict[str, dict[str, Any]] = {} names: dict[str, str] = {} for name, model_id, endpoint, npm in endpoint_rows: model_id = _nonempty_string(model_id, "OpenCode Go docs endpoint model id") route = OPENCODE_GO_ROUTE_BY_NPM.get(npm) if route is None: raise ValueError(f"OpenCode Go docs model {model_id} has unsupported package {npm!r}") expected_suffix = { "@ai-sdk/openai": "/responses", "@ai-sdk/openai-compatible": "/chat/completions", "@ai-sdk/anthropic": "/messages", }[npm] endpoint = _https_url(endpoint, f"OpenCode Go docs model {model_id}.endpoint") if endpoint != f"https://opencode.ai/zen/go/v1{expected_suffix}": raise ValueError( f"OpenCode Go docs model {model_id} has unexpected endpoint {endpoint}" ) if model_id in docs_by_id: raise ValueError(f"duplicate OpenCode Go docs endpoint row: {model_id}") docs_by_id[model_id] = { "id": model_id, "endpoint": {"name": name, "url": endpoint, "npm": npm}, } key = _opencode_zen_name_key(name) previous = names.get(key) if previous is not None and previous != model_id: raise ValueError(f"OpenCode Go docs model name {name!r} is ambiguous") names[key] = model_id for label, input_price, output_price, cache_read, cache_write, usage in pricing_rows: base_name = _opencode_pricing_base_name(label) pricing_model_id = names.get(_opencode_zen_name_key(base_name)) if pricing_model_id is None: # A price row may use a typography variant; preserve it only when # its endpoint identity is unambiguous. candidates = [ candidate for key, candidate in names.items() if key in _opencode_zen_name_key(base_name) or _opencode_zen_name_key(base_name) in key ] if len(set(candidates)) != 1: raise ValueError(f"OpenCode Go docs pricing model is unresolved: {label!r}") pricing_model_id = candidates[0] row: dict[str, Any] = { "label": label, "input": _opencode_zen_price(input_price, f"OpenCode Go docs {label}.input"), "output": _opencode_zen_price(output_price, f"OpenCode Go docs {label}.output"), "usage": _opencode_zen_price(usage, f"OpenCode Go docs {label}.usage"), } read_rate = _opencode_zen_price(cache_read, f"OpenCode Go docs {label}.cache_read") write_rate = _opencode_zen_price(cache_write, f"OpenCode Go docs {label}.cache_write") if read_rate is not None: row["cache_read"] = read_rate if write_rate is not None: row["cache_write"] = write_rate docs_by_id[pricing_model_id].setdefault("pricing", []).append(row) return docs_by_id def _opencode_go_catalog_record( metadata: Mapping[str, Any], docs_evidence: Mapping[str, Any] | None ) -> dict[str, Any]: model_id = _nonempty_string(metadata.get("id"), "OpenCode Go metadata model id") name = _nonempty_string(metadata.get("name"), f"OpenCode Go model {model_id}.name") description = _nonempty_string( metadata.get("description"), f"OpenCode Go model {model_id}.description" ) limit = _object(metadata.get("limit"), f"OpenCode Go model {model_id}.limit") context = _positive_int(limit.get("context"), f"OpenCode Go model {model_id}.limit.context") output = _positive_int(limit.get("output"), f"OpenCode Go model {model_id}.limit.output") modalities = _object(metadata.get("modalities"), f"OpenCode Go model {model_id}.modalities") inputs = _string_list( modalities.get("input"), f"OpenCode Go model {model_id}.modalities.input", allow_empty=False ) outputs = _string_list( modalities.get("output"), f"OpenCode Go model {model_id}.modalities.output", allow_empty=False, ) if "text" not in inputs or outputs != ["text"]: raise ValueError(f"OpenCode Go model {model_id} lacks the required text contract") tool_call = metadata.get("tool_call") if not isinstance(tool_call, bool): raise ValueError(f"OpenCode Go model {model_id}.tool_call must be boolean") structured = metadata.get("structured_output", False) if structured is not None and not isinstance(structured, bool): raise ValueError(f"OpenCode Go model {model_id}.structured_output is invalid") npm, _protocol_resolution = _opencode_go_protocol(metadata, docs_evidence) route_key = OPENCODE_GO_ROUTE_BY_NPM.get(npm) if route_key is None: raise ValueError(f"OpenCode Go model {model_id} has unsupported package {npm!r}") reasoning_levels, default_reasoning = opencode_zen_reasoning(metadata) record: dict[str, Any] = { "maker": infer_model_maker(model_id), "route": route_key, "upstream_id": model_id, "display_name": f"{name} via OpenCode Go", "description": description, "kind": "chat", "agent_compatible": tool_call, "context_window": context, "max_output_tokens": output, "reasoning_levels": reasoning_levels, "default_reasoning": default_reasoning, # The current Switchyard Go transports are text-only even when the # underlying model record advertises richer modalities. The complete # upstream modality set remains in fingerprint evidence. "modalities": ["text"], "output_modalities": ["text"], "tool_calling": tool_call, "parallel_tool_calls": False, "supports_reasoning_summaries": False, "structured_output": structured is True, "availability": "current", "capability_confidence": "opencode-maintained-catalog", "source": "models-dev-opencode-go", "availability_source": "opencode-go-models", "capability_source": "models-dev-opencode-go", "pricing_source": "opencode-go-docs-source", "inventory": "opencode-go", "resource_group": "opencode_go", } # Provider documentation is authoritative for the Go access product. # Time- or context-tiered tables remain exact evidence but cannot be # flattened into the catalog's scalar per-million estimate fields. if docs_evidence is not None: record.update(opencode_zen_catalog_pricing(docs_evidence)) return record def _opencode_go_protocol( metadata: Mapping[str, Any], docs_evidence: Mapping[str, Any] | None ) -> tuple[str, dict[str, Any]]: """Resolve Go's wire protocol and retain disagreements as evidence. OpenCode's Go endpoint table is the operator-owned transport contract. The Models.dev package identifies the client adapter OpenCode currently uses, but it can temporarily lag or lead the endpoint table. A disagreement must remain visible in the fingerprint; it must not make an explicitly documented endpoint impossible to represent. """ model_id = _nonempty_string(metadata.get("id"), "OpenCode Go metadata model id") provider_metadata = metadata.get("provider") models_dev_npm = ( "@ai-sdk/openai-compatible" if provider_metadata is None else _nonempty_string( _object(provider_metadata, f"OpenCode Go model {model_id}.provider").get("npm"), f"OpenCode Go model {model_id}.provider.npm", ) ) if models_dev_npm not in OPENCODE_GO_ROUTE_BY_NPM: raise ValueError( f"OpenCode Go model {model_id} has unsupported Models.dev package {models_dev_npm!r}" ) documented_npm: str | None = None if docs_evidence is not None: endpoint = _object( docs_evidence.get("endpoint"), f"OpenCode Go docs model {model_id}.endpoint" ) documented_npm = _nonempty_string( endpoint.get("npm"), f"OpenCode Go docs model {model_id}.endpoint.npm" ) selected_npm = documented_npm or models_dev_npm if selected_npm not in OPENCODE_GO_ROUTE_BY_NPM: raise ValueError( f"OpenCode Go model {model_id} has unsupported documented package {selected_npm!r}" ) return selected_npm, { "authority": "opencode-go-docs-source" if documented_npm else "models-dev-opencode-go", "models_dev_npm": models_dev_npm, "documented_npm": documented_npm, "selected_npm": selected_npm, "disagreement": documented_npm is not None and documented_npm != models_dev_npm, } def build_opencode_go_snapshot( listing_document: Any, models_dev_document: Any, docs_document: Any, *, as_of: str, retrieved_at: str, listing_sha256: str, models_dev_sha256: str, docs_sha256: str, listing_url: str, models_dev_url: str, docs_url: str, ) -> dict[str, Any]: """Join Go availability, Models.dev capabilities, and operator docs.""" as_of = _iso_date(as_of, "OpenCode Go snapshot as_of") _sha256(listing_sha256, "OpenCode Go listing SHA-256") _sha256(models_dev_sha256, "Models.dev response SHA-256") _sha256(docs_sha256, "OpenCode Go docs source SHA-256") _utc_timestamp(retrieved_at, "OpenCode Go retrieval timestamp") listing_url = _https_url(listing_url, "OpenCode Go models source URL") models_dev_url = _https_url(models_dev_url, "Models.dev source URL") docs_url = _https_url(docs_url, "OpenCode Go docs source URL") metadata_by_id = _opencode_go_metadata(models_dev_document) docs_by_id = _opencode_go_docs(docs_document) rows = _opencode_zen_listing_rows(listing_document) sources = dict(OPENCODE_GO_SOURCE_IDS) sources["opencode-go-models"] = listing_url sources["models-dev-opencode-go"] = models_dev_url sources["opencode-go-docs-source"] = docs_url seen_ids: set[str] = set() used: set[str] = set() records: dict[str, dict[str, Any]] = {} for row in sorted(rows, key=lambda value: str(value.get("id", ""))): live_evidence = _opencode_zen_live_evidence(row) model_id = str(live_evidence["id"]) if model_id in seen_ids: raise ValueError(f"duplicate OpenCode Go live model id: {model_id}") seen_ids.add(model_id) metadata = metadata_by_id.get(model_id) if metadata is None: # Keep the live ID visible in the comparison, but do not infer its # capabilities. The coverage marker makes the verification fail # with an actionable per-field report. route_key = "opencode_go_openai_chat" key = route_catalog_key(route_key, model_id, used) used.add(key) records[key] = { "catalog": { "maker": infer_model_maker(model_id), "route": route_key, "upstream_id": model_id, "display_name": f"Unverified live Go model {model_id}", "description": "Live ID without capability fingerprint evidence", "kind": "chat", "agent_compatible": False, "context_window": 0, "reasoning_levels": ["none"], "default_reasoning": "none", "modalities": ["text"], "output_modalities": ["text"], "tool_calling": False, "parallel_tool_calls": False, "supports_reasoning_summaries": False, "structured_output": False, "availability": "live-unverified", "capability_confidence": "id-only", "source": "opencode-go-models", "availability_source": "opencode-go-models", "capability_source": "opencode-go-models", "pricing_source": "opencode-go-models", "inventory": "opencode-go", "resource_group": "opencode_go", }, "evidence": { "live": live_evidence, "unverified_live_only": True, }, } continue if metadata.get("id") != model_id: raise ValueError(f"OpenCode Go Models.dev id mismatch for {model_id}") docs_evidence = docs_by_id.get(model_id) catalog_record = _opencode_go_catalog_record(metadata, docs_evidence) _selected_npm, protocol_resolution = _opencode_go_protocol(metadata, docs_evidence) key = route_catalog_key(str(catalog_record["route"]), model_id, used) used.add(key) records[key] = { "catalog": catalog_record, "evidence": { "docs": docs_evidence or {"id": model_id, "documented_endpoint": False}, "live": live_evidence, "models_dev": dict(sorted(metadata.items())), "protocol_resolution": protocol_resolution, }, } return build_inventory_snapshot( inventory="opencode-go", adapter="opencode_go_join", fingerprint_fields=FULL_FINGERPRINT_FIELDS, as_of=as_of, dynamic=True, sources=sources, discovery={"endpoint": listing_url}, captures=[ { "source": "opencode-go-models", "retrieved_at": retrieved_at, "response_sha256": listing_sha256, }, { "source": "models-dev-opencode-go", "retrieved_at": retrieved_at, "response_sha256": models_dev_sha256, }, { "source": "opencode-go-docs-source", "retrieved_at": retrieved_at, "response_sha256": docs_sha256, }, ], models=records, ) def codex_runtime_evidence(value: Mapping[str, Any]) -> dict[str, Any]: """Normalize only installed-Codex metadata relevant to route execution.""" row = _object(value, "Codex model metadata") slug = _nonempty_string(row.get("slug"), "Codex model slug") levels_raw = row.get("supported_reasoning_levels") if not isinstance(levels_raw, list) or not levels_raw: raise ValueError(f"Codex model {slug} lacks supported reasoning levels") levels: list[str] = [] for index, item in enumerate(levels_raw): effort = item.get("effort") if isinstance(item, Mapping) else item effort = _nonempty_string(effort, f"Codex model {slug} reasoning level {index}") if effort not in CODEX_REASONING_ORDER or effort in levels: raise ValueError(f"Codex model {slug} has invalid reasoning effort {effort!r}") levels.append(effort) default = _nonempty_string( row.get("default_reasoning_level"), f"Codex model {slug} default reasoning level" ) if default not in levels: raise ValueError(f"Codex model {slug} default reasoning level is unsupported") context = _positive_int(row.get("context_window"), f"Codex model {slug}.context_window") modalities = _string_list( row.get("input_modalities"), f"Codex model {slug}.input_modalities", allow_empty=False ) if set(modalities) - {"text", "image", "audio", "video", "file"}: raise ValueError(f"Codex model {slug} has unsupported input modalities") parallel = row.get("supports_parallel_tool_calls") if parallel is None: # The bundled source omits this derived installed-client field. Codex's # shell/apply-patch harness is the authoritative positive capability. parallel = bool( row.get("tool_mode") == "code_mode_only" or row.get("shell_type") or row.get("apply_patch_tool_type") ) if not isinstance(parallel, bool): raise ValueError(f"Codex model {slug}.supports_parallel_tool_calls is invalid") summaries = row.get("supports_reasoning_summaries") if summaries is None: summaries = "default_reasoning_summary" in row or row.get( "reasoning_summary_format" ) not in (None, "none") if not isinstance(summaries, bool): raise ValueError(f"Codex model {slug}.supports_reasoning_summaries is invalid") return { "slug": slug, "comp_hash": row.get("comp_hash"), "context_window": context, "input_modalities": modalities, "reasoning": {"levels": levels, "default": default, "summaries": summaries}, "tools": { "shell_type": row.get("shell_type"), "tool_mode": row.get("tool_mode"), "apply_patch_tool_type": row.get("apply_patch_tool_type"), "parallel_tool_calls": parallel, "search": row.get("supports_search_tool"), }, "structured_output": False, "visibility": row.get("visibility"), "supported_in_api": row.get("supported_in_api"), "multi_agent_version": row.get("multi_agent_version"), "service_tiers": row.get("service_tiers", []), } def _codex_model_rows(document: Any) -> list[dict[str, Any]]: root = _object(document, "Codex models document") rows = root.get("models") if not isinstance(rows, list) or not rows: raise ValueError("Codex models document.models must be a non-empty list") return [ _object(row, f"Codex models document.models[{index}]") for index, row in enumerate(rows) ] def build_codex_installed_snapshot( document: Any, reviewed_snapshot: Any, *, as_of: str, retrieved_at: str, response_sha256: str, source_url: str, ) -> dict[str, Any]: """Join a reviewed Codex catalog with the exact bundled client metadata.""" reviewed = validate_inventory_snapshot(reviewed_snapshot, expected_inventory="openai-codex") _sha256(response_sha256, "Codex models source SHA-256") _utc_timestamp(retrieved_at, "Codex models retrieval timestamp") source_url = _https_url(source_url, "Codex models source URL") rows = { str(evidence["slug"]): evidence for evidence in (codex_runtime_evidence(row) for row in _codex_model_rows(document)) } records: dict[str, dict[str, Any]] = {} for key, old_record in reviewed["models"].items(): catalog = dict(old_record["catalog"]) slug = str(catalog["upstream_id"]) evidence = rows.get(slug) if evidence is None: raise ValueError(f"reviewed Codex model is absent from bundled source: {slug}") catalog.update( { "capability_confidence": (f"codex-{APP_SERVER_PROTOCOL_CODEX_VERSION}-baseline"), "context_window": evidence["context_window"], "reasoning_levels": evidence["reasoning"]["levels"], "default_reasoning": evidence["reasoning"]["default"], "modalities": evidence["input_modalities"], "parallel_tool_calls": evidence["tools"]["parallel_tool_calls"], "supports_reasoning_summaries": evidence["reasoning"]["summaries"], "structured_output": evidence["structured_output"], } ) # The installed catalog does not expose a separate maximum output # limit; carrying an API-product limit here would be false precision. catalog.pop("max_output_tokens", None) records[key] = { "catalog": catalog, "evidence": { "codex_runtime": evidence, "verified_fingerprint_fields": list(FULL_FINGERPRINT_FIELDS), }, } sources = dict(reviewed["sources"]) sources["openai-codex-client-models"] = source_url return build_inventory_snapshot( inventory="openai-codex", adapter="codex_installed_models_join", fingerprint_fields=FULL_FINGERPRINT_FIELDS, as_of=as_of, dynamic=False, sources=sources, discovery=reviewed["discovery"], captures=[ { "source": "openai-codex-client-models", "retrieved_at": retrieved_at, "response_sha256": response_sha256, } ], models=records, )