Add support for GPT-6

[aesxu2345]

Proposed changes: model catalog, Astra entries, and error details

Following CONTRIBUTING.md, please advise on scope before we open a PR.

Our fork: https://github.com/aesxu2345/ChatMock
Full diff: upstream 5bd6e77 → fork bcbd73c

Key code changes

1. Catalog version and cache invalidation

chatmock/model_catalog.py:29:

-CODEX_MODELS_CLIENT_VERSION = "0.146.0"
+CODEX_MODELS_CLIENT_VERSION = "0.153.4"

Added in _load_cache():251–252 to reject metadata cached under another version:

if payload.get("client_version") != CODEX_MODELS_CLIENT_VERSION:
    return

2. Astra registration and selectable efforts

Added at chatmock/model_registry.py:90–96:

ModelSpec(
    public_id="gpt-6-astra",
    upstream_id="gpt-6-astra",
    aliases=("gpt6-astra", "gpt-6-astra-latest"),
    allowed_efforts=frozenset(("low", "medium", "high", "xhigh", "max", "ultra")),
    variant_efforts=("low", "medium", "high", "xhigh", "max"),
),

Listing rules added at lines 11–14:

ALWAYS_EXPOSE_REASONING_VARIANTS = frozenset(("gpt-6-astra",))
HIDDEN_MODEL_IDS = frozenset(("gpt-6-astra-ultra",))

This changes the default for Astra. ultra is hidden following reported upstream failures, not blocked for manual requests; no alternative mapping is claimed. Listed efforts are not all live-verified.

3. Merge remote and static model lists

chatmock/model_registry.py:258–291, list_public_models(): removed the early return after remote models. Static models are now appended with deduplication; existing remote effort metadata takes precedence. Key added loop:

for spec in _MODEL_SPECS:
    base_was_seen = spec.public_id in seen_ids
    append_model(spec.public_id)
    if expose_variants_for(spec.public_id):
        # Prefer account-specific reasoning metadata when the remote catalog
        # contains this model; static variants are only a compatibility fallback.
        if base_was_seen:
            continue
        for effort in spec.variant_efforts:
            append_model(f"{spec.public_id}-{effort}")

Affects both /v1/models and /api/tags. Prevents static entries disappearing, but may list models unavailable to an account.

4. Preserve useful upstream error messages

chatmock/routes_openai.py:74–98: new _upstream_error_message() handles nested errors, top-level messages and text. Added top-level extraction:

for key in ("message", "detail", "title"):
    value = body.get(key)
    if isinstance(value, str) and value.strip():
        return value.strip()

Used in chat_completions() at lines 286 and 296:

err = {"error": {"message": _upstream_error_message(upstream)}}

This is a Chat Completions change; Ollama error handling is unchanged.

Supporting changes and checks

• tests/test_routes.py:60–353: list/merge, cache, mapping and error tests; last full run 27 passed. Mocked results do not verify backend support. Live client checks remain pending.
• chatmock/fast_mode.py:81: error wording “GPT-5” → “GPT” only.
• README.md: pending local revision keeps changes confined to the model list; not yet in the linked fork commit. DOCKER.md: documents Astra listing behavior.
• No files deleted/moved, no entry points or public parameters renamed, no dependency or CLI/GUI/packaging changes.

Maintainer guidance requested

1. Should static supplementation be opt-in, and should Astra variants follow --expose-reasoning-models? We can keep default-changing behavior fork-only.
2. What catalog-version source, capability evidence, and unsupported-effort policy would you require?
3. Would you prefer separate catalog/cache and error-handling PRs? Please specify required client tests and error-redaction/route-consistency checks.