Summary

• normalize_model_name blindly rewrote dots to hyphens for every non-Bedrock model, corrupting non-Anthropic vendor IDs whose dots are version separators (gpt-5.4 → gpt-5-4, glm-4.7 → glm-4-7, z-ai/glm-5.1 → z-ai/glm-5-1, minimax-m2.5-free → minimax-m2-5-free).
• After the prefix-strip and Bedrock-detection guards, the rewrite is now gated on the model containing claude (case-insensitive). Claude callers are unaffected; non-Claude IDs survive intact even when preserve_dots=False.

The bug

The dot→hyphen rewrite is an Anthropic-Messages-API quirk: OpenRouter form claude-opus-4.6 must become Anthropic form claude-opus-4-6. The function applied it unconditionally to every model except Bedrock IDs, so a non-Claude model accidentally routed through the Anthropic adapter — typically because the caller defaulted to provider=anthropic after failing to infer the right provider — got its name mangled into a non-existent ID and 404'd at the API. The user-facing failure named a model they never typed (gpt-5-4), making the misconfiguration hard to recognize.

The reporter's repro (#17171):

1. webui session created with {"model": "gpt-5.4"} and no provider hint.
2. resolve_model_provider("gpt-5.4") returns (gpt-5.4, None, None) — no inference from the gpt-* shape.
3. resolve_runtime_provider(requested=None) falls back to the default-detected anthropic provider.
4. AIAgent is built with provider="anthropic", api_mode="anthropic_messages", model="gpt-5.4".
5. normalize_model_name("gpt-5.4", preserve_dots=False) returns "gpt-5-4".
6. Request goes to api.anthropic.com/v1/messages with model: gpt-5-4 → HTTP 404.

Same root cause underlies #16417 (custom anthropic_messages providers) and #13061 (zenmux + custom anthropic_messages). The opencode-zen subset (#7421) was previously mitigated by adding the provider to the _anthropic_preserve_dots() allowlist, but the underlying invariant — only Anthropic-shaped names need Anthropic-shaped normalization — was never enforced at the function itself.

The fix

Narrow the rewrite to Claude-shaped IDs:

if not preserve_dots:
    if _is_bedrock_model_id(model):
        return model
    if "claude" not in lower:        # ← new
        return model
    model = model.replace(".", "-")

• Claude users are unaffected (claude-opus-4.6 → claude-opus-4-6 still works, mixed case Claude-Opus-4.6 too).
• The explicit preserve_dots=True flag still short-circuits before the new check, so the documented contract for the alibaba / minimax / opencode-zen / zai / bedrock allowlist is preserved.
• The default preserve_dots=False path becomes safe for non-Claude IDs that get routed through the Anthropic adapter — they round-trip verbatim instead of being silently corrupted.

Contract Protected

Invariant: the dot→hyphen rewrite in normalize_model_name only fires when both (a) the caller has not opted into preserve_dots=True and (b) the model name shape is Claude-like (contains claude after the optional anthropic/ prefix strip).

Test plan

• Parametrized regression TestNormalizeModelNameNonClaudeInvariant::test_non_claude_model_ids_preserve_dots[…] for nine known-bad IDs from openai-codex fallback sends 'gpt-5-4' instead of 'gpt-5.4' to Codex backend, returning HTTP 404 #17171, [Bug]: custom anthropic_messages providers can rewrite model names by replacing '.' with '-' #16417, [Bug] normalize_model_name rewrites custom provider model IDs, breaking models with dots (e.g. z-ai/glm-5.1 via zenmux) #13061, Bug: opencode-zen incorrectly converts dots to hyphens in model names #7421.
• Negative invariant test_negative_claude_dotted_still_mangled: Claude IDs continue to be rewritten; mixed-case detection works.
• Reporter repro test_repro_17171_default_anthropic_provider_with_gpt_model: gpt-5.4 and anthropic/gpt-5.4 both round-trip.
• preserve_dots=True still authoritative for non-Claude callers (test_preserve_dots_true_remains_authoritative_for_non_claude).
• Updated test_minimax_provider.py::test_normalize_converts_without_preserve — its docstring previously labelled the behavior "broken for MiniMax"; it now documents the fix.
• tests/agent/test_bedrock_integration.py (55 tests) all pass — Bedrock detection still wins before the new Claude check.
• tests/agent/test_minimax_provider.py (41 tests) all pass.
• Regression guard: reverting the production change made 10/12 invariant tests fail with the expected gpt-5-4/MiniMax-M2-7/etc. outputs; restoring the fix flipped them back to green.

Related

Fixes #17171
Related: #16417, #13061, #7421 (mitigated earlier via _anthropic_preserve_dots() allowlist; this PR enforces the same outcome at the normalization layer for vendors not on that allowlist)

Positioning vs #17290

@vominh1919's #17290 (opened ~12 min after this one) is a narrower fix at the same line. Differences in scope:

• This PR uses a substring guard (if "claude" not in lower) applied after the anthropic/ prefix has been stripped, with an explicit re-lower() so the stripped name's case is captured. fix: narrow Anthropic adapter dot-mangling to Claude models only (#17171) #17290 keeps an _lower.startswith("anthropic/") arm in its guard which is unreachable at that point in the function.
• This PR adds a parametrized invariant (TestNormalizeModelNameNonClaudeInvariant, 12 cases) covering each of openai-codex fallback sends 'gpt-5-4' instead of 'gpt-5.4' to Codex backend, returning HTTP 404 #17171, [Bug]: custom anthropic_messages providers can rewrite model names by replacing '.' with '-' #16417, [Bug] normalize_model_name rewrites custom provider model IDs, breaking models with dots (e.g. z-ai/glm-5.1 via zenmux) #13061, Bug: opencode-zen incorrectly converts dots to hyphens in model names #7421. fix: narrow Anthropic adapter dot-mangling to Claude models only (#17171) #17290 ships no new tests and relies on the existing claude-* tests still passing.
• This PR replaces the stale test_normalize_converts_without_preserve in tests/agent/test_minimax_provider.py (which currently asserts the broken MiniMax-M2-7 output and labels it "broken for MiniMax") with a positive assertion that documents the fix. fix: narrow Anthropic adapter dot-mangling to Claude models only (#17171) #17290 leaves that test as-is, which would silently keep encoding the old bug as the expected behavior.

Either fix unblocks the reporter's repro on #17171; this PR additionally protects the invariant against future regressions and cleans up the stale test.