Summary

PR #655 fixes a spurious ModelProviderRegistry.default is deprecated DeprecationWarning that fires on every fresh-install bare DataDesigner() call. The fresh-install ~/.data-designer/model_providers.yaml ships three providers and no default: key, so resolve_model_provider_registry synthesises default=providers[0].name to satisfy check_implicit_default — and _warn_on_explicit_default then attributes a deprecation nudge to a user who never opted in.

The fix broadens the existing warnings.catch_warnings() suppression in DataDesigner.__init__ so it covers both library-synthesised default cases:

1. model_providers is None (caller passed nothing, library auto-fills from YAML)
2. default_provider_name is not None (YAML had a default: key — pre-existing case, already covered)

A new regression test (test_init_no_user_providers_no_yaml_default_stays_quiet) pins the bare-init quiet path against multi-provider YAML with no default: key, plus a behavioural assertion on first-wins resolution.

Findings

Correctness — looks good

• The library_synthesised_default = model_providers is None or default_provider_name is not None condition correctly captures both library-fill cases without leaking into user-driven entry points (model_providers=[multi] supplied directly, direct ModelProviderRegistry(default="x") construction). The behaviour matrix in the PR description matches the code.
• The pre-existing branch condition default_provider_name is not None is still subsumed — when model_providers is None and the YAML carries a default, both clauses are True (harmless OR), and get_default_provider_name()'s YAML deprecation still fires before the registry build (outside the suppression scope), so users still see the actionable YAML warning. Verified in test_init_yaml_default_emits_single_deprecation_warning (unchanged).
• get_default_provider_name() is called inside if model_providers is None: (line 164) — outside the catch_warnings() block, so the YAML-level warning remains visible in case 2. Suppression is correctly scoped to only the registry construction.

Code quality

• The 23-line block comment (lines 169–190) is dense but warranted — this is the second deprecation-suppression iteration in this exact spot (PR feat(models): deprecate implicit default provider routing #594 → PR fix(interface): silence registry-default deprecation when library auto-fills it #655) and the rationale is non-obvious. The comment explicitly names which entry points refactor: deprecate implicit default model provider routing #589 still targets, which is the right discipline given the inverted "we suppress to help the user" semantics. Suggestion (optional): the trailing two paragraphs (about hand-constructed multi-provider lists and direct ModelProviderRegistry(default=...)) duplicate the PR description's behaviour matrix; if length is a concern, this could be trimmed. But I'd lean toward keeping it — the next person who touches this code will thank you.
• library_synthesised_default is a one-shot local — could be inlined into the if. The named binding does aid readability and matches the comment's framing, so I'd keep it as is.
• Naming nit: British spelling "synthesised" appears in both code and comments. The codebase uses both spellings (e.g., "synthesise" in this file's older comments), so this is consistent — no change needed.

Test coverage — strong

• New test patches both get_default_providers and get_default_provider_name to simulate the fresh-install YAML shape (multi-provider, no default: → returns None). Correct simulation of the production code path.
• Asserts both the absence of the registry-level deprecation and that first-wins resolution still works (get_default_provider_name() == "yaml-first"). The behavioural pin guards against a future "fix" that suppresses the warning by hard-coding default=None, which would break first-wins.
• Docstring explicitly cross-references the counterpart test (test_init_user_supplied_providers_preserve_first_wins_over_yaml_default) — good navigability for the next maintainer.
• warnings.simplefilter("always", DeprecationWarning) is correctly used to defeat any inherited filter state, matching the convention of the sibling test at line 625.

Risks / blast radius — minimal

• Change is one boolean expression broadened from default_provider_name is not None to model_providers is None or default_provider_name is not None. No new code paths; suppression is still scoped to resolve_model_provider_registry(...) and only filters the exact deprecation message string.
• No public API surface changes; no behavioural change for callers who already pass model_providers= explicitly.
• Backwards-compatible — the suppression is strictly broader, so previously-quiet paths remain quiet.

Security — n/a

No auth, secret-handling, input-validation, or network surface touched. The test stubs use fake endpoints and PlaintextResolver() as the codebase convention.

Performance — n/a

No hot-path or import-time impact; the only added work is one boolean evaluation in __init__.

Minor observations (non-blocking)

• The comment block ends with "those are the entry points refactor: deprecate implicit default model provider routing #589 targets" — this is great context. Consider whether a one-line # See PR #655 for the broadened suppression reference might help future archaeology, but the in-code reference to # See PR #594 review already establishes the pattern.
• The behaviour matrix in the PR description is excellent and should arguably live alongside the code as a docstring or comment table — but that's a larger ergonomic improvement and out of scope here.

Verdict

Approve / LGTM (non-binding — automated review).

This is a tight, well-scoped fix for a high-visibility paper-cut (every fresh-install user hits this). The code change is one expression; the comment explaining why is appropriately defensive given this is the second iteration in this spot. The new regression test pins both the warning suppression and the first-wins behavioural contract that motivated #588, so the next refactor can't silently regress either property. The PR description, behaviour matrix, and test cross-references are exemplary.

No changes requested.

Greptile Summary

This PR fixes a spurious ModelProviderRegistry.default is deprecated warning that fired on every bare DataDesigner() construction when the default YAML ships multiple providers and no default: key — a case where resolve_model_provider_registry auto-synthesises the default on the user's behalf.

• data_designer.py: Broadens the existing warnings.catch_warnings() suppression from default_provider_name is not None to model_providers is None or default_provider_name is not None, silencing the registry-level warning whenever the library, not the user, is responsible for filling in default=.
• test_data_designer.py: Adds test_init_no_user_providers_no_yaml_default_stays_quiet to pin the quiet path — multi-provider YAML with no default: key — and verify that first-wins resolution still works correctly.

Important Files Changed

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["DataDesigner.__init__(model_providers=?)"] --> B{model_providers is None?}
    B -- Yes --> C["_resolve_model_providers(None)\n→ calls get_default_providers()"]
    B -- No --> D["_resolve_model_providers(model_providers)\ndefault_provider_name = None"]
    C --> E["get_default_provider_name()"]
    E -- "returns None\n(no YAML default:)" --> F["default_provider_name = None"]
    E -- "returns 'name'\n(YAML default: key set)" --> G["default_provider_name = 'name'\n(YAML DeprecationWarning already fired)"]
    F --> H["library_synthesised_default = True\n(model_providers is None)"]
    G --> H
    D --> I["library_synthesised_default = False"]
    H --> J["warnings.filterwarnings ignore\n'ModelProviderRegistry.default is deprecated'"]
    J --> K["resolve_model_provider_registry(providers, default_provider_name)"]
    I --> K
    K --> L{Multi-provider\n& no explicit default?}
    L -- Yes --> M["synthesise default=providers[0].name\n_warn_on_explicit_default fires"]
    L -- No --> N["Registry built normally"]
    M -- suppressed if library_synthesised_default --> O["✅ Quiet (library filled it in)"]
    M -- not suppressed if user-supplied --> P["⚠️ DeprecationWarning visible to user"]
    N --> Q["✅ Registry ready"]

_{Reviews (1): Last reviewed commit: "fix(interface): silence registry-default..." | Re-trigger Greptile}