[issue-7526][BE] fix: make Anthropic adaptive-thinking opt-out sync-proof

[Nimrod007]

Details

Follow-up to #7531. That PR opted the Anthropic adaptive-thinking models (claude-opus-4-7, claude-opus-4-8, claude-sonnet-5) out of sampling params using a per-constant flag in the AnthropicModelName enum:

CLAUDE_OPUS_4_8("claude-opus-4-8", false),

The enum constant list is regenerated by the recurring chore: sync provider model definitions job (#7088, #7315, #7582). That generator emits every model as a single-arg NAME("value") constant and has no notion of the extra capability argument. #7582 ran the sync ~18h after #7531 merged and silently dropped all three , false flags (reverting them to the single-arg constructor, which defaults supportsSamplingParams to true) — reopening #7526:

-    CLAUDE_OPUS_4_7("claude-opus-4-7", false),
-    CLAUDE_OPUS_4_8("claude-opus-4-8", false),
     ...
-    CLAUDE_SONNET_5("claude-sonnet-5", false);
+    CLAUDE_OPUS_4_7("claude-opus-4-7"),
+    CLAUDE_OPUS_4_8("claude-opus-4-8"),
+    CLAUDE_SONNET_5("claude-sonnet-5");

Current main still has the flags dropped, so online LLM-as-judge rules on these models 400 again with `temperature` is deprecated for this model whenever a temperature is present (API-created rules, which bypass the OPIK_6244 FE sanitizer — the exact case #7526/#7531 targeted).

Change

Move the opt-out out of the generated constant list into a hand-maintained ADAPTIVE_THINKING_MODELS set that supportsSamplingParams() consults. The constant list stays in the exact single-arg shape the generator emits, so future syncs can add/remove/reorder constants freely without ever clobbering the capability. Public API (supportsSamplingParams(String)) and behavior are unchanged; the field-backed Map + helper constructor are removed.

• AnthropicModelName — drop supportsSamplingParams enum field, the this(value, true) helper ctor, and the field-derived map; add ADAPTIVE_THINKING_MODELS set + doc explaining why it must live outside the generated list.
• AnthropicClientGenerator unchanged — the max_tokens / thinking / temperature-gating logic from [issue-7526][BE] fix: support Anthropic adaptive-thinking models as online-rule judges #7531 was never regressed (it lives outside the generated file).

Change checklist

• User facing
• Documentation update

Issues

• Re-resolves [Bug] Online rules (LLM-as-judge) fail with Anthropic adaptive-thinking models (Sonnet 5, Opus 4.7/4.8): temperature 400 + empty response, custom_parameters not forwarded #7526 (regressed by [NA] [BE][FE] chore: sync provider model definitions #7582 after [issue-7526][BE] fix: support Anthropic adaptive-thinking models as online-rule judges #7531).

Testing

• Existing AnthropicClientGeneratorTest.SamplingParamsCapability cases (assert false for the three adaptive models) pass again — they were effectively dead on main since test-compile is currently broken there, which is why the [NA] [BE][FE] chore: sync provider model definitions #7582 regression went undetected.
• New adaptiveThinkingModelsRemainEnumeratedAndRejectSamplingParams regression guard: asserts each adaptive model id still maps to a real enum constant (so the set can't drift when a sync renames/removes a constant) and reports no sampling-param support.
• Verified live on staging (opik-backend @ the regressed build): an API-created online-judge rule on claude-opus-4-8 with temperature 400s on temperature is deprecated; the same rule without temperature scores cleanly (confirming [issue-7526][BE] fix: support Anthropic adaptive-thinking models as online-rule judges #7531's max_tokens defaulting is intact and the temperature opt-out is the only regressed piece).
• Note: full mvn test-compile requires JDK 25 and is broken on pristine main (unrelated JDK-25/Lombok mismatch), so the suite was not run locally; the change is mechanical.

Documentation

No user-facing docs change required. Two follow-ups (out of scope): (1) the broken test-compile on main means these guards don't run in CI yet — worth fixing so this class of regression is caught; (2) the sync provider model definitions generator could be taught to preserve/emit capability flags, removing the need for a parallel set.

AI-WATERMARK

AI-WATERMARK: yes

• Tools: Claude Code
• Model(s): Claude Opus 4.8
• Scope: full change (implementation + test), authored under human direction and review
• Human verification: root cause reproduced end-to-end on staging; diff reviewed

🤖 Generated with Claude Code

[github-actions]

⏱️ pre-commit per-hook timing

Hook	Description	Result	Duration
☕ spotless — java backend	Format Java code	✅	3.98s
Total (1 ran)			3.98s

⏭️ 40 skipped (no matching files changed)

Hook	Description	Result
🐍 trim trailing whitespace — python sdk	Strip trailing whitespace	⏭️
🐍 fix end of files — python sdk	Ensure files end in a newline	⏭️
🐍 ruff — python sdk	Lint + autofix Python (ruff)	⏭️
🐍 ruff-format — python sdk	Format Python code (ruff)	⏭️
🐍 mypy — python sdk	Static type check	⏭️
🤖 trim trailing whitespace — optimizer	Strip trailing whitespace	⏭️
🤖 fix end of files — optimizer	Ensure files end in a newline	⏭️
🤖 check yaml — optimizer	Validate YAML syntax	⏭️
🤖 check json — optimizer	Validate JSON syntax	⏭️
🤖 check toml — optimizer	Validate TOML syntax	⏭️
🤖 check for added large files — optimizer	Block large files (>1MB)	⏭️
🔐 detect private key — optimizer	Block committed private keys	⏭️
🤖 check for merge conflicts — optimizer	Block merge-conflict markers	⏭️
🤖 check for case conflicts — optimizer	Block case-only name clashes	⏭️
🤖 pyupgrade — optimizer	Modernize Python syntax	⏭️
🤖 ruff — optimizer	Lint + autofix Python (ruff)	⏭️
🤖 ruff-format — optimizer	Format Python code (ruff)	⏭️
🤖 mypy — optimizer	Static type check	⏭️
📓 nbstripout — optimizer notebooks	Strip notebook output	⏭️
📝 markdownlint — optimizer	Lint Markdown	⏭️
🔤 codespell — optimizer	Fix common misspellings	⏭️
📊 radon cc — optimizer	Cyclomatic-complexity gate	⏭️
📊 radon raw — optimizer	Raw size metrics gate	⏭️
📊 xenon — optimizer	Fail on complexity thresholds	⏭️
📊 lizard — optimizer	Cyclomatic-complexity gate	⏭️
🧹 vulture — optimizer	Find dead code	⏭️
🛡️ trim trailing whitespace — guardrails	Strip trailing whitespace	⏭️
🛡️ fix end of files — guardrails	Ensure files end in a newline	⏭️
🛡️ ruff — guardrails	Lint + autofix Python (ruff)	⏭️
🛡️ ruff-format — guardrails	Format Python code (ruff)	⏭️
🛡️ mypy — guardrails	Static type check	⏭️
⚓ helm-docs	Regenerate Helm chart README	⏭️
block non-public FE plugins	Block non-public FE plugins	⏭️
🧪 pre-commit wrapper smoke tests	Self-test the wrapper scripts	⏭️
🌐 eslint — frontend	Lint + autofix JS/TS	⏭️
🌐 typecheck — frontend	Whole-project tsc type check	⏭️
📘 eslint — typescript sdk	Lint + autofix JS/TS	⏭️
📘 typecheck — typescript sdk	Whole-project tsc type check	⏭️
⚙️ actionlint — github workflows	Lint GitHub Actions workflows	⏭️
🐳 hadolint — dockerfiles	Lint Dockerfiles	⏭️

[andrescrz]

Let's polish this a bit, to prevent this issue from happening again.

[JetoPistola]

👋 Review summary

What looks good

• The core move — from a per-constant enum flag to a standalone ADAPTIVE_THINKING_MODELS set — is exactly the right shape for the problem. It decouples the capability from the generator-owned constant list, which is precisely what [NA] [BE][FE] chore: sync provider model definitions #7582 clobbered.
• The doc comment on the set is excellent: it captures why the set must live outside the generated list, with the [NA] [BE][FE] chore: sync provider model definitions #7088/[NA] [BE][FE] chore: sync provider model definitions #7315/[NA] [BE][FE] chore: sync provider model definitions #7582 history, so the next person touching the sync chore won't re-inline the flag.
• Public contract (supportsSamplingParams(String)) and behaviour are unchanged, and the old field/helper-ctor/derived-map are cleanly removed — no dead code left behind.
• Good instinct adding an explicit regression guard rather than trusting the existing capability cases (which, per the PR body, were dead on main).

Overall
Small, well-motivated, mechanical fix that re-resolves #7526 the durable way. It also underpins the sibling playground/inference-path fix in #7575, which consumes the same supportsSamplingParams contract — so the two land as a coherent pair. andrescrz has already flagged the substantive polish items (DRY the set against the enum values, use enum-backed ValueSource, value getter instead of toString(), single-membership check) — all worth doing, no need to repeat them here.

Inline comments: 1 question — on a residual coverage gap in the new regression guard (it protects against rename/removal of known adaptive models, but not against a sync adding a new adaptive model). Non-blocking.

🤖 Review posted via /review-github-pr

[github-actions]

TS SDK E2E Tests - Node 22

317 tests   315 ✅  13m 42s ⏱️
 38 suites    2 💤
  1 files      0 ❌

Results for commit bab46fc.

♻️ This comment has been updated with latest results.

[github-actions]

Backend Tests - Integration Group 13

 49 files   49 suites   5m 13s ⏱️
348 tests 346 ✅ 2 💤 0 ❌
343 runs  341 ✅ 2 💤 0 ❌

Results for commit bab46fc.

♻️ This comment has been updated with latest results.

[JetoPistola]

👋 Re-review summary (revision bab46fc)

The revision cleanly addresses every prior thread:

• DRY — ADAPTIVE_THINKING_MODEL_IDS now references CLAUDE_OPUS_4_7.value / CLAUDE_OPUS_4_8.value / CLAUDE_SONNET_5.value instead of duplicated literals, so a sync that renames/removes a referenced constant now fails compilation rather than drifting.
• Test drift — driven by @EnumSource(mode = INCLUDE, names = {…}), so renaming/removing an adaptive constant fails at test-discovery time; the set-scan was dropped.
• value getter — @Getter added; the test asserts via model.getValue(), not toString().
• Residual-gap doc — both the production and test doc comments now spell out the one remaining manual step: a newly synced adaptive model still has to be added to the set by hand (the durable fix being generator-emitted capability). That was my only open question — resolved.

Bonus: null/blank handling is now explicit (StringUtils.isBlank(...)) and covered by a dedicated @NullAndEmptySource + blank case, with identical semantics to the old getOrDefault(…, true).

No net-new findings; CI is fully green. Nicely done. ⚠️ This is not a formal approval — a human reviewer should still approve.

🤖 Review posted via /review-github-pr

[JetoPistola]

Addresses all review feedback (DRY via enum .value, @EnumSource drift guard, getValue(), residual-gap docs) plus explicit null/blank handling. Sound, sync-proof fix. CI green.

🤖 Approved via /review-github-pr follow-up