[i1hwan]

Summary

Add Claude Opus 4.7 compatibility to the OpenAI-compat → Claude OAuth pipeline, addressing all Anthropic breaking changes while preserving full Opus 4.6 backward compatibility.

Anthropic Breaking Changes Addressed

Breaking Change	Solution
thinking: {type:"enabled"} → 400 error	coerceThinkingForModel() converts to type:"adaptive" at both thinkingBudget and translator layers
temperature/top_p/top_k non-default → 400 error	Stripped in translator for rejectsSamplingParams models
thinking.display defaults to "omitted"	Auto-inject display:"summarized" so OpenCode sees thinking text
xhigh effort (4.7 only)	downgradeEffort() converts to max on unsupported models

Scope

OpenAI-compatible → Claude OAuth path only. Other paths (Claude-native passthrough, Claude→Claude, CC-compatible) are out of scope.

Actual Payload Diff (4.6 vs 4.7)

4.6 Provider Request (unchanged):

{ "thinking": { "type": "adaptive" }, "output_config": { "effort": "max" }, "temperature": 0.7 }

4.7 Provider Request (new):

{ "thinking": { "type": "adaptive", "display": "summarized" }, "output_config": { "effort": "max" } }

Changes: display field added, sampling params stripped.

Tests

• New: 26/26 (opus-4.7-compat.test.mjs)
• Regression: thinking-budget 30/30, provider-service 9/9, chatcore 43/43, translator 15/15, stream 16/16

Verification

• Oracle validation: main OpenAI→Claude path confirmed correct, no type:enabled leak
• Runtime payload trace: verified actual before/after transformations
• Baseline tag: apex-v1.0.0 for safe rollback

[i1hwan]

Accidentally opened against upstream. This is a private fork change.

[gemini-code-assist]

Code Review

This pull request introduces support for Claude Opus 4.7, featuring adaptive thinking and specific model constraints such as the removal of sampling parameters. It implements a lexical rewriting system for Claude OAuth requests to manage keyword and tag replacements across various content blocks. Furthermore, it enhances usage tracking by mapping cache and reasoning tokens to nested OpenAI-style details and adds a new settings tab for configuring forwarding keyword rules. Feedback identifies a redundant ternary operator in the thinking budget coercion logic and warns of potential double-rewriting issues in the iterative lexical replacement loop.

[gemini-code-assist]

The ternary operator here is logically redundant as it returns "high" regardless of whether the condition is truthy or falsy. This makes the code confusing and suggests a potential logic error or a leftover from a refactor. If the intention is to always default to "high" when output_config.effort is missing, a simple constant assignment would be clearer.

Suggested change

	const effort = result.output_config?.effort || result.thinking?.budget_tokens ? "high" : "high";
	const effort = "high";

[gemini-code-assist]

Applying multiple replaceAll calls in a loop can lead to unexpected results if a replacement string contains a keyword that matches a subsequent rule (double-rewriting). While the current default rules seem safe, this pattern is fragile for general-purpose lexical rewriting. Consider using a single-pass regex-based replacement or ensuring that rules are mutually exclusive.

[Copilot]

Pull request overview

Adds Claude Opus 4.7 compatibility to the OpenAI-compatible → Claude OAuth pipeline by enforcing Anthropic’s new request constraints (adaptive-only thinking, sampling param rejection) while keeping Opus 4.6 behavior intact, plus adds configurable lexical rewrite rules for the Claude forwarding lane.

Changes:

• Extend model specs + request translation to support Opus 4.7 constraints (adaptive-only thinking, effort downgrades, sampling param stripping, thinking.display injection).
• Introduce configurable “forwarding keyword rules” (tool/text/tag rewrites) with API + dashboard UI, and apply them in the proxy forwarding path.
• Improve usage accounting by preserving Claude cache token usage across streaming/non-streaming translations and promoting cache/reasoning fields into OpenAI *_details.

Reviewed changes

Copilot reviewed 41 out of 41 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
tests/unit/translator-resp-claude-to-openai.test.mjs	Adds coverage for cache token usage propagation in Claude→OpenAI translation (streaming + non-stream).
tests/unit/translator-openai-to-claude.test.mjs	Expands coverage for tool prefixing + lexical rewrites + thinking/effort promotion behaviors.
tests/unit/token-refresh-service.test.mjs	Updates Claude OAuth refresh contract assertions and adds rejection-path coverage.
tests/unit/token-health-check.test.mjs	Adds coverage for proxy unwrap helper used by refresh checks.
tests/unit/thinking-budget.test.mjs	Updates expectations for new default thinking budget behavior and Claude adaptive-effort promotion logic.
tests/unit/stream-utils.test.mjs	Adjusts expected total token accounting after usage merge changes.
tests/unit/provider-service.test.mjs	Adds target-format-aware thinking normalization coverage (Claude vs non-Claude).
tests/unit/opus-4.7-compat.test.mjs	New end-to-end-ish unit suite validating Opus 4.7 constraints across specs/budget/translation.
tests/unit/oauth-providers-config.test.mjs	Adds redirectUri assertion for Claude OAuth config.
tests/unit/claude-code-compatible-request.test.mjs	Updates effort handling to allow xhigh and validates lexical rewrites in CC-compatible requests.
tests/unit/chatcore-translation-paths.test.mjs	Adds coverage for lexical rewrites and thinking budget behavior in chatCore Claude paths.
tests/unit/cc-compatible-provider.test.mjs	Updates expected CC-compatible effort value (xhigh).
tests/unit/bootstrap-env.test.mjs	Adds coverage for loading OAuth client-id defaults from .env.example when .env is absent.
src/sse/handlers/chat.ts	Applies forwarding keyword + thinking budget settings during request handling.
src/shared/validation/settingsSchemas.ts	Introduces Zod schemas for forwarding keyword rules + thinking budget updates.
src/shared/constants/modelSpecs.ts	Adds Opus 4.6/4.7 model specs and helper functions for adaptive-only, effort, display, and sampling constraints.
src/lib/tokenHealthCheck.ts	Adds helper to unwrap proxy resolution result for refresh calls.
src/app/api/settings/thinking-budget/route.ts	Loads thinking budget settings on GET and switches schema import to settingsSchemas.
src/app/api/settings/forwarding-keywords/route.ts	New API to GET/PUT forwarding keyword rewrite configuration persisted in settings DB.
src/app/api/settings/tests/forwarding-keywords.test.ts	New Vitest coverage for the forwarding-keywords API route.
src/app/(dashboard)/dashboard/settings/page.tsx	Adds ForwardingKeywordsTab to the Advanced settings section.
src/app/(dashboard)/dashboard/settings/components/ThinkingBudgetTab.tsx	Extends adaptive mode UI to include “max” effort and clarifies adaptive behavior copy.
src/app/(dashboard)/dashboard/settings/components/ForwardingKeywordsTab.tsx	New UI tab for editing forwarding keyword rules as JSON with reload/reset/save flows.
scripts/bootstrap-env.mjs	Adds parsing of OAuth client-id defaults from .env.example and adjusts env layer priority.
open-sse/utils/usageTracking.ts	Promotes Claude flat cache tokens + reasoning_tokens into OpenAI nested *_details and prunes flat keys.
open-sse/utils/stream.ts	Changes usage extraction to merge rather than overwrite across events, preserving message_start usage.
open-sse/translator/response/claude-to-openai.ts	Captures cache usage on message_start and includes cache tokens in OpenAI prompt_tokens_details.
open-sse/translator/request/openai-to-claude.ts	Adds lexical rewrites, adaptive-only thinking coercion, effort downgrades, thinking.display injection, and sampling param stripping.
open-sse/translator/index.ts	Makes thinking normalization target-format-aware.
open-sse/translator/helpers/claudeHelper.ts	Adds Claude OAuth lexical rewrite helper for Claude-native payloads and tracks tool name mapping.
open-sse/services/tokenRefresh.ts	Adds Claude OAuth refresh request meta logging and aligns refresh request contract.
open-sse/services/thinkingBudget.ts	Shifts defaults to adaptive/max and adds Claude-specific adaptive thinking/effort promotion + adaptive-only coercion.
open-sse/services/provider.ts	Recognizes adaptive thinking in hasThinkingConfig and preserves thinking for Claude targets regardless of last role.
open-sse/services/claudeCodeCompatible.ts	Applies lexical rewrites to CC-compatible requests and expands supported effort values.
open-sse/handlers/responseTranslator.ts	Includes cache tokens when computing OpenAI prompt_tokens and emits prompt_tokens_details in non-stream responses.
open-sse/handlers/chatCore.ts	Applies thinking budget in Claude cache_control passthrough and applies lexical rewrite + toolNameMap handling for Claude.
open-sse/config/providerRegistry.ts	Adds claude-opus-4-7 to the Claude provider registry.
open-sse/config/forwardingKeywordRules.ts	New data-driven rewrite rule engine with defaults + normalization + per-lane rewrite helpers.
Dockerfile	Cleans standalone build output and ensures .env.example is included in runtime image.
.github/workflows/docker-publish.yml	Changes published image names/tags and limits platforms to linux/amd64.
.dockerignore	Expands ignored files for container build context (CI/config/docs metadata).