fix(provider): send enabled thinking for Claude Opus 4.5 across Anthropic, Vertex, Bedrock, and SAP

[rakshith1928]

Issue

Fixes #12546

Context

Why: KiloCode currently sends only the effort parameter for Claude Opus 4.5. Anthropic's adaptive-thinking documentation states that older models such as Claude Opus 4.5 do not support adaptive thinking and require thinking: { type: "enabled", budget_tokens } alongside effort. Sending effort alone does not enable extended thinking on Opus 4.5 because Anthropic requires thinking: { type: "enabled", budget_tokens } alongside effort.

Reference:
1.https://platform.claude.com/docs/en/build-with-claude/adaptive-thinking

2.anomalyco/opencode#38757 ("fix(provider): generalize Claude adaptive thinking") Thanks to them.

The Opus 5 routing portion of that upstream PR is tracked separately in #12544 and is intentionally not duplicated here.

Scope is restricted to Opus 4.5; no other Claude family is affected.

Implementation

Added two new helpers in packages/opencode/src/provider/transform.ts (after anthropicOmitsThinking):

• anthropicOpus45(apiId) — matches IDs containing opus-4-5 or opus-4.5.
• anthropicOpus45Effort(model, effort) — returns { thinking: { type: "enabled", budgetTokens: min(16_000, floor(output / 2 - 1)) }, effort }.

Then wired the helper into the three provider branches that previously mishandled Opus 4.5:

1. Anthropic / Vertex (case "@ai-sdk/anthropic": case "@ai-sdk/google-vertex/anthropic":): replaced the bare { effort } literal branch with a call to anthropicOpus45Effort(model, effort).
2. Bedrock (case "@ai-sdk/amazon-bedrock":): inserted a new branch before the generic api.id.includes("anthropic") legacy block. Emits { reasoningConfig: { type: "enabled", budgetTokens, maxReasoningEffort: effort } } to match Bedrock's adaptive-branch shape.
3. SAP AI Core (case "@jerome-benoit/sap-ai-provider-v2":): inserted a new branch before the generic { high, max } fallback. Uses snake_case budget_tokens (SAP's existing convention at that call site) and wraps each variant in modelParams via the existing wrapInSapModelParams helper.

The budget formula Math.min(16_000, Math.floor(model.limit.output / 2 - 1)) is identical to upstream PR #38757's anthropicOpus45Effort. The default mock model in tests has output: 64_000, so budgetTokens saturates to 16000 in regression tests.

Screenshots / Video

N/A — pure backend/provider transform change; no UI diff.

How to Test

Manual/local verification

• Select Claude Opus 4.5 and verify requests include the required thinking configuration alongside effort.
• Send a prompt that should trigger extended thinking; observe that thinking summary tokens are now returned and the API request payload contains thinking: { type: "enabled", budget_tokens } alongside effort (Anthropic/Vertex/SAP) or reasoningConfig: { type: "enabled", budgetTokens, maxReasoningEffort } (Bedrock).

Reviewer test steps

1. cd packages/opencode && bun test ./test/kilocode/transform-opus-4.5.test.ts — 5 cases across 4 providers (claude-opus-4-5-20251101, claude-opus-4.5-20251101 on Anthropic; claude-opus-4-5@default on Vertex; us.anthropic.claude-opus-4-5-20251101-v1:0 on Bedrock; anthropic--claude-opus-4-5-20251101 on SAP).
2. cd packages/opencode && bun test ./test/provider/transform.test.ts — confirm the existing opus 4.5 row in the shared parametrized array now passes with the new { thinking: { type: "enabled", budgetTokens: 16000 }, effort: "high" } shape and that all other rows (Opus 4.7/4.8, Sonnet 4.6, Sonnet 5, Fable 5) are unchanged.
3. bun run script/check-opencode-annotations.ts --worktree — confirm all shared-file edits are properly marked.

Blocked checks and substitute verification

• bun run typecheck from packages/opencode/: produces pre-existing errors in src/cli/cmd/run/footer.ts, src/cli/cmd/run/runtime.lifecycle.ts, src/kilocode/claw/autocomplete.tsx, src/kilocode/cli/cmd/tui/context/tui-config.tsx, etc., all arising from a @opentui/core 0.3.2 vs 0.3.4 version skew in node_modules. None of the errors touch packages/opencode/src/provider/transform.ts or its tests. Substitute verification was targeted typecheck output filtering, which shows zero errors mentioning transform.ts.

Checklist

• Issue linked above, or exception explained
• Tests/verification described (new transform-opus-4.5.test.ts for 4 providers; existing shared transform.test.ts row updated)
• Screenshots/video included for visual changes, or marked N/A
• Changeset considered for user-facing changes (.changeset/opus-45-enabled-thinking.md added)
• I personally reviewed the diff and can explain the changes, including any AI-assisted work.

Get in Touch

@TRAVIX26

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (4 files)

• .changeset/opus-45-enabled-thinking.md
• packages/opencode/src/provider/transform.ts
• packages/opencode/test/kilocode/transform-opus-4.5.test.ts
• packages/opencode/test/provider/transform.test.ts

Previous Review Summary (commit 9412dc4)

Current summary above is authoritative. Previous snapshots are kept for context only.

Previous review (commit 9412dc4)

Status: No Issues Found | Recommendation: Merge

Files Reviewed (4 files)

• .changeset/opus-45-enabled-thinking.md
• packages/opencode/src/provider/transform.ts
• packages/opencode/test/kilocode/transform-opus-4.5.test.ts
• packages/opencode/test/provider/transform.test.ts

---

_{Reviewed by claude-sonnet-5 · Input: 20 · Output: 2.8K · Cached: 394K}

_{Review guidance: REVIEW.md from base branch main}

[chrarnoldus]

thanks, but we'll inherit anomalyco/opencode#38757 when we merge the upstream