feat(cli): set prompt cache breakpoints on stable prefix and user query before environment details

[chrarnoldus]

Summary

• Explicit prompt cache breakpoints for GPT-5.6+: Added support for promptCacheBreakpoint: { mode: 'explicit' } on stable prompt prefixes (system prompt / developer instructions) and user queries for OpenAI, Azure, and Kilo Gateway models on GPT-5.6+ in @kilocode/cli (ProviderTransform.message).
• Improved user message cache breakpoint placement for all models with explicit caching: For all providers using content-level cache options (OpenAI, OpenRouter, Alibaba, etc.), cache breakpoints are now positioned on the last stable user prompt content part immediately preceding any trailing <environment_details> or synthetic context blocks. This prevents subsequent turns (which strip or regenerate ephemeral timestamps and open tabs) from invalidating the cached conversation history prefix across turns.
• Model gating: Kept pre-5.6 OpenAI models on automatic implicit caching without emitting inline breakpoint markers.
• Provider compatibility: Retained existing implicit mode behavior and existing provider breakpoints (Anthropic, Bedrock, etc.).

[kilo-code-bot]

Code Review Summary

Status: 2 Issues Found | Recommendation: Address before merge

Overview

Severity	Count
CRITICAL	0
WARNING	1
SUGGESTION	1

Issue Details (click to expand)

WARNING

File	Line	Issue
packages/opencode/src/provider/transform.ts	496	For kilo-gateway models the explicit breakpoint never leaves the client — the SDK reads providerOptions.openrouter, not openai/azure

SUGGESTION

File	Line	Issue
.changeset/openai-explicit-prompt-cache-breakpoints.md	2	Feature shipped as patch — repo convention is minor for new features

Files Reviewed (3 files)

• packages/opencode/src/provider/transform.ts - 1 issue
• packages/opencode/test/provider/transform.test.ts - 0 issues
• .changeset/openai-explicit-prompt-cache-breakpoints.md - 1 issue

Fix these issues in Kilo Cloud

Previous Review Summaries (4 snapshots, latest commit d3548d1)

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

Previous review (commit d3548d1)

Status: 2 Issues Found | Recommendation: Address before merge

Overview

Severity	Count
CRITICAL	0
WARNING	1
SUGGESTION	1

Issue Details (click to expand)

WARNING

File	Line	Issue
packages/opencode/src/provider/transform.ts	489	For kilo-gateway models the explicit breakpoint never leaves the client — the SDK reads providerOptions.openrouter, not openai/azure

SUGGESTION

File	Line	Issue
.changeset/openai-explicit-prompt-cache-breakpoints.md	2	Feature shipped as patch — repo convention is minor for new features

Files Reviewed (2 files)

• packages/opencode/src/provider/transform.ts - 1 issue
• .changeset/openai-explicit-prompt-cache-breakpoints.md - 1 issue

Fix these issues in Kilo Cloud

Previous review (commit 7c7c11e)

Status: 3 Issues Found | Recommendation: Address before merge

Overview

Severity	Count
CRITICAL	0
WARNING	1
SUGGESTION	2

Issue Details (click to expand)

WARNING

File	Line	Issue
packages/opencode/src/provider/transform.ts	491	For kilo-gateway models the explicit breakpoint never leaves the client — the SDK reads providerOptions.openrouter, not openai/azure

SUGGESTION

File	Line	Issue
packages/opencode/src/provider/transform.ts	391	part.synthetic can never be true at this layer — and the as any[] cast is what hides it
.changeset/openai-explicit-prompt-cache-breakpoints.md	2	Feature shipped as patch — repo convention is minor for new features

Files Reviewed (3 files)

• packages/opencode/src/provider/transform.ts - 2 issues
• .changeset/openai-explicit-prompt-cache-breakpoints.md - 1 issue
• packages/opencode/test/provider/transform.test.ts - no issues

Fix these issues in Kilo Cloud

Previous review (commit 435c029)

Status: No Issues Found | Recommendation: Merge

Both previous findings were addressed in the latest commits: openai-compatible-chat was removed from the inline-hint policy set, and the duplicated version gate was extracted into a shared supportsPromptCacheBreakpoint helper reused by all three call sites.

Files Reviewed (6 files)

• packages/llm/src/cache-policy.ts
• packages/llm/src/index.ts
• packages/llm/src/protocols/openai-chat.ts
• packages/llm/src/protocols/openai-responses.ts
• packages/llm/src/protocols/utils/openai-options.ts
• packages/opencode/src/provider/transform.ts

Previous review (commit 9dbf276)

Status: 2 Issues Found | Recommendation: Address before merge

Overview

Severity	Count
CRITICAL	0
WARNING	1
SUGGESTION	1

Issue Details (click to expand)

WARNING

File	Line	Issue
packages/llm/src/cache-policy.ts	46	openai-compatible-chat now receives inline cache hints, but the emission gate keys only on the model ID — compatible third-party endpoints with gpt-5.6+-named models may get prompt_cache_breakpoint fields and could 400 on the unknown field

SUGGESTION

File	Line	Issue
packages/llm/src/protocols/openai-responses.ts	33	supportsBreakpoint version gate duplicated verbatim in openai-chat.ts, openai-responses.ts, and transform.ts — consider a shared helper in the llm package

Files Reviewed (7 files)

• .changeset/openai-explicit-prompt-cache-breakpoints.md
• packages/llm/src/cache-policy.ts - 1 issue
• packages/llm/src/protocols/openai-chat.ts
• packages/llm/src/protocols/openai-responses.ts - 1 issue
• packages/llm/test/cache-policy.test.ts
• packages/opencode/src/provider/transform.ts
• packages/opencode/test/provider/transform.test.ts

Fix these issues in Kilo Cloud

---

_{Reviewed by kimi-k3 · Input: 93.7K · Output: 12.5K · Cached: 407.8K}

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

[passanger68]

I', afraid it needs fix. BTW - Azure is FIXED.