Problem

Dynamic per-channel/session sections (## Messaging, ## Group Chat Context, ## Reactions) were placed before the large static # Project Context block in buildAgentSystemPrompt(). This caused a full prefix cache miss on every request where channel context differed — the LLM had to reprocess the entire Project Context block from scratch.

Measured on a local Qwen3-Coder-30B agent (MLX, Apple M3 Ultra), as reported in openclaw#40256:

Fix

Reorder sections so all static content comes first, dynamic content comes last:

Before: [static] → [dynamic: Messaging/Reactions] → [Project Context] → [Runtime]
After:  [static] → [Project Context] → [Runtime]   → [dynamic: Messaging/Reactions]

The LLM reads the full system prompt regardless of order, so behaviour is identical — only caching changes.

Also moves the ## Workspace Files (injected) header to sit immediately before # Project Context, where it belongs semantically (previously it appeared before the dynamic sections, not the content it was announcing).

Changes

• src/agents/system-prompt.ts — reorder section assembly in buildAgentSystemPrompt()
• src/agents/system-prompt.e2e.test.ts — add ordering test so this cannot silently regress

Test

Added a test that asserts # Project Context always appears before ## Messaging, ## Group Chat Context, and ## Reactions:

it("places static Project Context before dynamic Messaging section for prefix cache stability", () => {
  // ...
  expect(messagingPos).toBeGreaterThan(projectContextPos);
  expect(groupChatPos).toBeGreaterThan(projectContextPos);
  expect(reactionsPos).toBeGreaterThan(projectContextPos);
});

All existing tests pass unchanged — none checked section ordering, only presence.

Closes openclaw#40256

• AI-assisted (Claude Sonnet 4.6)
• Fully tested — all unit tests pass, ordering test added
• No behaviour change — only section ordering, LLM sees identical content