Feature Request

Problem

The diffs plugin currently injects its agent guidance via prependContext in the before_prompt_build hook:

api.on("before_prompt_build", async () => ({
  prependContext: DIFFS_AGENT_GUIDANCE,
}));

prependContext prepends the guidance text to each individual user message turn. Since this text is constant and appears in the "new" portion of every prompt, it is re-sent and billed as fresh input tokens on every single turn — it cannot be cached at the system prompt level by providers like Anthropic.

For a session with 20 turns, the DIFFS_AGENT_GUIDANCE string (~175 tokens) is charged 20 times as new input, even though its content never changes.

Proposed Solution

Add a prependSystemContext (or appendSystemContext) return key to the before_prompt_build hook contract. Content returned via this key would be appended to the agent's system prompt rather than injected into each user turn.

The diffs plugin could then use:

api.on("before_prompt_build", async () => ({
  prependSystemContext: DIFFS_AGENT_GUIDANCE,
}));

This would allow the guidance to be cached as part of the system prompt prefix, eliminating the per-turn token overhead.

Impact

• Reduces input token cost for any session where the diffs plugin is active
• Particularly beneficial for long sessions or agents using expensive models (e.g. claude-sonnet-4-6)
• The pattern would also be useful for other plugins that inject static guidance today or in the future

Alternatives Considered

• Adding agent-level systemPromptSuffix config (checked schema — not currently supported)
• Manually disabling the plugin injection and copying guidance into SOUL.md/TOOLS.md — fragile and loses the automatic guidance on tool updates

Thanks for the great project!