Prompt Architecture

In Mister Morph's main loop, the only purpose of the prompt is to assemble a reasonable state for the Agent.

Prompt state is assembled from explicit files, policies, runtime metadata, and conversation history. There is no separate long-term or short-term memory injection.

In Mister Morph, these layers are assembled around agent/prompts/system.md.

Main Loop

Static Prompt Skeleton

This template roughly looks like this:

## Persona
{{ identity }}

## Available Skills
{{ skills }}

## Reference Format
{{ internal reference conventions }}

## Additional Policies
{{ large extra policy blocks }}

## Response Format
{{ expected output format }}

## Rules
{{ built-in rules and additional rules }}

Runtime Prompt

Once a concrete run starts, the runtime keeps adding current context.

Common sources include:

• local persona files such as persona/identity.yaml and persona/soul.md
• metadata for the skills enabled in this run
• extra policy blocks

This layer changes with the current task, current channel, and current local state. The final prompt used in CLI, Telegram, and Slack does not have to be identical.

Message Orchestration Outside the Prompt

After the final system prompt is ready, the main Agent still arranges a message stack in the request.

Runtime metadata is injected as a user-role JSON envelope under mister_morph_meta. Typical fields include trigger or correlation data when present, runtime clock fields such as now_utc / now_local / now_local_weekday, and host facts such as host_os.

The order can be understood like this:

[system] final system prompt
   ->
[user] context checkpoint (if any)
   ->
[user / assistant] individual history messages
   ->
[user] runtime metadata
   ->
[user] current message or raw task

Channel history preserves each message boundary. Inbound messages use user; replies sent by the current Agent use assistant and the Agent's response shape ({"type": "final", "output": ...}), so the model does not copy an inbound record as its answer. Other bots remain external participants. Compaction preserves runtime metadata without including it in the summary.

With caching enabled, the engine marks the end of history or the checkpoint on a request copy. Anthropic and OpenAI GPT-5.6 (Chat Completions and Responses) preserve this marker alongside the system marker. Cache hits still depend on the provider and an unchanged prefix.

Independent Prompts

Mister Morph also has another class of calls that does not first build the main Agent's full system prompt, and may not enter the multi-step tool loop at all. Instead, it builds a smaller dedicated prompt and makes a single llm.Chat call.

Independent prompts are used for:

• deciding how to step into a group chat
• task planning
• some narrow semantic judgment or matching tasks

Relationship to the Main Loop

You can think of the two paths like this:

Main Agent main loop
  -> full system prompt
  -> checkpoint / history / runtime metadata / current message
  -> can run multiple steps and call tools

Independent llm.Chat
  -> dedicated system prompt / user prompt
  -> single call
  -> solves one narrow problem only

For example, when the Agent decides how to enter a group chat, it checks:

• whether this message is addressing it
• whether it should interject now
• whether a lightweight response, such as an emoji, is enough

That decision happens before the main loop, so a smaller dedicated prompt is more suitable than the full main system prompt.