process-prompts

name	process-prompts
description	Process queued architecture prompts from the web viewer. Drains the prompt queue by claiming, executing, and acknowledging each prompt.
trigger	User says "process prompts", "check prompt queue", "run queued prompts", "drain prompt queue", or invokes /process-prompts.

/process-prompts — Process Architecture Prompt Queue

When to Use

Run this when the user has queued architecture change prompts from the arkive web viewer and wants the agent to execute them.

Flow

1. Check for pending prompts

If MCP is available: Call list_prompts with status filter pending. If no pending prompts, inform the user and stop.

If MCP is not available: Read arkive/prompt-queue.json directly. Filter entries with "status": "pending".

2. Process each prompt

Loop until no pending prompts remain:

1. Call claim_next_prompt — this atomically claims the oldest pending prompt and returns:
   • id — the prompt ID
   • prompt — the text of the architecture change request
   • targetServices — array of service IDs this prompt is scoped to
2. Read arkive/architecture.json (or architecture://session-context) to understand current state
3. Execute the prompt:
   • If it requests adding/modifying services → call update_architecture
   • If it requests an architectural decision → call record_decision
   • If it requests code changes → make the changes, then update architecture
4. Call ack_prompt with the prompt id and outcome:
   • "done" — if successfully executed
   • "failed" — if the prompt cannot be executed (explain why in the message)
5. Move to the next prompt

3. Report summary

After draining the queue, tell the user:

• How many prompts were processed
• How many succeeded vs failed
• What architectural changes were made

MUST DO

• Process prompts one at a time in FIFO order — do not skip or reorder
• Always ack_prompt after each prompt, even if it failed
• Update arkive/architecture.json after each prompt that changes the architecture
• Append a history entry for each architectural change

MUST NOT

• Do NOT process prompts without claiming them first (prevents race conditions)
• Do NOT leave prompts in claimed state — always ack as done or failed
• Do NOT batch multiple prompts into a single architecture update — process individually so each has its own ack
• Do NOT execute prompts that are ambiguous or dangerous without asking the user first