axom-dispatch

Axom Subagent Dispatch

When the problem is complex and context is thin: Dispatch a subagent to gather context in parallel while you solve.

When to Dispatch

• Initial memory search returned thin results
• Creative flow needs dreams + long_term + web combined
• You need distilled context before the combine step

Dispatch Flow

1. Dispatch subagent first (runs in parallel).

   • Cursor: Use /axom-agent Gather context for [topic]. Return context envelope. (installed to ~/.cursor/agents/ via make agents)
   • Generic: Pass axom-agent.md as context. Use mcp_task with attachments: [docs/agents/subagent/axom-agent.md]

2. Main agent: discover, read, analyze while subagent runs

3. At combine step: Wait for subagent result (with timeout ~60–120s)

4. If timeout: Proceed with partial context. Store: "Subagent timed out; proceeded with [what we had]."

Context Envelope (from Subagent)

{
  "memory_hits": [...],
  "dreams": [...],
  "web_snippets": [...],
  "discover_summary": "...",
  "distilled_summary": "2–4 sentences: key patterns, wild ideas, relevant facts."
}

Early Dispatch = Parallel Work

Dispatch at the start of your chain so the subagent runs while you:

• Run axom_mcp_discover
• Read relevant files
• Run axom_mcp_analyze

By the combine step, subagent result may already be ready.

Timeout Guidance

• Default: 60–120 seconds
• On timeout: proceed with partial context
• Store in memory: "Subagent timed out; proceeded with [summary of what you had]"

Chain Failure Recovery

If a chain fails mid-way, store partial result + error in memory so the next agent knows where it stopped:

axom_mcp_memory(
  action="write",
  name="partial_[task]_[YYYYMMDD]",
  content="TASK|STEPS_COMPLETED|ERROR|REFLECTION: ...",
  memory_type="short_term"
)

Reference

• Cursor subagent: ~/.cursor/agents/axom-agent.md — invoke via /axom-agent (installed by make agents)
• Subagent: Axom Context Gatherer — generic spec for mcp_task
• axom-memory
• axom-reflex