064-loop: autonomous goal-directed research

[ahoward]

Summary

brane loop — autonomous research cycle. Give brane a goal, it reflects on gaps, searches the web, digests findings, and repeats until convergent.

Why

This is the capstone. An agent with digest, ask, and storm can consume, query, and explore — but a human has to drive it. Loop makes brane self-directed: "learn everything about X" becomes a single command. Bunny's loop does this with web search + digest in a cycle. We do the same but write into the structured graph.

What

CLI

brane loop "understand JWT refresh token rotation patterns"
brane loop "map the competitive landscape for vector databases" --rounds 5
brane loop --resume <id>              # resume a paused loop
brane loop list                        # show all loops with status

Flags

• --rounds <n> — max rounds (default: 5)
• --resume <id> — resume a previous loop
• --dry-run / -n — show what would happen without writing
• --json / -j — JSON output

Loop Cycle (per round)

1. Reflect — LLM examines current knowledge vs goal, identifies gaps

   { gaps: string[], queries: string[], assessment: string, converging: boolean }

2. Search — execute web searches for identified queries
3. Fetch — load top URLs from search results
4. Digest — feed fetched content through digest pipeline
5. Journal — record what was searched, fetched, digested this round

State Persistence

Loop state is stored so loops can be resumed:

{
  id: string,
  goal: string,
  status: "running" | "converged" | "paused" | "max_rounds",
  rounds_completed: number,
  search_history: string[],  // avoid re-searching
  created_at: string,
  updated_at: string
}

Convergence

Loop stops when:

• LLM signals converging: true for 2 consecutive rounds
• Max rounds reached
• No new search results found

Web Search

Uses existing LLM tool-use infrastructure (Claude with WebSearch tool). Search history prevents duplicate fetches across rounds.

MCP Tool

loop: { goal: string, rounds?: number, dry_run?: boolean }
loop_resume: { id: string }
loop_list: {}

Depends On

• 059-digest: consume URLs, files, directories, stdin into knowledge graph #82 (059-digest) — loop digests fetched content
• 060-ask: conversational synthesis over accumulated knowledge #83 (060-ask) — loop's reflect phase is essentially an ask
• Web search capability (Claude tool use or equivalent)

Acceptance

• Single goal → multi-round autonomous research
• Web search finds relevant sources
• Each round digests new findings into graph
• Convergence detection stops the loop
• State persists for resume
• Search history prevents duplicate work
• Round journals record what happened
• MCP tools exposed
• Spike test passes