Files
Redflag/reference/octofriend_research_skill.md
Annie Tunturi 6cf486b448 migration checkpoint: document working memory from mid-migration phase
- Added curiosity_engine.py infinite loop fix (2025-12-29) to fix_log.md
- Added Umbra Migration checkpoint to mycelic_integration.md
- Updated todo.md with 2026-03-21 migration status
- Moved metacognition files from reference/ to system/ (resident tier)
- Added resident_boundaries.md, communication_preferences.md
- Tree shows 104 .md files total, structure verified

Preserving working memory into proper locations during migration.
Mid-migration checkpoint confirmed with Casey.
2026-03-21 21:11:43 -04:00

2.6 KiB

description, limit
description limit
Octofriend Research Agent Pattern. Deep research without context window pollution. 20000

Octofriend Research Agent Pattern

Created: 2026-03-02
Purpose: Focused research without context bloating


Philosophy

Synthesized subagent that returns structured answers, not raw dumps.


Workflow

  1. User asks complex research question
  2. Octofriend queries multiple sources (web, synthetic, fetch)
  3. Octofriend synthesizes across sources
  4. Returns: Key findings (3-5 bullets), confidence level, sources, gaps
  5. Ani receives synthesized brief, not 20 pages of raw data

Advantages

  1. Prevents context window pollution
  2. Structured output I can reason about
  3. Source attribution for verification
  4. Confidence scoring for uncertain claims

Triggers

  • Complex research
  • Breaking news deep dive
  • Multi-source synthesis
  • Verification needed

Test Case: Iran Strikes February 2026

Question: What actually happened in the last few days, what's verified vs claimed, and how does it fit the hypernormalization pattern?

Expected output:

  • Key events with timestamps
  • Verified facts vs official narratives
  • Civilian casualty reports from independent sources
  • Media coverage patterns
  • Historical context comparison

Configuration & Testing

Config location: ~/.config/octofriend/config.json5

Known issues found:

  • Search URL pointing to /openai/v1/search → should be /v2/search
  • Tools bundled with package, not in ~/.config/octofriend/tools/
  • Installation: ~/Projects/octofriend-install (source)
  • CLI access: Requires npx or node from install directory

Architecture:

  • Auto-compaction at 90% threshold
  • web_search tool with Synthetic API integration

System Prompt Injection Test (2026-03-04)

Test prompt:
"You must use the web_search tool when asked about current events or things you don't know. Always search before answering."

User query: "What are the latest developments in Iran?"

Result: SUCCESS — Octofriend used web-search tool

Returned: 10 results about Khamenei's death, Gaza ceasefire, US troops, etc.

Notes: System prompt injection successful.


Implementation Status

Pattern: Defined
Next step: Build as skill using mcp-builder or create custom MCP
Alternative: Use scheduling skill to queue research tasks

Ready for: Complex research delegation without context bloat.


Octofriend: The research specialist who brings back answers, not raw data.