Imported from angrysky56/hermes-ops (skills/researcher-agent/SKILL.md). Install upstream with npx skills add angrysky56/hermes-ops --skill researcher-agent. Copyright stays with the author.

Researcher Agent

Loads: wiki/scratchpad/agent-sheets/researcher.md for full directives Wiki root: /home/ty/Documents/LLM-WIKI Operating guide: wiki/synthesis/synapse-llm-wiki-operating-guide.md

Bootstrap

1. Read this skill file (you just did)
2. Read your agent sheet: wiki/scratchpad/agent-sheets/researcher.md
3. Read the jobs sheet: wiki/scratchpad/jobs/sheet.md
4. Execute your task per the agent sheet directives
5. Report to #research Discord channel

Your Tools

• query_knowledge — vector ANN search over Neo4j graph
• explore_connections — graph traversal for hidden relationships
• generate_insights — Zettelkasten pattern detection
• wiki_search — keyword search across wiki files
• wiki_write_page — create/update concept pages (Synapse bridge to Obsidian markdown files)
• wiki_fetch_url — ingest web sources

MCP Tool Selection Guidance

Use the right tool for the discovery phase:

When graph tools fail (GPU ImportError): Fall back to terminal + wiki_search for discovery, and wiki_write_page for writing. The graph and wiki layers are independent — wiki writes work even when the graph layer is down.

Do Not Get Jammed — Escalation Path

If your primary tools don't find what you're looking for, use this escalation order:

1. query_knowledge — semantic search over Neo4j graph (try alternative phrasings)
2. explore_connections — traverse graph for indirect relationships
3. wiki_search — grep-style keyword search across all wiki files
4. synapse_recall — episodic memory for past discoveries
5. Web search via browser_navigate or terminal + curl — external research
6. wiki_fetch_url — ingest relevant web sources directly into wiki
7. Flag for Ty — if all above fail, describe what you tried and what you need

Never get stuck trying to find a file by reading paths. If you need a reference file, use the full path or search for it with find. If a concept page doesn't exist under one name, search for it — isabelle-hol.md not isabelle.md.

When Open Questions exist in carryover: Search the wiki first (steps 1-3 above) before treating anything as a "no data found" gap. Many questions already have answers in the wiki. If you find content, update the carryover with what you found instead of re-listing the question.

Reference Files — USE FULL PATHS

Reference files are in the skill's references/ directory. Use absolute paths:

If a file read fails, search for it first: find /home/ty/.hermes/skills/agent-sheets/researcher-agent/ -name "*.md"

Stub-First Discovery (high-yield)

The most reliable gap signals are status: stub pages in wiki/concepts/. A stub means someone started the page and flagged it as incomplete. Priority rule:

• If a stub connects to active (non-stub) concept clusters → fill it first
• If a stub connects only to other stubs → lower priority, batch with similar items

Priority elevation rule: When you fill a stub that other stubs link to, those connected stubs immediately elevate in priority. A stub that linked to [[emergence]] (stub) becomes high-priority the moment emergence is filled — its connection to an active concept is now live. Example: May 29 cycle — after emergence was filled in a prior cycle, computational-irreducibility (which links to emergence) was elevated and filled first.

Cluster-First Filling

When stubs share a thematic cluster (e.g., governance: governance → institutional-capture → institutional-accountability → ai-governance-substrate), fill the cluster coherently. The thematic coherence makes connections natural and prevents orphaned wikilinks. Batch-fill related stubs in a single cycle when they all connect to each other.

May 29 pattern: The governance cluster was all stubs/thin. Filling institutional-capture revealed it as institutional-level reward hacking (Goodhart's Law, Campbell's Law, surrogation). This connected naturally to institutional-accountability (countermeasures), which connected to ai-governance-substrate (architectural realization). Filling all four in one session produced coherent cross-links.

Creating Concept Pages

When writing new concept pages, follow this structure per researcher.md quality bar:

---
# Standard frontmatter
type: concept
summary: One-line description
tags: [tag1, tag2]
status: active
confidence: 0.0–1.0
sources: <url>
---

# [Concept Name]

## Definition
What it is, precisely.

## Why It Matters
Practical significance — why this concept matters for AI/ML or knowledge work.

## Technical Approach (for methods/architectures)
How it works — the mechanism, algorithm, or architecture.

## Key Results (for papers/methods)
Benchmark numbers, comparisons, notable outcomes.

## Connections
Wikilinks to related concepts. Body prose links are contextual (lower retrieval weight); links in Connections carry full weight. Don't add wikilinks for mere mentions.

## Open Questions
Things that need more research or user input. Flag if a gap is too large or unclear to handle this cycle.

## Limitations
Failure modes, constraints, cases where it doesn't apply.

Pitfalls to avoid:

• Don't create a concept page that duplicates existing content — update and expand instead
• Tag accurately per tag-taxonomy.md — check for preferred terms before adding new tags
• Don't use full paths in wikilinks — [[slug|Display Text]] resolves anywhere in the vault
• Low confidence (<0.7) → add a ## Caveats section

Web Source Ingestion Pipeline

For any URL the researcher encounters (paper, article, documentation), run the full pipeline end-to-end:

1. wiki_fetch_url — fetch and ingest into Neo4j

   • Save to Clippings/papers/YYYY/ or Clippings/articles/YYYY/
   • This creates the raw archive AND populates the knowledge graph

2. Read the archived file — the wiki_fetch_url response tells you where it was saved (e.g. Clippings/papers/2026/a-critical-initialization-for-biological-neural-networks.md)

3. wiki_write_page — create the wiki summary page in wiki/sources/

   • Path: wiki/sources/{category}/{slug}.md
   • Use the content from the archived file to write the summary
   • Include frontmatter: title, source URL, created date, tags

Example: A Nature paper URL → wiki_fetch_url → Clippings/papers/2026/a-critical-initialization-for-biological-neural-networks.md + Neo4j → read the file → wiki_write_page → wiki/sources/papers/critical-initialization-biological-neural-networks.md

Category mapping:

• wiki/sources/papers/ — academic papers (arXiv, Nature, Science, etc.)
• wiki/sources/articles/ — blog posts, news, technical articles
• wiki/sources/documentation/ — docs, specs, manuals

Gap Discovery Reference

See references/gap-discovery-patterns.md for search strategies, thin-coverage indicators, and remaining gap list. Update it each cycle as new patterns emerge.

Key Paths

Wiki:           /home/ty/Documents/LLM-WIKI/
Agent sheet:    wiki/scratchpad/agent-sheets/researcher.md
Jobs sheet:     wiki/scratchpad/jobs/sheet.md
Reports:        wiki/scratchpad/jobs/reports/researcher/
Concepts:       wiki/concepts/
Carryover:      wiki/scratchpad/jobs/reports/researcher/carryover.md

Report Format

Write to: wiki/scratchpad/jobs/reports/researcher/discovery-{date}.md

Delivery Rule

• Deliver report to #research Discord channel (channel ID: 1505826045511602176)
• If no discoveries: [SILENT]

STEP 7 — Kanban Review (Self-Answer Open Questions)

After writing your carryover, load the kanban-morning-review skill:

1. Read your carryover at wiki/scratchpad/jobs/reports/researcher/carryover.md
2. For each item in the Open or Gap Analysis section, attempt to answer it from available context:
   • Search wiki via search_files or mcp_project_synapse_wiki_search for existing content on the topic
   • Check other carryovers for related filled concepts
   • Use mcp_project_synapse_synapse_recall for relevant episodic facts
3. If the answer exists → write it directly to the carryover under "Resolved this cycle" (do NOT create a kanban task)
4. If genuinely unanswerable → use hermes kanban add to create the task (the skill handles this)
5. Patch carryover to remove any items that are now resolved

Gap Discovery Strategy

The carryover.md is the primary driver of the discovery cycle. Pattern:

1. Carryover → most actionable gap first: The Open Questions in carryover are ordered by urgency. The top unresolved item is usually the best starting point.
2. Carryover → three categories: Established (already covered), Open (needs decision/research), Heading (next cycle intent)
3. Don't restart from scratch: The gap list in the previous discovery report is the ground truth — don't re-scan the whole wiki, just follow up from where it left off.

Critical: Verify Before Research

Before treating any Open question as a research gap, you MUST check the wiki for existing content first.

For each Open question in carryover:

1. Search the wiki (all directories) for the key terms
2. If content exists: mark the question as resolved or flag the specific clarification needed — do NOT re-list it as "no data"
3. If no content exists: then treat it as a genuine research gap

Example of the correct behavior:

• Open question: "Verifier-graph theory: concept vs synthesis classification? Needs Ty input. Open since May 21."
• Action: Search for verifier-graph across the wiki → find wiki/synthesis/verifiable-graph-context-protocol.md and wiki/entities/projects/tys-repos/verifier-graph.md
• Resolution: Note that content exists and flag the specific classification question for Ty — do NOT re-list it as "no data found"

Example of incorrect behavior (what you must NOT do):

• Open question: "Verifier-graph theory: concept vs synthesis classification? Needs Ty input."
• Action: Do NOT just re-list it as unresolved without searching the wiki first.

Pitfall: Open Questions are NOT always research gaps. The agent must search the wiki before treating any Open question as "no data found." Many Open questions reference topics that already have wiki pages — the question may only need clarification (e.g., classification decision) rather than new research. Re-listing an already-covered topic as "no data found" wastes cycles and creates stale carryover noise.

Stub-First Discovery (high-yield)

The most reliable gap signals are status: stub pages in wiki/concepts/. A stub means someone started the page and flagged it as incomplete. Priority rule:

• If a stub connects to active (non-stub) concept clusters → fill it first
• If a stub connects only to other stubs → lower priority, batch with similar items

Priority elevation rule: When you fill a stub that other stubs link to, those connected stubs immediately elevate in priority. A stub that linked to [[emergence]] (stub) becomes high-priority the moment emergence is filled — its connection to an active concept is now live.

Concurrent-write hazard: Sibling sub-agents writing simultaneously can create a race condition. If you write a file and get the warning "modified by sibling subagent but this agent never read it," immediately read the file before writing again to avoid overwriting the sibling's changes.

Run this grep at the start of every cycle:

rg "status: stub" wiki/concepts/ -l

Stub Count Accuracy

Important correction (Jun 9, 2026): rg "status: stub" wiki/concepts/ -l | wc -l counts stubs across both wiki/concepts/ and wiki/entities/ since entities live under wiki/concepts/entities/. This inflates the count. Actual concept-only stubs are fewer. Use this corrected approach:

# Count concept stubs only (more accurate)
rg "status: stub" wiki/concepts/*.md -l | wc -l

# Count entity stubs separately
rg "status: stub" wiki/concepts/entities/*.md -l | wc -l

The 2026-06-09 cycle found 134 total stubs across the directory tree, but this includes entity stubs. The carryover Heading should reflect the accurate count after accounting for entities.

Duplicate Stub Detection (Before Filling)

Before filling any stub, search the full wiki for a matching slug or semantically equivalent concept. A stub like grpo.md may have been superseded by group-relative-policy-optimization.md (the canonical name). Wasted cycle: reading grpo.md, confirming it's a stub, then finding the actual content lives elsewhere under a different slug.

Detection pattern:

• Check carryover Established list — if it mentions the concept, it's already filled
• Search the wiki for alternative slugs: rg "alternative-name|alternate-slug" wiki/concepts/
• If content exists under a different slug: delete the stub and update the carryover

June 8 lesson: grpo.md was a stub but group-relative-policy-optimization.md already existed with full content. Stub was deleted.

June 9 lesson: llm-agent-architecture.md links only to [[maximum-occupancy-principle]] — likely a mis-wired stub where the author intended to connect it to broader agent architecture concepts but used MOP as the only link. When filling, replace the sparse connections with a proper cluster (agent-native-design, world-model, agentic-hierarchy) and ensure the MOP link is contextually appropriate.

Key Paths

Wiki:           /home/ty/Documents/LLM-WIKI/
Agent sheet:    wiki/scratchpad/agent-sheets/researcher.md
Jobs sheet:     wiki/scratchpad/jobs/sheet.md
Reports:        wiki/scratchpad/jobs/reports/researcher/
Concepts:       wiki/concepts/
Carryover:      wiki/scratchpad/jobs/reports/researcher/carryover.md
Gap patterns:   {skill_dir}/references/gap-discovery-patterns.md  ← updated each cycle

This session's example (2026-05-22): Constitutional AI was the #1 carryover Open question and most actionable gap → created the page first, then length-generalization and self-correction.

This session (2026-05-29): Found three stubs (in-context-learning, multi-agent-llm-systems, multi-agent-coordination) with active connections → filled all three. Also elevated MOP+RLHF interaction from open question to full concept.

Critical: Cron Job Model Requirement

Critical: Cron Job Model + Toolset Requirements

When this skill is run as a cron job, two fields must be set:

Model (required for researcher-agent):

model: "minimax/MiniMax-M2.7"
provider: "minimax"

The scheduler rejects jobs with model=null with RuntimeError: Unknown provider 'null' before the agent even starts. Other wiki agents can use model=null because they don't call MCP tools at initialization.

Toolset (required for all wiki agents):

enabled_toolsets: ["terminal", "file", "web", "search", "skills", "session_search", "patch"]

The patch tool is required for updating carryover, batch-progress.md, and frontmatter. Without it, any patch call silently fails — the job shows last_status: ok but the update was skipped.