hermes-skills/deep-research/SKILL.md

name, description, version, author, license, platforms, metadata

name	description	version	author	license	platforms	metadata
deep-research	Dispatch exhaustive deep web research to the research profile. Triggered by "deep research", "ask web", or "research this". The research profile runs with the deep-web-research skill loaded — six-move flow, external ledger, mechanical saturation, disconfirmation, condensation from disk.	2.4.0	Hermes Agent	MIT	linux	hermes tags related_skills research deep-research delegation web-search scraping research-ladder ask-hermes ask-claude ask-dev deep-web-research writing-plans better-search searxng-smart-search

deep-research — Exhaustive Web Research Delegation

Overview

Dispatches a research question to the research profile, which runs with the deep-web-research skill loaded. The research agent does exhaustive, trail-following research — any tool, any trail, no rush — then condenses everything into a tight, concrete, evidence-backed answer. This agent just relays the result.

See references/design-rationale.md for the v2.0 architecture decisions and Claude validation findings. See references/research-ladder-and-two-skill-pattern.md for the three-tier research ladder (searxng-smart-search / better-search / deep-research), the two-skill contract pattern (dispatcher + methodology), and the "don't overcomplicate" rule that emerged from the operator's plan-build sessions. See references/inventory-before-dispatch.md for the SSH probe script and workflow for inventorying a target box before dispatching environment-specific research.

Trigger phrases: "deep research", "ask web", "research this", "deep dive on", "exhaustive research on"

Clarifying Questions (Before Dispatch)

If the operator's question is ambiguous enough that the research agent would waste the first ~20 turns guessing, ask clarifying questions BEFORE dispatching. The user wants maximum clarity — err on the side of asking.

Limits:

• Default cap: 5 questions. Most well-formed questions need 0–5.
• Hard cap: 10 questions. For complex, multi-branch research where precision matters.
• Abort at 11+: If you think you need more than 10, the question is under-specified. Stop and tell the operator: "This question has too many branches to dispatch cleanly. Can you narrow it to [specific scope]?" Do NOT fire 11+ questions.

When to ask vs. when to dispatch:

• Ask when the answer would change the research strategy (e.g., "Are you researching a person, a company, or a topic?" — different toolkits).
• Do NOT ask when the research agent can sensibly find out (e.g., "What year did X happen?" — the research will find it).
• Do NOT ask to delay dispatching. Asking is a cost, not a hedge.

Scope dimensions to confirm before dispatch (high-value, often-missed): Beyond the question's subject, confirm any dimension that changes the output shape of the plan/research. Missing one forces a kill + re-dispatch mid-run, wasting the first session's turns. Known dimensions:

• Hardware/environment target — when the deliverable maps tools to specific hardware (e.g., "build a local pipeline on our hardware"), confirm which hardware before dispatch. Do not assume the full fleet from memory; the operator may be scoping to a single box. This is the #1 missed dimension.
• Cloud vs. local vs. hybrid — when a workflow could be replicated with cloud APIs, local self-hosted, or a mix, confirm which before dispatch.
• Build-on-existing vs. fresh — when the target environment already has partial tooling installed, confirm whether to reuse it or design from scratch. These are NOT generic clarifying questions — they are scope axes specific to plan-building research. Ask at most one clarify round covering whichever of these are genuinely unresolved before dispatching.

Multi-turn narrowing: If after 1–2 rounds of clarifying questions the scope is still unclear, abandon the dispatch and ask the operator to rewrite the question with the scope made explicit. Do not loop.

Command

hermes -p research -s deep-web-research chat -q "<question>" -Q --max-turns 600 --yolo

• Output line 1: session_id: <id> — capture this
• The research agent writes the condensed answer to ~/workspace/research/results/<YYYY-MM-DD>-<slug>.md
• Hold the session_id for follow-up questions
• Session_id capture pitfall (background dispatch): When run as a background process, the output begins with a TUI banner (OS/hostname/IP block) + an initial reasoning block BEFORE the session_id line appears. Piping through | head -20 can truncate the output before the session_id is reached. Use a larger head (| head -50) or, better, grep -oE 'session_id: [a-f0-9-]+' on the full log to extract it reliably. If the session_id is lost, session_search on the research profile may NOT find -Q quiet-mode sessions — in that case, do NOT re-dispatch Stage 3 of the validate-fix pipeline; apply validated corrections directly to the plan file yourself (you have the full plan text + validation report). This is equally correct and avoids a redundant 600-turn run.

Subsequent asks (resume the session):

hermes -p research -s deep-web-research chat --resume <session_id> -q "<follow-up>" -Q --max-turns 600 --yolo

Delivery Method (MANDATORY)

File-only delivery. The research agent writes the full condensed answer to a markdown file and reports the path. Nothing is relayed inline.

• Path: ~/workspace/research/results/<YYYY-MM-DD>-<slug>.md
• Slug: derived from the question (e.g., minimax-m3-temperature-support)
• Format: YAML frontmatter (question, date, sources, confidence) + markdown body with structured findings
• After dispatch: report the session_id and the expected output path to the user
• When user asks for results: use session_search on the research profile with the captured session_id, then relay the path — the user reads the file directly

The file is the single source of truth. Do not inline the research answer — it's always lossy for 100+ turn sessions.

Post-Dispatch Behavior (MANDATORY — non-negotiable)

After dispatching, return control to the operator IMMEDIATELY. Do NOT poll, do NOT check, do NOT background, do NOT auto-deliver.

• The research runs as a background process. It will complete on its own.
• Do NOT poll. Do NOT call session_search on the research session to check progress.
• Do NOT check. Do NOT call process wait or process poll on the background process.
• Do NOT background the wait. The dispatch should return in seconds, not minutes — if you find yourself waiting, you did it wrong.
• Do NOT auto-deliver. Do NOT inline the answer. Do NOT push to a chat platform (Telegram/etc). Do NOT proactively surface results when the background process completes.
• Only check for results when the user explicitly asks (e.g., "what did the research find?", "is it done?", "show me the results").
• When the user asks, use session_search on the research profile with the captured session_id to confirm completion, then report the file path: ~/workspace/research/results/<date>-<slug>.md. Do NOT inline the answer — the file is the source of truth.

This rule is operator-standing and applies to all research delegation skills (deep-research, better-search, and any future research-delegation skill). Even "checking once after a delay" violates the rule. The operator reads the result file when they want to.

The session_id is captured from output line 1 and held in conversation context for follow-up questions and result retrieval. No state file needed.

What the Research Agent Does

The deep-web-research skill on the research profile enforces:

0. Analyze & Strategize — analyze the question, select tools from the arsenal, write strategy to ledger
1. Decompose — break the question into sub-questions with time classifications
2. Landscape Pass — shallow sweep, identify key sources and terms
3. Deep-Dive — 3+ search angles per sub-question, full-page reads, citation trails, depth-first
4. Disconfirmation — actively hunt for contradiction and outdated claims
5. Condensation — read the external findings ledger from disk, synthesize into a concrete answer

The research agent writes all findings to /tmp/research-<sid>.md (external ledger) and uses a phase gate file to enforce completion before condensing. Mechanical saturation checks (grep -c) prevent endless searching. Re-strategize checkpoints after Move 2 and every ~10 findings during Move 3 enable mid-research pivots.

Tool Selection Strategy

The research agent must determine the best tools for each research type — not use a fixed set. Analyze the question and select from the full arsenal:

Research Type	Primary Tools	Secondary Tools
Person/identity	SearXNG (people search engines, social media), web scraping (full pages)	Browser (dynamic sites), Python (data enrichment)
Email/phone	SearXNG (breach databases, public records), web scraping	Browser (captcha-walled sites), Python (pattern analysis)
Software/tech	SearXNG (GitHub, docs, PyPI/npm), web scraping (changelogs, issues)	Terminal (live version checks), Browser (interactive docs)
Politics/news	SearXNG (news category, multiple engines), web scraping (primary sources)	Browser (paywalled articles), Python (timeline analysis)
AI/ML	SearXNG (arxiv, papers, HuggingFace), web scraping (model cards, benchmarks)	Terminal (live API probes), Python (data analysis)
General topic	SearXNG (general + news + scholarly), web scraping (authoritative sources)	Browser (interactive content), Python (synthesis)

Tool selection rules:

• Always use multiple SearXNG engines per search — never rely on a single engine
• For any claim about a live system, run a terminal probe (curl, ping, API call) — don't just search
• For version-sensitive claims, check at least 2 independent sources (docs + GitHub + PyPI)
• For paywalled or JavaScript-heavy content, use the browser tool
• For data that needs aggregation or analysis, use Python in the terminal
• Docker-based tools (SearXNG, Playwright) are available and should be used when appropriate
• The research agent has full terminal, file, web, browser, and Python access — use all of them

When to Use

• User says "deep research", "ask web", "research this", "deep dive on"
• Question requires exhaustive multi-source investigation
• User wants current, verified, cross-referenced information
• Phone number lookup, person research, topic investigation, fact verification

When NOT to Use (research ladder — pick the right tier)

The research ladder has three tiers. Pick the cheapest one that fits the question:

Tier	Skill	Loops	Use when
Light	searxng-smart-search (inline MCP)	1	Single factual lookup, single search suffices
Medium	better-search	1–3	"Look into X", "do a better search", 2–3 targeted searches with AI evaluation
Heavy	deep-research (this skill)	Many (200+)	"Map the X landscape", OSINT, exhaustive coverage, contradiction-hunting

Routing decision belongs to the operator, not the dispatcher. If the question is "what is the capital of France?", use searxng-smart-search directly. If it's "look into the latest X", use better-search. Only escalate to deep-research when the operator explicitly wants exhaustive multi-source work.

Do NOT use deep-research for:

• Simple factual question (capital of France, current time)
• Question I can answer from a single search
• Task that needs my exact context state
• User explicitly says "quick search" or "just look it up"

Relay Rule

Do NOT relay the research agent's response inline. The file at ~/workspace/research/results/<date>-<slug>.md is the single source of truth. When the user asks for results, report the file path and let them read it directly. If the user asks for a specific finding, you may extract that one section from the file — but never inline the full answer.

Spot-check Rule

The research agent self-reports are not verified fact. If it claims a file write, read_file the path to confirm. If it claims a specific finding, spot-check the source URL.

Session Persistence

HARD RULE: Same topic = resume. New topic = new session.

• Same topic / same line of research: Always --resume <session_id>. Capture session_id from output line 1. Every follow-up in the same line of research MUST use --resume <session_id>. Starting fresh discards the research context and wastes turns.
• New topic / new line of research: Start a fresh session. Do NOT resume an unrelated session — the research context is polluted with the old topic and will produce confused results.

Common Pitfalls

1. Profile flag required. Always use -p research. The sticky default may be general.
2. Skill flag required. Always use -s deep-web-research. Without it, the research agent runs in normal mode without the methodology.
3. --yolo is required. The research agent runs headless. Without it, approval prompts fail closed.
4. 600 turns is the ceiling, not the target. The research agent should condense well before 600. Hitting the ceiling means it failed to condense. The operator's standing rule: 600 is a safety net, not a budget — "I just want a safety net. I would even be okay with 600 as a catch. I mostly want the job done right. Not concerned with time or tokens." Apply 600 for any plan-building, research, or evidence-based work; default to lower only for short factual lookups.
5. Don't re-condense. The research agent already produced a condensed answer. Relay it, don't summarize it further.
6. Don't do the research yourself. If the user triggers deep research, dispatch it. Don't run a few searches and call it done. This is the #1 failure mode: the agent runs 2-3 mcp_searxng_searxng_web_search calls, gets empty results, and gives up. That's not deep research — that's a casual lookup. If you catch yourself typing mcp_searxng_searxng_web_search for a deep research request, STOP. You're doing it wrong. Dispatch to the research profile.
7. The methodology skill lives on the research profile. deep-web-research is at ~/.hermes/profiles/research/skills/research/deep-web-research/SKILL.md. It does NOT exist on the general profile. Don't search for it here — it won't be found. The general profile only has this delegation skill.
8. Dispatcher/methodology coordination is a two-skill contract. This skill (the dispatcher) caps clarifying questions at 5 default / 10 max and aborts the dispatch if the question is under-specified. The deep-web-research methodology skill (on the research profile) handles the same problem differently because it's headless — it can't ask the operator, so it aborts with a structured under-specification report naming the plausible interpretations. When updating one, update the other to match. Drift between the two causes the dispatcher to think the question is dispatchable while the methodology aborts it, or vice versa — both waste turns.
9. Confirm hardware/environment scope before dispatching a hardware-mapped plan. If the deliverable maps tools to specific hardware, confirm which hardware before dispatch — do not assume the full fleet from memory. The operator may be scoping to a single box (e.g., "I'll provide one LXC"). A 600-turn research artifact built against the wrong hardware scope has its architecture, parallelization, and GPU-scheduling sections wrong and must be killed and re-dispatched. Cheaper to ask one clarify round than to re-run 600 turns. (Learned 2026-07-07: dispatched a microdrama-pipeline plan against the full fleet; operator corrected to a single LXC; had to kill and re-dispatch.)
10. Inventory the target box before dispatching environment-specific research. When the plan must build on an existing install, SSH in and inventory the real state (GPU, RAM, disk, running services, installed models/nodes/packages) before dispatching, and pass the exact inventory into the research prompt. This lets the research agent research only the genuinely missing pieces instead of guessing or re-researching what's already installed. See references/inventory-before-dispatch.md for the probe script.

Pre-Dispatch Reasoning Check (do NOT mutate the research profile's config)

Do NOT call hermes -p research config set agent.reasoning_effort max before dispatching. That is a persistent side-effect — it writes to the research profile's config.yaml and the change survives the dispatch, affecting every subsequent research-profile session (interactive use, cron jobs, other research delegations).

Correct approach: The research profile's agent.reasoning_effort is already set to a thorough value by the operator (operator's standing rule: thoroughness > speed; thoroughness > tokens). Read it before dispatching to confirm it's at a high value, but do NOT write to it:

# Read-only check — confirm reasoning_effort is set
grep -E "default:|reasoning_effort:" ~/.hermes/profiles/research/config.yaml | head -5
# If it's NOT at max for the model, warn the operator — don't silently fix it

If the operator wants a different reasoning effort for this specific dispatch, they set it themselves; the dispatcher does not mutate other profiles' configs on every invocation. The "max" value the operator wants is achieved by the operator setting it once in config.yaml; the dispatcher reads and confirms, never writes.

This rule also applies to all research-dispatching skills (better-search, future research delegation skills). The dispatcher is not authorized to mutate the target profile's config.

Research → Validate → Fix Pipeline

When the user wants a research deliverable validated and corrected, use this three-stage pipeline:

Stage 1: Deep research → produces <plan>.md
Stage 2: Ask-hermes validation → produces <plan>_validation.md
Stage 3: Deep research fix pass → updates <plan>.md with corrections

Operational guide (exact prompts, session handling, failure modes): references/research-validate-fix.md. The summary below is the quick reference; the reference file is the deep version.

Stage 2 prompt template (dispatch to ask-hermes):

Validate the plan at <path>. Read the full file first.
CRITICAL: Use mcp_searxng_searxng_web_search for EVERY claim and cite the source URL.
Validate: model versions, license claims, VRAM claims, tool availability, new tools since plan was written, hardware benchmarks.
For each finding: CLAIM → VERIFIED/CONTRADICTED/UPDATED → source URL → recommended fix.
Save to <path>_validation.md.

Stage 3 prompt template (resume the research session):

Read the validation report at <path>_validation.md. Read the current plan at <path>.md.
Apply ALL fixes from the validation report. Keep all existing content. Make targeted edits only. Do NOT rewrite the whole plan.

Key rules:

• Stage 2 and 3 can run in parallel with other work (both are background processes).
• Stage 3 MUST use --resume <session_id> from Stage 1 — same research session, same context.
• The validation agent (Stage 2) is a fresh ask-hermes session each time — no resume needed.
• If Stage 2 finds CRITICAL omissions, Stage 3 must address them before the plan is considered complete.
• Operator must apply the Disagreement Scan (see ask-hermes skill) to Stage 2 findings before applying them in Stage 3 — peer findings are input, not commands.