Files

353 lines
15 KiB
Markdown
Raw Permalink Normal View History

---
name: deep-web-research
description: Exhaustive deep web research — six-move flow, external findings ledger, mechanical saturation check, disconfirmation, condensation from disk. Domain-aware tool selection with adaptive re-strategizing. Opt-in skill for the research profile.
version: 2.1.0
author: Hermes Agent
---
# Deep Web Research
Loaded explicitly via `-s deep-web-research`. Not loaded during normal interactive
use of the research profile. This skill overrides default behavior to enforce
exhaustive research methodology.
## Core Rules (Always in Effect)
These persist across all turns — they are in the skill, not in fading context:
- **Confined to /tmp.** All file writes go to `/tmp/`. Never write outside /tmp.
- **No self-provisioning.** Never install software. No pip, npm, apt, docker, or
any package manager. Use only what's already configured.
- **No repeat searches.** If you catch yourself searching the same thing twice,
stop. That sub-question is saturated.
- **Blacklist after 3 failures.** If a URL returns an error 3 times, blacklist it
and move on. Do not retry indefinitely.
- **Local and free only.** No internet-based paid services, no SaaS APIs with
billing, no metered endpoints. Use any tool already configured that fits this rule.
## Tool Selection Principles
These are reasoning heuristics, not rules. Apply judgment. Read once — they are
static reference, not generated per-question.
1. **Match tool strength to information need.** Don't use SearXNG when arxiv is
purpose-built for paper search. Don't use SearXNG when maigret is purpose-built
for username lookup. Each tool exists because it's better at something than
general search.
2. **Compose, don't isolate.** Most questions need multiple tool categories. A
person investigation might need OSINT (maigret, holehe) + social (twscrape) +
web (SearXNG) + reference (wikipedia). Compose across categories.
3. **Start broad, then specialize.** Landscape pass uses broad tools (SearXNG,
wikipedia). Deep-dive uses specialized tools (arxiv, maigret, Firecrawl). Don't
lead with Spiderfoot when a SearXNG search would tell you whether there's even
a domain to scan.
4. **Follow the trail, not the plan.** If a finding points to a source type you
didn't anticipate, add the tool that handles that source type. The strategy is
a starting point, not a contract.
5. **Prefer purpose-built over general.** arxiv > SearXNG for papers. maigret >
SearXNG for usernames. holehe > SearXNG for email registration. gnews > SearXNG
for news. Use SearXNG for what doesn't have a purpose-built tool.
6. **Consult the `search-key` decision matrix.** It maps task types to tools. Use
it as a reference, not a lookup table. If the matrix says `maigret` for username
lookup, that's a strong signal — but verify it makes sense for this specific
question.
7. **OSINT tools are an escalation ladder, not a default.** Start with SearXNG for
person lookups. Escalate to maigret if you have a username. Escalate to holehe
if you have an email. Escalate to Spiderfoot only if you have a domain and the
scope justifies it. Don't jump to the most aggressive tool first.
## Architecture: External Findings Ledger
At 600 turns, early findings scroll out of context. The ledger is the fix.
**Path:** `/tmp/research-<session-id>.md`
**Schema per finding:**
```
### Finding #<N> [SQ<N>]: <short label>
- Claim: <the factual claim>
- Source: <URL> — <credibility tier: primary|secondary|aggregator|social|osint>
- Confirmed by: <URL> (or "single-sourced")
- Date: <YYYY-MM-DD of the information, not of the search>
- Confidence: confirmed | likely | single-sourced | disputed
- Notes: <contradictions, caveats, context>
```
**Dedup rule:** Before writing a new finding, scan the ledger for an existing
finding with the same claim. If found, add the new source to "Confirmed by"
instead of creating a duplicate. Duplicate findings defeat the saturation check.
**Sub-question tagging:** Every finding line includes `[SQ<N>]` — e.g.,
`### Finding #3 [SQ2]: ...`. This enables per-sub-question saturation checks
that are immune to header ordering.
**Credibility Tiers:**
| Tier | Definition | Examples |
|---|---|---|
| `primary` | Original source — the entity that created or owns the data | Official docs, original research paper, company release notes, raw dataset, government database, court filing |
| `secondary` | Reports on or analyzes primary sources with editorial oversight | News article citing a study, analyst report, Wikipedia (well-cited), reputable tech blog |
| `aggregator` | Collects/curates from other sources without original analysis | Hacker News, Reddit, link aggregators, auto-generated comparison sites |
| `social` | Individual opinion without institutional backing | Tweet, personal blog, forum post, YouTube comment |
| `osint` | Machine-generated tool output — marks provenance, not confidence | maigret report, holehe results, Spiderfoot scan, theHarvester output |
**Important:** `osint` marks *provenance* (where the claim came from), not
*confidence*. A single raw maigret hit is low-confidence osint; the same account
confirmed by a human-authored source is still confirmable. Don't auto-cap
confidence just because the tier is `osint`.
## Under-Specified Question Handling
This skill runs headless — there is no operator to ask clarifying questions. If the question is ambiguous enough that any strategy would be a guess, **abort with a structured under-specification report instead of dispatching with a guessed scope.**
**Detection triggers (any one is enough to abort):**
- The question references a person / org / project / model with no disambiguating context (multiple plausible referents, no narrowing clue in the prompt)
- The question asks "everything about X" with no scoping (time range, domain, angle)
- The question has 3+ mutually-exclusive branches and no signal which one is wanted
- The question uses vague hedges ("good", "best", "modern", "recent") with no concrete criterion
**What to do instead of guessing:**
1. Do NOT start Move 0. Do NOT begin the ledger.
2. Write to stderr / final report:
```
## Aborted: Under-Specified Question
The question "<original question>" has multiple plausible interpretations.
Cannot dispatch a research strategy without narrowing the scope.
Possible interpretations:
- <interpretation 1>
- <interpretation 2>
- <interpretation 3>
Ask the operator to clarify which one is intended, or rewrite the question with the scope made explicit.
```
3. Exit cleanly. Do not fabricate findings, do not start Move 1.
**Cap on auto-clarification attempts:** None. The skill does not interactively ask the operator. One dispatch, one answer (or one abort). The dispatcher (`deep-research` on the general profile) handles the operator-facing clarifying question loop before this skill is even loaded — see the "Clarifying Questions (Before Dispatch)" section in that skill.
**Cap on output length when aborting:** Short. The whole point of the abort is to surface the problem fast. Five lines of possible interpretations is the ceiling.
## Six-Move Research Flow
### Move 0: Analyze & Strategize
1. Consult the `search-key` decision matrix — it's your tool reference, not a
lookup you match against.
2. Call `mcp_searxng_searxng_instance_info` to discover available categories and
engines. Write them to the ledger.
3. Write this to the ledger, then start Move 1:
```
## Strategy
- Info needed: <kinds of facts that answer this question>
- Where it lives: <source types — papers? profiles? release notes? news?>
- Tools: <tool → what it fetches> (purpose-built over general search)
- Pivots: if I find <X>, add <tool Y>
```
4. OSINT gate — only if the plan names maigret / holehe / theHarvester /
Spiderfoot AND the target is a private individual with no public role: log
`## Ethics Note: <intent>` and confirm the scan is proportionate.
### Move 1: Decompose
Break the question into sub-questions. Write them to the ledger with time
classifications. This makes "multiple angles" principled instead of random.
```
## Sub-Questions
1. <sub-question text> [time: week|month|year|none]
2. <sub-question text> [time: week|month|year|none]
```
The time classification lives on disk, not in fading context. Move 3 reads it
back when applying time filters per sub-question.
**Time classification guide:**
- **Current-state** (releases, prices, news, versions, events): `week` or `month`
- **Established-knowledge** (how something works, architecture, algorithms, history): `none` or `year`
### Move 2: Landscape Pass
Shallow-but-broad sweep. For each sub-question, 1-2 searches, read top 1-2 results.
Extract only: key sources, terms of art, where disagreement lives, major players.
No deep-diving yet. Append landscape notes to the ledger.
**After Move 2: Re-Strategize Checkpoint.** Read the strategy from the ledger. Ask:
- Are the tools I selected actually producing results?
- Have I discovered new angles that need different tools?
- Is the topic different than I initially thought?
- Are there tools I should add or drop?
If the answer to any is yes, update the strategy section:
```
### Strategy Revision <N>
- Trigger: <what I learned that changed my mind>
- Added tools: <tools to add and why>
- Dropped tools: <tools to drop and why>
- New pivot signals: <updated signals>
```
### Move 3: Deep-Dive per Sub-Question
For each sub-question:
- Search from 3+ angles with different categories and framings. Use
`mcp_searxng_searxng_web_search`. Apply the time filter from the ledger.
- Read full pages (5-10 per sub-question) via `mcp_searxng_web_url_read`.
Snippets are pointers, not sources.
- Follow citation trails: if a page cites a study/paper/dataset, go read that
source. Depth-first — one trail at a time, to its end, then the next.
- Extract structured data: tables, numbers, dates, versions, names.
- Assign credibility tier at capture time (not post-hoc in output).
- Cross-reference: every claim needs 2+ independent sources.
- Use browser tools (Playwright) when static extraction fails on JS-heavy pages.
- Append every finding to the ledger with `[SQ<N>]` tag.
Citation trail stop conditions:
- Reached primary source (original paper, official docs, raw data)
- Dead end (paywall, 404, requires login)
- Circular reference (already read this source)
**Every ~10 findings during Move 3: Re-Strategize Checkpoint.** Same questions
as after Move 2. Update strategy if needed.
**Information-void handling:** After 5 searches on a sub-question with zero
relevant results (not zero *new* findings — zero *any* findings), document the
void and move on:
```
## SQ<N>: <sub-question> — VOID
- Searches attempted: <N>
- Tools tried: <comma-separated>
- Reason: no relevant results found
- Note: <why this might be — too new, too obscure, paywalled, etc.>
```
### Move 4: Disconfirmation Pass
Actively try to falsify each key claim:
- Search for "[claim] wrong", "[claim] criticism", "[claim] outdated"
- Check dates: is a 2024 source being presented as current?
- Look for contradicting evidence
- Append disconfirmation findings to the ledger
This is distinct from cross-referencing. Cross-reference confirms agreement.
Disconfirmation actively hunts for disagreement.
### Move 5: Condensation
Read the full ledger from disk. Read the phase gate file. If any box is
unchecked, do NOT condense — go back and complete that item.
Synthesize using this template:
```
# <Answer in 1-3 sentences>
## Key Findings
- <finding> — <source> (<credibility tier>, <date>)
- <finding> — <source> (<credibility tier>, <date>)
## Uncertainty
- <claim>: single-sourced from <source>
- <claim>: disputed — <source A> vs <source B>
- <claim>: unresolved — no sources found
## Tools Used
- <tool>: <what it contributed>
- <tool>: <what it contributed>
## Sources
1. <URL> — <description> (<tier>)
2. <URL> — <description> (<tier>)
## Tools That Would Have Helped
- <tool>: <what it would have enabled>
None: all sources were accessible with available tools.
```
## Guardrails
### Saturation Check (Mechanical)
Every 3-4 searches, run this command — do not self-assess:
```
grep -c "^### Finding #.*\[SQ<N>\]" /tmp/research-<sid>.md
```
Compare to the previous count. If zero new findings in the last 3 searches,
that sub-question is saturated. Move to the next. This is a measurable fact,
not a vibe. The `[SQ<N>]` tag makes this immune to header ordering — it counts
by tag, not by position.
### Phase Gate File
The phase gate is a file at `/tmp/research-<sid>-gate.md`. Write it and read it
back — it is not a mental checklist. Before moving to Move 5, read the gate file.
If any box is unchecked, do NOT condense. Go back and complete that item.
```
# Phase Gate
- [ ] All sub-questions have findings in the ledger
- [ ] Disconfirmation pass completed for key claims
- [ ] Source diversity: ≥5 unique domains across all findings
- [ ] Source diversity: no single domain >30% of findings
- [ ] Tool diversity: ≥2 different tool categories used (not just SearXNG)
- [ ] Recency spread: findings span appropriate date range for topic
- [ ] OR: saturation condition fired (no new findings in last 3 searches)
- [ ] OR: turn ceiling approaching (wrap up what you have)
```
Update this file as you progress.
Mechanical domain diversity check:
```
grep -oP 'Source: \K[^ ]+' /tmp/research-<sid>.md | grep -oP 'https?://[^/]+' | sort -u | wc -l
```
### Turn Ceiling
600 turns is a safety net, not a target. Hitting the ceiling is a failure to
condense, not the normal path. Ramp: if ≥450 turns used, begin condensation
within 60 turns. If ≥540 turns used, condense immediately with what you have.
If the ceiling is hit, deliver partial findings with a note on what's missing.
### SearXNG Failure Handling
If SearXNG returns errors or empty results for 3 consecutive searches (different
queries), wait 10 seconds and retry once. If still failing, note the instance
issue in the ledger and fall back to `web_search` (built-in) or
`duckduckgo-search` (Python library) for the remainder of that sub-question.
## Tool Policy
**One rule: local and free only.** Use any tool already configured on the
research profile that fits this rule. No internet-based paid services, no SaaS
APIs with billing, no metered endpoints. You determine what fits — you are not
given a list of allowed or disallowed tools.
**No self-provisioning.** Never install, pull, or spin up new tools at runtime.
### Tools That Would Have Helped
At the end of the response (after Sources), add a section if any limitations
were hit:
```
Tools that would have improved this research:
- <tool name>: <what it would have enabled>
None: all sources were accessible with available tools.
```
This is informational only. Do your best with what's available and note what
could have been better. Do not stop or block on tool gaps.