Research
SkillAI & modelsUse when current external facts, source-backed comparisons, investigations, or recent community evidence are needed.
Available today. Use it from your connected AI after setup.
No other account needed.
Connect ahel once, and every AI you use reads what you have installed.
Then ask your AI: use the Research skill
What this skill tells your AI
The instructions your AI receives, as published by lev-os/agents in skills/research/SKILL.md and read by ahel’s review.
Use lev find or native retrieval for local code, documents, tasks and sessions.
For external research, select one method, print its route, and follow its stages.
Only this root is the research entry point; child methods load conditionally.
Select the method
research --sherlock and research --last30 are skill invocation syntax.
They are not lev research CLI flags. Inspect live help before runtime calls.
| Request | Load | Accepted output |
|---|---|---|
Explicit --sherlock, dossier, OSINT, multi-round investigation or resume | Sherlock | Source-bearing investigation state, deductions, gaps and local dashboard |
Explicit --last30, last 30 days, recent community testimony or engagement | Last30 | Dated platform evidence, coverage, source URLs and provider gaps |
| Ordinary external question | Classifier, then its selected node | Source-backed results from the fixed routine |
| Aviation training resources/corpus | Pilot training | Sourced learning resources; no corpus acceptance claim |
Explicit selectors take precedence. If both child selectors are supplied, resolve which deliverable is required before spending provider calls. A date mentioned as background alone does not select last30. A code investigation alone remains local retrieval. Ordinary "deep research" uses the deep routine.
Execute the ordinary routine
- Read the selected routine under
nodes/; domain routines load domain contracts. These domain contracts are not native CLI strategies. Follow fixed stages and fallback order. - For deep/max/academic, load prompt and evidence contracts and prepare objective, audience, scope, context, subquestions, evidence and output.
- Run
lev research --help. Implementation isplugins/research; compatibility routing throughlev timetraveldoes not change the northbound command. Strategies/providers are owned byplugins/research/src/strategy.ts. - Inspect planned prompts with
lev research "<query>" -s <strategy> --prompt-plan-onlywhen appropriate. Printresearch-route: <routine> | lev research -s <strategy>. - Execute
lev research "<query>" -s <strategy>. Default deep prompt planning and AutoWiki raw intake remain in place. Use--no-wikifyfor a no-intake run. Raw outputs are evidence, not promoted guidance. - If the CLI errors, lacks adapters or yields no URLs, immediately use available internet-search tools following the fallback contract. Print the failure and fallback; do not substitute an unsourced answer.
- Return source URLs, findings, uncertainty and remaining gaps. A completed process, synthesis without URLs or fixture success does not prove live research.
The existing --flow=deep-research route is a prompt-compilation/recursive-search
POC. Do not describe it as execution of arbitrary graph nodes. Sherlock's method
requires its own verified graph route and reports node outcomes.
Acceptance
Preserve source provenance and observation/inference distinctions. Show requested versus observed provider coverage and partial failures. Budget or scope expansion requires a human decision. Save only artifacts authorized by the selected method; local dashboard creation does not authorize external publication. Evaluation requirements and current evidence live in qualification.
Load tool catalog only when inspecting capability availability. Catalog discovery does not authorize changing provider order.
Signals
- GitHub stars
- 22
- Forks
- 2
- Last commit
- Sep 2026
- Hacker News mentions
- 20
Advanced
- Catalog kind
- skill
- Gateway key
research-lev-os- Source
- github.com/lev-os/agents