Research

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 read its selected routine file (step 1)	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

1. When the classifier selects a routine, read nodes/<routine>.md (domain routines: domain contracts) before applying or adapting its method, including when backend or network tools are unavailable. These domain contracts are not native CLI strategies. Follow fixed stages and fallback order.
2. For deep/max/academic, load prompt and evidence contracts and prepare objective, audience, scope, context, subquestions, evidence and output.
3. Run lev research --help. Implementation is plugins/research; compatibility routing through lev timetravel does not change the northbound command. Strategies/providers are owned by plugins/research/src/strategy.ts.
4. Inspect planned prompts with lev research "<query>" -s <strategy> --prompt-plan-only when appropriate. Print research-route: <routine> | lev research -s <strategy>.
5. Execute lev research "<query>" -s <strategy>. Default deep prompt planning and AutoWiki raw intake remain in place. Use --no-wikify for a no-intake run. Raw outputs are evidence, not promoted guidance.
6. 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.
7. Return source URLs, findings, uncertainty and remaining gaps. A completed process, synthesis without URLs or fixture success does not prove live research.

Quick and deep runs alike keep a short source ledger in the returned answer. Mark each source discovered (URL or search snippet only), inspected (content actually fetched and read) or failed/unavailable, with its URL or locator, access time, publication time (unknown when the source shows none) and how it was obtained. Access time records retrieval only. Order sources by publication, observation or validity time only where each source states that time; otherwise weigh authority and applicability, not recency. Tie each factual claim to the inspected sources that support it. A URL or fetched search snippet is discovery, not full-page verification: label a claim resting only on discovered sources as unverified. A publisher or authority taken only from a URL, domain or other locator is inference: label it so, and call a publisher identified only when inspected content shows it. Such inference may still inform authority and applicability. Record tool failures and the step 6 fallbacks as ledger gaps.

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.