Known Problem Hint Research

Use research to turn uncertainty into a concrete next test. Start as soon as an external fact or unfamiliar technique could unblock the task; do not wait for repeated failures, a named attack, a complete fingerprint, or a fixed number of local hypotheses.

Identify the missing clue

Use the available artifact, error, trace, equation, API, version or observed behavior. Inspect enough local context to avoid searching a misread symptom. Separate observations from guesses and note relevant attempts already made; unknown fields can stay unknown.

State the question that would change the next action: what the behavior means, which precondition is missing, whether a known issue matches, or which technique can test the clue. If classification is the blocker, research to identify the problem family first. Define a discriminating check when one becomes possible.

Use systematic-debugging for the local root-cause process. Use deep-research-generic for a broader investigation or comparison. A routine edit with a known solution does not need a research detour.

Search where the answer is likely to live

1. Check relevant supplied/local docs and available connected sources. For version-specific behavior, search upstream documentation, code, issues, commits and release notes. For a missing construction or bound, seek the original paper, specification or implementation. Technical discussions and writeups can supply leads; follow their evidence.
2. Build focused queries from distinctive observations. Start with the exact error or symbol plus product/version when known, then try the underlying behavior, synonyms or a suspected mechanism. Do not require the user to know the technical name before research can begin.
3. Choose an available search provider or site-native search. Tavily is an option when exposed; inspect its current schema. Use supported domain/date filters only when helpful. If results are empty, remove accidental constraints or switch source families instead of repeating the query. Strip secrets and irrelevant private identifiers before sending a query externally.
4. Read the promising source through a connector, fetch or extraction tool; use browser automation when relevant content is missing or interactive. Check full context, affected version and preconditions. Generated summaries and search snippets alone do not establish the hint.
5. Follow citations, issue links or fixing commits while they resolve the specific uncertainty. Stop collecting sources once there is a sufficiently grounded next experiment. Avoid a site-wide crawl or literature survey for one blocker.