twg-agentic-search

Use together with the root twg skill. Exact command grammar comes from live twg help, especially twg help describe "rovo search" when filter or output options matter.

CLI launcher fallback

Run twg <command>. On shell command not found, use $HOME/.local/bin/twg (macOS/Linux) / $env:LOCALAPPDATA\Programs\twg\bin\twg.exe (PowerShell), then tell user to add that directory to PATH. Do not treat auth or command errors as PATH failures.

Workflow

1. Classify the request as fuzzy or cross-product internal research. Prefer this skill when the source is unclear, current company knowledge is needed, or the answer may span Confluence, Jira, Drive, Slack, Bitbucket, GitHub, or other Rovo-connected apps. When the user supplies an exact key, URL, title, or named canonical source, preserve that anchor: resolve and verify the returned identity before considering similarly named results. A similar title is not a substitute for the requested source.
2. Confirm or infer the Atlassian site. Ask only when no configured or explicit site is available and the ambiguity would change the search.
3. For unanchored work or knowledge, begin with Jira and Confluence built-ins. Expand to Drive, SharePoint, Bitbucket, GitHub, or another connector only when the request names that source or the current evidence leaves a material gap. If availability is unknown, run twg rovo list-apps -o json; reuse readiness/auth and do not list apps before every query, start setup/login, or fabricate an empty result when unavailable.
4. Start with one query combining the topic and requested artifact or decision. Add a focused canonical/authoritative or recent refinement when results mix scopes, lack primary sources, or miss the requested time signal. Continue only while it can change identity, scope, authority, or freshness. Do not fan out exact-title searches for every returned candidate. Follow canonical, moved, or superseding source relationships while retaining requested semantics.
5. Choose filters deliberately. Default to Confluence and Jira built-ins for official/internal knowledge. Broaden to Slack, Google Drive, SharePoint, Bitbucket, GitHub, or other connectors only when useful and available. Use app, type, recency, owner/contributor/assignee/reporter/status, title-only, label/space, and site filters when they narrow evidence without hiding likely answers.
6. Search with bounded output:

twg rovo search "<query>" --output json --output-summary auto --agent-fields @compact
twg rovo search "<query>" --output json --output-summary auto --agent-fields @evidence

Use @compact to shortlist candidates and @evidence when snippets, URLs, and provenance need more detail.

Evidence Rules

• Treat search snippets as candidates, not facts.
• Treat ranking, recency, activity volume, and similarity as discovery signals, not proof of authority, identity, validity, or currentness. Verify the selected source's key/URL/title and canonical or supersession relationship from hydrated evidence before replacing an explicit anchor.
• Hydrate a small, diverse primary-source set before final claims. For document discovery, cover distinct roles such as requirements, architecture/design, and current delivery rather than redundant pages. Use product-native commands such as twg confluence content get, twg jira workitem get, twg jira workitem query, twg bb pull-requests get <id>, twg bb repo get, or the relevant product command for the result URL/type. When a native key or URL is known, hydrate it directly; use another search or a document-body fetch only when the native result lacks a material field or narrative needed for the answer.
• For document or PRD discovery, select a small, diverse set across needed roles. Hydrate one per role; add another only for a material conflict, missing field, current-delivery question, or requested relationship. Stop when roles, current delivery, and conflicts are supported or unavailable.
• Prefer official spaces, owned project pages, current Jira issues, and recent decision records over personal drafts or stale chat mentions, unless the user explicitly asked for informal signal.
• Compare hydrated evidence for conflicts, recency, ownership, and authority.
• Use one authoritative source when it fully supports the claim. Treat the source-count limit as a default, not a quota: add a narrow follow-up only for a material unresolved conflict, missing field, or requested relationship.
• Do not rerun an @evidence search for a candidate that can be hydrated through its product-native URL or ID, or refetch one source under another projection merely to increase context. Use a different native projection when it supplies a missing material field or relationship, and explain the remaining gap. Call out ACL gaps, unavailable connectors, low recall, and unresolved contradictions instead of flattening them into a single claim.

Output

Lead with the answer or best-supported conclusion. Cite hydrated titles/URLs and include the source app, date or status when available, and why each source was trusted. Separate confirmed facts, likely interpretations, conflicts, and gaps.