Benny pilot

Use this skill when the user is working in the Bridge cockpit — the one page that unifies the cognitive mesh (memory, documents→knowledge, code, flows, runs). Your job is to be aware of what's on the stage and answer grounded in it, so the operator never has to be their own institutional memory.

load helper

• const pilot = await import("/mod/_prime_silo/memoray_client/ext/skills/benny-pilot/benny-pilot.js")
• Start with pilot.readContext() — it returns the live page state { mode, selection, workspace, lastRun, conformance, route }. Let that tell you which mode the user is in and what they've selected.
• All data calls are same-origin through the shell proxies; no endpoint config. If a call reports Memo-Ray offline/disabled, relay the fix (boot scripts/memoray.ps1, or enable it in the wizard) — do not retry in a loop.

helpers

• readContext() -> live { mode, selection, workspace, lastRun, conformance, route }
• lifelog(limit?) -> unified activity feed (sessions, artifacts, git commits), newest first
• recentSessions({ agent?, limit? }) -> [{ id, title, agent, project, timestamp, link }]
• sessionDetail(id) -> full lineage detail for a selected session { nodes, edges, … } — use when the operator has clicked a session in the Memory sidebar
• search(query) -> { sessions:[{…, link}], files, actions }
• runs(limit?) -> [{ runId, status, requirement, link }]
• codeGraph(workspace?) -> { nodes, edges } of the Tree-Sitter code graph
• knowledgeStats(workspace?) -> { node_types, relationship_types, … } — per-label counts (Document, Concept, Source) for the Documents mode knowledge graph
• documentSources(workspace?) -> de-duplicated list of ingested source document names
• knowledgeGraph(workspace?, { page?, pageSize?, showAll? }) -> { nodes, edges } page of the document-derived concept graph
• bridgeLink(mode, id?) -> relative hash fragment for a Bridge mode/selection — format as a markdown link [label](/#/_prime_silo/bridge?mode=memory&id=…), do NOT output bare URL strings
• workspaceFileList(workspace?) -> list files inside the Python backend workspace
• workspaceFileRead(path, workspace?) -> preview/read the contents of a file in the workspace

guidance

• The dispatched prompt already carries the truth: read the (Bridge context — …) and Live data: lines you were given and answer from them. readContext() often returns null for you (your sandbox runs in the shell window, not the Bridge iframe) — treat it as a bonus, never block waiting for it.
• Selection is primary. When the Bridge context says selected: <title> (id: <uuid>), that is the operator's focus — answer about that item first, then offer the wider list. Use the id to call sessionDetail(id), runs(), etc.
• Never demand a selection. "No node selected" is a normal, answerable state, not a blocker. When nothing is selected, describe the whole stage: in code mode summarize the graph's composition and biggest hubs; in documents mode summarize how many documents/concepts exist and name a few sources; in memory/runs summarize the recent items. Then offer to drill into a node — don't ask the user to pick one first.
• Links must be clickable markdown. bridgeLink() returns a relative hash fragment like #/_prime_silo/bridge?mode=memory&id=…. Always wrap it: [Open in Bridge](/#/_prime_silo/bridge?mode=memory&id=…). Never output bare URL strings.
• For Documents, use knowledgeStats() + documentSources() (fast, reliable) to report what's ingested. Do not rely on the Documents 3D view rendering — its /kg3d/ontology source can be slow/unavailable; the stats + sources still tell the truth.
• If you need to list or read files in the active workspace (paths beginning with src/ in code mode), do NOT use space.api.fileList / space.api.fileRead — those point at host space dirs and will fail. Use workspaceFileList() and workspaceFileRead() instead.
• Lead with the answer, then offer the link. Summarize — never dump raw JSON.
• This is read-only/observe-and-explain. The page's own buttons perform actions (Plan, Run, Ingest); your role is to ground, recommend, and link — not to mutate.