Talking to Weave
the front door to Fabric's intelligence.

Weave knows the ticket, character, agent and screen you're looking at, so "what's the confidence here?" or "why did this hold?" resolve against your current context — you don't have to repeat yourself.

Identity and evidence-basis questions receive direct answers instead of unrelated dashboard summaries. Operational metrics are filtered and labeled by the selected agent and workspace; ordinary questions do not receive global run or health telemetry unless it is relevant to what you asked.

Ask “show me your learnings” for selected-agent outcomes, corrections and governed candidate states. Ask “in a mindmap” or “graphical view” for a native visual map of the evaluation → pilot → promotion → rollback lifecycle. These are live, scoped measurements—not model claims.

Settings → Weave shows context pressure and the exact active turns. Pin decisions that must remain verbatim, compact older unpinned turns into a bounded summary, archive finished context until it is explicitly restored, or permanently forget selected items. These lifecycle changes are durable and audited; archive restore fails closed if its stored digest no longer matches.

Useful cloud results become inert learning candidates. A candidate must pass deterministic server-side checks, accumulate a measured pilot of at least five samples with 80% success and zero regressions, and be explicitly promoted before Weave may reuse it as advisory context. Every transition is digest-bound and audited, and promoted learning can be rolled back. Learning never changes model weights, agents, validators, permissions or approval gates.

The Conversational Orchestrator resolves social dialogue, follow-ups, corrections, scope and requested format before action routing. An authoritative task ledger keeps each Main or branched conversation’s objective, subject, requested format, verified evidence references, decisions, unresolved requirements and next safe step—never credentials, raw prompts or full answers. Branches inherit typed context explicitly and can merge it back without losing their audit trail. Free-form turns receive the smallest relevant Weave specialists plus up to two enabled domain playbooks matched from the repository. Vision, Data, Ticket Investigation and Clarification specialists now join the existing knowledge, code, operations, planning and visual skills. The enforcing Answer Critic repairs or replaces empty, unscoped, unresolved, wrong-format or clearly off-topic model answers before display.

Each answer can expand Sources & skills to show the exact skill set and bounded RAG, Graphify or attachment evidence used. Weave also records privacy-bounded skill effectiveness observations—selection, answer-quality issues and repairs—so skill routing can be evaluated from outcomes instead of intuition.

Weave first separates the user's goal from its subject, then traverses a declarative goal → subject → capability → intent graph. The capability directory resolves specific questions before generic commands—for example, “explain Fabric” selects product architecture while “explain the failed run” selects operational evidence. Every selected route includes a bounded, inspectable path instead of depending on ordered keyword exceptions.

Attachments are owner-scoped, short-lived and size/count limited. Their paths are never disclosed to the browser, and a file is never treated as executable input.

Say “load a local file” (common misspellings are accepted) to open the native file picker immediately, or use Attach, drag and drop, or paste. The picker is opened directly from your gesture so browser security does not block it.

If Weave can't answer from real signals, it says so and points you at the screen or source — it doesn't invent an answer. Free-form reasoning is capability-routed when Fabric AI is on — it runs on a local model or, for the reason/code lanes, a cloud specialist under budget, degrading to local automatically; with Fabric AI off, Weave stays on the deterministic knowledge/health/queue answers.

When a code question names a file or symbol, Weave can use the governed Code Intelligence seam with Graphify as its first provider. Only bounded, revision-bound findings enter the answer as derived advisory evidence; raw graph artifacts never enter the prompt, and structural evidence never overrides compilation, tests, policy or approval.

Enabled repository skills can be composed into a Weave answer when their name and description strongly match the request. This extends Weave with team-specific procedures without granting those playbooks execution authority; action and approval boundaries remain unchanged.

In Settings → Knowledge · RAG, choose Auto to rebuild the local Graphify AST index when Git HEAD changes, or Manual to keep updates operator-triggered. The status shows the indexed and current revisions. Graph refresh is local, bounded to one build at a time, and has no LLM or API cost.