Changelog

​

Breaking

• Signals are 7 types now, all candidates. The old signal catalog is replaced. Every signal — verifier · causal · human · behavioral · tool_use · guardrail · canary — lands as a candidate and must be human-confirmed to become a fix target. Nothing auto-confirms. See Signals · sources.
• Fault track is retired. Runtime errors, cost stops, latency outliers, SDK silence, and PII / banned-word matches are no longer signals — they surface on trace errors, simulation cost views, /signals/sdk-health, or the audit log. Cases table gains data.fragile for regression tracking.
• signals.report is removed from the SDK. Send app outcomes as a dislike on nora.feedback with a comment — the confirm queue is the single funnel now.
• Self-serve billing is gone. No plan picker, no Polar portal, no subscribe / cancel / change-plan verbs. Nora is sold as an enterprise license granted per workspace; existing workspaces are grandfathered. The Plans & pricing page has been removed from Settings. See Usage.
• --model on flows.run / run_flow / nora flows run is a JSON object. Per-agent map {"agent-id-or-name":"model-id"}, not a single string. Only listed agents are affected; others keep their configured model.

​

New

• Clusters only form on confirmation. Confirming a signal spawns the cluster + case in the background (drawer flips immediately); the signal list’s cluster column live-updates without a page refresh, with a short 군집 배정 중 badge while it lands. Empty clusters never accumulate.
• Guardrail block · confirm splits two ways. Guardrail (F) signals ask for 정상 차단 (block was correct → case) or 오발 (block was over-eager → excluded with reason guardrail_misfire and a link to the rule).
• Case input is three columns. Text / voice / image — voice and image use transcript fallback so the answer key stays comparable. Any trace with a live signal (candidate or confirmed) is protected from deletion with a confirmation dialog.
• Flow-wide turn and token budgets. New Max turns and Max tokens on Flow settings apply on top of per-agent caps; the run stops on whichever hits first. See Flow settings · Limits.
• nora flows run — CLI now documents the flow twin of pipelines invoke. PAT-authenticated, per-agent --model map, --test to run the DRAFT.

​

Improved

• Signal drawer stays live. Open a row and the drawer’s status / cluster columns update themselves — no reload needed while you work through the queue.
• Case comment auto-fill. If a trace already carried a comment or before → after correction, the confirm form pre-fills instead of asking you to retype.
• failure.code accuracy. OpenAI-compat mid-stream errors are now detected and reported as provider failures, not swallowed as empty output; LLM-provider errors no longer misclassify as run.tool_failed.

​

Fixed

• run_flow model override silently dropped — fixed. CLI flows run --model now parses as JSON object matching the MCP schema.
• Modal / drawer <Select> z-index — every dropdown that rendered behind its parent modal is now on top.
• Retrieval on partial vector-store failures — partial results are preserved instead of the whole page falling through.
• Wiki note scope on edit-after-restore — a restored note now keeps its original scope when you edit it.
• PromptModal now syncs the input when initialValue changes; the “Add source” picker’s stray Clear button is gone; folder-move toast no longer double-fires.

​

New

• SDK on npm — @conscience-technology/nora-sdk is now published. npm install @conscience-technology/nora-sdk works from any Node backend; the pre-publish workaround is retired.
• Claude Code connection — connect your Claude Code account from Profile → Connect Claude Code. Paste the one-time seed command, credentials land in your OS keychain, and Nora agents can run under your Claude Code subscription. See Agent block · Mode.
• Supervisor agent — a new agent role that spawns and coordinates sub-workers per task. Sub-workers inherit the parent’s tools and retrieval scope, and each sub-task can carry its own as_of retrieval window. Configure spawn model and budget in the agent block’s Model tab.
• Nora desktop app — run the platform locally as a native app (macOS · Windows · Linux). Tauri-based WebView with tray menu (Show / Stop / Restart / Quit), Settings panel (license · ports · Reset), and full uninstall.
• On-prem deployment — three shapes:
  • Docker Compose — five-service one-command deployment.
  • Helm chart for Kubernetes (nora-onprem).
  • Air-gap tarball + offline installer for isolated networks.
• Causal graph as retrieval structure — causal graphs are traversed automatically when the query looks causal, rather than being a separate scoring dial. Agents bind graphs at attachment time and verify answers against them (see Causal graph · verification).
• Fault-track simulation — a new simulation kind that re-runs detected faults through the current pipeline. Adopted improvements now render as a commit-style diff per system layer.
• Case-level feedback in simulations — click a failed case in a round to flip its verdict (right / wrong) and leave a note; feedback saves back on the cluster and is available to the next round.
• Cluster governance — folders can act as a cluster surface. Rename a cluster with an audit trail; delete is refused (renames are the sanctioned path); imports warn on rows without a score. See Signals · clusters.
• Signals as rate — signals now render as an occurrence rate (발생률) rather than a raw count, with severity per source shown per row.
• Trace approvals inline — pending-approval traces get a badge; approve or reject inline from the trace drawer without a page jump.
• Flow token budgets in Settings — cap total tokens per run at the flow level, in addition to per-agent caps.
• Bare variable syntax — {{name}} is now the canonical form for trigger and tool variables; {{vars.<name>}} still resolves as a legacy alias. Variables must be declared on the flow to be substituted — undeclared names on the request body silently drop with a trace warning instead of 400-ing.
• Canvas shortcuts — F fits every block to the viewport, Cmd+D duplicates the selection without touching your clipboard, Cmd+S is a force-save pulse that no-ops when nothing changed since the last press. See Workflow · Canvas.
• OpenAI-compat provider — new adapter in the LLM bridge speaks OpenAI-shape to any self-hosted endpoint (vLLM · TGI · llama.cpp).
• GPT-6 in the model catalog — the gpt-6 line is now selectable in the agent builder with real provider pricing.

​

Improved

• Knowledge — Shift-select ranges in the document list; folder right-click menu unified with the sidebar; folder moves preserve nested structure; the move-destination picker now includes ancestor-prefix folders. Foundry runs kicked off from a document detail now start in the background with a toast, so the page doesn’t block.
• Retrieval — vector search fans out across every wired backend (fixes “just indexed but doesn’t show up”); result hits carry document_name so agents can cite by filename; scope and preset edits flush before search so the change lands immediately.
• Memory — Wiki note rename fixed; excluded notes no longer resurrect on the next write; MCP responses expose space config and note timestamps; accessScopeKey is restricted to the fixed vocabulary user_id · flow_id · thread_id.
• Simulations — search debounce (250ms); round scores refresh immediately on completion; correction cases connect back to the cluster feedback loop; feedback correction merged into the comment section (no separate widget).
• Trace — filter modal dropdowns no longer render behind the modal; TraceDrawer closes on outside click; K / J shortcuts work while a Korean IME is active; table cells get proper padding.
• Workflow canvas — Space-pan no longer opens the block panel when clicking through; workspace-switch flicker gone.
• Collaboration — a user’s own second tab no longer locks them read-only on the first tab.
• Layout — minimum-width restriction removed for narrow screens; sidebar bottom padding trimmed.

​

Fixed

• Foundry — Chunk block click no longer flashes the entire screen white.
• Knowledge document detail — section-toggle flash removed; long extracted text renders fold-first so the initial paint stays fast.
• Anthropic first-turn — first-turn LLM errors (e.g. balance exceeded) were being swallowed as empty output; now surfaced with a clear message.
• MCP — delete_document now returns not-found instead of generic 500 on missing id; get_active_workspace no longer returns an empty workspace after session reset; list_documents has a size ceiling with a field whitelist; new create_flow tool, and create_agent fails clearly when the target slug doesn’t exist.

​

New

• @conscience-technology/nora-sdk — TypeScript client for calling Nora flows, capturing feedback, and embedding the improvement queue in your own product. Full SDK reference with install, auth, client options, and per-surface method pages (Execution, Feedback, Transparency, Improvements & versions, Simulations, Resources, Errors).
• Developers tab — the former CLI tab is now Developers, housing both the CLI reference and the SDK reference side by side. Existing /cli/* URLs are unchanged.

​

Improved

• Retrieval, Memory, and Foundry docs got a clarity pass — new screenshots, refined descriptions, and consistent terminology across English and Korean.

​

New CLI command groups

• nora knowledge — folder + access-rule management (folders list / add / delete, access list / set).
• nora clause-lineage — review Foundry’s proposed clause revision history: browse pending queue, approve / reject / bulk-approve.

​

Expanded CLI surfaces

• nora documents — now covers the full library CRUD: upload, upload-large, upload-folder, create-external, update (with regulatory dates for time-aware retrieval), chunks, chunk-text, extracted-text, extract-structure, delete.
• nora pipelines — added invoke (programmatic API trigger call), rotate-secret (inbound-token rotation), palette (list palette IDs), and block (config schema for one palette id).
• nora retrieval — new type-profile sub-group (show / set / backfill / custom) for type-aware retrieval, plus new search flags (--routing-mode, --target-type, --target-entity, --expand-window, --reserved-per-type).
• nora causal — added instance CRUD (graphs), pipeline bindings, evidence trails, and atomic edge CRUD alongside the folder-based extraction workflow.
• nora agents tools — per-attachment budget and failure-handling policy (list / set-policy / clear-policy).
• Block locks — lock / unlock added consistently across agents, tools, actions, and triggers to freeze blocks against MCP / CLI edits.
• nora agents update — new --set-output-rules and --set-retry flags.
• nora tools update — new --set-retry-safe and --set-dedupe-key-template flags.

​

Foundry

• Six-node model made explicit across all pages: Trigger → Source → Clean → Normalize → Enrich → Output.
• Payload vs Signal triggers — Webhook and API triggers hand their body in as input and don’t need a Source or Output node; Schedule and Watch still do. Documented in Triggers.
• DB picker + SQL editor shell unified across Source · Database, Enrich · DB Lookup, and Output · DB Upsert.
• Chunk typing profile — inline create / edit right from the Chunk / Clause split node inspector.
• Table fix — deterministic stitch + split always runs; optional LLM refine model for the ambiguous residual.
• Run logs drawer — right pane split into Overview (default) vs Stage detail (after clicking a stage), with Trigger metadata / Variables override / Stage summary / Folder documents on the Overview and a Config · Output · New output · Artifacts · Error scroll layout on Stage detail. Readable / Raw JSON toggle persists across sessions.

​

Memory

• IA restructured around the four paradigms — Wiki (human-authored notes), Index (retrieval-shaped nodes + optional graph), Session (short-term conversation buffer + rolling summary), Evolution (diagnostic memory).
• Attachment page for connecting memory spaces into agents.

​

Knowledge

• New Foundry output page — how Refine documents proposals surface for review.

​

MCP for coding agents

• MCP guide reworked around Claude Code as the primary host (also documented for Cursor and Claude Desktop). See MCP: coding agents.

​

Operate

• New Clusters section under Operate — how failure clusters group signals and drive fixes.
• Signals got a Revalidation page and a clearer “from signal to fix” walkthrough.

​

Workflow

• Agent block Mode clarified: API (hosted provider) vs Internal (workspace-hosted endpoint), with pointers to Internal resources.
• Trigger block: body mapping for webhooks, token generation, and schedule fields (preset vs cron) documented in full.
• Action vs Tool distinction sharpened; approval flow for high-stakes tools written out.

​

Settings

• New Internal resources page (workspace-hosted inference endpoints, databases registered for CLI / MCP use).
• Providers / model routing refresh to reflect the current UI.

​

Korean documentation

• Full Korean translation across CLI (Setup, Flows, Blocks, Data & knowledge, Reliability, Workspace, Reference) and Settings (Data, Notifications, Billing).

​

CLI reference

• Initial CLI reference pages published for every command group: nora auth, nora flows, nora agents, nora tools, nora actions, nora triggers, nora edges, nora vars, nora schema, nora pipelines, nora documents, nora memory, nora retrieval, nora causal, nora traces, nora feedback, nora approvals, nora providers, nora models, nora billing, nora connectors, nora foundry.

​

Docs site