@superatomai/sdk-node 0.0.39-mds → 0.0.40-dsp

package/dist/index.d.mts

@@ -1,15 +1,929 @@
 import { z } from 'zod';
 import Anthropic from '@anthropic-ai/sdk';
 
+/**
+ * StreamBuffer - Buffered streaming utility for smoother text delivery
+ * Batches small chunks together and flushes at regular intervals
+ */
+type StreamCallback = (chunk: string) => void;
+/**
+ * StreamBuffer class for managing buffered streaming output
+ * Provides smooth text delivery by batching small chunks
+ */
+declare class StreamBuffer {
+    private buffer;
+    private flushTimer;
+    private callback;
+    private fullText;
+    constructor(callback?: StreamCallback);
+    /**
+     * Check if the buffer has a callback configured
+     */
+    hasCallback(): boolean;
+    /**
+     * Get all text that has been written (including already flushed)
+     */
+    getFullText(): string;
+    /**
+     * Write a chunk to the buffer
+     * Large chunks or chunks with newlines are flushed immediately
+     * Small chunks are batched and flushed after a short interval
+     *
+     * @param chunk - Text chunk to write
+     */
+    write(chunk: string): void;
+    /**
+     * Flush the buffer immediately
+     * Call this before tool execution or other operations that need clean output
+     */
+    flush(): void;
+    /**
+     * Internal flush implementation
+     */
+    private flushNow;
+    /**
+     * Clean up resources
+     * Call this when done with the buffer
+     */
+    dispose(): void;
+}
+
+/**
+ * ToolExecutorService - Handles execution of SQL queries and external tools
+ * Extracted from BaseLLM.generateTextResponse for better separation of concerns
+ */
+
+/**
+ * External tool definition
+ */
+interface ExternalTool {
+    id: string;
+    name: string;
+    description?: string;
+    /** Tool type: "source" = routed through SourceAgent, "direct" = called directly by MainAgent */
+    toolType?: 'source' | 'direct';
+    /** Full untruncated schema for source agent (all columns visible) */
+    fullSchema?: string;
+    /** Schema size tier: small (≤50 tables), medium (51-200), large (201-500), very_large (500+) */
+    schemaTier?: string;
+    /** Schema search function for very_large tier — keyword search over entities */
+    schemaSearchFn?: (keywords: string[]) => string;
+    /** `signal` fires when the user hits Stop mid-query — passed through to the backend's SourceTool.fn. */
+    fn: (input: any, signal?: AbortSignal) => Promise<any>;
+    limit?: number;
+    outputSchema?: any;
+    executionType?: 'immediate' | 'deferred';
+    userProvidedData?: any;
+    params?: Record<string, any>;
+}
+/**
+ * Executed tool tracking info
+ */
+interface ExecutedToolInfo {
+    id: string;
+    name: string;
+    params: any;
+    result: {
+        _totalRecords: number;
+        _recordsShown: number;
+        _metadata?: any;
+        _sampleData: any[];
+        /** Bounded summary over the FULL fetched result (complete structure). */
+        _summary?: any;
+        /** Up to MAIN_AGENT_COMPLETE_ROWS rows — the complete result when small. */
+        _mainAgentRows?: any[];
+    };
+    outputSchema?: any;
+    sourceSchema?: string;
+    sourceType?: string;
+}
+
+/**
+ * Script Flow Types
+ *
+ * Defines interfaces for the script-based query architecture:
+ * - ScriptRecipe: metadata for matching, validation, and quality tracking
+ * - ScriptResult: output from executing a script
+ * - ScriptMatch: result from the LLM-based script matcher
+ */
+/**
+ * Recipe metadata stored alongside each script.
+ * Used for matching, validation, and quality tracking.
+ */
+interface ScriptRecipe {
+    /** Unique script identifier */
+    id: string;
+    /** Version number (incremented on regeneration) */
+    version: number;
+    /** Human-readable name (e.g., "Revenue by Dimension") */
+    name: string;
+    /** Natural language description of what this script does */
+    intentDescription: string;
+    /** Keyword tags for quick filtering */
+    tags: string[];
+    /** Source tool IDs this script queries (e.g., ["mssql-abc123_query"]) */
+    sourceIds: string[];
+    /** Table names used (for future schema drift detection) */
+    tables: string[];
+    /** Parameter definitions — what can vary */
+    parameters: ScriptParameter[];
+    /** The script function body as a string. Loaded from disk (scripts-store/<fileBase>.ts). */
+    scriptBody: string;
+    /**
+     * On-disk filename stem for the body: scripts-store/<fileBase>.ts.
+     * Editable in the IDE. Decided at authoring time (slug of `name`, with a
+     * short id suffix on collision) and stable across promotion.
+     */
+    fileBase?: string;
+    /** sha256 of the on-disk body — lets the runtime detect manual edits. */
+    bodyHash?: string;
+    /** Project scope (single-VM deployments may leave this undefined). */
+    projectId?: string;
+    /** Times this script was used successfully */
+    successCount: number;
+    /** Times this script failed */
+    failureCount: number;
+    /** ISO timestamp of last usage */
+    lastUsed: string;
+    /** Original user question that created this script */
+    createdFrom: string;
+    /** ISO timestamp */
+    createdAt: string;
+    /** ISO timestamp */
+    updatedAt: string;
+    /**
+     * `recipe.id` of the parent this script was forked from.
+     * Undefined for root scripts (those written from scratch by MainAgent).
+     * See backend/docs/SCRIPT-FLOW-FORK-ADAPT.md.
+     */
+    parentId?: string;
+    /** 0 for root scripts; `parent.forkDepth + 1` for forks. Capped at 3. */
+    forkDepth?: number;
+    /**
+     * Brief description of what this fork changed vs its parent
+     * (sourced from the matcher's `modificationHint`).
+     */
+    forkReason?: string;
+    /**
+     * Validated component specs captured at authoring time. On a tier-high
+     * replay these are rebound to fresh queryIds deterministically — no
+     * component-generation LLM call, and the rendered columns can't drift from
+     * what was validated when the script was authored. Absent on recipes
+     * authored before this landed; those fall back to LLM component generation.
+     * See backend/docs/SCRIPT-COMPONENT-CONSISTENCY.md.
+     */
+    components?: ScriptComponentSpec[];
+    /**
+     * What the user explicitly asked the output to LOOK like, set by a display
+     * edit ("show that as a bar chart", "make the bars horizontal").
+     *
+     * Deliberately SEPARATE from `components`. Those are validated bindings that
+     * must be cleared on a data edit — after "break it down by brand" the axis
+     * keys are wrong, and after "use the mode" the stored title still says
+     * "Average…". A rendering CHOICE, by contrast, is still valid afterwards.
+     * Keeping them in one field meant every data edit silently discarded the
+     * user's chart type.
+     *
+     * Fed to the component generator as a hint whenever specs are (re)generated,
+     * so the chosen rendering comes back even after the SQL changes.
+     */
+    displayPreference?: ScriptDisplayPreference;
+    /**
+     * Lifecycle stage of this recipe on disk.
+     * - 'draft':    written by MainAgent's write_script during a turn; filtered out
+     *               of FTS results (status='verified' only) so the matcher never picks it.
+     *               Filename is suffixed with `turnId` to keep concurrent turns
+     *               from clobbering each other's drafts.
+     * - 'verified': promoted after `execute_script` succeeded; the matcher sees it.
+     *               Filename drops the turn suffix unless a verified file with
+     *               the same slug already exists (collision case keeps the suffix).
+     *
+     * Recipes loaded from disk without this field default to 'verified' so
+     * existing scripts keep working unchanged.
+     */
+    status?: 'draft' | 'verified';
+    /**
+     * Per-turn unique suffix used for draft filenames (e.g. `1714745623-x9k2`).
+     * Set when the draft is saved; carried until the recipe is promoted.
+     */
+    turnId?: string;
+    /**
+     * Last execution error captured by `recordDraftError` while the recipe was
+     * still a draft. Lets users open the draft .json file and see why it failed
+     * without grepping logs. Cleared on promotion to 'verified'.
+     */
+    lastError?: {
+        phase: 'compile' | 'runtime' | 'timeout' | 'ipc';
+        message: string;
+        at: string;
+        attempt: number;
+    };
+    /** userId who committed the most recent edit (audit trail — recipes are project-scoped). */
+    editedBy?: string;
+    /** The instruction that produced the current version. */
+    editedFrom?: string;
+    /**
+     * Prior versions, newest last. The body itself is archived on disk as
+     * `<fileBase>.v<version>.ts`; this records what changed and who did it.
+     */
+    history?: ScriptVersionRecord[];
+}
+/**
+ * A durable, user-stated rendering choice for a recipe. Survives data edits.
+ */
+interface ScriptDisplayPreference {
+    /** Component types the user's chosen rendering resolved to, e.g. ["DynamicBarChart"]. */
+    componentTypes: string[];
+    /** The instruction itself — carries nuance the types can't ("horizontal", "sorted descending"). */
+    instruction: string;
+    /** ISO timestamp of the display edit that set this. */
+    at: string;
+}
+/** One superseded version of a recipe body (see ScriptStore.commitEdit). */
+interface ScriptVersionRecord {
+    /** Version number this record superseded (i.e. the OLD version). */
+    version: number;
+    /** ISO timestamp of the edit that superseded it. */
+    at: string;
+    /** userId who made the edit, when known. */
+    by?: string;
+    /** The edit instruction that caused the supersede. */
+    instruction?: string;
+    /** One-line summary of what changed, for the version picker. */
+    changeSummary?: string;
+    /** sha256 of the superseded body — pairs with `<fileBase>.v<version>.ts`. */
+    bodyHash?: string;
+}
+interface ScriptParameter {
+    /** Parameter name (used in script body as params.name) */
+    name: string;
+    /** Parameter type */
+    type: 'string' | 'number' | 'date' | 'date_range' | 'enum' | 'boolean';
+    /** Whether this parameter is required */
+    required: boolean;
+    /** Default value if not provided */
+    default?: any;
+    /** For enum type — maps user-facing values to internal values */
+    enumValues?: Record<string, string>;
+    /** Human-readable description (used in the matcher LLM prompt) */
+    description: string;
+}
+/**
+ * A reusable component binding captured when a script is authored. Stored on
+ * the recipe so tier-high replays rebuild components deterministically (rebind
+ * to fresh queryIds) instead of re-running the component-picker LLM.
+ */
+interface ScriptComponentSpec {
+    /** Registered component name (e.g. "DynamicBarChart") — matched against the available component library. */
+    componentType: string;
+    /** `executedQuery.sourceId` to bind to (e.g. a tool id or 'computed:_final'), 'federation' for a cross-source component, or 'markdown' for a content-only narrative block (no data source). */
+    sourceRef: string;
+    /** Present only when sourceRef === 'federation' — the DuckDB SQL to re-execute on replay. */
+    federationSql?: string;
+    /** Present only when sourceRef === 'markdown' — the narrative text to render on replay (markdown has no data source, so its content must be persisted). */
+    content?: string;
+    title?: string;
+    description?: string;
+    /** Validated axis/value keys + aggregation — all referencing real columns of the bound source. */
+    config: Record<string, any>;
+}
+/**
+ * Result from executing a script via ScriptRunner.
+ */
+interface ScriptResult {
+    /** Whether the script executed successfully */
+    success: boolean;
+    /** Combined data from all queries */
+    data: any[];
+    /** Individual query results tracked during execution */
+    executedQueries: ScriptQueryResult[];
+    /** Error message if failed */
+    error?: string;
+    /**
+     * Where in the lifecycle the error occurred. Lets MainAgent's fix-loop
+     * decide between "rewrite the whole draft" (compile) and "patch the
+     * specific line" (runtime).
+     */
+    errorPhase?: 'compile' | 'runtime' | 'timeout' | 'ipc';
+    /** Total execution time in milliseconds */
+    executionTimeMs: number;
+}
+/**
+ * A single query executed during script runtime.
+ * Tracked by ScriptContext for component generation and debugging.
+ */
+interface ScriptQueryResult {
+    /** Source tool ID */
+    sourceId: string;
+    /** Human-readable source name */
+    sourceName: string;
+    /** The SQL that was executed */
+    sql: string;
+    /** Result data rows */
+    data: any[];
+    /** Number of rows returned */
+    count: number;
+    /** Total rows that matched before limit (if available) */
+    totalCount?: number;
+    /** Query execution time in milliseconds */
+    executionTimeMs: number;
+    /**
+     * True for rows that did NOT come from a real SQL execution — either a
+     * ctx.emit() dataset or the synthesized "computed:_final" entry that
+     * carries the script's post-JS returned data. The component generator
+     * uses this to route the resulting component through the script_dataset
+     * sentinel toolId so the frontend resolves it via the queryCache short-circuit.
+     */
+    virtual?: boolean;
+}
+/**
+ * Match tier returned by the LLM script matcher.
+ *
+ * - 'high': the script answers the question directly; only parameter values
+ *   may differ. The runtime replays it with extracted params (cheapest path).
+ * - 'near': the script answers a STRUCTURALLY similar question but needs
+ *   body modification (different metric, dimension, table, filter shape).
+ *   The runtime forks the parent and adapts the body via MainAgent's normal
+ *   write_script + execute_script loop — no SourceAgent dispatch needed.
+ *   See backend/docs/SCRIPT-FLOW-FORK-ADAPT.md for the full design.
+ * - 'edit': the user is INSTRUCTING a change to the script that produced the
+ *   previous answer (not asking a new question). Only reachable when the turn
+ *   carries a ScriptBinding — without one the matcher coerces it to 'none', so
+ *   a prompt regression can't turn this into a loose similarity path. The
+ *   runtime runs MainAgent in edit mode and commits the result in place.
+ *   See backend/docs/SCRIPT-EDIT-DESIGN.md.
+ * - 'none': no script is relevant; full agent flow runs.
+ */
+type MatchTier = 'high' | 'near' | 'edit' | 'none';
+/**
+ * Which recipe produced the answer the user is currently looking at.
+ *
+ * Set whenever a turn's answer came from a script (replay, fresh authoring, or
+ * a committed edit) and carried on the UIBlock + saved conversation row. Two
+ * consumers:
+ *  1. The matcher — the 'edit' tier is ONLY reachable when a binding exists, so
+ *     "use mode instead of mean" resolves to a concrete script instead of being
+ *     matched on keywords it shares with no script name.
+ *  2. Cache invalidation — after an edit commits, conversations bound to that
+ *     recipeId must be dropped, or the exact-match cache replays the pre-edit
+ *     answer and the edit looks like a no-op.
+ *
+ * See backend/docs/SCRIPT-EDIT-DESIGN.md § Component 1.
+ */
+interface ScriptBinding {
+    recipeId: string;
+    /** Params the script ran with — the edit's starting point. */
+    params: Record<string, any>;
+    /** Recipe name at bind time (matcher catalog + user-facing confirmation). */
+    name: string;
+    /** Columns the last run returned — grounds the editor without a re-query. */
+    columns?: string[];
+    /**
+     * The question that produced this answer. Required for disambiguation when a
+     * thread ran several scripts — without it the candidates are just names and
+     * the matcher cannot resolve "use the mode for the WSP one".
+     */
+    userPrompt?: string;
+}
+/**
+ * Result from the LLM-based script matcher.
+ *
+ * For `tier: 'high'`, `extractedParams` carries the values to pass to the
+ * existing script. For `tier: 'near'`, `gaps` and `modificationHint` describe
+ * what the fork-author needs to change in the parent body.
+ */
+interface ScriptMatch {
+    /** The matched script recipe */
+    recipe: ScriptRecipe;
+    /** Match tier — see MatchTier docs */
+    tier: MatchTier;
+    /** Similarity score (0-1, derived from LLM tier) */
+    similarity: number;
+    /**
+     * Legacy confidence level. Mirrors `tier === 'high'`/`'near'` for now;
+     * kept so existing callers compile while we migrate to tier-based logic.
+     */
+    confidence: 'high' | 'medium';
+    /** Parameters extracted from the user question by the LLM (tier='high') */
+    extractedParams?: Record<string, any>;
+    /**
+     * Proper nouns the question named, VERBATIM, with the parameter each fills.
+     *
+     * The matcher must NOT resolve these itself. Observed live: it substituted the
+     * customer name "PREMER ENRGES PHOTOOTAIC" into an id parameter believing that
+     * "the runtime will resolve the customer name to its PartyId(s)" — nothing on
+     * the replay path does, and `WHERE BillingPartyId IN (PREMER ENRGES
+     * PHOTOOTAIC)` reached SQL. Reported here, the runtime resolves them before
+     * replay and rejects the match when one cannot be resolved.
+     */
+    mentions?: Array<{
+        text: string;
+        param?: string;
+    }>;
+    /** Business concepts the question relies on (e.g. "outstanding balance"). */
+    entityTypes?: string[];
+    /** What the user question needs that the parent doesn't cover (tier='near') */
+    gaps?: string[];
+    /** One-sentence description of the change the fork-author should make (tier='near') */
+    modificationHint?: string;
+    /**
+     * Which HALF of the recipe the edit targets (tier='edit'). A recipe has two
+     * independently editable halves:
+     *  - 'data'    — the scriptBody: how the rows are produced (aggregation,
+     *                filters, joins, grouping). Runs MainAgent in edit mode.
+     *  - 'display' — the component specs: how those SAME rows are shown (chart
+     *                type, orientation, columns, labels). Replays the proven SQL
+     *                and regenerates the specs — never authors a script.
+     * Defaults to 'data' when the matcher omits it.
+     */
+    editTarget?: 'data' | 'display';
+    /**
+     * Self-contained restatement of the change the user asked for (tier='edit').
+     * MUST have pronouns/deixis resolved ("this", "it", "that column") — the edit
+     * prompt never sees the conversation history, so an unresolved instruction is
+     * unusable downstream.
+     */
+    editInstruction?: string;
+    /** Why the matcher made this choice (for logs and telemetry) */
+    reasoning?: string;
+}
+
+/**
+ * Multi-Agent Architecture Types
+ *
+ * Defines interfaces for the hierarchical agent system:
+ * - Main Agent: ONE LLM.streamWithTools() call with source agent tools
+ * - Source Agents: independent agents that query individual data sources
+ *
+ * The main agent sees only source summaries. When it calls a source tool,
+ * the SourceAgent runs independently (own LLM, own retries) and returns clean data.
+ */
+
+/**
+ * Per-entity detail: name, row count, and column names.
+ * Gives the main agent enough context to route to the right source.
+ */
+interface EntityDetail {
+    /** Entity name (table, sheet, endpoint) */
+    name: string;
+    /** Approximate row count */
+    rowCount?: number;
+    /** Column/field names */
+    columns: string[];
+    /** Entity-level semantic summary (what the table means) — for main-agent routing. */
+    summary?: string;
+}
+/**
+ * Representation of a data source for the main agent.
+ * Contains entity names WITH column names so the LLM can route accurately.
+ */
+interface SourceSummary {
+    /** Source ID (matches tool ID prefix) */
+    id: string;
+    /** Human-readable source name */
+    name: string;
+    /** Source type: postgres, excel, rest_api, etc. */
+    type: string;
+    /** Brief description of what data this source contains */
+    description: string;
+    /** Detailed entity info with column names for routing */
+    entityDetails: EntityDetail[];
+    /** The tool ID associated with this source */
+    toolId: string;
+}
+/**
+ * What a source agent returns after querying its data source.
+ * The main agent uses this to analyze and compose the final response.
+ */
+interface SourceAgentResult {
+    /** Source ID */
+    sourceId: string;
+    /** Source name */
+    sourceName: string;
+    /** Whether the query succeeded */
+    success: boolean;
+    /** Result data rows */
+    data: any[];
+    /** Metadata about the query execution */
+    metadata: SourceAgentMetadata;
+    /** Tool execution info for the last successful query (backward compat) */
+    executedTool: ExecutedToolInfo;
+    /** All successful tool executions (primary + follow-up queries) */
+    allExecutedTools?: ExecutedToolInfo[];
+    /** Error message if failed */
+    error?: string;
+}
+interface SourceAgentMetadata {
+    /** Total rows that matched the query (before limit) */
+    totalRowsMatched: number;
+    /** Rows actually returned (after limit) */
+    rowsReturned: number;
+    /** Whether the result was truncated by the row limit */
+    isLimited: boolean;
+    /** The query/params that were executed */
+    queryExecuted?: string;
+    /** Execution time in milliseconds */
+    executionTimeMs: number;
+}
+/**
+ * A pre-built, multi-step UI flow registered with the SDK.
+ *
+ * When the main agent decides a user's question matches a workflow's whenToUse
+ * trigger, it picks the workflow instead of running source agents / generating
+ * dashboard components. The LLM extracts the workflow's required props from the
+ * prompt (using `propsSchema` as the tool input_schema) and the SDK returns the
+ * workflow component directly — no analysis text, no chart generation. The
+ * frontend renders the registered workflow component with the LLM-extracted
+ * props.
+ */
+interface WorkflowDescriptor {
+    /** Unique workflow id (used as the LLM tool name) */
+    id: string;
+    /** Component name on the frontend (matches the registered React component) */
+    name: string;
+    /** Short human-readable description of what this workflow does */
+    description: string;
+    /**
+     * 1–2 sentence trigger condition. The LLM uses this to decide if the
+     * user's prompt matches this workflow. Be specific — e.g.
+     * "User wants to *initiate* an inventory transfer (review + submit POs),
+     * not just see analysis or charts."
+     */
+    whenToUse: string;
+    /**
+     * JSON-schema-style description of the props the workflow needs. Becomes
+     * the LLM tool's input_schema, so the model fills these from the prompt.
+     * Use the same shape as `params` on direct tools — string descriptors with
+     * an optional "(optional)" suffix.
+     *
+     * Example:
+     * ```
+     * {
+     *   selectedStore: 'object — { id, name } of the source branch',
+     *   minROI: 'number (optional) — only show transfers with ROI ≥ this',
+     * }
+     * ```
+     */
+    propsSchema: Record<string, string>;
+    /**
+     * Optional: static prop defaults merged with LLM-extracted props before
+     * the component is returned. Useful for things like the embedded
+     * `externalTool` config that the workflow uses to fetch its own data.
+     */
+    defaultProps?: Record<string, any>;
+}
+/**
+ * The workflow selection captured during a routing call.
+ * Set on AgentResponse when the LLM picks a workflow tool.
+ */
+interface SelectedWorkflow {
+    /** Component name (matches WorkflowDescriptor.name) */
+    name: string;
+    /** Props extracted from the prompt + merged with workflow.defaultProps */
+    props: Record<string, any>;
+}
+/**
+ * Set when this turn applies a user-directed edit to an existing script instead
+ * of authoring a new one. MainAgent keeps the SAME harness (tools, loop,
+ * write_script/execute_script verification) and swaps only the system prompt —
+ * `agent-main-edit` instead of `agent-main`.
+ *
+ * The two framings contradict each other on whether to query a source first, so
+ * they must never be resident in one rendered prompt.
+ * See backend/docs/SCRIPT-EDIT-DESIGN.md § D3.
+ */
+interface EditContext {
+    /** Recipe being edited — the shadow draft records it as parentId. */
+    recipeId: string;
+    parentName: string;
+    parentBody: string;
+    /** Self-contained restatement of the change (from the matcher). */
+    instruction: string;
+    /** Columns the last run returned — grounds the edit without a re-query. */
+    lastResultColumns?: string[];
+    /** Rendered parameter list of the parent, for the prompt. */
+    parentParams?: string;
+}
+/**
+ * The complete response from the multi-agent system.
+ * Contains everything needed for text display + component generation.
+ */
+interface AgentResponse {
+    /** Generated text response (analysis of the data) */
+    text: string;
+    /** All executed tools across all source agents (for component generation) */
+    executedTools: ExecutedToolInfo[];
+    /** Individual results from each source agent */
+    sourceResults: SourceAgentResult[];
+    /**
+     * Populated when MainAgent wrote AND successfully executed a script during its turn.
+     * Caller (agent-user-response.ts) persists it via ScriptStore.save().
+     * Absent when MainAgent didn't write one (trivial question / all attempts failed).
+     */
+    savedScript?: AgentWrittenScript;
+    /**
+     * Set when the LLM routed the question to a registered workflow component.
+     * When present, the upstream caller should skip component generation and
+     * return this workflow as the response.
+     */
+    workflow?: SelectedWorkflow;
+    /**
+     * Validated component specs the agent authored via `render_components`.
+     * When present the caller assembles the dashboard from these instead of
+     * running the component-generation LLM. Absent when the agent didn't call
+     * the tool — the caller then falls back to `generateScriptComponents`.
+     * See backend/docs/COMPONENT-GENERATION-V2.md §2.2.
+     */
+    componentSpecs?: ScriptComponentSpec[];
+    /** Layout title/description from the same `render_components` call. */
+    componentLayout?: {
+        title: string;
+        description?: string;
+    };
+}
+/**
+ * A script MainAgent authored + verified during its turn. Shape aligns with
+ * what ScriptStore.save() needs — minus store-assigned fields (id, timestamps, counts).
+ */
+interface AgentWrittenScript {
+    /**
+     * `ScriptRecipe.id` of the draft that was authored + verified during this turn.
+     * The caller passes this to `ScriptStore.promoteToVerified(recipeId, …)` to
+     * flip the draft to verified status and (when possible) drop the turn-suffix
+     * from its filename.
+     */
+    recipeId: string;
+    name: string;
+    intentDescription: string;
+    tags: string[];
+    parameters: Array<{
+        name: string;
+        type: 'string' | 'number' | 'date' | 'date_range' | 'enum' | 'boolean';
+        required: boolean;
+        default?: any;
+        enumValues?: Record<string, string>;
+        description: string;
+    }>;
+    scriptBody: string;
+    /** Source IDs referenced by the script (extracted from ctx.query calls) */
+    sourceIds: string[];
+    /** Tables referenced in the script's SQL (regex-extracted) */
+    tables: string[];
+    /** Executed queries from the verified run — fed to component generation */
+    executedQueries: Array<{
+        sourceId: string;
+        sourceName: string;
+        sql: string;
+        data: any[];
+        count: number;
+        totalCount?: number;
+        executionTimeMs: number;
+        /**
+         * True for synthetic entries (ctx.emit datasets, the computed:_final
+         * post-JS data). The component generator routes virtual sources through
+         * the script_dataset sentinel toolId so the frontend resolves them via
+         * queryCache instead of attempting to re-execute SQL.
+         */
+        virtual?: boolean;
+    }>;
+}
+/**
+ * Debug trace for a single chat_agent turn — which KB/schema nodes were
+ * retrieved and considered, and which script (if any) produced the answer.
+ * Persisted separately from the answer itself (fusion-5's
+ * `conversation_traces` table, written via `saveTrace`), never sent through
+ * `analyticsClient` (data-residency — see design doc §5).
+ */
+interface RunTrace {
+    version: 1;
+    kbNodes: {
+        /**
+         * Always-on nodes included regardless of the question (getGlobalKnowledgeBase).
+         * `content` is the node's full text AS IT WAS at retrieval time — captured so a
+         * later edit/delete of the KB node doesn't erase what the LLM actually saw.
+         */
+        global: Array<{
+            kbId: string;
+            title: string;
+            content?: string;
+        }>;
+        /** Semantically matched for this specific question (getKnowledgeBase). Same `content` reasoning as above. */
+        query: Array<{
+            kbId: string;
+            title: string;
+            similarity?: number;
+            content?: string;
+        }>;
+    };
+    schemaNodes: {
+        /** MainAgent's cross-source prefilter — which sources were even considered. */
+        routing: Array<{
+            sourceId: string;
+            sourceName?: string;
+            similarity: number;
+        }>;
+        /** SourceAgent-scope per-source table resolution (MainAgent.preResolveSchema). */
+        preResolve: Array<{
+            sourceId: string;
+            table?: string;
+            description?: string;
+            similarity: number;
+        }>;
+        /**
+         * The exact {{SOURCE_SUMMARIES}} text injected into MainAgent's system prompt
+         * (agent-prompt-builder.ts's formatSummariesForPrompt output) — every
+         * registered source's catalog entry, not retrieval-specific, captured once per
+         * turn. This is what the main agent actually saw when deciding which source(s)
+         * to call.
+         */
+        mainAgentCatalog?: string;
+        /**
+         * One entry per SourceAgent dispatch (a source called twice in one turn gets
+         * two entries) — the exact {{FULL_SCHEMA}} text injected into that
+         * SourceAgent's own system prompt, however it was sourced, plus any
+         * search_schema follow-up lookups it made mid-conversation when the schema
+         * shown wasn't enough.
+         */
+        sourceAgentSchemas: Array<{
+            sourceId: string;
+            sourceName?: string;
+            /**
+             * The exact intent text MainAgent wrote for this dispatch (before the
+             * resolved-entities block gets appended by code — see
+             * buildResolvedEntitiesBlock) — what MainAgent actually asked this
+             * source agent to go find.
+             */
+            intent: string;
+            /** How `schemaText` was sourced — see SourceAgent.buildPrompt's fullSchema fallback chain. */
+            method: 'preresolved-embedding' | 'tool-full-schema' | 'tool-description' | 'none';
+            schemaText: string;
+            /** In call order — each search_schema tool call this SourceAgent made, and the exact result text handed back to its own LLM. */
+            searchSchemaLookups: Array<{
+                keywords: string[];
+                result: string;
+            }>;
+        }>;
+    };
+    /** The script that produced this answer, if any. Null for a general (no-tool) answer. */
+    script: {
+        /** True when an existing recipe was replayed unchanged (ScriptMatcher tier='high'). */
+        reused: boolean;
+        recipeId?: string | null;
+        body?: string | null;
+        /** Set when reused — the matcher's reasoning for picking this recipe. */
+        reasoning?: string | null;
+    } | null;
+    /**
+     * `resolve_entities` tool activity this turn (see main-agent.ts's
+     * RESOLVE_ENTITIES_TOOL_DEF / handleResolveEntities) — names/identifiers the
+     * user mentioned, resolved to database ids before any SQL was written.
+     * Absent/empty entirely for a project with no entity-search collection
+     * registered, or a turn that named nothing.
+     */
+    entityResolution: {
+        /**
+         * The entity-type catalog text injected into MainAgent's system prompt
+         * (loadEntityCatalog's output) — what let the LLM name a concept instead
+         * of guessing the project's vocabulary. Captured once per turn.
+         */
+        catalogText?: string;
+        /**
+         * One entry per `resolve_entities` tool call this turn (a turn can call it
+         * more than once) — the FULL raw response from the entity-search
+         * collection, not just the trimmed {mention,entityType,instanceId,
+         * displayName} MainAgent keeps for itself to build the dispatch block.
+         * `score`/`ambiguous`/`alternatives` only exist here, nowhere else.
+         */
+        calls: Array<{
+            mentions: Array<{
+                text: string;
+                entityType?: string;
+            }>;
+            entityTypes: string[];
+            resolved: Array<{
+                mention: string;
+                entityType: string;
+                instanceId: string;
+                displayName: string;
+                attrs?: Record<string, any>;
+                score?: number;
+                ambiguous?: boolean;
+                alternatives?: Array<{
+                    instanceId: string;
+                    displayName: string;
+                    score: number;
+                }>;
+            }>;
+            unresolved: string[];
+            entityMap: Record<string, any>;
+            unmatchedEntityTypes?: Array<{
+                requested: string;
+                candidates?: string[];
+            }>;
+            /** Set when the collection call itself failed (fails open — see handleResolveEntities). */
+            error?: string;
+        }>;
+    };
+    /**
+     * The prior-turn conversation context injected into this turn's prompt as
+     * CONVERSATION_HISTORY — exactly what each flow already builds for itself
+     * (chat: thread.getConversationContext(); report/dashboard: their own
+     * equivalents), captured as-is. Every call site sets this unconditionally
+     * (empty string '' when there's no prior context — new thread, or history
+     * trimmed to nothing), so an empty turn is visibly "no history" in the
+     * saved trace rather than a silently missing key. Optional only because
+     * traces saved before this field existed won't have it at all.
+     */
+    conversationHistory?: string;
+    /**
+     * The exact {{USER_ACCESS_CONTEXT}} text injected into MainAgent's system
+     * prompt — the per-user authorization config (sa-api users.config)
+     * rendered by buildUserAccessContext, as-is. Captured once (MainAgent's
+     * prompt is what SourceAgent's identical value is derived from — no need
+     * to duplicate it per source dispatch). Set unconditionally (empty string
+     * '' when there's no config for this user), so "no restrictions" is
+     * visible in the saved trace rather than a silently missing key. Optional
+     * only because traces saved before this field existed won't have it.
+     */
+    userAccessContext?: string;
+}
+/**
+ * Configuration for the multi-agent system.
+ * Controls limits, models, and behavior.
+ */
+interface AgentConfig {
+    /** Max rows shown to the UI preview / inlined per source (default: 10) */
+    maxRowsPerSource: number;
+    /**
+     * Max rows a source query may FETCH from the DB server-side (default: 2000).
+     * Decoupled from what the main agent is shown: the full result is fetched and
+     * summarized (bounded), but only a small/complete slice enters LLM context.
+     * This lets small lookups (benchmark maps) arrive COMPLETE without letting
+     * large results blow up context.
+     */
+    maxRowsFetched: number;
+    /** Model for the main agent (routing + analysis in one LLM call) */
+    mainAgentModel: string;
+    /** Model for source agent query generation */
+    sourceAgentModel: string;
+    /** API key for LLM calls */
+    apiKey?: string;
+    /** Max retry attempts per source agent */
+    maxRetries: number;
+    /** Max tool calling iterations for the main agent loop */
+    maxIterations: number;
+    /** Global knowledge base context (static, same for all users/questions — cached in system prompt) */
+    globalKnowledgeBase?: string;
+    /** Per-request knowledge base context (user-specific + query-matched — dynamic, not cached) */
+    knowledgeBaseContext?: string;
+    /** Per-user authorization config (sa-api users.config), rendered to text — see buildUserAccessContext */
+    userAccessContext?: string;
+    /** Collections registry (ChromaDB search hooks) for embedding-based schema + source search */
+    collections?: any;
+    /** Optional project ID for scoping embedding searches */
+    projectId?: string;
+    /**
+     * Optional debug-trace accumulator for this turn (Phase 1, chat_agent only).
+     * When present, MainAgent-scope retrieval call sites (e.g. preResolveSchema)
+     * push their structured results into it. Absent for callers that don't care
+     * about trace capture (e.g. headless flows) — every push site must treat
+     * this as optional and never let a capture failure affect the real answer.
+     */
+    trace?: RunTrace;
+}
+/**
+ * Default agent configuration
+ */
+declare const DEFAULT_AGENT_CONFIG: AgentConfig;
+
 /**
  * Unified UIBlock structure for database storage
- * Used in both bookmarks and user-conversations tables
+ * Used in both bookmarks and user-conversations tables.
+ *
+ * `analysis` always holds whatever real narration exists — the full answer on
+ * success, or whatever was actually streamed before a failure (may be empty).
+ * `error` is a dedicated field, null on success — it can be a plain string OR
+ * a structured object/array (e.g. the agent's raw `errors` list), whatever
+ * shape the failure naturally has; it's never coerced into `analysis`.
  */
 interface DBUIBlock {
     id: string;
     component: Record<string, any> | null;
     analysis: string | null;
     user_prompt: string;
+    error?: unknown | null;
+    /**
+     * The script recipe that produced this answer, when one did. Read back via
+     * `conversation-history.exactMatch` so an edit follow-up still has a target
+     * after a reload (the in-memory thread is gone by then), and indexed so
+     * committing an edit can purge every cached answer bound to the recipe.
+     */
+    scriptBinding?: {
+        recipeId: string;
+        params?: Record<string, any>;
+        name?: string;
+        columns?: string[];
+    };
 }
 
 /**
@@ -160,6 +1074,7 @@ declare const DSLRendererPropsSchema$1: z.ZodObject<{
             dependencies?: string[] | undefined;
         } | undefined;
         props?: Record<string, any> | undefined;
+        data?: Record<string, any> | undefined;
         render?: any;
         states?: Record<string, any> | undefined;
         methods?: Record<string, {
@@ -170,7 +1085,6 @@ declare const DSLRendererPropsSchema$1: z.ZodObject<{
             fn: string;
             deps?: string[] | undefined;
         }[] | undefined;
-        data?: Record<string, any> | undefined;
         pages?: {
             id: string;
             name: string;
@@ -192,6 +1106,7 @@ declare const DSLRendererPropsSchema$1: z.ZodObject<{
             dependencies?: string[] | undefined;
         } | undefined;
         props?: Record<string, any> | undefined;
+        data?: Record<string, any> | undefined;
         render?: any;
         states?: Record<string, any> | undefined;
         methods?: Record<string, {
@@ -202,7 +1117,6 @@ declare const DSLRendererPropsSchema$1: z.ZodObject<{
             fn: string;
             deps?: string[] | undefined;
         }[] | undefined;
-        data?: Record<string, any> | undefined;
         pages?: {
             id: string;
             name: string;
@@ -228,6 +1142,7 @@ declare const DSLRendererPropsSchema$1: z.ZodObject<{
             dependencies?: string[] | undefined;
         } | undefined;
         props?: Record<string, any> | undefined;
+        data?: Record<string, any> | undefined;
         render?: any;
         states?: Record<string, any> | undefined;
         methods?: Record<string, {
@@ -238,7 +1153,6 @@ declare const DSLRendererPropsSchema$1: z.ZodObject<{
             fn: string;
             deps?: string[] | undefined;
         }[] | undefined;
-        data?: Record<string, any> | undefined;
         pages?: {
             id: string;
             name: string;
@@ -264,6 +1178,7 @@ declare const DSLRendererPropsSchema$1: z.ZodObject<{
             dependencies?: string[] | undefined;
         } | undefined;
         props?: Record<string, any> | undefined;
+        data?: Record<string, any> | undefined;
         render?: any;
         states?: Record<string, any> | undefined;
         methods?: Record<string, {
@@ -274,7 +1189,6 @@ declare const DSLRendererPropsSchema$1: z.ZodObject<{
             fn: string;
             deps?: string[] | undefined;
         }[] | undefined;
-        data?: Record<string, any> | undefined;
         pages?: {
             id: string;
             name: string;
@@ -355,6 +1269,7 @@ declare const DSLRendererPropsSchema: z.ZodObject<{
             dependencies?: string[] | undefined;
         } | undefined;
         props?: Record<string, any> | undefined;
+        data?: Record<string, any> | undefined;
         render?: any;
         states?: Record<string, any> | undefined;
         methods?: Record<string, {
@@ -365,7 +1280,6 @@ declare const DSLRendererPropsSchema: z.ZodObject<{
             fn: string;
             deps?: string[] | undefined;
         }[] | undefined;
-        data?: Record<string, any> | undefined;
     }, {
         id: string;
         name?: string | undefined;
@@ -379,6 +1293,7 @@ declare const DSLRendererPropsSchema: z.ZodObject<{
             dependencies?: string[] | undefined;
         } | undefined;
         props?: Record<string, any> | undefined;
+        data?: Record<string, any> | undefined;
         render?: any;
         states?: Record<string, any> | undefined;
         methods?: Record<string, {
@@ -389,7 +1304,6 @@ declare const DSLRendererPropsSchema: z.ZodObject<{
             fn: string;
             deps?: string[] | undefined;
         }[] | undefined;
-        data?: Record<string, any> | undefined;
     }>;
     data: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodAny>>;
     context: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodAny>>;
@@ -407,6 +1321,7 @@ declare const DSLRendererPropsSchema: z.ZodObject<{
             dependencies?: string[] | undefined;
         } | undefined;
         props?: Record<string, any> | undefined;
+        data?: Record<string, any> | undefined;
         render?: any;
         states?: Record<string, any> | undefined;
         methods?: Record<string, {
@@ -417,7 +1332,6 @@ declare const DSLRendererPropsSchema: z.ZodObject<{
             fn: string;
             deps?: string[] | undefined;
         }[] | undefined;
-        data?: Record<string, any> | undefined;
     };
     data?: Record<string, any> | undefined;
     context?: Record<string, any> | undefined;
@@ -435,6 +1349,7 @@ declare const DSLRendererPropsSchema: z.ZodObject<{
             dependencies?: string[] | undefined;
         } | undefined;
         props?: Record<string, any> | undefined;
+        data?: Record<string, any> | undefined;
         render?: any;
         states?: Record<string, any> | undefined;
         methods?: Record<string, {
@@ -445,7 +1360,6 @@ declare const DSLRendererPropsSchema: z.ZodObject<{
             fn: string;
             deps?: string[] | undefined;
         }[] | undefined;
-        data?: Record<string, any> | undefined;
     };
     data?: Record<string, any> | undefined;
     context?: Record<string, any> | undefined;
@@ -777,6 +1691,14 @@ declare const ToolSchema: z.ZodObject<{
             description: string;
         }[];
     }>>;
+    /** Cache policy. `false` = never cache (live data, write ops). Mirrors HTTP `Cache-Control: no-store`. */
+    cache: z.ZodOptional<z.ZodUnion<[z.ZodLiteral<false>, z.ZodObject<{
+        ttlMs: z.ZodOptional<z.ZodNumber>;
+    }, "strip", z.ZodTypeAny, {
+        ttlMs?: number | undefined;
+    }, {
+        ttlMs?: number | undefined;
+    }>]>>;
 }, "strip", z.ZodTypeAny, {
     id: string;
     params: Record<string, string>;
@@ -793,6 +1715,9 @@ declare const ToolSchema: z.ZodObject<{
             description: string;
         }[];
     } | undefined;
+    cache?: false | {
+        ttlMs?: number | undefined;
+    } | undefined;
 }, {
     id: string;
     params: Record<string, string>;
@@ -809,6 +1734,9 @@ declare const ToolSchema: z.ZodObject<{
             description: string;
         }[];
     } | undefined;
+    cache?: false | {
+        ttlMs?: number | undefined;
+    } | undefined;
 }>;
 type Tool$1 = z.infer<typeof ToolSchema>;
 type CollectionOperation = 'getMany' | 'getOne' | 'query' | 'mutation' | 'updateOne' | 'deleteOne' | 'createOne';
@@ -994,7 +1922,7 @@ declare const KbNodesRequestPayloadSchema: z.ZodObject<{
         content?: string | undefined;
     }>>;
 }, "strip", z.ZodTypeAny, {
-    operation: "create" | "getOne" | "update" | "delete" | "getAll" | "search" | "getByCategory" | "getByUser" | "getCategories" | "getTags";
+    operation: "create" | "getOne" | "search" | "update" | "delete" | "getAll" | "getByCategory" | "getByUser" | "getCategories" | "getTags";
     data?: {
         id?: number | undefined;
         type?: "query" | "user" | "global" | undefined;
@@ -1017,7 +1945,7 @@ declare const KbNodesRequestPayloadSchema: z.ZodObject<{
         content?: string | undefined;
     } | undefined;
 }, {
-    operation: "create" | "getOne" | "update" | "delete" | "getAll" | "search" | "getByCategory" | "getByUser" | "getCategories" | "getTags";
+    operation: "create" | "getOne" | "search" | "update" | "delete" | "getAll" | "getByCategory" | "getByUser" | "getCategories" | "getTags";
     data?: {
         id?: number | undefined;
         type?: "query" | "user" | "global" | undefined;
@@ -1301,279 +2229,310 @@ declare class ReportManager {
     reportExists(reportId: string): boolean;
     /**
      * Get report count
-     * @returns Number of reports
-     */
-    getReportCount(): number;
-}
-
-/**
- * StreamBuffer - Buffered streaming utility for smoother text delivery
- * Batches small chunks together and flushes at regular intervals
- */
-type StreamCallback = (chunk: string) => void;
-/**
- * StreamBuffer class for managing buffered streaming output
- * Provides smooth text delivery by batching small chunks
- */
-declare class StreamBuffer {
-    private buffer;
-    private flushTimer;
-    private callback;
-    private fullText;
-    constructor(callback?: StreamCallback);
-    /**
-     * Check if the buffer has a callback configured
-     */
-    hasCallback(): boolean;
-    /**
-     * Get all text that has been written (including already flushed)
-     */
-    getFullText(): string;
-    /**
-     * Write a chunk to the buffer
-     * Large chunks or chunks with newlines are flushed immediately
-     * Small chunks are batched and flushed after a short interval
-     *
-     * @param chunk - Text chunk to write
-     */
-    write(chunk: string): void;
-    /**
-     * Flush the buffer immediately
-     * Call this before tool execution or other operations that need clean output
-     */
-    flush(): void;
-    /**
-     * Internal flush implementation
-     */
-    private flushNow;
-    /**
-     * Clean up resources
-     * Call this when done with the buffer
-     */
-    dispose(): void;
-}
-
-/**
- * ToolExecutorService - Handles execution of SQL queries and external tools
- * Extracted from BaseLLM.generateTextResponse for better separation of concerns
- */
-
-/**
- * External tool definition
- */
-interface ExternalTool {
-    id: string;
-    name: string;
-    description?: string;
-    /** Tool type: "source" = routed through SourceAgent, "direct" = called directly by MainAgent */
-    toolType?: 'source' | 'direct';
-    /** Full untruncated schema for source agent (all columns visible) */
-    fullSchema?: string;
-    /** Schema size tier: small (≤50 tables), medium (51-200), large (201-500), very_large (500+) */
-    schemaTier?: string;
-    /** Schema search function for very_large tier — keyword search over entities */
-    schemaSearchFn?: (keywords: string[]) => string;
-    fn: (input: any) => Promise<any>;
-    limit?: number;
-    outputSchema?: any;
-    executionType?: 'immediate' | 'deferred';
-    userProvidedData?: any;
-    params?: Record<string, any>;
-}
-/**
- * Executed tool tracking info
- */
-interface ExecutedToolInfo {
-    id: string;
-    name: string;
-    params: any;
-    result: {
-        _totalRecords: number;
-        _recordsShown: number;
-        _metadata?: any;
-        _sampleData: any[];
-    };
-    outputSchema?: any;
-    sourceSchema?: string;
-    sourceType?: string;
+     * @returns Number of reports
+     */
+    getReportCount(): number;
 }
 
 /**
- * Multi-Agent Architecture Types
+ * ScriptRecipeStore — injected metadata backend for the script flow.
  *
- * Defines interfaces for the hierarchical agent system:
- * - Main Agent: ONE LLM.streamWithTools() call with source agent tools
- * - Source Agents: independent agents that query individual data sources
+ * The SDK is standalone (no DB dependency). The backend implements this
+ * interface over Postgres (full-text search + atomic counters) and injects it
+ * via `collections['script-recipes']`, exactly like `collections['source-embeddings']`.
+ * `ScriptStore` consumes it for all METADATA operations while keeping the
+ * executable body on disk as scripts-store/<fileBase>.ts.
  *
- * The main agent sees only source summaries. When it calls a source tool,
- * the SourceAgent runs independently (own LLM, own retries) and returns clean data.
+ * All metadata rows are plain JSON (no scriptBody — that lives on disk).
+ * See backend/docs/SCRIPT-FLOW-SCALING-ISSUES.md (#1, #3, #7).
  */
 
-/**
- * Per-entity detail: name, row count, and column names.
- * Gives the main agent enough context to route to the right source.
- */
-interface EntityDetail {
-    /** Entity name (table, sheet, endpoint) */
-    name: string;
-    /** Approximate row count */
-    rowCount?: number;
-    /** Column/field names */
-    columns: string[];
-}
-/**
- * Representation of a data source for the main agent.
- * Contains entity names WITH column names so the LLM can route accurately.
- */
-interface SourceSummary {
-    /** Source ID (matches tool ID prefix) */
+/** One recipe's metadata as stored in Postgres (mirrors the script_recipes table). */
+interface ScriptRecipeMetaRow {
     id: string;
-    /** Human-readable source name */
+    projectId?: string | null;
+    version: number;
     name: string;
-    /** Source type: postgres, excel, rest_api, etc. */
-    type: string;
-    /** Brief description of what data this source contains */
-    description: string;
-    /** Detailed entity info with column names for routing */
-    entityDetails: EntityDetail[];
-    /** The tool ID associated with this source */
-    toolId: string;
-}
-/**
- * What a source agent returns after querying its data source.
- * The main agent uses this to analyze and compose the final response.
- */
-interface SourceAgentResult {
-    /** Source ID */
-    sourceId: string;
-    /** Source name */
-    sourceName: string;
-    /** Whether the query succeeded */
-    success: boolean;
-    /** Result data rows */
-    data: any[];
-    /** Metadata about the query execution */
-    metadata: SourceAgentMetadata;
-    /** Tool execution info for the last successful query (backward compat) */
-    executedTool: ExecutedToolInfo;
-    /** All successful tool executions (primary + follow-up queries) */
-    allExecutedTools?: ExecutedToolInfo[];
-    /** Error message if failed */
-    error?: string;
+    intentDescription: string;
+    tags: string[] | null;
+    createdFrom: string | null;
+    sourceIds: string[] | null;
+    tables: string[] | null;
+    parameters: ScriptParameter[] | null;
+    components?: ScriptComponentSpec[] | null;
+    displayPreference?: ScriptDisplayPreference | null;
+    fileBase: string;
+    bodyHash?: string | null;
+    successCount: number;
+    failureCount: number;
+    lastUsed: string | null;
+    parentId?: string | null;
+    forkDepth?: number | null;
+    forkReason?: string | null;
+    status: 'draft' | 'verified' | string;
+    turnId?: string | null;
+    editedBy?: string | null;
+    editedFrom?: string | null;
+    history?: ScriptVersionRecord[] | null;
+    lastError?: {
+        phase: 'compile' | 'runtime' | 'timeout' | 'ipc';
+        message: string;
+        at: string;
+        attempt: number;
+    } | null;
+    createdAt?: string | null;
+    updatedAt?: string | null;
 }
-interface SourceAgentMetadata {
-    /** Total rows that matched the query (before limit) */
-    totalRowsMatched: number;
-    /** Rows actually returned (after limit) */
-    rowsReturned: number;
-    /** Whether the result was truncated by the row limit */
-    isLimited: boolean;
-    /** The query/params that were executed */
-    queryExecuted?: string;
-    /** Execution time in milliseconds */
-    executionTimeMs: number;
+interface ScriptRecipeStore {
+    /** FTS shortlist of healthy verified recipes for the matcher (metadata only). */
+    search(params: {
+        prompt: string;
+        projectId?: string;
+        limit?: number;
+    }): Promise<ScriptRecipeMetaRow[]>;
+    /** Fetch one recipe by id (any status). */
+    getById(id: string): Promise<ScriptRecipeMetaRow | null>;
+    /** Count healthy verified recipes (drives the "any scripts?" gate). */
+    count(params?: {
+        projectId?: string;
+    }): Promise<number>;
+    /** Insert or update a recipe row (keyed by id). */
+    upsert(row: ScriptRecipeMetaRow): Promise<void>;
+    /** Atomically bump counters / last-used. */
+    updateStats(id: string, patch: {
+        successDelta?: number;
+        failureDelta?: number;
+        lastUsed?: string;
+    }): Promise<void>;
+    /** Flip a draft to verified, applying provenance + optional fork lineage. */
+    promote(id: string, patch: {
+        sourceIds: string[];
+        tables: string[];
+        fileBase?: string;
+        parentId?: string;
+        forkDepth?: number;
+        forkReason?: string;
+        components?: ScriptComponentSpec[];
+    }): Promise<ScriptRecipeMetaRow | null>;
+    /**
+     * Commit a verified edit onto an EXISTING recipe: bump `version`, replace the
+     * body-bearing metadata, append a history record, reset health counters, and
+     * clear the component specs (they were validated against the old shape).
+     * Returns the updated row, or null when the target is gone.
+     * See backend/docs/SCRIPT-EDIT-DESIGN.md § Component 5.
+     */
+    commitEdit?(id: string, patch: {
+        name?: string;
+        intentDescription?: string;
+        tags?: string[];
+        parameters?: ScriptParameter[];
+        bodyHash: string;
+        /** New on-disk stem when the edit renamed the script; omitted otherwise. */
+        fileBase?: string;
+        sourceIds?: string[];
+        tables?: string[];
+        editedBy?: string;
+        editedFrom?: string;
+        historyEntry: {
+            version: number;
+            at: string;
+            by?: string;
+            instruction?: string;
+            changeSummary?: string;
+            bodyHash?: string;
+        };
+    }): Promise<ScriptRecipeMetaRow | null>;
+    /** Stamp a draft's last execution error. */
+    recordDraftError(id: string, err: {
+        phase: string;
+        message: string;
+        attempt: number;
+        at: string;
+    }): Promise<void>;
+    /** Delete a recipe row (body file removed separately). */
+    remove(id: string): Promise<void>;
+    /** True if `fileBase` is taken by a different recipe in this project. */
+    fileBaseTaken(fileBase: string, excludeId: string, projectId?: string): Promise<boolean>;
 }
+/** Pull the injected store off the collections bag (or null if not wired). */
+declare function resolveScriptRecipeStore(collections: any): ScriptRecipeStore | null;
+
 /**
- * A pre-built, multi-step UI flow registered with the SDK.
+ * ScriptStore — Postgres metadata + on-disk body for script recipes.
  *
- * When the main agent decides a user's question matches a workflow's whenToUse
- * trigger, it picks the workflow instead of running source agents / generating
- * dashboard components. The LLM extracts the workflow's required props from the
- * prompt (using `propsSchema` as the tool input_schema) and the SDK returns the
- * workflow component directly — no analysis text, no chart generation. The
- * frontend renders the registered workflow component with the LLM-extracted
- * props.
+ * Split of responsibilities:
+ *   - METADATA  → injected `ScriptRecipeStore` (Postgres FTS + atomic counters),
+ *                 resolved from `collections['script-recipes']`.
+ *   - BODY      → scripts-store/<fileBase>.ts, editable in your IDE. Written
+ *                 atomically (temp + rename); `bodyHash` (sha256) detects edits.
+ *
+ * The old "read every file every turn + send the whole catalog to the LLM"
+ * matcher is gone — matching is `store.search(prompt)` (FTS shortlist). The
+ * draft/verified filename dance is gone too: `status` is a DB column and the
+ * file keeps a stable `<fileBase>.ts` name across promotion.
+ *
+ * When no metadata store is injected, the store degrades to a safe no-op
+ * (count 0 → script flow disabled) instead of crashing.
+ *
+ * See backend/docs/SCRIPT-FLOW-SCALING-ISSUES.md.
  */
-interface WorkflowDescriptor {
-    /** Unique workflow id (used as the LLM tool name) */
-    id: string;
-    /** Component name on the frontend (matches the registered React component) */
+
+interface SaveDraftInput {
+    /** Reuse an existing draft (retry); omit to mint a new one. */
+    recipeId?: string;
+    /** Per-turn unique suffix, stable across retries within the turn. */
+    turnId: string;
     name: string;
-    /** Short human-readable description of what this workflow does */
-    description: string;
-    /**
-     * 1–2 sentence trigger condition. The LLM uses this to decide if the
-     * user's prompt matches this workflow. Be specific — e.g.
-     * "User wants to *initiate* an inventory transfer (review + submit POs),
-     * not just see analysis or charts."
-     */
-    whenToUse: string;
-    /**
-     * JSON-schema-style description of the props the workflow needs. Becomes
-     * the LLM tool's input_schema, so the model fills these from the prompt.
-     * Use the same shape as `params` on direct tools — string descriptors with
-     * an optional "(optional)" suffix.
-     *
-     * Example:
-     * ```
-     * {
-     *   selectedStore: 'object — { id, name } of the source branch',
-     *   minROI: 'number (optional) — only show transfers with ROI ≥ this',
-     * }
-     * ```
-     */
-    propsSchema: Record<string, string>;
-    /**
-     * Optional: static prop defaults merged with LLM-extracted props before
-     * the component is returned. Useful for things like the embedded
-     * `externalTool` config that the workflow uses to fetch its own data.
-     */
-    defaultProps?: Record<string, any>;
+    intentDescription: string;
+    tags: string[];
+    parameters: ScriptParameter[];
+    scriptBody: string;
+    createdFrom: string;
+    /**
+     * Set when this draft is a SHADOW of an existing recipe being edited. The
+     * draft is verified independently and merged back onto the parent via
+     * `commitEdit`, so the working script is never clobbered by an edit that
+     * turns out not to run. See backend/docs/SCRIPT-EDIT-DESIGN.md § D5.
+     */
+    parentId?: string;
 }
-/**
- * The workflow selection captured during a routing call.
- * Set on AgentResponse when the LLM picks a workflow tool.
- */
-interface SelectedWorkflow {
-    /** Component name (matches WorkflowDescriptor.name) */
-    name: string;
-    /** Props extracted from the prompt + merged with workflow.defaultProps */
-    props: Record<string, any>;
+interface PromoteToVerifiedInput {
+    sourceIds: string[];
+    tables: string[];
+    parentId?: string;
+    forkDepth?: number;
+    forkReason?: string;
+    components?: ScriptComponentSpec[];
+}
+interface ScriptStoreOptions {
+    /** Explicit metadata store, or resolved from `collections['script-recipes']`. */
+    store?: ScriptRecipeStore | null;
+    collections?: any;
+    /** Body directory (defaults to <cwd>/scripts-store). */
+    baseDir?: string;
+    /** Project scope stamped on every row. */
+    projectId?: string;
 }
 /**
- * The complete response from the multi-agent system.
- * Contains everything needed for text display + component generation.
+ * Normalize a scriptBody into the on-disk form (strip a leading comment block,
+ * ensure `export async function getData`). Exported for MainAgent.
  */
-interface AgentResponse {
-    /** Generated text response (analysis of the data) */
-    text: string;
-    /** All executed tools across all source agents (for component generation) */
-    executedTools: ExecutedToolInfo[];
-    /** Individual results from each source agent */
-    sourceResults: SourceAgentResult[];
+declare function normalizeScriptBody(scriptBody: string): string;
+declare class ScriptStore {
+    private store;
+    private storeDir;
+    private projectId?;
+    constructor(opts?: ScriptStoreOptions);
+    /** Whether a metadata store is wired (matcher / authoring are gated on this). */
+    hasStore(): boolean;
+    /** Number of healthy verified recipes (gates the script-matching path). */
+    count(): Promise<number>;
+    /**
+     * FTS shortlist for the matcher (metadata only — bodies are loaded lazily by
+     * `get()` once the LLM picks one). Returns verified, healthy recipes ranked
+     * by relevance.
+     */
+    search(prompt: string, limit?: number): Promise<ScriptRecipe[]>;
+    /** Fetch one recipe by id with its body loaded from disk. */
+    get(id: string): Promise<ScriptRecipe | null>;
+    /** Create or update a recipe (metadata upsert + body write when changed). */
+    save(recipe: ScriptRecipe): Promise<void>;
+    /**
+     * Persist (or update) a draft. Within a turn, retries that pass the same
+     * `recipeId` overwrite the same row + file; a fresh `recipeId` mints a new
+     * draft. The body is visible at scripts-store/<fileBase>.ts immediately.
+     */
+    saveDraft(input: SaveDraftInput): Promise<ScriptRecipe>;
+    /** Stamp a draft's last execution error (metadata only). */
+    recordDraftError(recipeId: string, err: {
+        phase: 'compile' | 'runtime' | 'timeout' | 'ipc';
+        message: string;
+        attempt: number;
+    }): Promise<void>;
+    /**
+     * Promote a successfully-executed draft into a verified script.
+     * The on-disk body already exists at <fileBase>.ts (written at write_script
+     * time) and keeps its name — only the DB row flips status + provenance.
+     */
+    promoteToVerified(recipeId: string, input: PromoteToVerifiedInput): Promise<ScriptRecipe | null>;
+    /**
+     * Commit a user-directed edit: merge a VERIFIED shadow draft back onto the
+     * recipe it was editing, as version N+1 under the SAME recipe id.
+     *
+     * Keeping the id stable is the point of the whole feature — every cached
+     * conversation, `script_dataset` regeneration descriptor and persisted
+     * component spec already points at it, so the correction applies retroactively
+     * to replays instead of stranding them on the old body.
+     *
+     * Steps (see backend/docs/SCRIPT-EDIT-DESIGN.md § Component 5):
+     *  1. archive the current body as `<fileBase>.v<version>.ts`
+     *  2. write the edited body to the original fileBase (atomic)
+     *  3. version++, copy name/description/params, append a history record
+     *  4. reset health counters — the pre-edit failure history is stale
+     *  5. clear component specs — they were validated against the OLD shape
+     *  6. delete the shadow draft (row + file)
+     *
+     * Returns the updated recipe, or null when the commit could not be applied
+     * (caller should then fall back to treating the draft as a new script).
+     */
+    commitEdit(targetId: string, draft: {
+        recipeId: string;
+        name?: string;
+        intentDescription?: string;
+        tags?: string[];
+        parameters?: ScriptParameter[];
+        scriptBody: string;
+        sourceIds?: string[];
+        tables?: string[];
+    }, meta: {
+        instruction?: string;
+        changeSummary?: string;
+        editedBy?: string;
+    }): Promise<ScriptRecipe | null>;
+    /**
+     * Drop a draft (row + body file). MainAgent calls this at end-of-turn when a
+     * draft was authored but never verified — failed drafts are never matched, so
+     * deleting them immediately avoids unbounded accumulation (#5). No-op if the
+     * recipe isn't a draft (so a promoted/verified script is never removed here).
+     */
+    discardDraft(recipeId: string): Promise<void>;
+    /** Delete a recipe (row + body file). */
+    delete(id: string): Promise<void>;
+    /** Record a successful execution (atomic counter bump). */
+    recordSuccess(id: string): Promise<void>;
+    /** Record a failed execution (atomic counter bump). */
+    recordFailure(id: string): Promise<void>;
+    /** Absolute path to the .ts body for a recipe (used by the runner/MainAgent). */
+    getScriptPath(recipe: ScriptRecipe): string;
+    private removeById;
+    private rowToRecipe;
+    private recipeToRow;
+    /** slug of name, with a short id suffix when the bare slug is already taken. */
+    private computeFileBase;
+    private toSlug;
+    private hash;
+    private bodyPath;
+    private readBody;
+    /** Directory holding superseded bodies. Dot-prefixed so IDEs/`ls` hide it. */
+    private get archiveDir();
+    /**
+     * Archive a superseded body as `.versions/<fileBase>.v<n>.ts`.
+     *
+     * Kept out of the main store directory on purpose — see commitEdit step 1.
+     * To roll back: copy the file back over `scripts-store/<fileBase>.ts`.
+     */
+    private writeArchive;
     /**
-     * Set when the LLM routed the question to a registered workflow component.
-     * When present, the upstream caller should skip component generation and
-     * return this workflow as the response.
+     * Move a recipe's archived versions to a new prefix when its fileBase changes,
+     * so all versions of one recipe stay grouped. Without this, two renames would
+     * scatter a single recipe's history across three prefixes in `.versions/` with
+     * nothing linking them back to the live script.
      */
-    workflow?: SelectedWorkflow;
-}
-/**
- * Configuration for the multi-agent system.
- * Controls limits, models, and behavior.
- */
-interface AgentConfig {
-    /** Max rows a source agent can return (default: 50) */
-    maxRowsPerSource: number;
-    /** Model for the main agent (routing + analysis in one LLM call) */
-    mainAgentModel: string;
-    /** Model for source agent query generation */
-    sourceAgentModel: string;
-    /** API key for LLM calls */
-    apiKey?: string;
-    /** Max retry attempts per source agent */
-    maxRetries: number;
-    /** Max tool calling iterations for the main agent loop */
-    maxIterations: number;
-    /** Global knowledge base context (static, same for all users/questions — cached in system prompt) */
-    globalKnowledgeBase?: string;
-    /** Per-request knowledge base context (user-specific + query-matched — dynamic, not cached) */
-    knowledgeBaseContext?: string;
+    private renameArchives;
+    /** Atomic body write (temp + rename) so concurrent reads never see a partial file. */
+    private writeBody;
+    private unlinkBody;
 }
-/**
- * Default agent configuration
- */
-declare const DEFAULT_AGENT_CONFIG: AgentConfig;
 
 /**
  * Main Agent (Orchestrator)
@@ -1595,7 +2554,60 @@ declare class MainAgent {
     private workflows;
     private config;
     private streamBuffer;
-    constructor(externalTools: ExternalTool[], config: AgentConfig, streamBuffer?: StreamBuffer, workflows?: WorkflowDescriptor[]);
+    /**
+     * Optional: when provided, MainAgent exposes the `write_script` /
+     * `execute_script` tools to the LLM and persists drafts to disk via the
+     * store. Headless callers (alert analyzer, metric resolver) omit these to
+     * suppress script authoring entirely — drafts would otherwise leak onto
+     * disk with no caller to promote or clean them up.
+     */
+    private scriptStore;
+    private turnId;
+    private createdFromPrompt;
+    private scriptState;
+    /** Answer-flow component catalog (filtered + projected by the caller). */
+    private componentCatalog;
+    /** Specs accepted by `render_components` this turn; empty until it succeeds. */
+    /**
+     * Entities resolved this turn, appended verbatim to every source dispatch.
+     *
+     * Not left to the model to copy into its intent prose: observed live, the
+     * MainAgent wrote names instead of ids on the first dispatch AND re-spelled
+     * 'F D C LIMITED' as 'FDC Limited', so the source agent fell back to
+     * `PartyName LIKE '%FDC%'` and found 2 of 3 customers. Carrying the ids
+     * structurally makes that impossible rather than instruction-dependent.
+     */
+    private resolvedEntities;
+    private componentSpecs;
+    private componentLayout;
+    private renderComponentAttempts;
+    /**
+     * Fork mode — set when this turn is adapting a near-matching parent script.
+     * In fork mode there is no legitimate "answer with bare text" outcome: the
+     * only correct first move is a tool call (write_script, or a source tool for
+     * schema discovery). We therefore force tool use on the first LLM iteration
+     * so the model can't end its turn with a bare "I'll adapt…" preamble and zero
+     * tool calls. Never set on the fresh-authoring / general-question path.
+     */
+    private forkMode;
+    /**
+     * Edit mode — set when this turn applies a user-directed change to an
+     * existing script. Swaps the system prompt to `agent-main-edit` and stamps
+     * the shadow draft's parentId. Like fork mode there is no legitimate
+     * "answer with bare text" outcome, so tool use is forced on iteration 1.
+     * See backend/docs/SCRIPT-EDIT-DESIGN.md § Component 4.
+     */
+    private editContext;
+    /**
+     * Per-turn cancellation signal (user hit "Stop"). Set at the top of
+     * handleQuestion and read by the tool handler, the SourceAgent dispatch, and
+     * the script subprocess so an abort tears down every layer of the turn.
+     */
+    private abortSignal?;
+    constructor(externalTools: ExternalTool[], config: AgentConfig, scriptStore?: ScriptStore, turnId?: string, streamBuffer?: StreamBuffer, workflows?: WorkflowDescriptor[], forkMode?: boolean, editContext?: EditContext, componentCatalog?: Component[]);
+    /** True when the turn is applying a user-directed script edit. */
+    private get editMode();
+    private get scriptingEnabled();
     /**
      * Handle a user question using the multi-agent system.
      *
@@ -1606,7 +2618,66 @@ declare class MainAgent {
      * 4. Direct tools → fn() called directly with LLM params → returns data
      * 5. Generates final analysis text
      */
-    handleQuestion(userPrompt: string, apiKey?: string, conversationHistory?: string, streamCallback?: (chunk: string) => void): Promise<AgentResponse>;
+    handleQuestion(userPrompt: string, apiKey?: string, conversationHistory?: string, streamCallback?: (chunk: string) => void, signal?: AbortSignal): Promise<AgentResponse>;
+    /**
+     * Tool definition, with the catalog rendered from the registered `Metadata`
+     * the caller passed in. Generated rather than hand-written so it cannot drift
+     * from what the frontend actually ships.
+     */
+    private buildRenderComponentsToolDef;
+    /**
+     * Validate the agent's component picks against the REAL columns of each bound
+     * dataset. All-or-nothing: any error rejects the whole call and returns the
+     * findings, because a partially-accepted dashboard makes the retry ambiguous.
+     */
+    private handleRenderComponents;
+    private handleWriteScript;
+    private handleExecuteScript;
+    /**
+     * Build the AgentWrittenScript payload the caller will hand to
+     * `ScriptStore.promoteToVerified()`. Only returned when a verified
+     * successful execution is on record.
+     */
+    private buildSavedScript;
+    private normalizeParameterList;
+    /**
+     * Use the schema embedding collection to pre-select relevant tables for
+     * this source + intent. Returns a formatted schema block if confidence is
+     * high (top match ≥ 0.55 and ≥3 candidates), otherwise null.
+     *
+     * When this returns a block, we can skip the SourceAgent's `search_schema`
+     * loop and reduce iteration budget. When it returns null, the SourceAgent
+     * falls back to the existing LLM-driven keyword search (same as today).
+     */
+    /**
+     * Ids + verbatim names for everything resolved this turn, appended by CODE to
+     * each source dispatch so neither omission nor re-spelling by the model can
+     * lose them. Empty string when nothing was resolved, so intents are untouched
+     * for questions that name no entities.
+     */
+    private buildResolvedEntitiesBlock;
+    /** True when the backend registered the entity-search collection. */
+    private get entitySearchEnabled();
+    /**
+     * Compact catalog (type + gloss) for the system prompt. This is what lets the
+     * LLM NAME a concept instead of guessing at our vocabulary — it can infer
+     * that a question is about unpaid money, but not that we call that
+     * `outstanding-balance`. Full `resolution` detail is fetched by the tool.
+     *
+     * Deliberately NOT cached in-process. The token cost is already handled by
+     * prompt caching (the system prompt carries a cache_control breakpoint), and
+     * re-reading a tiny PK-ordered table costs ~nothing. Fetching fresh does not
+     * weaken the prompt cache either: unchanged data renders byte-identical text,
+     * so the prefix still hits. A process cache would only buy a staleness window
+     * in which a newly indexed entity is invisible to the agent.
+     */
+    private loadEntityCatalog;
+    /**
+     * Never throws: a resolver outage must degrade to "nothing resolved" and
+     * leave the agent exactly as it was, not take the whole answer down.
+     */
+    private handleResolveEntities;
+    private preResolveSchema;
     /**
      * Execute a direct tool — call fn() with LLM-provided params, no SourceAgent.
      */
@@ -1672,7 +2743,45 @@ interface LLMOptions {
     temperature?: number;
     topP?: number;
     apiKey?: string;
+    baseURL?: string;
     partial?: (chunk: string) => void;
+    /**
+     * Per-request cancellation. When the caller aborts this signal (user hit
+     * "Stop"), the underlying provider request is cancelled and the call throws
+     * a RequestAbortedError. Threaded into the provider `messages.create` request
+     * options and checked between tool-loop iterations. Currently honored on the
+     * Anthropic path (the agent flow's default provider).
+     */
+    signal?: AbortSignal;
+    /**
+     * Forces a tool call on the FIRST iteration of streamWithTools only
+     * (subsequent iterations revert to auto). Used by fork mode to stop the
+     * model from ending its turn with a bare "I'll adapt the script…" preamble
+     * and zero tool calls. `{ type: 'any' }` lets the model pick which tool
+     * (write_script in the common case, a source tool for schema discovery);
+     * `{ type: 'tool', name }` pins a specific tool. Honored on both the
+     * Anthropic path and the OpenAI/OpenRouter path (mapped to OpenAI's
+     * tool_choice: 'required' / a named function).
+     */
+    firstIterationToolChoice?: {
+        type: 'any';
+    } | {
+        type: 'tool';
+        name: string;
+    };
+    /**
+     * Internal — set only by the OpenRouter wrappers when the target is a Claude
+     * model. Tells the OpenAI-wire path to emit Anthropic `cache_control`
+     * breakpoints (OpenRouter forwards them to Anthropic for prompt caching).
+     * Never set for direct OpenAI/Groq calls, so their requests are unchanged.
+     */
+    _openrouterClaudeCaching?: boolean;
+    /**
+     * Internal — OpenRouter provider-routing preferences (forwarded as the
+     * `provider` body field). Set by the OpenRouter wrappers to steer routing to
+     * a fast backend (e.g. {sort:'throughput'}). Never set for direct OpenAI/Groq.
+     */
+    _openrouterProvider?: Record<string, unknown>;
 }
 interface Tool {
     name: string;
@@ -1694,6 +2803,16 @@ declare class LLM {
      * @returns Normalized system prompt for Anthropic API
      */
     private static _normalizeSystemPrompt;
+    /**
+     * Strip unpaired UTF-16 surrogates from every text field of a message set.
+     *
+     * A lone surrogate (from mid-pair string slicing or corrupt source data)
+     * serializes to a bare `\udXXX` escape that strict JSON parsers — including
+     * the one on Anthropic's API — reject with "no low surrogate in string",
+     * failing the whole request. Sanitizing here, at the single boundary every
+     * provider call flows through, guarantees no request can carry one.
+     */
+    private static _sanitizeMessages;
     /**
      * Log cache usage metrics from Anthropic API response
      * Shows cache hits, costs, and savings
@@ -1710,11 +2829,46 @@ declare class LLM {
      * "claude-sonnet-4-5" → ["anthropic", "claude-sonnet-4-5"] (default)
      */
     private static _parseModel;
+    /**
+     * Map an Anthropic model id (e.g. "claude-sonnet-4-5-20250929") to the OpenRouter slug
+     * (e.g. "claude-sonnet-4.5"). OpenRouter slugs drop the date suffix and use dotted versions.
+     */
+    private static _toOpenRouterSlug;
+    /**
+     * Per-provider proxy base URL. Returns `${SUPERATOM_LLM_PROXY_URL}/<provider>`
+     * when our Cloudflare LLM proxy is configured, else undefined (→ talk to the
+     * provider directly, legacy behaviour). An explicit options.baseURL (e.g.
+     * OpenRouter) always wins and is never overridden. See backend/docs/llm-proxy.md.
+     */
+    private static _proxyBaseURL;
+    private static _openrouterOptions;
+    private static _isRetryableProviderError;
+    private static _withOpenrouterRetry;
+    private static _openrouterText;
+    private static _openrouterStream;
+    private static _openrouterStreamWithTools;
+    /**
+     * Build an Anthropic client. Routes through our Cloudflare LLM proxy when
+     * SUPERATOM_LLM_PROXY_URL is set (each client ships a per-client proxy key as
+     * ANTHROPIC_API_KEY and never holds the real key); otherwise talks to
+     * api.anthropic.com directly. See backend/docs/llm-proxy.md.
+     */
+    private static _anthropicClient;
+    /** True when OpenRouter is configured as a fail-open fallback for Claude. */
+    private static _openrouterAvailable;
+    /** Remap an Anthropic model id to the OpenRouter model path for fail-open. */
+    private static _anthropicFallbackModel;
     private static _anthropicText;
     private static _anthropicStream;
     private static _anthropicStreamWithTools;
     private static _groqText;
     private static _groqStream;
+    /**
+     * Gemini request options carrying the proxy base URL, or undefined → talk to
+     * generativelanguage.googleapis.com directly. The Google SDK takes baseUrl as a
+     * per-model request option, not a constructor arg. See backend/docs/llm-proxy.md.
+     */
+    private static _geminiRequestOptions;
     private static _geminiText;
     private static _geminiStream;
     /**
@@ -1723,8 +2877,28 @@ declare class LLM {
      */
     private static _cleanSchemaForGemini;
     private static _geminiStreamWithTools;
+    /** True for Anthropic/Claude model ids — gates OpenRouter prompt caching. */
+    private static _isClaudeModel;
+    /**
+     * Build the OpenAI-wire system message. For OpenRouter + Claude
+     * (cacheClaude=true) it emits content parts carrying Anthropic
+     * `cache_control` breakpoints (preserving any the caller set, else marking
+     * the last block), so OpenRouter forwards them to Anthropic for prompt
+     * caching. Otherwise it returns a plain flattened string — unchanged for
+     * direct OpenAI/Groq.
+     */
+    private static _openaiSystemMessage;
+    /**
+     * Split an OpenAI-wire usage object. `prompt_tokens` INCLUDES cached tokens,
+     * so we subtract them out (Anthropic-style: input excludes cache reads) and
+     * report cached separately — this makes calculateCost price cache reads at
+     * the discounted rate and reflects OpenRouter prompt-cache savings in logs.
+     */
+    private static _openaiUsage;
     private static _openaiText;
     private static _openaiStream;
+    /** Map the Anthropic-style firstIterationToolChoice to OpenAI's tool_choice. */
+    private static _openaiToolChoice;
     private static _openaiStreamWithTools;
     /**
      * Parse JSON string, handling markdown code blocks and surrounding text
@@ -1826,6 +3000,13 @@ declare class UIBlock {
     private textResponse;
     private actions;
     private createdAt;
+    /**
+     * Which script recipe produced this answer, when a script did. Read on the
+     * NEXT turn so the user can say "use mode instead" and have the matcher
+     * resolve it to a concrete script (the `edit` tier is unreachable without it).
+     * See backend/docs/SCRIPT-EDIT-DESIGN.md § Component 1.
+     */
+    private scriptBinding;
     /**
      * Creates a new UIBlock instance
      * @param userQuestion - The user's question or input
@@ -1920,6 +3101,14 @@ declare class UIBlock {
     /**
      * Get creation timestamp
      */
+    /**
+     * Bind this block to the script recipe that produced its answer.
+     */
+    setScriptBinding(binding: Record<string, any> | null): void;
+    /**
+     * The script recipe bound to this block, if any.
+     */
+    getScriptBinding(): Record<string, any> | null;
     getCreatedAt(): Date;
     /**
      * Convert UIBlock to JSON-serializable object
@@ -1987,6 +3176,32 @@ declare class Thread {
      * @param currentUIBlockId - ID of current UIBlock to exclude from context (optional)
      * @returns Formatted conversation history string
      */
+    /**
+     * The script recipe bound to the most recent completed UIBlock — i.e. the
+     * script behind the answer the user is currently looking at. Drives the
+     * matcher's `edit` tier: without it, "use mode instead" has no target and
+     * falls through to regeneration.
+     * See backend/docs/SCRIPT-EDIT-DESIGN.md § Component 1.
+     */
+    getActiveScriptBinding(currentUIBlockId?: string): Record<string, any> | null;
+    /**
+     * The recent script-backed answers in this thread, newest first — the
+     * candidate set a user-directed edit can target.
+     *
+     * Returning several rather than only the newest is what lets an instruction
+     * name its own target ("use the mode for the WSP one"). With a single
+     * candidate every edit lands on the most recent script, which silently edits
+     * the wrong recipe whenever the user meant an earlier one.
+     *
+     * Deduped by recipeId (newest occurrence wins) so a long editing session on
+     * one script doesn't crowd out the others. Each entry carries the question
+     * that produced it — without that the candidates are indistinguishable.
+     *
+     * In-memory only: dies with the process. The caller falls back to the
+     * persisted bindings when this comes back empty.
+     * See backend/docs/SCRIPT-EDIT-DESIGN.md § Component 1.
+     */
+    getScriptBindings(limit?: number, currentUIBlockId?: string): Record<string, any>[];
     getConversationContext(limit?: number, currentUIBlockId?: string): string;
     /**
      * Convert Thread to JSON-serializable object
@@ -2649,7 +3864,7 @@ declare abstract class BaseLLM {
      * This helps provide intelligent suggestions for follow-up queries
      * For general/conversational questions without components, pass textResponse instead
      */
-    generateNextQuestions(originalUserPrompt: string, component?: Component | null, componentData?: Record<string, unknown>, apiKey?: string, conversationHistory?: string, textResponse?: string): Promise<string[]>;
+    generateNextQuestions(originalUserPrompt: string, component?: Component | null, componentData?: Record<string, unknown>, apiKey?: string, conversationHistory?: string, textResponse?: string, signal?: AbortSignal): Promise<string[]>;
 }
 
 interface AnthropicLLMConfig extends BaseLLMConfig {
@@ -2712,7 +3927,8 @@ declare const openaiLLM: OpenAILLM;
  * Query Cache — Two mechanisms:
  *
  * 1. `cache` (query string → result data) — TTL-based with max size, for avoiding re-execution
- *    of recently validated queries. LRU eviction when max size exceeded.
+ *    of recently validated queries. True LRU eviction: reads bubble entries to the back via
+ *    delete+re-set so the oldest *unused* entry is evicted, not the oldest *inserted*.
  *
  * 2. Encrypted queryId tokens — SQL is encrypted into the queryId itself (self-contained).
  *    No server-side storage needed for SQL mappings. The token is decrypted on each request.
@@ -2739,11 +3955,16 @@ declare class QueryCache {
      */
     getTTL(): number;
     /**
-     * Store query result in data cache
+     * Store query result in data cache.
+     * If the key already exists, it's removed first so the re-insert places it
+     * at the back of the iteration order (LRU). Eviction only fires when adding
+     * a genuinely new key past the size limit.
      */
     set(query: string, data: any): void;
     /**
-     * Get cached result if exists and not expired
+     * Get cached result if exists and not expired.
+     * On hit, re-inserts the entry so it moves to the back of the Map's
+     * iteration order — turning FIFO eviction into true LRU.
      */
     get(query: string): any | null;
     /**
@@ -2849,6 +4070,187 @@ declare class DashboardConversationHistory {
 }
 declare const dashboardConversationHistory: DashboardConversationHistory;
 
+/**
+ * Whole-dashboard generation via Pi, a terminal coding agent — as opposed to
+ * DASH_COMP_REQ's single-widget-at-a-time flow. Runs Pi in-process via its
+ * SDK (createAgentSession), not as a subprocess: no shell, no argument
+ * quoting, no stdin/stdout piping, none of the Windows-specific subprocess
+ * issues that came with spawning the `pi` CLI directly.
+ *
+ * Called from sdk-nodejs/src/dashboardAgent/index.ts (DASHBOARD_AGENT_REQ),
+ * which owns the generic streaming/abort machinery (mirrors USER_PROMPT_REQ)
+ * and passes `signal`/`onProgress` alongside the normal params — this stays
+ * within CollectionHandler's loose (params) => Promise<result> typing, no
+ * change needed to that shared type.
+ *
+ * This mechanism is generic and reusable across any deployment. What's
+ * genuinely project-specific — where AGENTS.md lives, which model to use —
+ * is supplied via `DashboardAgentCollectionConfig`, with defaults sensible
+ * enough that most callers don't need to override them (see below). The
+ * data-source tool list and the dashboard's current state both come from
+ * things sdk-nodejs already exposes generically: `sdk.getTools()` (whatever
+ * this deployment registered via `sdk.setTools()`) and `sdk.callCollection
+ * ('dashboards', 'query', ...)` (whatever this deployment already registered
+ * under that name/shape) — no per-deployment callback needed for either.
+ * Same convention on the way out: after a successful run, the prompt and the
+ * full response text are handed to `sdk.callCollection('dashboard-agent-
+ * conversations', 'create', ...)` if this deployment has registered one —
+ * skipped silently otherwise, since conversation history is optional.
+ *
+ * Pi verifies every query against the live database itself (via whatever
+ * local tool-execution bridge the deployment exposes, e.g. an HTTP bridge
+ * on localhost), but does NOT persist the result itself — it writes the
+ * finished DSL to an absolute path inside `runtimeDir`, told to it explicitly
+ * in the prompt, and stops there. This handler reads that file after the run
+ * finishes and returns its content as `dashboard` in the result. The caller
+ * (frontend) is the one that actually saves it, via whatever authenticated
+ * create/update path any other dashboard edit goes through — Pi has no user
+ * session/auth context of its own, so persistence shouldn't happen from
+ * inside it.
+ *
+ * Session persistence: the FIRST call for a dashboardId pays the full cost
+ * (explore KB, discover schema, plan, verify, build). Every call after that
+ * resumes the same session file (SessionManager.open) so Pi has everything
+ * it already learned — it only needs to reason about the new, smaller ask,
+ * not rediscover the whole dashboard from scratch. The session's file path
+ * (AgentSession.sessionFile) is captured right after creation and persisted
+ * in a small local file, keyed by dashboardId, inside `runtimeDir`.
+ */
+interface DashboardAgentCollectionConfig {
+    /**
+     * Working directory Pi runs from — must contain AGENTS.md. This is a
+     * version-controlled prompt file, so `cwd` is expected to live somewhere
+     * like a `.prompts/` folder alongside the deployment's other prompts.
+     * Default: `<process.cwd()>/.prompts/dashboard-agent` — the same
+     * process.cwd()-based convention PromptLoader already uses for the main
+     * agent's prompts, which needs no explicit override in the common case
+     * (the backend process's own cwd already is its project root).
+     */
+    cwd?: string;
+    /**
+     * Where drafts/, the session-id map, and dashboard.log get written —
+     * separate from `cwd` deliberately, so this deployment's runtime state
+     * (regenerated per session, safe to gitignore) doesn't sit inside the
+     * same folder as the version-controlled AGENTS.md prompt.
+     * Default: `<process.cwd()>/.pi-dashboard-agent-runtime`.
+     */
+    runtimeDir?: string;
+    /** Model provider (default: process.env.PI_AGENT_PROVIDER || 'openrouter'). */
+    provider?: string;
+    /** Model id (default: process.env.PI_AGENT_MODEL || 'anthropic/claude-sonnet-4.5'). */
+    model?: string;
+    /**
+     * true (default): every request starts a brand-new pi session, with the 2
+     * most recent prior responses (if any) injected into the prompt as
+     * context — bounded cost per request, but pi re-explores schema/KB facts
+     * it already verified in an earlier turn on this same dashboard.
+     * false: resumes the same session file across requests on a given
+     * dashboard — pi keeps everything it already learned, but context (and
+     * cost) grows unbounded across turns (one observed turn: 2M+ cache-read
+     * tokens after a handful of edits on the same dashboard).
+     * Default: process.env.PI_AGENT_FRESH_SESSION !== 'false'.
+     */
+    freshSession?: boolean;
+}
+declare function registerDashboardAgentCollection(sdk: SuperatomSDK, config?: DashboardAgentCollectionConfig): void;
+
+/**
+ * ScriptMatcher — LLM-Based Script Matching + Parameter Extraction
+ *
+ * Uses ONE LLM call to:
+ * 1. Pick the best matching script from the library (or "none")
+ * 2. Extract parameter values from the user question
+ *
+ * Why LLM over embeddings:
+ * - Embeddings capture topic similarity ("overstock" ≈ "inventory" ≈ "revenue")
+ *   but can't distinguish structurally different questions about the same domain
+ * - LLM understands that "overstock by warehouse" needs a different script than
+ *   "revenue by warehouse" even though they're semantically close
+ * - One call does both matching AND parameter extraction
+ *
+ * When script library grows past ~50, add an embedding pre-filter
+ * (ChromaDB narrows to top 10 → LLM picks from those 10).
+ */
+
+declare class ScriptMatcher {
+    private store;
+    constructor(store: ScriptStore);
+    /**
+     * Find the best matching script for a user question.
+     * Uses ONE LLM call that picks the script AND extracts parameters.
+     * Returns null if no script matches.
+     */
+    match(userPrompt: string, apiKey?: string, model?: string, signal?: AbortSignal, 
+    /**
+     * Recent script-backed answers in this thread, newest first. Presence of at
+     * least one is what makes the `edit` tier reachable at all (see the guards
+     * below), and handing over SEVERAL is what lets an instruction name its own
+     * target instead of always hitting the most recent script.
+     */
+    activeBindings?: ScriptBinding[], 
+    /**
+     * Recent conversation turns, most recent last. The gate needs this to tell
+     * "this reply is an edit instruction for the active script" apart from
+     * "this reply is answering a clarifying question about a DIFFERENT,
+     * unresolved topic that never became a script" — a bare parameter-shaped
+     * reply ("April 2025 to March 2026") is textually indistinguishable
+     * between those two cases without seeing what the assistant just asked.
+     * Once this tier is decided, nothing downstream re-checks it — the edit
+     * path hands MainAgent a system prompt that explicitly instructs it to
+     * trust the premise and not treat the turn as a new question — so this
+     * gate is the only place that can catch a mismatch.
+     */
+    conversationHistory?: string): Promise<ScriptMatch | null>;
+    /**
+     * Build the script catalog string for the LLM prompt.
+     * Each script gets: index, ID, name, description, and parameter definitions.
+     */
+    private buildScriptCatalog;
+    /**
+     * The recent script-backed answers in this thread — the bounded set an edit
+     * may target. Rendered as its own prompt section (never merged into the
+     * ranked catalog) so the `edit` rules have an unambiguous referent set, and
+     * numbered newest-first so the prompt's "prefer the most recent when the
+     * instruction is ambiguous" tie-break has something to point at.
+     *
+     * Each entry carries the QUESTION that produced it plus the columns it
+     * returned — that is what lets the matcher resolve "use the mode for the WSP
+     * one" instead of blindly taking the newest.
+     */
+    private buildActiveScriptBlock;
+}
+
+/**
+ * ScriptRunner — Execute scripts in an isolated tsx subprocess.
+ *
+ * The subprocess approach replaces the earlier `new Function()` eval and gives us:
+ *   - Real sandbox (separate process, SIGKILL on timeout).
+ *   - Real TypeScript (tsx transpiles on the fly).
+ *   - npm imports available to scripts (clustering, stats, geo, etc.).
+ *
+ * Protocol: NDJSON over the child's stdin/stdout. See script-ipc.ts + backend/docs/SCRIPT-FLOW-IMPLEMENTATION.md.
+ */
+
+interface RunScriptOptions {
+    /** Data sources the script is allowed to query via ctx.query */
+    externalTools: ExternalTool[];
+    /** Optional — for propagating per-query UI progress to the user */
+    streamBuffer?: StreamBuffer;
+    /** Override the wall-clock timeout (default `SCRIPT_TIMEOUT_MS`, 60s). */
+    timeoutMs?: number;
+    /**
+     * Per-turn cancellation signal. When the user hits "Stop" mid-run, the child
+     * process group is SIGKILLed and the run resolves as an aborted failure (the
+     * caller is already unwinding, so the result is discarded).
+     */
+    signal?: AbortSignal;
+}
+/**
+ * Execute a recipe by spawning a tsx child on the script's .ts file.
+ * `scriptPath` is the absolute path to the saved `.ts` body.
+ */
+declare function runScript(recipe: ScriptRecipe, scriptPath: string, params: Record<string, any>, options: RunScriptOptions): Promise<ScriptResult>;
+
 type MessageTypeHandler = (message: IncomingMessage) => void | Promise<void>;
 declare class SuperatomSDK {
     private ws;
@@ -2884,6 +4286,8 @@ declare class SuperatomSDK {
     private lastPong;
     private readonly PING_INTERVAL_MS;
     private readonly PONG_TIMEOUT_MS;
+    private pendingOutbox;
+    private readonly MAX_OUTBOX_SIZE;
     constructor(config: SuperatomSDKConfig);
     /**
      * Initialize PromptLoader and load prompts into memory
@@ -2928,6 +4332,20 @@ declare class SuperatomSDK {
      * Does NOT throw on closed connections — callers can check the return value if needed.
      */
     send(message: Message): boolean;
+    /**
+     * Queue a message that couldn't be delivered because the socket was down,
+     * to be resent once it reconnects. Drops the oldest entry once the bound is
+     * hit — an outage long enough to fill this queue means the oldest queued
+     * responses are for requests the caller has likely already given up on.
+     */
+    private queuePendingMessage;
+    /**
+     * Resend everything queued while the socket was down, now that it's back
+     * up. Uses this.ws.send() directly (not send()) so a message that fails
+     * again goes back through queuePendingMessage() rather than being silently
+     * dropped a second time.
+     */
+    private flushPendingOutbox;
     /**
      * Register a message handler to receive all messages
      */
@@ -2967,6 +4385,14 @@ declare class SuperatomSDK {
      */
     private handlePong;
     private storeComponents;
+    /**
+     * The live, frontend-registered component catalog (name, type, description,
+     * and full prop schema per component) — the same authoritative source
+     * DASH_COMP_REQ's LLM prompt is built from. Exposed so other integrations
+     * (e.g. the dashboard-agent script bridge) can read real component
+     * contracts instead of maintaining a separate, driftable hand-written copy.
+     */
+    getComponents(): Component[];
     /**
      * Set tools for the SDK instance
      */
@@ -2975,6 +4401,16 @@ declare class SuperatomSDK {
      * Get the stored tools
      */
     getTools(): Tool$1[];
+    /**
+     * Call a registered collection operation in-process — no WebSocket
+     * round-trip, since the caller is already running inside this same SDK
+     * instance. Lets SDK-internal features (e.g. the dashboard agent) reuse
+     * whatever collection a deployment has already registered (e.g.
+     * 'dashboards'.'query') by name/convention, instead of requiring a
+     * separate callback purely to re-expose data a collection already serves.
+     * Throws if the collection or operation isn't registered.
+     */
+    callCollection<TResult = any>(collectionName: string, operation: string, params?: any): Promise<TResult>;
     /**
      * Register workflow components for the SDK instance.
      *
@@ -3018,4 +4454,4 @@ declare class SuperatomSDK {
     getConversationSimilarityThreshold(): number;
 }
 
-export { type Action, type AgentConfig, type AgentResponse, BM25L, type BM25LOptions, type BaseLLMConfig, CONTEXT_CONFIG, type CapturedLog, CleanupService, type CollectionHandler, type CollectionOperation, type DBUIBlock, DEFAULT_AGENT_CONFIG, type DatabaseType, type HybridSearchOptions, type IncomingMessage, type KbNodesQueryFilters, type KbNodesRequestPayload, LLM, type LLMUsageEntry, type LogLevel, MainAgent, type Message, type ModelStrategy, type OutputField, type RerankedResult, STORAGE_CONFIG, type SelectedWorkflow, SuperatomSDK, type SuperatomSDKConfig, type TaskType, Thread, ThreadManager, type Tool$1 as Tool, type ToolOutputSchema, UIBlock, UILogCollector, type User, UserManager, type UsersData, type WorkflowDescriptor, anthropicLLM, dashboardConversationHistory, geminiLLM, groqLLM, hybridRerank, llmUsageLogger, logger, openaiLLM, queryCache, rerankChromaResults, rerankConversationResults, userPromptErrorLogger };
+export { type Action, type AgentConfig, type AgentResponse, BM25L, type BM25LOptions, type BaseLLMConfig, CONTEXT_CONFIG, type CapturedLog, CleanupService, type CollectionHandler, type CollectionOperation, type DBUIBlock, DEFAULT_AGENT_CONFIG, type DashboardAgentCollectionConfig, type DatabaseType, type HybridSearchOptions, type IncomingMessage, type KbNodesQueryFilters, type KbNodesRequestPayload, LLM, type LLMUsageEntry, type LogLevel, MainAgent, type Message, type ModelStrategy, type OutputField, type RerankedResult, STORAGE_CONFIG, type ScriptComponentSpec, ScriptMatcher, type ScriptParameter, type ScriptRecipe, type ScriptRecipeMetaRow, type ScriptRecipeStore, type ScriptResult, ScriptStore, type ScriptStoreOptions, type SelectedWorkflow, SuperatomSDK, type SuperatomSDKConfig, type TaskType, Thread, ThreadManager, type Tool$1 as Tool, type ToolOutputSchema, UIBlock, UILogCollector, type User, UserManager, type UsersData, type WorkflowDescriptor, anthropicLLM, dashboardConversationHistory, geminiLLM, groqLLM, hybridRerank, llmUsageLogger, logger, normalizeScriptBody, openaiLLM, queryCache, registerDashboardAgentCollection, rerankChromaResults, rerankConversationResults, resolveScriptRecipeStore, runScript, userPromptErrorLogger };