@arbidocs/client 0.3.152 → 0.3.154

package/dist/index.d.cts

@@ -3454,7 +3454,7 @@ interface components {
             WEB_SEARCH_ENABLED: boolean;
             /**
              * Run Code Enabled
-             * @description Enable the run_code tool (sandboxed code execution). (Sandbox tuning lives in the developer-only RunCode config.)
+             * @description Enable the spawn_coding_agent tool (sandboxed code execution). (Sandbox tuning lives in the developer-only RunCode config.)
              * @default true
              */
             RUN_CODE_ENABLED: boolean;
@@ -3622,7 +3622,7 @@ interface components {
             AGENT_MAX_ITERATIONS: number;
             /**
              * Agent Turn Credit Budget
-             * @description Maximum AI credits (1 credit = $0.01) a turn may spend, shared across the main agent and every deep_research/run_code/sub_agent child it spawns — priced per call using each call's own model, since different stages run different models. Enforced as a meter on the shared run budget alongside tool calls and tokens, so nearing it also nudges the agent to converge (70%), blocks new sub-agent spawns (90%), and offers the pause-and-extend prompt before forcing an answer. Default is half FREE_AI_CREDITS (100) — a free account's entire grant, so one turn can't consume it all. 0 = uncapped.
+             * @description Maximum AI credits (1 credit = $0.01) a turn may spend, shared across the main agent and every deep_research/spawn_coding_agent/sub_agent child it spawns — priced per call using each call's own model, since different stages run different models. Enforced as a meter on the shared run budget alongside tool calls and tokens, so nearing it also nudges the agent to converge (70%), blocks new sub-agent spawns (90%), and offers the pause-and-extend prompt before forcing an answer. Default is half FREE_AI_CREDITS (100) — a free account's entire grant, so one turn can't consume it all. 0 = uncapped.
              * @default 50
              */
             AGENT_TURN_CREDIT_BUDGET: number;
@@ -3661,8 +3661,6 @@ interface components {
              * @description System prompt for the agent. Dynamic context is appended automatically.
              * @default Answer any question the user asks. You have access to workspace documents — use them when relevant. When no documents are relevant, answer from your own knowledge. Always provide a helpful answer.
              *
-             *     IMPORTANT: Never say "I don't have information in the available documents" or refuse to answer. If documents are not relevant, just answer the question directly from your own knowledge. If the user asks about a specific workspace document, call read_document or get_table_of_contents — never answer from memory.
-             *
              *     When using documents, retrieve actual document content — do not rely on metadata or headings alone. Cross-check information across sources. Not all assertions in sources are factual — some are arguments by a party. Attribute claims to their source where parties disagree.
              *
              *     DOCUMENT INDEX:
@@ -3681,7 +3679,7 @@ interface components {
              *     7. Never search for meta-queries like "summary", "overview", or "all documents" — search for the actual subject matter.
              *
              *     FORMATTING:
-             *     - Start your answer directly with the content — never open with a description of what you searched for or read (e.g. never write "Searching for X", "Reading Y", "I found Z documents" as the first lines of an answer).
+             *     - Start your answer directly with the content.
              *     - Use markdown: headings, bold, bullet points, numbered lists
              *     - Main conclusion first, followed by supporting points
              *     - Bullet points for lists of discrete items
@@ -3690,8 +3688,7 @@ interface components {
              *
              *     TOOL CALLING:
              *     - Your job is to gather the context the question needs, then answer. Do not give a final answer while there is relevant context you have not yet retrieved — but material already in your context counts as retrieved, and fetching it a second time is not progress.
-             *     - Always invoke tools using the tool-calling mechanism. NEVER write text that describes or simulates a tool call (e.g. never write "Searching for X...", "Reading document Y...", "Calling search_inside_documents" as prose — actually call the tool).
-             *     - If you intend to search or read, issue the tool call immediately. Do not narrate your intent first.
+             *     - Always invoke tools using the tool-calling mechanism. Say what a call is doing in its `focus` argument.
              *
              *     EFFICIENCY:
              *     1. Before each step, review the results of prior steps. Adapt your approach based on what you learned — if a call was rejected or returned poor results, adjust your strategy.
@@ -3724,7 +3721,7 @@ interface components {
             WEB_SEARCH_ENABLED: boolean;
             /**
              * Run Code Enabled
-             * @description Enable the run_code tool (sandboxed code execution). (Sandbox tuning lives in the developer-only RunCode config.)
+             * @description Enable the spawn_coding_agent tool (sandboxed code execution). (Sandbox tuning lives in the developer-only RunCode config.)
              * @default true
              */
             RUN_CODE_ENABLED: boolean;
@@ -3849,7 +3846,7 @@ interface components {
             AGENT_MAX_ITERATIONS: number;
             /**
              * Agent Turn Credit Budget
-             * @description Maximum AI credits (1 credit = $0.01) a turn may spend, shared across the main agent and every deep_research/run_code/sub_agent child it spawns — priced per call using each call's own model, since different stages run different models. Enforced as a meter on the shared run budget alongside tool calls and tokens, so nearing it also nudges the agent to converge (70%), blocks new sub-agent spawns (90%), and offers the pause-and-extend prompt before forcing an answer. Default is half FREE_AI_CREDITS (100) — a free account's entire grant, so one turn can't consume it all. 0 = uncapped.
+             * @description Maximum AI credits (1 credit = $0.01) a turn may spend, shared across the main agent and every deep_research/spawn_coding_agent/sub_agent child it spawns — priced per call using each call's own model, since different stages run different models. Enforced as a meter on the shared run budget alongside tool calls and tokens, so nearing it also nudges the agent to converge (70%), blocks new sub-agent spawns (90%), and offers the pause-and-extend prompt before forcing an answer. Default is half FREE_AI_CREDITS (100) — a free account's entire grant, so one turn can't consume it all. 0 = uncapped.
              * @default 50
              */
             AGENT_TURN_CREDIT_BUDGET: number;
@@ -3894,7 +3891,7 @@ interface components {
              *       "AGENT_MAX_RUN_TOKENS": 1500000,
              *       "AGENT_MAX_SUBAGENT_SPAWNS": 4,
              *       "AGENT_HISTORY_CHAR_THRESHOLD": 8000,
-             *       "AGENT_SYSTEM_PROMPT": "Answer any question the user asks. You have access to workspace documents — use them when relevant. When no documents are relevant, answer from your own knowledge. Always provide a helpful answer.\n\nIMPORTANT: Never say \"I don't have information in the available documents\" or refuse to answer. If documents are not relevant, just answer the question directly from your own knowledge. If the user asks about a specific workspace document, call read_document or get_table_of_contents — never answer from memory.\n\nWhen using documents, retrieve actual document content — do not rely on metadata or headings alone. Cross-check information across sources. Not all assertions in sources are factual — some are arguments by a party. Attribute claims to their source where parties disagree.\n\nDOCUMENT INDEX:\nThe \"Workspace Documents\" table in your system prompt lists every document with its Id, Title, Date, and Pages. The Subject field is a brief navigation label only — it does not contain enough information to answer questions about a document's content. Use this table directly when the user asks to list, catalogue, or tabulate documents — do NOT call any tools for that. Only use tools when you need to read or search document content.\n\nREADING TOOL RESULTS:\nEvery retrieval result opens with a CAPTION line naming the call that produced it, what it contains, and what it does not (for example \"FULL CONTENT … complete verbatim text, all 10 pages\" or \"DIGEST … verbatim page text is NOT part of this result\"). The caption also tells you when you already retrieved that material earlier in this turn. Read it before choosing your next call — it is the record of what you already hold.\n\nDOCUMENT STRATEGY, in order:\n1. Listing or cataloguing documents: use the document index already in your prompt — no tool calls needed.\n2. Getting oriented in a document you have not read: call get_table_of_contents FIRST. It is cheap, returns the heading structure with page references, and tells you which pages actually matter. Do this before any read_document on a document longer than a few pages, rather than reading blind and then discovering the structure.\n3. Specific questions across several documents: search_inside_documents to locate the passages, then read_document on the pages it points to.\n4. Reading a document: read_document with page ranges for the sections you need, or with no ranges to load the complete text. A read with no ranges returns the WHOLE document — once you have it there is nothing further to retrieve from that document. Do not read it again, do not re-read ranges out of it, and do not ask for its outline afterwards.\n5. Prefer digest=true from the outset. It is the normal retrieval mode: findings with citations, compact context. Decide against it only when the task needs the exact words — translation, quotation, close reading of style or drafting. Choosing digest=false IS that decision, and it is final for that document: digest is a way of retrieving, not a way of processing text you already hold, so once you have loaded a document in full a digest of it is refused. Reason over the text you have.\n6. Open-ended investigative queries (\"find interesting things\", \"what's notable\", \"what stands out\", \"explore X\", \"what should I know about\", \"any findings\", \"anything unusual\", \"key insights\"): ALWAYS call search_inside_documents with multiple diverse queries covering different aspects of the documents before answering. The document index is never sufficient for investigative answers — you must retrieve actual content.\n7. Never search for meta-queries like \"summary\", \"overview\", or \"all documents\" — search for the actual subject matter.\n\nFORMATTING:\n- Start your answer directly with the content — never open with a description of what you searched for or read (e.g. never write \"Searching for X\", \"Reading Y\", \"I found Z documents\" as the first lines of an answer).\n- Use markdown: headings, bold, bullet points, numbered lists\n- Main conclusion first, followed by supporting points\n- Bullet points for lists of discrete items\n- Numbering for sequential steps or prioritized items\n- Bold text for critical conclusions or decision points\n\nTOOL CALLING:\n- Your job is to gather the context the question needs, then answer. Do not give a final answer while there is relevant context you have not yet retrieved — but material already in your context counts as retrieved, and fetching it a second time is not progress.\n- Always invoke tools using the tool-calling mechanism. NEVER write text that describes or simulates a tool call (e.g. never write \"Searching for X...\", \"Reading document Y...\", \"Calling search_inside_documents\" as prose — actually call the tool).\n- If you intend to search or read, issue the tool call immediately. Do not narrate your intent first.\n\nEFFICIENCY:\n1. Before each step, review the results of prior steps. Adapt your approach based on what you learned — if a call was rejected or returned poor results, adjust your strategy.\n2. Call multiple tools in parallel when a query has several aspects.\n3. NEVER repeat a tool call with the same arguments, and never re-retrieve material you already hold — the caption on each result tells you what that is.\n4. Use targeted queries and narrow page ranges per retrieval call.\n5. Stop when you have sufficient evidence — do not over-research."
+             *       "AGENT_SYSTEM_PROMPT": "Answer any question the user asks. You have access to workspace documents — use them when relevant. When no documents are relevant, answer from your own knowledge. Always provide a helpful answer.\n\nWhen using documents, retrieve actual document content — do not rely on metadata or headings alone. Cross-check information across sources. Not all assertions in sources are factual — some are arguments by a party. Attribute claims to their source where parties disagree.\n\nDOCUMENT INDEX:\nThe \"Workspace Documents\" table in your system prompt lists every document with its Id, Title, Date, and Pages. The Subject field is a brief navigation label only — it does not contain enough information to answer questions about a document's content. Use this table directly when the user asks to list, catalogue, or tabulate documents — do NOT call any tools for that. Only use tools when you need to read or search document content.\n\nREADING TOOL RESULTS:\nEvery retrieval result opens with a CAPTION line naming the call that produced it, what it contains, and what it does not (for example \"FULL CONTENT … complete verbatim text, all 10 pages\" or \"DIGEST … verbatim page text is NOT part of this result\"). The caption also tells you when you already retrieved that material earlier in this turn. Read it before choosing your next call — it is the record of what you already hold.\n\nDOCUMENT STRATEGY, in order:\n1. Listing or cataloguing documents: use the document index already in your prompt — no tool calls needed.\n2. Getting oriented in a document you have not read: call get_table_of_contents FIRST. It is cheap, returns the heading structure with page references, and tells you which pages actually matter. Do this before any read_document on a document longer than a few pages, rather than reading blind and then discovering the structure.\n3. Specific questions across several documents: search_inside_documents to locate the passages, then read_document on the pages it points to.\n4. Reading a document: read_document with page ranges for the sections you need, or with no ranges to load the complete text. A read with no ranges returns the WHOLE document — once you have it there is nothing further to retrieve from that document. Do not read it again, do not re-read ranges out of it, and do not ask for its outline afterwards.\n5. Prefer digest=true from the outset. It is the normal retrieval mode: findings with citations, compact context. Decide against it only when the task needs the exact words — translation, quotation, close reading of style or drafting. Choosing digest=false IS that decision, and it is final for that document: digest is a way of retrieving, not a way of processing text you already hold, so once you have loaded a document in full a digest of it is refused. Reason over the text you have.\n6. Open-ended investigative queries (\"find interesting things\", \"what's notable\", \"what stands out\", \"explore X\", \"what should I know about\", \"any findings\", \"anything unusual\", \"key insights\"): ALWAYS call search_inside_documents with multiple diverse queries covering different aspects of the documents before answering. The document index is never sufficient for investigative answers — you must retrieve actual content.\n7. Never search for meta-queries like \"summary\", \"overview\", or \"all documents\" — search for the actual subject matter.\n\nFORMATTING:\n- Start your answer directly with the content.\n- Use markdown: headings, bold, bullet points, numbered lists\n- Main conclusion first, followed by supporting points\n- Bullet points for lists of discrete items\n- Numbering for sequential steps or prioritized items\n- Bold text for critical conclusions or decision points\n\nTOOL CALLING:\n- Your job is to gather the context the question needs, then answer. Do not give a final answer while there is relevant context you have not yet retrieved — but material already in your context counts as retrieved, and fetching it a second time is not progress.\n- Always invoke tools using the tool-calling mechanism. Say what a call is doing in its `focus` argument.\n\nEFFICIENCY:\n1. Before each step, review the results of prior steps. Adapt your approach based on what you learned — if a call was rejected or returned poor results, adjust your strategy.\n2. Call multiple tools in parallel when a query has several aspects.\n3. NEVER repeat a tool call with the same arguments, and never re-retrieve material you already hold — the caption on each result tells you what that is.\n4. Use targeted queries and narrow page ranges per retrieval call.\n5. Stop when you have sufficient evidence — do not over-research."
              *     }
              */
             Agents: components['schemas']['AgentsConfig'];
@@ -4339,17 +4336,17 @@ interface components {
             context?: components['schemas']['TokenBudgetContext'] | null;
             /**
              * Parent Id
-             * @description Output-item id this nests under (sub-agent steps).
+             * @description Output-item id this item nests under. Two cases, one rule — it always points at the item that OWNS this one. On a sub-agent lane header (the item carrying `subagent_status: 'running'`) it is the function_call that spawned the child, so a lane can be folded under its delegation call. On every other item carrying a `subagent` handle it is that lane's header, so a child's steps are attributable to THAT child even while other children interleave on the same stream. Absent on top-level items.
              */
             parent_id?: string | null;
             /**
              * Subagent
-             * @description Sub-agent handle for live-lane grouping.
+             * @description Sub-agent handle (a task-derived slug like 'apple-inc') this item belongs to. Stable for the child's whole life and identical on its header, its own steps and its closing marker.
              */
             subagent?: string | null;
             /**
              * Subagent Status
-             * @description Sub-agent lane lifecycle marker.
+             * @description Lane lifecycle marker: 'running' opens the lane (the header — `content` carries the child's task), 'done'/'failed' closes it (`outcome` carries the run summary, `content` the child's synthesis). Absent on the child's own intermediate steps, so `!subagent_status` selects exactly the rows to render inside an expanded lane.
              */
             subagent_status?: ('running' | 'done' | 'failed') | null;
             /**
@@ -7705,6 +7702,20 @@ interface components {
          *     The terminal message of a response is the answer; an earlier message (one
          *     followed by a function_call) is narration the model emitted alongside a tool
          *     call. Which is which is an ordering fact the client reads, not a backend gate.
+         *
+         *     **When is a message known to be the answer?** On ``output_item.done`` always
+         *     (``arbi.step == "answering"``). On ``output_item.added`` — the open, before any
+         *     text — ONLY when the backend genuinely knows, which is when the generation was
+         *     made with no tools available and so cannot resolve into a tool call: a forced
+         *     wind-down (budget/wall-clock), a goal-loop answer released whole, or the
+         *     no-AnswerEvent fallback.
+         *
+         *     On an ordinary generation it is NOT knowable at open time and ``arbi`` is
+         *     absent. That is not an oversight: the model emits its prose preamble and its
+         *     tool-call block in ONE completion, so at the first token the two are
+         *     indistinguishable, and the loop only branches on ``output.tool_calls`` once the
+         *     completion returns. A client must therefore read an absent ``arbi`` as
+         *     **unresolved**, never as "narration" — the same item may settle as the answer.
          */
         OutputMessage: {
             /**
@@ -9413,7 +9424,7 @@ interface components {
         };
         /**
          * RunCodeConfig
-         * @description run_code sandbox tuning (developer-only — image/network/limits are infra,
+         * @description spawn_coding_agent sandbox tuning (developer-only — image/network/limits are infra,
          *     never user-editable). The on/off lives on AgentsConfig.RUN_CODE_ENABLED.
          */
         RunCodeConfig: {
@@ -9425,7 +9436,7 @@ interface components {
             IMAGE: string;
             /**
              * Timeout Seconds
-             * @description Maximum execution time per run_code call (seconds).
+             * @description Maximum execution time per spawn_coding_agent call (seconds).
              * @default 60
              */
             TIMEOUT_SECONDS: number;
@@ -9442,23 +9453,6 @@ interface components {
              */
             NETWORK: string;
         };
-        /**
-         * RunCodeDetail
-         * @description Detail for a run_code tool call.
-         */
-        RunCodeDetail: {
-            /**
-             * Tool
-             * @default run_code
-             * @constant
-             */
-            tool: 'run_code';
-            /**
-             * Task Length
-             * @description Length of the task description
-             */
-            task_length: number;
-        };
         /**
          * SSESchemas
          * @description Container for all SSE event schemas — drives OpenAPI component generation.
@@ -9511,7 +9505,7 @@ interface components {
             detail_read_url?: components['schemas']['ReadUrlDetail'] | null;
             detail_create_plan?: components['schemas']['CreatePlanDetail'] | null;
             detail_save_skill?: components['schemas']['SaveSkillDetail'] | null;
-            detail_run_code?: components['schemas']['RunCodeDetail'] | null;
+            detail_spawn_coding_agent?: components['schemas']['SpawnCodingAgentDetail'] | null;
             detail_create_artifact?: components['schemas']['CreateArtifactDetail'] | null;
             detail_ask_user?: components['schemas']['AskUserDetail'] | null;
             detail_goal_progress?: components['schemas']['GoalProgressDetail'] | null;
@@ -9858,6 +9852,23 @@ interface components {
             /** Reason */
             reason: string;
         };
+        /**
+         * SpawnCodingAgentDetail
+         * @description Detail for a spawn_coding_agent tool call.
+         */
+        SpawnCodingAgentDetail: {
+            /**
+             * Tool
+             * @default spawn_coding_agent
+             * @constant
+             */
+            tool: 'spawn_coding_agent';
+            /**
+             * Task Length
+             * @description Length of the task description
+             */
+            task_length: number;
+        };
         /**
          * StepUsage
          * @description Per-step token usage from a single LLM call.
@@ -10483,7 +10494,7 @@ interface components {
             total_cost?: number | null;
             /**
              * Credits Spent
-             * @description AI credits spent so far this turn (1 credit = $0.01), priced per LLM call using each call's own model — unlike total_cost above (single-rate, main query LLM only), this is correct across a turn that mixes models (agent LLM, deep_research/run_code child LLMs, evaluator, judge). Updates live as the turn progresses, not just a final total.
+             * @description AI credits spent so far this turn (1 credit = $0.01), priced per LLM call using each call's own model — unlike total_cost above (single-rate, main query LLM only), this is correct across a turn that mixes models (agent LLM, deep_research/spawn_coding_agent child LLMs, evaluator, judge). Updates live as the turn progresses, not just a final total.
              */
             credits_spent?: number | null;
             /**

package/dist/index.d.ts

@@ -3454,7 +3454,7 @@ interface components {
             WEB_SEARCH_ENABLED: boolean;
             /**
              * Run Code Enabled
-             * @description Enable the run_code tool (sandboxed code execution). (Sandbox tuning lives in the developer-only RunCode config.)
+             * @description Enable the spawn_coding_agent tool (sandboxed code execution). (Sandbox tuning lives in the developer-only RunCode config.)
              * @default true
              */
             RUN_CODE_ENABLED: boolean;
@@ -3622,7 +3622,7 @@ interface components {
             AGENT_MAX_ITERATIONS: number;
             /**
              * Agent Turn Credit Budget
-             * @description Maximum AI credits (1 credit = $0.01) a turn may spend, shared across the main agent and every deep_research/run_code/sub_agent child it spawns — priced per call using each call's own model, since different stages run different models. Enforced as a meter on the shared run budget alongside tool calls and tokens, so nearing it also nudges the agent to converge (70%), blocks new sub-agent spawns (90%), and offers the pause-and-extend prompt before forcing an answer. Default is half FREE_AI_CREDITS (100) — a free account's entire grant, so one turn can't consume it all. 0 = uncapped.
+             * @description Maximum AI credits (1 credit = $0.01) a turn may spend, shared across the main agent and every deep_research/spawn_coding_agent/sub_agent child it spawns — priced per call using each call's own model, since different stages run different models. Enforced as a meter on the shared run budget alongside tool calls and tokens, so nearing it also nudges the agent to converge (70%), blocks new sub-agent spawns (90%), and offers the pause-and-extend prompt before forcing an answer. Default is half FREE_AI_CREDITS (100) — a free account's entire grant, so one turn can't consume it all. 0 = uncapped.
              * @default 50
              */
             AGENT_TURN_CREDIT_BUDGET: number;
@@ -3661,8 +3661,6 @@ interface components {
              * @description System prompt for the agent. Dynamic context is appended automatically.
              * @default Answer any question the user asks. You have access to workspace documents — use them when relevant. When no documents are relevant, answer from your own knowledge. Always provide a helpful answer.
              *
-             *     IMPORTANT: Never say "I don't have information in the available documents" or refuse to answer. If documents are not relevant, just answer the question directly from your own knowledge. If the user asks about a specific workspace document, call read_document or get_table_of_contents — never answer from memory.
-             *
              *     When using documents, retrieve actual document content — do not rely on metadata or headings alone. Cross-check information across sources. Not all assertions in sources are factual — some are arguments by a party. Attribute claims to their source where parties disagree.
              *
              *     DOCUMENT INDEX:
@@ -3681,7 +3679,7 @@ interface components {
              *     7. Never search for meta-queries like "summary", "overview", or "all documents" — search for the actual subject matter.
              *
              *     FORMATTING:
-             *     - Start your answer directly with the content — never open with a description of what you searched for or read (e.g. never write "Searching for X", "Reading Y", "I found Z documents" as the first lines of an answer).
+             *     - Start your answer directly with the content.
              *     - Use markdown: headings, bold, bullet points, numbered lists
              *     - Main conclusion first, followed by supporting points
              *     - Bullet points for lists of discrete items
@@ -3690,8 +3688,7 @@ interface components {
              *
              *     TOOL CALLING:
              *     - Your job is to gather the context the question needs, then answer. Do not give a final answer while there is relevant context you have not yet retrieved — but material already in your context counts as retrieved, and fetching it a second time is not progress.
-             *     - Always invoke tools using the tool-calling mechanism. NEVER write text that describes or simulates a tool call (e.g. never write "Searching for X...", "Reading document Y...", "Calling search_inside_documents" as prose — actually call the tool).
-             *     - If you intend to search or read, issue the tool call immediately. Do not narrate your intent first.
+             *     - Always invoke tools using the tool-calling mechanism. Say what a call is doing in its `focus` argument.
              *
              *     EFFICIENCY:
              *     1. Before each step, review the results of prior steps. Adapt your approach based on what you learned — if a call was rejected or returned poor results, adjust your strategy.
@@ -3724,7 +3721,7 @@ interface components {
             WEB_SEARCH_ENABLED: boolean;
             /**
              * Run Code Enabled
-             * @description Enable the run_code tool (sandboxed code execution). (Sandbox tuning lives in the developer-only RunCode config.)
+             * @description Enable the spawn_coding_agent tool (sandboxed code execution). (Sandbox tuning lives in the developer-only RunCode config.)
              * @default true
              */
             RUN_CODE_ENABLED: boolean;
@@ -3849,7 +3846,7 @@ interface components {
             AGENT_MAX_ITERATIONS: number;
             /**
              * Agent Turn Credit Budget
-             * @description Maximum AI credits (1 credit = $0.01) a turn may spend, shared across the main agent and every deep_research/run_code/sub_agent child it spawns — priced per call using each call's own model, since different stages run different models. Enforced as a meter on the shared run budget alongside tool calls and tokens, so nearing it also nudges the agent to converge (70%), blocks new sub-agent spawns (90%), and offers the pause-and-extend prompt before forcing an answer. Default is half FREE_AI_CREDITS (100) — a free account's entire grant, so one turn can't consume it all. 0 = uncapped.
+             * @description Maximum AI credits (1 credit = $0.01) a turn may spend, shared across the main agent and every deep_research/spawn_coding_agent/sub_agent child it spawns — priced per call using each call's own model, since different stages run different models. Enforced as a meter on the shared run budget alongside tool calls and tokens, so nearing it also nudges the agent to converge (70%), blocks new sub-agent spawns (90%), and offers the pause-and-extend prompt before forcing an answer. Default is half FREE_AI_CREDITS (100) — a free account's entire grant, so one turn can't consume it all. 0 = uncapped.
              * @default 50
              */
             AGENT_TURN_CREDIT_BUDGET: number;
@@ -3894,7 +3891,7 @@ interface components {
              *       "AGENT_MAX_RUN_TOKENS": 1500000,
              *       "AGENT_MAX_SUBAGENT_SPAWNS": 4,
              *       "AGENT_HISTORY_CHAR_THRESHOLD": 8000,
-             *       "AGENT_SYSTEM_PROMPT": "Answer any question the user asks. You have access to workspace documents — use them when relevant. When no documents are relevant, answer from your own knowledge. Always provide a helpful answer.\n\nIMPORTANT: Never say \"I don't have information in the available documents\" or refuse to answer. If documents are not relevant, just answer the question directly from your own knowledge. If the user asks about a specific workspace document, call read_document or get_table_of_contents — never answer from memory.\n\nWhen using documents, retrieve actual document content — do not rely on metadata or headings alone. Cross-check information across sources. Not all assertions in sources are factual — some are arguments by a party. Attribute claims to their source where parties disagree.\n\nDOCUMENT INDEX:\nThe \"Workspace Documents\" table in your system prompt lists every document with its Id, Title, Date, and Pages. The Subject field is a brief navigation label only — it does not contain enough information to answer questions about a document's content. Use this table directly when the user asks to list, catalogue, or tabulate documents — do NOT call any tools for that. Only use tools when you need to read or search document content.\n\nREADING TOOL RESULTS:\nEvery retrieval result opens with a CAPTION line naming the call that produced it, what it contains, and what it does not (for example \"FULL CONTENT … complete verbatim text, all 10 pages\" or \"DIGEST … verbatim page text is NOT part of this result\"). The caption also tells you when you already retrieved that material earlier in this turn. Read it before choosing your next call — it is the record of what you already hold.\n\nDOCUMENT STRATEGY, in order:\n1. Listing or cataloguing documents: use the document index already in your prompt — no tool calls needed.\n2. Getting oriented in a document you have not read: call get_table_of_contents FIRST. It is cheap, returns the heading structure with page references, and tells you which pages actually matter. Do this before any read_document on a document longer than a few pages, rather than reading blind and then discovering the structure.\n3. Specific questions across several documents: search_inside_documents to locate the passages, then read_document on the pages it points to.\n4. Reading a document: read_document with page ranges for the sections you need, or with no ranges to load the complete text. A read with no ranges returns the WHOLE document — once you have it there is nothing further to retrieve from that document. Do not read it again, do not re-read ranges out of it, and do not ask for its outline afterwards.\n5. Prefer digest=true from the outset. It is the normal retrieval mode: findings with citations, compact context. Decide against it only when the task needs the exact words — translation, quotation, close reading of style or drafting. Choosing digest=false IS that decision, and it is final for that document: digest is a way of retrieving, not a way of processing text you already hold, so once you have loaded a document in full a digest of it is refused. Reason over the text you have.\n6. Open-ended investigative queries (\"find interesting things\", \"what's notable\", \"what stands out\", \"explore X\", \"what should I know about\", \"any findings\", \"anything unusual\", \"key insights\"): ALWAYS call search_inside_documents with multiple diverse queries covering different aspects of the documents before answering. The document index is never sufficient for investigative answers — you must retrieve actual content.\n7. Never search for meta-queries like \"summary\", \"overview\", or \"all documents\" — search for the actual subject matter.\n\nFORMATTING:\n- Start your answer directly with the content — never open with a description of what you searched for or read (e.g. never write \"Searching for X\", \"Reading Y\", \"I found Z documents\" as the first lines of an answer).\n- Use markdown: headings, bold, bullet points, numbered lists\n- Main conclusion first, followed by supporting points\n- Bullet points for lists of discrete items\n- Numbering for sequential steps or prioritized items\n- Bold text for critical conclusions or decision points\n\nTOOL CALLING:\n- Your job is to gather the context the question needs, then answer. Do not give a final answer while there is relevant context you have not yet retrieved — but material already in your context counts as retrieved, and fetching it a second time is not progress.\n- Always invoke tools using the tool-calling mechanism. NEVER write text that describes or simulates a tool call (e.g. never write \"Searching for X...\", \"Reading document Y...\", \"Calling search_inside_documents\" as prose — actually call the tool).\n- If you intend to search or read, issue the tool call immediately. Do not narrate your intent first.\n\nEFFICIENCY:\n1. Before each step, review the results of prior steps. Adapt your approach based on what you learned — if a call was rejected or returned poor results, adjust your strategy.\n2. Call multiple tools in parallel when a query has several aspects.\n3. NEVER repeat a tool call with the same arguments, and never re-retrieve material you already hold — the caption on each result tells you what that is.\n4. Use targeted queries and narrow page ranges per retrieval call.\n5. Stop when you have sufficient evidence — do not over-research."
+             *       "AGENT_SYSTEM_PROMPT": "Answer any question the user asks. You have access to workspace documents — use them when relevant. When no documents are relevant, answer from your own knowledge. Always provide a helpful answer.\n\nWhen using documents, retrieve actual document content — do not rely on metadata or headings alone. Cross-check information across sources. Not all assertions in sources are factual — some are arguments by a party. Attribute claims to their source where parties disagree.\n\nDOCUMENT INDEX:\nThe \"Workspace Documents\" table in your system prompt lists every document with its Id, Title, Date, and Pages. The Subject field is a brief navigation label only — it does not contain enough information to answer questions about a document's content. Use this table directly when the user asks to list, catalogue, or tabulate documents — do NOT call any tools for that. Only use tools when you need to read or search document content.\n\nREADING TOOL RESULTS:\nEvery retrieval result opens with a CAPTION line naming the call that produced it, what it contains, and what it does not (for example \"FULL CONTENT … complete verbatim text, all 10 pages\" or \"DIGEST … verbatim page text is NOT part of this result\"). The caption also tells you when you already retrieved that material earlier in this turn. Read it before choosing your next call — it is the record of what you already hold.\n\nDOCUMENT STRATEGY, in order:\n1. Listing or cataloguing documents: use the document index already in your prompt — no tool calls needed.\n2. Getting oriented in a document you have not read: call get_table_of_contents FIRST. It is cheap, returns the heading structure with page references, and tells you which pages actually matter. Do this before any read_document on a document longer than a few pages, rather than reading blind and then discovering the structure.\n3. Specific questions across several documents: search_inside_documents to locate the passages, then read_document on the pages it points to.\n4. Reading a document: read_document with page ranges for the sections you need, or with no ranges to load the complete text. A read with no ranges returns the WHOLE document — once you have it there is nothing further to retrieve from that document. Do not read it again, do not re-read ranges out of it, and do not ask for its outline afterwards.\n5. Prefer digest=true from the outset. It is the normal retrieval mode: findings with citations, compact context. Decide against it only when the task needs the exact words — translation, quotation, close reading of style or drafting. Choosing digest=false IS that decision, and it is final for that document: digest is a way of retrieving, not a way of processing text you already hold, so once you have loaded a document in full a digest of it is refused. Reason over the text you have.\n6. Open-ended investigative queries (\"find interesting things\", \"what's notable\", \"what stands out\", \"explore X\", \"what should I know about\", \"any findings\", \"anything unusual\", \"key insights\"): ALWAYS call search_inside_documents with multiple diverse queries covering different aspects of the documents before answering. The document index is never sufficient for investigative answers — you must retrieve actual content.\n7. Never search for meta-queries like \"summary\", \"overview\", or \"all documents\" — search for the actual subject matter.\n\nFORMATTING:\n- Start your answer directly with the content.\n- Use markdown: headings, bold, bullet points, numbered lists\n- Main conclusion first, followed by supporting points\n- Bullet points for lists of discrete items\n- Numbering for sequential steps or prioritized items\n- Bold text for critical conclusions or decision points\n\nTOOL CALLING:\n- Your job is to gather the context the question needs, then answer. Do not give a final answer while there is relevant context you have not yet retrieved — but material already in your context counts as retrieved, and fetching it a second time is not progress.\n- Always invoke tools using the tool-calling mechanism. Say what a call is doing in its `focus` argument.\n\nEFFICIENCY:\n1. Before each step, review the results of prior steps. Adapt your approach based on what you learned — if a call was rejected or returned poor results, adjust your strategy.\n2. Call multiple tools in parallel when a query has several aspects.\n3. NEVER repeat a tool call with the same arguments, and never re-retrieve material you already hold — the caption on each result tells you what that is.\n4. Use targeted queries and narrow page ranges per retrieval call.\n5. Stop when you have sufficient evidence — do not over-research."
              *     }
              */
             Agents: components['schemas']['AgentsConfig'];
@@ -4339,17 +4336,17 @@ interface components {
             context?: components['schemas']['TokenBudgetContext'] | null;
             /**
              * Parent Id
-             * @description Output-item id this nests under (sub-agent steps).
+             * @description Output-item id this item nests under. Two cases, one rule — it always points at the item that OWNS this one. On a sub-agent lane header (the item carrying `subagent_status: 'running'`) it is the function_call that spawned the child, so a lane can be folded under its delegation call. On every other item carrying a `subagent` handle it is that lane's header, so a child's steps are attributable to THAT child even while other children interleave on the same stream. Absent on top-level items.
              */
             parent_id?: string | null;
             /**
              * Subagent
-             * @description Sub-agent handle for live-lane grouping.
+             * @description Sub-agent handle (a task-derived slug like 'apple-inc') this item belongs to. Stable for the child's whole life and identical on its header, its own steps and its closing marker.
              */
             subagent?: string | null;
             /**
              * Subagent Status
-             * @description Sub-agent lane lifecycle marker.
+             * @description Lane lifecycle marker: 'running' opens the lane (the header — `content` carries the child's task), 'done'/'failed' closes it (`outcome` carries the run summary, `content` the child's synthesis). Absent on the child's own intermediate steps, so `!subagent_status` selects exactly the rows to render inside an expanded lane.
              */
             subagent_status?: ('running' | 'done' | 'failed') | null;
             /**
@@ -7705,6 +7702,20 @@ interface components {
          *     The terminal message of a response is the answer; an earlier message (one
          *     followed by a function_call) is narration the model emitted alongside a tool
          *     call. Which is which is an ordering fact the client reads, not a backend gate.
+         *
+         *     **When is a message known to be the answer?** On ``output_item.done`` always
+         *     (``arbi.step == "answering"``). On ``output_item.added`` — the open, before any
+         *     text — ONLY when the backend genuinely knows, which is when the generation was
+         *     made with no tools available and so cannot resolve into a tool call: a forced
+         *     wind-down (budget/wall-clock), a goal-loop answer released whole, or the
+         *     no-AnswerEvent fallback.
+         *
+         *     On an ordinary generation it is NOT knowable at open time and ``arbi`` is
+         *     absent. That is not an oversight: the model emits its prose preamble and its
+         *     tool-call block in ONE completion, so at the first token the two are
+         *     indistinguishable, and the loop only branches on ``output.tool_calls`` once the
+         *     completion returns. A client must therefore read an absent ``arbi`` as
+         *     **unresolved**, never as "narration" — the same item may settle as the answer.
          */
         OutputMessage: {
             /**
@@ -9413,7 +9424,7 @@ interface components {
         };
         /**
          * RunCodeConfig
-         * @description run_code sandbox tuning (developer-only — image/network/limits are infra,
+         * @description spawn_coding_agent sandbox tuning (developer-only — image/network/limits are infra,
          *     never user-editable). The on/off lives on AgentsConfig.RUN_CODE_ENABLED.
          */
         RunCodeConfig: {
@@ -9425,7 +9436,7 @@ interface components {
             IMAGE: string;
             /**
              * Timeout Seconds
-             * @description Maximum execution time per run_code call (seconds).
+             * @description Maximum execution time per spawn_coding_agent call (seconds).
              * @default 60
              */
             TIMEOUT_SECONDS: number;
@@ -9442,23 +9453,6 @@ interface components {
              */
             NETWORK: string;
         };
-        /**
-         * RunCodeDetail
-         * @description Detail for a run_code tool call.
-         */
-        RunCodeDetail: {
-            /**
-             * Tool
-             * @default run_code
-             * @constant
-             */
-            tool: 'run_code';
-            /**
-             * Task Length
-             * @description Length of the task description
-             */
-            task_length: number;
-        };
         /**
          * SSESchemas
          * @description Container for all SSE event schemas — drives OpenAPI component generation.
@@ -9511,7 +9505,7 @@ interface components {
             detail_read_url?: components['schemas']['ReadUrlDetail'] | null;
             detail_create_plan?: components['schemas']['CreatePlanDetail'] | null;
             detail_save_skill?: components['schemas']['SaveSkillDetail'] | null;
-            detail_run_code?: components['schemas']['RunCodeDetail'] | null;
+            detail_spawn_coding_agent?: components['schemas']['SpawnCodingAgentDetail'] | null;
             detail_create_artifact?: components['schemas']['CreateArtifactDetail'] | null;
             detail_ask_user?: components['schemas']['AskUserDetail'] | null;
             detail_goal_progress?: components['schemas']['GoalProgressDetail'] | null;
@@ -9858,6 +9852,23 @@ interface components {
             /** Reason */
             reason: string;
         };
+        /**
+         * SpawnCodingAgentDetail
+         * @description Detail for a spawn_coding_agent tool call.
+         */
+        SpawnCodingAgentDetail: {
+            /**
+             * Tool
+             * @default spawn_coding_agent
+             * @constant
+             */
+            tool: 'spawn_coding_agent';
+            /**
+             * Task Length
+             * @description Length of the task description
+             */
+            task_length: number;
+        };
         /**
          * StepUsage
          * @description Per-step token usage from a single LLM call.
@@ -10483,7 +10494,7 @@ interface components {
             total_cost?: number | null;
             /**
              * Credits Spent
-             * @description AI credits spent so far this turn (1 credit = $0.01), priced per LLM call using each call's own model — unlike total_cost above (single-rate, main query LLM only), this is correct across a turn that mixes models (agent LLM, deep_research/run_code child LLMs, evaluator, judge). Updates live as the turn progresses, not just a final total.
+             * @description AI credits spent so far this turn (1 credit = $0.01), priced per LLM call using each call's own model — unlike total_cost above (single-rate, main query LLM only), this is correct across a turn that mixes models (agent LLM, deep_research/spawn_coding_agent child LLMs, evaluator, judge). Updates live as the turn progresses, not just a final total.
              */
             credits_spent?: number | null;
             /**

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@arbidocs/client",
-  "version": "0.3.152",
+  "version": "0.3.154",
   "description": "TypeScript SDK for the ARBI API — zero-knowledge auth, E2E encryption, and type-safe REST client",
   "type": "module",
   "main": "dist/index.cjs",