@cognipeer/agent-sdk 0.8.6 → 0.8.7

package/dist/index.cjs

@@ -420,7 +420,8 @@ function createTool({
   approvalDefaults,
   maxExecutionsPerRun,
   cache,
-  retry
+  retry,
+  retention
 }) {
   const execute = async (input) => func(input);
   const toolRecord = {
@@ -451,6 +452,9 @@ function createTool({
   if (retry !== void 0) {
     toolRecord.retry = retry;
   }
+  if (retention !== void 0) {
+    toolRecord.retention = retention;
+  }
   toolRecord.__source = toolRecord.__source || "smart";
   toolRecord.__impl = func;
   return toolRecord;
@@ -808,28 +812,6 @@ function normalizeUsage(raw) {
   };
   return normalized;
 }
-function recordUsage(state, modelName, normalized) {
-  const usageState = state.usage || { perRequest: [], totals: {} };
-  const id = `${Date.now()}_${Math.random().toString(36).slice(2, 8)}`;
-  const cachedInput = normalized.prompt_tokens_details?.cached_tokens;
-  usageState.perRequest.push({
-    id,
-    modelName,
-    usage: normalized,
-    timestamp: (/* @__PURE__ */ new Date()).toISOString(),
-    turn: usageState.perRequest.length + 1,
-    cachedInput
-  });
-  const previous = usageState.totals[modelName] || { input: 0, output: 0, total: 0, cachedInput: 0 };
-  usageState.totals[modelName] = {
-    input: previous.input + (Number(normalized.prompt_tokens) || 0),
-    output: previous.output + (Number(normalized.completion_tokens) || 0),
-    total: previous.total + (Number(normalized.total_tokens) || 0),
-    cachedInput: previous.cachedInput + (Number(cachedInput) || 0)
-  };
-  state.usage = usageState;
-  return id;
-}
 function num(...vals) {
   for (const v of vals) {
     if (v === 0) return 0;
@@ -1972,6 +1954,270 @@ function estimatePayloadBytes(value) {
   }
 }
 
+// src/smart/toolResponses.ts
+init_utilTokens();
+
+// src/utils/content.ts
+function contentToString(content) {
+  try {
+    if (content == null) return "";
+    if (typeof content === "string") return content;
+    if (Array.isArray(content)) {
+      return content.map((p) => {
+        if (!p || typeof p !== "object") return String(p);
+        if (p.type === "text") return String(p.text ?? "");
+        if (p.type === "image_url") {
+          const img = p.image_url;
+          const url = typeof img === "string" ? img : img?.url;
+          const detail = typeof img === "string" ? void 0 : img?.detail;
+          return `[image:${url || "unknown"}${detail ? ` detail=${detail}` : ""}]`;
+        }
+        return JSON.stringify(p);
+      }).join("\n");
+    }
+    return JSON.stringify(content);
+  } catch {
+    return String(content);
+  }
+}
+function mergeContentsToString(contents) {
+  return contents.map((c) => contentToString(c)).join("\n");
+}
+function safeStringify(value, pretty = false) {
+  if (value == null) return "";
+  if (typeof value === "string") return value;
+  try {
+    return JSON.stringify(value, null, pretty ? 2 : void 0) ?? String(value);
+  } catch {
+    return String(value);
+  }
+}
+function extractMessageText2(message) {
+  if (!message) return "";
+  const content = message.content;
+  if (typeof content === "string") return content;
+  if (Array.isArray(content)) {
+    return content.map((part) => typeof part === "string" ? part : part?.text ?? part?.content ?? "").join("");
+  }
+  return "";
+}
+
+// src/smart/toolResponses.ts
+var CONTROL_PLANE_TOOL_NAMES = [
+  "response",
+  "manage_todo_list",
+  "get_tool_response",
+  "ask_user_question",
+  "open_skill",
+  "bind_skill_tools",
+  "search_skills"
+];
+var DELEGATION_TOOL_NAMES = [
+  "delegate_to",
+  "spawn_subagent",
+  "spawn_subagents_parallel"
+];
+var INPUT_PROTECTED_TOOL_NAMES = /* @__PURE__ */ new Set([
+  ...CONTROL_PLANE_TOOL_NAMES,
+  ...DELEGATION_TOOL_NAMES
+]);
+var CONTROL_PLANE_OUTPUT_DEFAULTS = new Set(CONTROL_PLANE_TOOL_NAMES);
+var DEFAULT_MAX_TOOL_INPUT_FIELD_CHARS = 2e3;
+var DEFAULT_TOOL_INPUT_DIGEST_HEAD_CHARS = 200;
+var TOOL_INPUT_DIGEST_KEY = "__digest";
+function summarizeObject(value) {
+  if (value == null) return "null";
+  if (typeof value === "string") {
+    return value.length > 240 ? `${value.slice(0, 240)}...` : value;
+  }
+  if (Array.isArray(value)) {
+    const preview = value.slice(0, 3).map((entry) => safeStringify(entry).slice(0, 120)).join(" | ");
+    return `array(length=${value.length}) ${preview}`.trim();
+  }
+  if (typeof value === "object") {
+    const keys = Object.keys(value);
+    const preview = keys.slice(0, 6).map((key) => `${key}=${safeStringify(value[key]).slice(0, 80)}`).join("; ");
+    return `object(keys=${keys.length}) ${preview}`.trim();
+  }
+  return safeStringify(value);
+}
+function classifyToolResponse(toolName, output, config) {
+  const criticalTools = new Set(config.toolResponses.criticalTools);
+  if (criticalTools.has(toolName)) return "critical";
+  const serialized = safeStringify(output);
+  const tokenCount = countApproxTokens(serialized);
+  if (tokenCount >= config.toolResponses.maxToolResponseTokens || serialized.length >= config.toolResponses.maxToolResponseChars) {
+    return "verbose";
+  }
+  return "informative";
+}
+function applyToolResponseHardCap(toolName, output, executionId, config) {
+  const serialized = safeStringify(output);
+  const tokenCount = countApproxTokens(serialized);
+  const classification = classifyToolResponse(toolName, output, config);
+  if (classification !== "verbose") {
+    return { classification, content: serialized, rawOutput: output, tokenCount, truncated: false };
+  }
+  const cap = config.toolResponses.maxToolResponseChars;
+  const head = serialized.slice(0, cap);
+  const note = `
+... [TRUNCATED ${serialized.length - cap} chars by hard cap. Use get_tool_response with executionId="${executionId}" to fetch the full payload.]`;
+  return {
+    classification,
+    content: head + note,
+    rawOutput: output,
+    tokenCount,
+    truncated: true
+  };
+}
+function collectToolRetentionDeclarations(tools) {
+  const declarations = /* @__PURE__ */ new Map();
+  if (!Array.isArray(tools)) return declarations;
+  for (const tool of tools) {
+    const spec = tool?.retention;
+    const name = typeof tool?.name === "string" ? tool.name : "";
+    if (!name || !spec || typeof spec !== "object") continue;
+    declarations.set(name, spec);
+  }
+  return declarations;
+}
+function resolveSummarizationRetention(toolName, config, declarations) {
+  const criticalTools = new Set(config.toolResponses.criticalTools);
+  if (criticalTools.has(toolName)) return "keep_full";
+  const twoAxis = config.toolResponses.retentionByTool?.[toolName]?.output;
+  if (twoAxis) return twoAxis;
+  const legacy = config.toolResponses.toolResponseRetentionByTool[toolName];
+  if (legacy) return legacy;
+  const declared = declarations?.get(toolName)?.output;
+  if (declared) return declared;
+  if (CONTROL_PLANE_OUTPUT_DEFAULTS.has(toolName)) return "keep_full";
+  return config.toolResponses.defaultPolicy;
+}
+function resolveInputRetention(toolName, config, declarations) {
+  if (INPUT_PROTECTED_TOOL_NAMES.has(toolName)) return "keep";
+  const criticalTools = new Set(config.toolResponses.criticalTools);
+  if (criticalTools.has(toolName)) return "keep";
+  const twoAxis = config.toolResponses.retentionByTool?.[toolName]?.input;
+  if (twoAxis) return twoAxis;
+  const declared = declarations?.get(toolName)?.input;
+  if (declared) return declared;
+  return config.toolResponses.defaultInputPolicy;
+}
+function shortSha256(value) {
+  return crypto.createHash("sha256").update(value, "utf8").digest("hex").slice(0, 16);
+}
+function isDigestMarker(value) {
+  return Boolean(
+    value && typeof value === "object" && !Array.isArray(value) && TOOL_INPUT_DIGEST_KEY in value
+  );
+}
+function digestToolInputValue(value, options) {
+  const maxFieldChars = options?.maxFieldChars ?? DEFAULT_MAX_TOOL_INPUT_FIELD_CHARS;
+  const headChars = options?.headChars ?? DEFAULT_TOOL_INPUT_DIGEST_HEAD_CHARS;
+  const executionId = options?.executionId;
+  const digestedPaths = [];
+  let reclaimedChars = 0;
+  const walk = (node, path2) => {
+    if (typeof node === "string") {
+      if (node.length <= maxFieldChars) return node;
+      digestedPaths.push(path2 || "(root)");
+      const marker = {
+        chars: node.length,
+        sha256: shortSha256(node),
+        head: node.slice(0, headChars)
+      };
+      if (executionId) {
+        marker.recover = `get_tool_response executionId="${executionId}" part="input"`;
+      }
+      const replacement = { [TOOL_INPUT_DIGEST_KEY]: marker };
+      reclaimedChars += node.length - safeStringify(replacement).length;
+      return replacement;
+    }
+    if (Array.isArray(node)) {
+      let changed = false;
+      const next = node.map((entry, index) => {
+        const walked = walk(entry, path2 ? `${path2}[${index}]` : `[${index}]`);
+        if (walked !== entry) changed = true;
+        return walked;
+      });
+      return changed ? next : node;
+    }
+    if (node && typeof node === "object") {
+      if (isDigestMarker(node)) return node;
+      let changed = false;
+      const next = {};
+      for (const [key, entry] of Object.entries(node)) {
+        const walked = walk(entry, path2 ? `${path2}.${key}` : key);
+        if (walked !== entry) changed = true;
+        next[key] = walked;
+      }
+      return changed ? next : node;
+    }
+    return node;
+  };
+  const nextValue = walk(value, "");
+  return { value: nextValue, digestedPaths, reclaimedChars: Math.max(0, reclaimedChars) };
+}
+function digestToolInputArguments(argumentsJson, options) {
+  const unchanged = { arguments: argumentsJson, digestedPaths: [], reclaimedChars: 0 };
+  if (typeof argumentsJson !== "string" || argumentsJson.length === 0) return unchanged;
+  let parsed;
+  try {
+    parsed = JSON.parse(argumentsJson);
+  } catch {
+    return unchanged;
+  }
+  const digested = digestToolInputValue(parsed, options);
+  if (digested.digestedPaths.length === 0) return unchanged;
+  const serialized = safeStringify(digested.value);
+  if (serialized.length >= argumentsJson.length) return unchanged;
+  return {
+    arguments: serialized,
+    digestedPaths: digested.digestedPaths,
+    reclaimedChars: argumentsJson.length - serialized.length
+  };
+}
+function renderRetainedToolMessage(args) {
+  const { policy, rawOutput, toolName, toolCallId, executionId, prebuiltSummary } = args;
+  if (policy === "keep_full") {
+    return typeof rawOutput === "string" ? rawOutput : safeStringify(rawOutput);
+  }
+  const summary = prebuiltSummary || summarizeObject(rawOutput);
+  const refs = [
+    `toolName=${toolName}`,
+    toolCallId ? `toolCallId=${toolCallId}` : null,
+    executionId ? `executionId=${executionId}` : null
+  ].filter(Boolean).join("; ");
+  const refId = executionId || toolCallId || "";
+  if (policy === "keep_structured") {
+    return `STRUCTURED_TOOL_RESPONSE [${refs}]
+Preview: ${summary}
+Use get_tool_response with executionId "${refId}" to fetch the full payload when a specific field is needed.`;
+  }
+  if (policy === "summarize_archive") {
+    return `ARCHIVED_TOOL_RESPONSE [${refs}]
+Summary: ${summary}
+Use get_tool_response with executionId "${refId}" to fetch the full payload.`;
+  }
+  return `DROPPED_TOOL_RESPONSE [${refs}]
+Use get_tool_response with executionId "${refId}" only if you must recover the original payload.`;
+}
+function validateToolArgs(tool, args) {
+  const schema = tool.schema;
+  if (!schema || typeof schema.parse !== "function") {
+    return { ok: true, value: args };
+  }
+  try {
+    return { ok: true, value: schema.parse(args) };
+  } catch (error) {
+    if (error instanceof zod.ZodError) {
+      const message = error.issues.map((issue) => `${issue.path.join(".") || "root"}: ${issue.message}`).join("; ");
+      return { ok: false, message };
+    }
+    return { ok: false, message: error instanceof Error ? error.message : String(error) };
+  }
+}
+
 // src/smart/runtimeConfig.ts
 var DEFAULT_CRITICAL_TOOLS = ["response", "manage_todo_list", "get_tool_response"];
 var BASE_DEFAULTS = {
@@ -2022,6 +2268,11 @@ var BASE_DEFAULTS = {
     maxToolResponseTokens: 8e3,
     defaultPolicy: "summarize_archive",
     toolResponseRetentionByTool: {},
+    // Argument digesting is opt-in: the default never touches a tool's request.
+    defaultInputPolicy: "keep",
+    retentionByTool: {},
+    maxToolInputFieldChars: DEFAULT_MAX_TOOL_INPUT_FIELD_CHARS,
+    maxToolInputDigestHeadChars: DEFAULT_TOOL_INPUT_DIGEST_HEAD_CHARS,
     criticalTools: [...DEFAULT_CRITICAL_TOOLS],
     schemaValidation: "strict"
   },
@@ -2076,7 +2327,8 @@ function buildProfile(overrides) {
       ...BASE_DEFAULTS.toolResponses,
       ...overrides.toolResponses,
       criticalTools: overrides.toolResponses?.criticalTools ?? [...DEFAULT_CRITICAL_TOOLS],
-      toolResponseRetentionByTool: overrides.toolResponses?.toolResponseRetentionByTool ?? {}
+      toolResponseRetentionByTool: overrides.toolResponses?.toolResponseRetentionByTool ?? {},
+      retentionByTool: overrides.toolResponses?.retentionByTool ?? {}
     },
     contextPilot: mergeContextPilot(BASE_DEFAULTS.contextPilot, overrides.contextPilot)
   };
@@ -2230,6 +2482,14 @@ function normalizeSmartAgentOptions(opts) {
         ...customProfile.toolResponses?.toolResponseRetentionByTool || {},
         ...opts.toolResponses?.toolResponseRetentionByTool || {}
       },
+      // Two-axis map merges the same way, so a caller can layer a small override
+      // on top of a profile's table instead of restating it.
+      retentionByTool: {
+        ...preset.toolResponses.retentionByTool,
+        ...customProfile.toolResponses?.retentionByTool || {},
+        ...opts.toolResponses?.retentionByTool || {}
+      },
+      defaultInputPolicy: opts.toolResponses?.defaultInputPolicy ?? customProfile.toolResponses?.defaultInputPolicy ?? preset.toolResponses.defaultInputPolicy,
       criticalTools: opts.toolResponses?.criticalTools ?? customProfile.toolResponses?.criticalTools ?? preset.toolResponses.criticalTools
     },
     contextPilot: mergeContextPilot(mergeContextPilot(preset.contextPilot, customProfile.contextPilot), opts.contextPilot)
@@ -2414,53 +2674,6 @@ function detectVolatileContent(text, limit = 20) {
 
 // src/smart/contextPilot/jsonCrusher.ts
 init_utilTokens();
-
-// src/utils/content.ts
-function contentToString(content) {
-  try {
-    if (content == null) return "";
-    if (typeof content === "string") return content;
-    if (Array.isArray(content)) {
-      return content.map((p) => {
-        if (!p || typeof p !== "object") return String(p);
-        if (p.type === "text") return String(p.text ?? "");
-        if (p.type === "image_url") {
-          const img = p.image_url;
-          const url = typeof img === "string" ? img : img?.url;
-          const detail = typeof img === "string" ? void 0 : img?.detail;
-          return `[image:${url || "unknown"}${detail ? ` detail=${detail}` : ""}]`;
-        }
-        return JSON.stringify(p);
-      }).join("\n");
-    }
-    return JSON.stringify(content);
-  } catch {
-    return String(content);
-  }
-}
-function mergeContentsToString(contents) {
-  return contents.map((c) => contentToString(c)).join("\n");
-}
-function safeStringify(value, pretty = false) {
-  if (value == null) return "";
-  if (typeof value === "string") return value;
-  try {
-    return JSON.stringify(value, null, pretty ? 2 : void 0) ?? String(value);
-  } catch {
-    return String(value);
-  }
-}
-function extractMessageText2(message) {
-  if (!message) return "";
-  const content = message.content;
-  if (typeof content === "string") return content;
-  if (Array.isArray(content)) {
-    return content.map((part) => typeof part === "string" ? part : part?.text ?? part?.content ?? "").join("");
-  }
-  return "";
-}
-
-// src/smart/contextPilot/jsonCrusher.ts
 var CONTEXT_PILOT_JSON_MARKER_KEY = "__contextPilotDropped";
 function findCompressibleArray(output, minItems) {
   if (Array.isArray(output)) {
@@ -2968,7 +3181,24 @@ function createAgentCoreNode(opts) {
       messageList: messagesWithResponse
     });
     if (normalized) {
-      recordUsage(state, modelName, normalized);
+      const usageState = state.usage || { perRequest: [], totals: {} };
+      const id = `${Date.now()}_${Math.random().toString(36).slice(2, 8)}`;
+      const turn = usageState.perRequest.length + 1;
+      const timestamp = (/* @__PURE__ */ new Date()).toISOString();
+      const cachedInputTok = normalized.prompt_tokens_details?.cached_tokens;
+      usageState.perRequest.push({ id, modelName, usage: normalized, timestamp, turn, cachedInput: cachedInputTok });
+      const inputTok = normalized.prompt_tokens;
+      const outputTok = normalized.completion_tokens;
+      const totalTok = normalized.total_tokens;
+      const key = modelName;
+      const agg = usageState.totals[key] || { input: 0, output: 0, total: 0, cachedInput: 0 };
+      usageState.totals[key] = {
+        input: agg.input + (Number(inputTok) || 0),
+        output: agg.output + (Number(outputTok) || 0),
+        total: agg.total + (Number(totalTok) || 0),
+        cachedInput: agg.cachedInput + (Number(cachedInputTok) || 0)
+      };
+      state.usage = usageState;
     }
     return { messages: messagesWithResponse, usage: state.usage };
   };
@@ -3012,100 +3242,6 @@ function extractToolItems(output) {
   });
 }
 
-// src/smart/toolResponses.ts
-init_utilTokens();
-function summarizeObject(value) {
-  if (value == null) return "null";
-  if (typeof value === "string") {
-    return value.length > 240 ? `${value.slice(0, 240)}...` : value;
-  }
-  if (Array.isArray(value)) {
-    const preview = value.slice(0, 3).map((entry) => safeStringify(entry).slice(0, 120)).join(" | ");
-    return `array(length=${value.length}) ${preview}`.trim();
-  }
-  if (typeof value === "object") {
-    const keys = Object.keys(value);
-    const preview = keys.slice(0, 6).map((key) => `${key}=${safeStringify(value[key]).slice(0, 80)}`).join("; ");
-    return `object(keys=${keys.length}) ${preview}`.trim();
-  }
-  return safeStringify(value);
-}
-function classifyToolResponse(toolName, output, config) {
-  const criticalTools = new Set(config.toolResponses.criticalTools);
-  if (criticalTools.has(toolName)) return "critical";
-  const serialized = safeStringify(output);
-  const tokenCount = countApproxTokens(serialized);
-  if (tokenCount >= config.toolResponses.maxToolResponseTokens || serialized.length >= config.toolResponses.maxToolResponseChars) {
-    return "verbose";
-  }
-  return "informative";
-}
-function applyToolResponseHardCap(toolName, output, executionId, config) {
-  const serialized = safeStringify(output);
-  const tokenCount = countApproxTokens(serialized);
-  const classification = classifyToolResponse(toolName, output, config);
-  if (classification !== "verbose") {
-    return { classification, content: serialized, rawOutput: output, tokenCount, truncated: false };
-  }
-  const cap = config.toolResponses.maxToolResponseChars;
-  const head = serialized.slice(0, cap);
-  const note = `
-... [TRUNCATED ${serialized.length - cap} chars by hard cap. Use get_tool_response with executionId="${executionId}" to fetch the full payload.]`;
-  return {
-    classification,
-    content: head + note,
-    rawOutput: output,
-    tokenCount,
-    truncated: true
-  };
-}
-function resolveSummarizationRetention(toolName, config) {
-  const criticalTools = new Set(config.toolResponses.criticalTools);
-  if (criticalTools.has(toolName)) return "keep_full";
-  const byTool = config.toolResponses.toolResponseRetentionByTool[toolName];
-  return byTool || config.toolResponses.defaultPolicy;
-}
-function renderRetainedToolMessage(args) {
-  const { policy, rawOutput, toolName, toolCallId, executionId, prebuiltSummary } = args;
-  if (policy === "keep_full") {
-    return typeof rawOutput === "string" ? rawOutput : safeStringify(rawOutput);
-  }
-  const summary = prebuiltSummary || summarizeObject(rawOutput);
-  const refs = [
-    `toolName=${toolName}`,
-    toolCallId ? `toolCallId=${toolCallId}` : null,
-    executionId ? `executionId=${executionId}` : null
-  ].filter(Boolean).join("; ");
-  const refId = executionId || toolCallId || "";
-  if (policy === "keep_structured") {
-    return `STRUCTURED_TOOL_RESPONSE [${refs}]
-Preview: ${summary}
-Use get_tool_response with executionId "${refId}" to fetch the full payload when a specific field is needed.`;
-  }
-  if (policy === "summarize_archive") {
-    return `ARCHIVED_TOOL_RESPONSE [${refs}]
-Summary: ${summary}
-Use get_tool_response with executionId "${refId}" to fetch the full payload.`;
-  }
-  return `DROPPED_TOOL_RESPONSE [${refs}]
-Use get_tool_response with executionId "${refId}" only if you must recover the original payload.`;
-}
-function validateToolArgs(tool, args) {
-  const schema = tool.schema;
-  if (!schema || typeof schema.parse !== "function") {
-    return { ok: true, value: args };
-  }
-  try {
-    return { ok: true, value: schema.parse(args) };
-  } catch (error) {
-    if (error instanceof zod.ZodError) {
-      const message = error.issues.map((issue) => `${issue.path.join(".") || "root"}: ${issue.message}`).join("; ");
-      return { ok: false, message };
-    }
-    return { ok: false, message: error instanceof Error ? error.message : String(error) };
-  }
-}
-
 // src/utils/pendingToolCalls.ts
 function selectPendingToolCalls(messages) {
   if (!Array.isArray(messages) || messages.length === 0) return [];
@@ -5362,7 +5498,10 @@ function createAgent(opts) {
         }
       }
     }
-    if (soManager && !state.ctx?.__finalizedDueToStructuredOutput) {
+    const structuredFinalizeBlocked = Boolean(
+      pausedStage || state.ctx?.__awaitingApproval || state.ctx?.__awaitingUserQuestion || state.ctx?.__cancelled || state.ctx?.__needsSummarization || state.ctx?.__limitBreached || state.ctx?.__guardrailBlocked
+    );
+    if (soManager && !state.ctx?.__finalizedDueToStructuredOutput && !structuredFinalizeBlocked) {
       const lastForParse = state.messages[state.messages.length - 1];
       if (lastForParse?.role === "assistant") {
         const assistantText = extractMessageText2(lastForParse);
@@ -5379,10 +5518,11 @@ function createAgent(opts) {
         }
       }
     }
-    if (soManager && soManager.strategy.kind === "tool_based" && !state.ctx?.__finalizedDueToStructuredOutput) {
+    if (soManager && soManager.strategy.kind === "tool_based" && !state.ctx?.__finalizedDueToStructuredOutput && !structuredFinalizeBlocked) {
       const maxPostLoopRetries = soManager.maxRetries;
       for (let postRetry = 0; postRetry < maxPostLoopRetries; postRetry++) {
         if (state.ctx?.__finalizedDueToStructuredOutput) break;
+        if (state.ctx?.__awaitingApproval || state.ctx?.__awaitingUserQuestion || state.ctx?.__cancelled || state.ctx?.__limitBreached) break;
         const last = state.messages[state.messages.length - 1];
         const lastHasToolCalls = Array.isArray(last?.tool_calls) && last.tool_calls.length > 0;
         if (!lastHasToolCalls) {
@@ -5751,7 +5891,24 @@ function createAgentNode(opts) {
       messageList: messagesWithResponse
     });
     if (normalized) {
-      recordUsage(state, modelName, normalized);
+      const usageState = state.usage || { perRequest: [], totals: {} };
+      const id = `${Date.now()}_${Math.random().toString(36).slice(2, 8)}`;
+      const turn = usageState.perRequest.length + 1;
+      const timestamp = (/* @__PURE__ */ new Date()).toISOString();
+      const cachedInputTok = normalized.prompt_tokens_details.cached_tokens;
+      usageState.perRequest.push({ id, modelName, usage: normalized, timestamp, turn, cachedInput: cachedInputTok });
+      const inputTok = normalized.prompt_tokens;
+      const outputTok = normalized.completion_tokens;
+      const totalTok = normalized.total_tokens;
+      const key = modelName;
+      const agg = usageState.totals[key] || { input: 0, output: 0, total: 0, cachedInput: 0 };
+      usageState.totals[key] = {
+        input: agg.input + (Number(inputTok) || 0),
+        output: agg.output + (Number(outputTok) || 0),
+        total: agg.total + (Number(totalTok) || 0),
+        cachedInput: agg.cachedInput + (Number(cachedInputTok) || 0)
+      };
+      state.usage = usageState;
     }
     return { messages: messagesWithResponse, usage: state.usage };
   };
@@ -5779,7 +5936,7 @@ function collectRecentTurns(messages, lastTurnsToKeep) {
       if (turnCount >= lastTurnsToKeep) break;
     }
   }
-  if (countAssistantTurns && collected.length > 0 && collected[0].role !== "user") {
+  if (collected.length > 0) {
     const firstUserMsg = body.find((m) => m.role === "user");
     if (firstUserMsg && !collected.includes(firstUserMsg)) {
       collected.unshift(firstUserMsg);
@@ -5811,12 +5968,18 @@ function renderMemoryBlock(facts) {
     ...facts.map((fact) => `- ${fact.key}: ${fact.value}`)
   ].join("\n");
 }
+function isClampProtected(message, index, firstUserIndex) {
+  if (message.role === "system") return true;
+  if (index === firstUserIndex) return true;
+  return false;
+}
 function clampToBudget(messages, maxContextTokens) {
   let working = [...messages];
   while (working.length > 2) {
     const tokenCount = countApproxTokens(working.map(extractMessageText2).join("\n"));
     if (tokenCount <= maxContextTokens) return working;
-    const firstNonSystem = working.findIndex((message, index) => !(index === 0 && message.role === "system"));
+    const firstUserIndex = working.findIndex((message) => message.role === "user");
+    const firstNonSystem = working.findIndex((message, index) => !isClampProtected(message, index, firstUserIndex));
     if (firstNonSystem < 0) break;
     const target = working[firstNonSystem];
     if (target.role === "assistant" && Array.isArray(target.tool_calls) && target.tool_calls.length > 0) {
@@ -6195,7 +6358,6 @@ ${canonicalFacts.length > 0 ? canonicalFacts.map((fact) => `- ${fact.key}: ${fac
         outputTokensActual = normalized.completion_tokens;
         totalTokensActual = normalized.total_tokens;
         cachedInputTokens = normalized.prompt_tokens_details?.cached_tokens;
-        recordUsage(state, getModelName(model) || "unknown_model", normalized);
       }
     } catch (err) {
       durationMs = Date.now() - startTime;
@@ -6255,9 +6417,42 @@ ${canonicalFacts.length > 0 ? canonicalFacts.map((fact) => `- ${fact.key}: ${fac
         break;
       }
     }
+    const retentionDeclarations = collectToolRetentionDeclarations(state.agent?.tools);
+    const digestableToolCallIds = /* @__PURE__ */ new Set();
+    for (const m of messages) {
+      if (m.role !== "tool" || isSyntheticSummaryMessage(m)) continue;
+      const toolCallId = m.tool_call_id;
+      if (!toolCallId || !toolCallIdToAssistantIdx.has(toolCallId)) continue;
+      if (protectedToolCallIds.has(toolCallId)) continue;
+      const toolName = m.name || "";
+      if (resolveInputRetention(toolName, resolved, retentionDeclarations) !== "digest") continue;
+      digestableToolCallIds.add(toolCallId);
+    }
     const archivedExecutions = [...state.toolHistoryArchived || []];
     const archivedIds = new Set(archivedExecutions.map((entry) => entry?.executionId).filter(Boolean));
     const newMessages = messages.map((m) => {
+      if (m.role === "assistant" && Array.isArray(m.tool_calls) && m.tool_calls.length > 0) {
+        if (m.reasoning?.blocks?.length) return m;
+        let changed = false;
+        const nextToolCalls = m.tool_calls.map((tc) => {
+          const args = tc?.function?.arguments;
+          if (!tc?.id || !digestableToolCallIds.has(tc.id)) return tc;
+          if (typeof args !== "string") return tc;
+          const historyEntry = findToolHistoryEntry(state, tc.id, tc.function?.name);
+          const digested = digestToolInputArguments(args, {
+            maxFieldChars: resolved.toolResponses.maxToolInputFieldChars,
+            headChars: resolved.toolResponses.maxToolInputDigestHeadChars,
+            executionId: historyEntry?.executionId ?? tc.id
+          });
+          if (digested.digestedPaths.length === 0) return tc;
+          changed = true;
+          return {
+            ...tc,
+            function: { ...tc.function, arguments: digested.arguments }
+          };
+        });
+        return changed ? { ...m, tool_calls: nextToolCalls } : m;
+      }
       if (m.role === "tool") {
         if (isSyntheticSummaryMessage(m)) {
           return m;
@@ -6393,7 +6588,9 @@ var TOOL_RESPONSE_RECOVERY_MARKERS = [
   /STRUCTURED_TOOL_RESPONSE\s*\[/,
   /SUMMARIZED_TOOL_RESPONSE\b/,
   /DROPPED_TOOL_RESPONSE\s*\[/,
-  /Use get_tool_response with executionId/i
+  /Use get_tool_response with executionId/i,
+  // Field-level ARGUMENT digest marker (see smart/toolResponses digestToolInputValue).
+  /"__digest"\s*:/
 ];
 var todoStatusSchema = zod.z.enum(["not-started", "in-progress", "completed", "blocked"]);
 var todoWriteItemSchema = zod.z.object({
@@ -6485,6 +6682,24 @@ function flattenRecoveryContent(content) {
   }
   return "";
 }
+function flattenToolCallArguments(message) {
+  const toolCalls = message?.tool_calls;
+  if (!Array.isArray(toolCalls)) {
+    return "";
+  }
+  return toolCalls.map((call) => {
+    const args = call?.function?.arguments;
+    if (typeof args === "string") return args;
+    if (args && typeof args === "object") {
+      try {
+        return JSON.stringify(args);
+      } catch {
+        return "";
+      }
+    }
+    return "";
+  }).filter(Boolean).join("\n");
+}
 function hasToolResponseRecoveryReference(messages, executionId) {
   if (!Array.isArray(messages) || messages.length === 0) {
     return false;
@@ -6493,7 +6708,10 @@ function hasToolResponseRecoveryReference(messages, executionId) {
     if (message?.role === "system") {
       return false;
     }
-    const content = flattenRecoveryContent(message?.content);
+    const content = [
+      flattenRecoveryContent(message?.content),
+      flattenToolCallArguments(message)
+    ].filter(Boolean).join("\n");
     if (!content) {
       return false;
     }
@@ -6673,12 +6891,15 @@ function createManageTodoTool(stateRef) {
 function createGetToolResponseTool(stateRef) {
   const getTool = createTool({
     name: "get_tool_response",
-    description: "RETRIEVE the full output of a tool execution whose response was archived, dropped, truncated, or later compacted in conversation history. This tool is only useful when the visible transcript already shows a reduced tool-response marker for the referenced executionId or tool_call_id. Use it only when a marker such as 'ARCHIVED_TOOL_RESPONSE', 'STRUCTURED_TOOL_RESPONSE', 'SUMMARIZED_TOOL_RESPONSE', 'DROPPED_TOOL_RESPONSE', or an explicit truncation note points you to a missing field you still need. Do not call it to re-fetch a normal tool result that is already present in context.",
-    schema: zod.z.object({ executionId: zod.z.string().describe("Tool execution id or original tool_call_id") }),
+    description: `RETRIEVE the full input or output of a tool execution that was reduced in conversation history. This tool is only useful when the visible transcript already shows a reduced marker for the referenced executionId or tool_call_id. Use it when a marker such as 'ARCHIVED_TOOL_RESPONSE', 'STRUCTURED_TOOL_RESPONSE', 'SUMMARIZED_TOOL_RESPONSE', 'DROPPED_TOOL_RESPONSE', an explicit truncation note, or a '__digest' descriptor inside a past tool call's arguments points you to something you still need. Pass part="input" to recover arguments that were digested (their small fields are still visible; only the large payload was replaced), or part="output" (the default) to recover an archived result. Do not call it to re-fetch a normal tool result that is already present in context.`,
+    schema: zod.z.object({
+      executionId: zod.z.string().describe("Tool execution id or original tool_call_id"),
+      part: zod.z.enum(["input", "output"]).optional().describe('Which side of the exchange to recover. Defaults to "output".')
+    }),
     maxExecutionsPerRun: 8,
-    func: async ({ executionId }) => {
+    func: async ({ executionId, part }) => {
       if (!hasToolResponseRecoveryReference(stateRef.messages, executionId)) {
-        return "Execution not recoverable from the visible transcript. Use this only when the conversation already shows a reduced tool-response marker for that executionId.";
+        return "Execution not recoverable from the visible transcript. Use this only when the conversation already shows a reduced tool-response or __digest marker for that executionId.";
       }
       const matchesExecution = (t) => t?.executionId === executionId || t?.tool_call_id === executionId;
       let execution = stateRef.toolHistory?.find((t) => matchesExecution(t));
@@ -6686,6 +6907,9 @@ function createGetToolResponseTool(stateRef) {
         execution = stateRef.toolHistoryArchived?.find((t) => matchesExecution(t));
       }
       if (execution) {
+        if (part === "input") {
+          return execution.args ?? "Arguments were not recorded for this execution.";
+        }
         return execution.rawOutput || execution.output;
       }
       const ccrStore = stateRef.ctx?.__contextPilot?.ccrStore;
@@ -11025,10 +11249,14 @@ exports.BaseProvider = BaseProvider;
 exports.BedrockProvider = BedrockProvider;
 exports.CCRStore = CCRStore;
 exports.CONTEXT_PILOT_JSON_MARKER_KEY = CONTEXT_PILOT_JSON_MARKER_KEY;
+exports.CONTROL_PLANE_TOOL_NAMES = CONTROL_PLANE_TOOL_NAMES;
+exports.DEFAULT_MAX_TOOL_INPUT_FIELD_CHARS = DEFAULT_MAX_TOOL_INPUT_FIELD_CHARS;
 exports.DEFAULT_PROFILE_CONFIGS = DEFAULT_PROFILE_CONFIGS;
 exports.DEFAULT_SKILL_POLICY = DEFAULT_SKILL_POLICY;
 exports.DEFAULT_SUBAGENT_POLICY = DEFAULT_SUBAGENT_POLICY;
 exports.DEFAULT_SUBAGENT_TOOL_NAMES = DEFAULT_SUBAGENT_TOOL_NAMES;
+exports.DEFAULT_TOOL_INPUT_DIGEST_HEAD_CHARS = DEFAULT_TOOL_INPUT_DIGEST_HEAD_CHARS;
+exports.DELEGATION_TOOL_NAMES = DELEGATION_TOOL_NAMES;
 exports.DedupTracker = DedupTracker;
 exports.GuardrailPhase = GuardrailPhase;
 exports.InMemoryMemoryStore = InMemoryMemoryStore;
@@ -11038,6 +11266,7 @@ exports.OpenAIProvider = OpenAIProvider;
 exports.ProviderError = ProviderError;
 exports.SMALL_TIER_SKILL_POLICY = SMALL_TIER_SKILL_POLICY;
 exports.StructuredOutputManager = StructuredOutputManager;
+exports.TOOL_INPUT_DIGEST_KEY = TOOL_INPUT_DIGEST_KEY;
 exports.ToolBasedStrategy = ToolBasedStrategy;
 exports.VertexProvider = VertexProvider;
 exports.agentVerdictRule = agentVerdictRule;
@@ -11050,6 +11279,7 @@ exports.canSpawn = canSpawn;
 exports.captureSnapshot = captureSnapshot;
 exports.codePresenceRule = codePresenceRule;
 exports.cognipeerSink = cognipeerSink;
+exports.collectToolRetentionDeclarations = collectToolRetentionDeclarations;
 exports.composeToolSets = composeToolSets;
 exports.compressDiffOutput = compressDiffOutput;
 exports.compressJsonOutput = compressJsonOutput;
@@ -11093,6 +11323,8 @@ exports.dedupeToolsByName = dedupeToolsByName;
 exports.defaultTokenCounter = defaultTokenCounter;
 exports.detectTextFormat = detectTextFormat;
 exports.detectVolatileContent = detectVolatileContent;
+exports.digestToolInputArguments = digestToolInputArguments;
+exports.digestToolInputValue = digestToolInputValue;
 exports.emptyUsage = emptyUsage;
 exports.evaluateGuardrails = evaluateGuardrails;
 exports.extractLatestUserQuery = extractLatestUserQuery;
@@ -11120,10 +11352,12 @@ exports.readMemoryFacts = readMemoryFacts;
 exports.regexRule = regexRule;
 exports.resolveAvailableSkills = resolveAvailableSkills;
 exports.resolveAvailableSubagents = resolveAvailableSubagents;
+exports.resolveInputRetention = resolveInputRetention;
 exports.resolveMemoryStore = resolveMemoryStore;
 exports.resolveReasoning = resolveReasoning;
 exports.resolveStrategy = resolveStrategy;
 exports.resolveSubagentTools = resolveSubagentTools;
+exports.resolveSummarizationRetention = resolveSummarizationRetention;
 exports.resolveToolApprovalState = resolveToolApprovalState;
 exports.resolveUserQuestionState = resolveUserQuestionState;
 exports.restoreSnapshot = restoreSnapshot;
@@ -11134,6 +11368,7 @@ exports.seedChildMessages = seedChildMessages;
 exports.selectTopIndicesInOrder = selectTopIndicesInOrder;
 exports.setTokenCounter = setTokenCounter;
 exports.startStreamingSession = startStreamingSession;
+exports.summarizeObject = summarizeObject;
 exports.tokenize = tokenize;
 exports.traceSessionToOtlp = traceSessionToOtlp;
 exports.validateReasoningConfig = validateReasoningConfig;