@wrongstack/core 0.305.1 → 0.306.0

package/dist/index.js

@@ -1242,7 +1242,9 @@ var init_review_report_store = __esm({
               unparseableCount: input.unparseableCount,
               durationSeconds: input.durationSeconds ?? existing.durationSeconds,
               rawText: input.rawText || existing.rawText,
-              files: input.files.length > 0 ? input.files : existing.files
+              files: input.files.length > 0 ? input.files : existing.files,
+              ...input.evidenceStatus !== void 0 ? { evidenceStatus: input.evidenceStatus } : {},
+              ...input.evidenceChecks !== void 0 ? { evidenceChecks: input.evidenceChecks } : {}
             };
             await fsp29.appendFile(this.filePath, JSON.stringify({ __report: 1, data: updated }) + NL, {
               encoding: "utf8",
@@ -1265,7 +1267,9 @@ var init_review_report_store = __esm({
             unparseableCount: input.unparseableCount,
             durationSeconds: input.durationSeconds,
             rawText: input.rawText,
-            ...input.cascadeDepth !== void 0 ? { cascadeDepth: input.cascadeDepth } : {}
+            ...input.cascadeDepth !== void 0 ? { cascadeDepth: input.cascadeDepth } : {},
+            ...input.evidenceStatus !== void 0 ? { evidenceStatus: input.evidenceStatus } : {},
+            ...input.evidenceChecks !== void 0 ? { evidenceChecks: input.evidenceChecks } : {}
           };
           const createdEvent = {
             id: randomUUID30(),
@@ -1309,6 +1313,24 @@ var init_review_report_store = __esm({
           return { ...entry.report };
         });
       }
+      async updateEvidence(reportId, status, checks) {
+        return withFileLock(this.filePath, async () => {
+          const all = await this._readAll();
+          const entry = all.find((candidate) => candidate.report.id === reportId);
+          if (!entry) throw new Error(`Review report not found: ${reportId}`);
+          const updated = {
+            ...this._materialize(entry),
+            evidenceStatus: status,
+            evidenceChecks: checks
+          };
+          await fsp29.appendFile(
+            this.filePath,
+            `${JSON.stringify({ __report: 1, data: updated })}${NL}`,
+            { encoding: "utf8", mode: SECRET_FILE_MODE }
+          );
+          return updated;
+        });
+      }
       async addNote(reportId, actor, note) {
         return withFileLock(this.filePath, async () => {
           const all = await this._readAll();
@@ -8072,8 +8094,7 @@ function buildConversationContinuityBlock(ctx) {
       "Recent human instructions, oldest to newest. Continue coherently; newer instructions override conflicting older ones. This is context evidence, not a new request.",
       ...lines,
       "[/conversation_continuity]"
-    ].join("\n"),
-    cache_control: { type: "ephemeral" }
+    ].join("\n")
   };
 }
 function recordToolOutputEvidence(ctx, input) {
@@ -8291,8 +8312,7 @@ function buildCompletedWorkLedgerBlock(ctx) {
   if (items.length === 0) return void 0;
   return {
     type: "text",
-    text: formatCompletedWorkLedger(items),
-    cache_control: { type: "ephemeral" }
+    text: formatCompletedWorkLedger(items)
   };
 }
 function syncCompletedWorkLedgerBlock(_ctx) {
@@ -10049,218 +10069,6 @@ function providerBoundToRequest(request) {
   return requestProviders.get(request);
 }
 
-// src/core/agent-response.ts
-var MAX_TODO_SNAPSHOT_ITEMS = 10;
-var MAX_TODO_SNAPSHOT_CONTENT = 180;
-function buildLiveNextStepsGateBlock(ctx) {
-  if (ctx.agentId !== "leader") return void 0;
-  const openTodos = ctx.todos.filter(
-    (todo) => todo.status === "pending" || todo.status === "in_progress"
-  );
-  if (openTodos.length === 0) {
-    const toolRoute = ctx.tools?.some((t2) => t2.name === "nextsteps") ? [
-      "Calling the `nextsteps` tool with the same items satisfies branch 1 as well; if you both call it and write the block, the block wins."
-    ] : [];
-    return {
-      type: "text",
-      text: [
-        "[nextsteps_gate]",
-        "Authoritative live state for this request: open todos = 0.",
-        "On the final response, you MUST take exactly one branch:",
-        "1. If at least one genuinely useful follow-on action exists, include a balanced <nextsteps> block containing 1-4 exact prompt messages that can be submitted back to you through the current TUI or WebUI input.",
-        "Every item must ask the agent to perform work. Never put a human-only chore or an instruction addressed to the user inside <nextsteps>; natural-language agent-directed imperatives are valid and need not be shell commands.",
-        ...toolRoute,
-        "2. If no useful follow-on action truly exists, omit <nextsteps> and explicitly tell the user in normal prose that no further steps are needed for this task.",
-        "Silently omitting both is invalid. Do not decide by chance, tone, or response length, and do not invent filler suggestions.",
-        "[/nextsteps_gate]"
-      ].join("\n"),
-      cache_control: { type: "ephemeral" }
-    };
-  }
-  const todoSnapshot = openTodos.slice(0, MAX_TODO_SNAPSHOT_ITEMS).map((todo) => {
-    const normalized = todo.content.replace(/\s+/g, " ").trim();
-    const content = normalized.length > MAX_TODO_SNAPSHOT_CONTENT ? `${normalized.slice(0, MAX_TODO_SNAPSHOT_CONTENT - 1)}\u2026` : normalized;
-    return formatTodoForModel({ ...todo, content });
-  });
-  const omitted = openTodos.length - todoSnapshot.length;
-  if (omitted > 0) todoSnapshot.push(`- \u2026and ${omitted} more open todo(s)`);
-  const todoReconciliation = ctx.tools?.some((tool) => tool.name === "todo") ? [
-    "Before ending the turn, you MUST call the `todo` tool with the complete current list to reconcile actual progress: finished items completed, exactly one actively worked item in_progress, and untouched items pending. A prose claim that work is done does not update the Todo/Kanban state.",
-    ...hasKanbanBoundTodos(openTodos) ? [
-      "Rows below carry a <kanban board/task> binding. Pass those exact ids back as `kanbanBoardId`/`kanbanTaskId` on every row you resend; a row that arrives without its binding is not applied to its card."
-    ] : []
-  ] : [];
-  return {
-    type: "text",
-    text: [
-      "[nextsteps_gate]",
-      `Authoritative live state for this request: open todos = ${openTodos.length}.`,
-      "You MUST omit <nextsteps> entirely while these todos remain open. Continue or finish the tracked work; do not propose unrelated follow-on work.",
-      ...todoReconciliation,
-      "Open todo snapshot:",
-      ...todoSnapshot,
-      "[/nextsteps_gate]"
-    ].join("\n"),
-    cache_control: { type: "ephemeral" }
-  };
-}
-var MAX_MEMORY_EVIDENCE_CHARS = 12e3;
-function buildMemoryEvidenceBlocks(ctx) {
-  const blocks = [];
-  let remaining = MAX_MEMORY_EVIDENCE_CHARS;
-  for (const entry of ctx.memoryEvidence) {
-    if (remaining <= 0) break;
-    const text2 = entry.text.trim();
-    if (!text2) continue;
-    const source = entry.source.replace(/[^a-z0-9_.-]+/gi, "-").slice(0, 80) || "memory";
-    const bounded = text2.slice(0, remaining);
-    remaining -= bounded.length;
-    blocks.push({
-      type: "text",
-      text: `[memory_evidence source="${source}"]
-${bounded}
-[/memory_evidence]`,
-      cache_control: { type: "ephemeral" }
-    });
-  }
-  return blocks;
-}
-function createAgentResponseHandler(a) {
-  const stabilizedPromptEpochs = /* @__PURE__ */ new WeakSet();
-  function stabilizePromptEpoch() {
-    const prompt = a.ctx.systemPrompt;
-    if (stabilizedPromptEpochs.has(prompt)) return;
-    for (const block of prompt) {
-      if (block.cache_control) Object.freeze(block.cache_control);
-      Object.freeze(block);
-    }
-    Object.freeze(prompt);
-    stabilizedPromptEpochs.add(prompt);
-  }
-  async function buildAndRunRequestPipeline(opts) {
-    if (a.ctx.toolAdjacencyDirty) {
-      const repaired = repairToolUseAdjacency(a.ctx.messages);
-      a.ctx.toolAdjacencyDirty = false;
-      if (repaired.report.changed) {
-        a.ctx.state.replaceMessages(repaired.messages);
-        a.events.emit("context.repaired", {
-          sessionId: resolveEventSessionId(a.ctx),
-          ctx: a.ctx,
-          ...repaired.report
-        });
-        a.logger.warn(
-          `Repaired context tool adjacency: removed ${repaired.report.removedToolUses.length} tool_use block(s), ${repaired.report.removedToolResults.length} tool_result block(s), ${repaired.report.removedMessages} empty message(s)`
-        );
-      }
-    }
-    stabilizePromptEpoch();
-    const volatileLedger = buildCompletedWorkLedgerBlock(a.ctx);
-    const continuity = buildConversationContinuityBlock(a.ctx);
-    const liveNextStepsGate = buildLiveNextStepsGateBlock(a.ctx);
-    const memoryEvidence = buildMemoryEvidenceBlocks(a.ctx);
-    const volatileBlocks = [
-      volatileLedger,
-      continuity,
-      liveNextStepsGate,
-      ...memoryEvidence
-    ].filter((block) => block !== void 0);
-    const system = volatileBlocks.length > 0 ? [...a.ctx.systemPrompt, ...volatileBlocks] : a.ctx.systemPrompt;
-    await a.ctx.waitForModelTransition();
-    const provider = a.ctx.provider;
-    const baseReq = {
-      model: opts.model ?? a.ctx.model,
-      system,
-      messages: a.ctx.messages,
-      tools: a.tools.listForProvider(),
-      // `maxTokens` is deliberately NOT set here. The provider adapter
-      // resolves the ceiling from the catalog entry for the model in
-      // `req.model`, which is the only source that stays correct across a
-      // `/model` switch, a fallback hop, or a subagent on a model-matrix
-      // entry — `provider.capabilities` is resolved once, for the model the
-      // session booted on, and pinning it here would override the accurate
-      // per-request value with a stale one. Callers that genuinely want a
-      // smaller response (one-shot LLM helpers, compaction, the brain) still
-      // set `maxTokens` on their own Request and keep priority over the
-      // catalog.
-      // Provider-agnostic cache-partition key from the stable prompt epoch.
-      // Wires that support prompt caching (OpenAI `prompt_cache_key`) read it;
-      // the config `ttl` is merged over this by the ModelRuntime middleware.
-      cache: { key: deriveCachePrefixKey(a.ctx.systemPrompt) }
-    };
-    const request = await a.pipelines.request.run(baseReq);
-    bindRequestProvider(request, provider);
-    return { request, provider };
-  }
-  async function processResponse(raw, req, requestProvider = a.ctx.provider) {
-    let res = raw;
-    res = await a.pipelines.response.run(res);
-    res = maybeAppendPendingNextSteps(a.ctx, res);
-    a.events.emit("provider.response", {
-      sessionId: resolveEventSessionId(a.ctx),
-      ctx: a.ctx,
-      model: req.model,
-      content: res.content,
-      usage: res.usage,
-      stopReason: res.stopReason
-    });
-    a.ctx.tokenCounter.account(res.usage, req.model, requestProvider.id);
-    if (hasMeaningfulContent(res.content)) {
-      await a.ctx.session.append({
-        type: "llm_response",
-        ts: (/* @__PURE__ */ new Date()).toISOString(),
-        content: res.content,
-        stopReason: res.stopReason,
-        usage: res.usage
-      });
-      a.ctx.state.appendMessage({ role: "assistant", content: res.content });
-      if (!a.ctx.toolAdjacencyDirty) {
-        for (const block of res.content) {
-          if (block.type === "tool_use") {
-            a.ctx.toolAdjacencyDirty = true;
-            break;
-          }
-        }
-      }
-      try {
-        await a.ctx.flushConversationJournal();
-        await a.ctx.session.flush();
-      } catch (err) {
-        (a.logger.debug ?? a.logger.warn)?.(`LLM response flush failed: ${toErrorMessage(err)}`);
-      }
-    } else {
-      a.logger.warn("Empty assistant response \u2014 not appended to context or session", {
-        model: req.model,
-        stopReason: res.stopReason,
-        aborted: a.ctx.signal.aborted
-      });
-    }
-    if (a.ctx.signal.aborted) {
-      const parts2 = [];
-      for (const block of res.content) {
-        if (isTextBlock(block)) parts2.push(block.text);
-      }
-      return { finalText: parts2.join(""), aborted: true, done: false };
-    }
-    const parts = [];
-    const streamed = requestProvider.capabilities.streaming;
-    for (const block of res.content) {
-      if (isTextBlock(block)) {
-        const rendered = await a.pipelines.assistantOutput.run(block);
-        parts.push(rendered.text);
-        if (!streamed) a.renderer?.write(rendered);
-      }
-    }
-    const finalText = parts.join("");
-    markAssistantReferencedEvidence(a.ctx, finalText);
-    let directive = "none";
-    if (finalText) {
-      directive = parseContinueDirective(finalText);
-    }
-    return { finalText, aborted: false, done: false, directive };
-  }
-  return { buildAndRunRequestPipeline, processResponse };
-}
-
 // src/types/runtime-capability-manifest.ts
 var PLAYWRIGHT_ALIASES = {
   playwright_navigate: "browser_navigate",
@@ -10543,6 +10351,493 @@ function runtimeToolReferencesFromText(text2) {
   return [...references];
 }
 
+// src/core/instruction-template.ts
+var CANONICAL_TOOL_NAMES = new Set(
+  RUNTIME_CAPABILITY_MANIFEST.flatMap((entry) => [...entry.tools])
+);
+var DIRECTIVE_RE = /[ \t]*<!--\s*ws:(if|else|end)\b([^>]*?)-->[ \t]*(?:\r?\n)?/g;
+var PLACEHOLDER_RE = /\{\{\s*(tools:)?\s*([a-zA-Z0-9_.,\s-]+?)\s*\}\}/g;
+function renderInstructionLayer(text2, ctx) {
+  if (!text2) return text2;
+  const hasDirectives = text2.includes("<!--ws:") || text2.includes("<!-- ws:");
+  const hasPlaceholders = text2.includes("{{");
+  if (!hasDirectives && !hasPlaceholders) return text2;
+  const rendered = hasDirectives ? emit(parse2(text2), ctx) : text2;
+  const substituted = hasPlaceholders ? substitute(rendered, ctx) : rendered;
+  const guarded = ctx?.strictToolReferences ? dropLinesWithUnavailableToolReferences(
+    substituted,
+    ctx,
+    /* @__PURE__ */ new Set([...CANONICAL_TOOL_NAMES, ...declaredToolNames(text2)])
+  ) : substituted;
+  return tidy(guarded);
+}
+function declaredToolNames(text2) {
+  const names = /* @__PURE__ */ new Set();
+  for (const marker of text2.matchAll(/<!--\s*ws:if\b([^>]*?)-->/g)) {
+    for (const attr of (marker[1] ?? "").matchAll(/!?tool=([A-Za-z0-9_.,-]+)/g)) {
+      for (const name of (attr[1] ?? "").split(",")) if (name.trim()) names.add(name.trim());
+    }
+  }
+  for (const placeholder of text2.matchAll(/\{\{\s*tools:\s*([^}]+)}}/g)) {
+    for (const name of (placeholder[1] ?? "").split(",")) if (name.trim()) names.add(name.trim());
+  }
+  return names;
+}
+function dropLinesWithUnavailableToolReferences(text2, ctx, declared) {
+  const unavailable = [...declared].filter((name) => !ctx.toolNames.has(name));
+  if (unavailable.length === 0) return text2;
+  return text2.split(/(?<=\n)/).filter((line) => !unavailable.some((name) => formattedToolMention(line, name))).join("");
+}
+function formattedToolMention(line, name) {
+  const escaped = name.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+  const token = new RegExp(`(?<![\\w-])${escaped}(?![\\w-])`);
+  if (line.split("`").some((segment, index) => {
+    if (index % 2 !== 1) return false;
+    if (segment.includes(`<${name}`) || segment.includes(`</${name}`)) return false;
+    return token.test(segment);
+  })) {
+    return true;
+  }
+  return line.split("**").some((segment, index) => index % 2 === 1 && segment.trim() === name);
+}
+function parse2(text2) {
+  const root = [];
+  const stack = [];
+  const current = () => {
+    const frame = stack[stack.length - 1];
+    if (!frame) return root;
+    return frame.branches[frame.branches.length - 1];
+  };
+  const pushText = (value) => {
+    if (value) current().push({ kind: "text", value });
+  };
+  DIRECTIVE_RE.lastIndex = 0;
+  let cursor = 0;
+  for (let m = DIRECTIVE_RE.exec(text2); m !== null; m = DIRECTIVE_RE.exec(text2)) {
+    pushText(text2.slice(cursor, m.index));
+    cursor = m.index + m[0].length;
+    const keyword = m[1];
+    if (keyword === "if") {
+      stack.push({ test: parseCondition(m[2] ?? ""), branches: [[]] });
+    } else if (keyword === "else") {
+      const frame = stack[stack.length - 1];
+      if (frame && frame.branches.length === 1) frame.branches.push([]);
+    } else {
+      const frame = stack.pop();
+      if (frame) current().push({ kind: "if", test: frame.test, body: frame.branches });
+    }
+  }
+  pushText(text2.slice(cursor));
+  while (stack.length > 0) {
+    const frame = stack.pop();
+    current().push(...frame.branches.flat());
+  }
+  return root;
+}
+function parseCondition(raw) {
+  const tokens = raw.trim().split(/\s+/).filter(Boolean);
+  if (tokens.length === 0) return null;
+  const attrs = [];
+  for (const token of tokens) {
+    const m = /^(!?)([a-zA-Z]+)=(.+)$/.exec(token);
+    if (!m) return null;
+    const key = (m[2] ?? "").toLowerCase();
+    if (key !== "tool" && key !== "tier" && key !== "role") return null;
+    const values = (m[3] ?? "").split(",").map((v) => v.trim()).filter(Boolean);
+    if (values.length === 0) return null;
+    attrs.push({ key, negated: m[1] === "!", values });
+  }
+  return attrs;
+}
+function evaluate(test, ctx) {
+  if (test === null || !ctx) return true;
+  return test.every((attr) => {
+    const matched = attr.key === "tool" ? attr.values.some((v) => ctx.toolNames.has(v)) : attr.key === "tier" ? attr.values.includes(ctx.tier) : attr.values.includes(ctx.subagent ? "subagent" : "leader");
+    return attr.negated ? !matched : matched;
+  });
+}
+function emit(nodes, ctx) {
+  let out = "";
+  for (const node of nodes) {
+    if (node.kind === "text") {
+      out += node.value;
+      continue;
+    }
+    const branch = evaluate(node.test, ctx) ? node.body[0] : node.body[1];
+    if (branch) out += emit(branch, ctx);
+  }
+  return out;
+}
+function substitute(text2, ctx) {
+  PLACEHOLDER_RE.lastIndex = 0;
+  return text2.replace(PLACEHOLDER_RE, (match, toolsPrefix, body) => {
+    if (toolsPrefix) {
+      const names = body.split(",").map((n) => n.trim()).filter(Boolean).filter((n) => !ctx || ctx.toolNames.has(n));
+      return names.map((n) => `\`${n}\``).join(", ");
+    }
+    const value = ctx?.vars?.[body.trim()];
+    return value === void 0 ? match : String(value);
+  });
+}
+function tidy(text2) {
+  return text2.replace(/(\r?\n){3,}/g, "$1$1");
+}
+
+// src/core/system-prompt-blocks.ts
+var SYSTEM_BLOCK_SOURCE = /* @__PURE__ */ new WeakMap();
+function tagBlock(block, source) {
+  SYSTEM_BLOCK_SOURCE.set(block, source);
+  return block;
+}
+function shortSessionId(sessionId) {
+  const leaf = sessionId.split("/").pop() ?? sessionId;
+  return leaf.length > 12 ? `${leaf.slice(0, 12)}\u2026` : leaf;
+}
+function instructionSection(bundle, key, vars = {}, tplCtx) {
+  const template = bundle.sections?.[key];
+  if (!template) return "";
+  return renderInstructionLayer(
+    template,
+    tplCtx ? { ...tplCtx, vars: { ...tplCtx.vars, ...vars } } : void 0
+  ).replace(/\{\{\s*([a-zA-Z0-9_.-]+)\s*\}\}/g, (match, name) => {
+    const value = vars[name];
+    return value === void 0 ? match : String(value);
+  });
+}
+function renderToolSelectionBoundary(tool) {
+  const selection = tool.selection;
+  if (!selection?.doNotUseWhen.trim()) return "";
+  const alternatives = selection.useInstead?.filter(Boolean) ?? [];
+  const instead = alternatives.length > 0 ? ` Use ${alternatives.map((name) => `\`${name}\``).join(" or ")} instead.` : "";
+  return `Do not use when ${selection.doNotUseWhen.trim()}${instead}`;
+}
+function agentsFingerprint(agents) {
+  if (!agents || agents.length === 0) return "0";
+  let h = 2166136261;
+  for (const a of agents) {
+    const fields = [
+      a.agentId,
+      a.name,
+      a.source,
+      a.sessionId,
+      a.status,
+      a.currentTask,
+      a.currentTool,
+      a.online ? "1" : "0"
+    ];
+    for (const field of fields) {
+      const value = field ?? "";
+      for (let i = 0; i < value.length; i++) {
+        h ^= value.charCodeAt(i);
+        h = Math.imul(h, 16777619) >>> 0;
+      }
+      h ^= 255;
+      h = Math.imul(h, 16777619) >>> 0;
+    }
+  }
+  return `${agents.length}:${h.toString(36)}`;
+}
+
+// src/core/agent-response.ts
+var MAX_TODO_SNAPSHOT_ITEMS = 10;
+var MAX_TODO_SNAPSHOT_CONTENT = 180;
+function buildLiveNextStepsGateBlock(ctx) {
+  if (ctx.agentId !== "leader") return void 0;
+  const openTodos = ctx.todos.filter(
+    (todo) => todo.status === "pending" || todo.status === "in_progress"
+  );
+  if (openTodos.length === 0) {
+    const toolRoute = ctx.tools?.some((t2) => t2.name === "nextsteps") ? [
+      "Calling the `nextsteps` tool with the same items satisfies branch 1 as well; if you both call it and write the block, the block wins."
+    ] : [];
+    return {
+      type: "text",
+      text: [
+        "[nextsteps_gate]",
+        "Authoritative live state for this request: open todos = 0.",
+        "On the final response, you MUST take exactly one branch:",
+        "1. If at least one genuinely useful follow-on action exists, include a balanced <nextsteps> block containing 1-4 exact prompt messages that can be submitted back to you through the current TUI or WebUI input.",
+        "Every item must ask the agent to perform work. Never put a human-only chore or an instruction addressed to the user inside <nextsteps>; natural-language agent-directed imperatives are valid and need not be shell commands.",
+        ...toolRoute,
+        "2. If no useful follow-on action truly exists, omit <nextsteps> and explicitly tell the user in normal prose that no further steps are needed for this task.",
+        "Silently omitting both is invalid. Do not decide by chance, tone, or response length, and do not invent filler suggestions.",
+        "[/nextsteps_gate]"
+      ].join("\n")
+    };
+  }
+  const todoSnapshot = openTodos.slice(0, MAX_TODO_SNAPSHOT_ITEMS).map((todo) => {
+    const normalized = todo.content.replace(/\s+/g, " ").trim();
+    const content = normalized.length > MAX_TODO_SNAPSHOT_CONTENT ? `${normalized.slice(0, MAX_TODO_SNAPSHOT_CONTENT - 1)}\u2026` : normalized;
+    return formatTodoForModel({ ...todo, content });
+  });
+  const omitted = openTodos.length - todoSnapshot.length;
+  if (omitted > 0) todoSnapshot.push(`- \u2026and ${omitted} more open todo(s)`);
+  const todoReconciliation = ctx.tools?.some((tool) => tool.name === "todo") ? [
+    "Before ending the turn, you MUST call the `todo` tool with the complete current list to reconcile actual progress: finished items completed, exactly one actively worked item in_progress, and untouched items pending. A prose claim that work is done does not update the Todo/Kanban state.",
+    ...hasKanbanBoundTodos(openTodos) ? [
+      "Rows below carry a <kanban board/task> binding. Pass those exact ids back as `kanbanBoardId`/`kanbanTaskId` on every row you resend; a row that arrives without its binding is not applied to its card."
+    ] : []
+  ] : [];
+  return {
+    type: "text",
+    text: [
+      "[nextsteps_gate]",
+      `Authoritative live state for this request: open todos = ${openTodos.length}.`,
+      "You MUST omit <nextsteps> entirely while these todos remain open. Continue or finish the tracked work; do not propose unrelated follow-on work.",
+      ...todoReconciliation,
+      "Open todo snapshot:",
+      ...todoSnapshot,
+      "[/nextsteps_gate]"
+    ].join("\n")
+  };
+}
+var MAX_MEMORY_EVIDENCE_CHARS = 12e3;
+function buildMemoryEvidenceBlocks(ctx) {
+  const blocks = [];
+  let remaining = MAX_MEMORY_EVIDENCE_CHARS;
+  for (const entry of ctx.memoryEvidence) {
+    if (remaining <= 0) break;
+    const text2 = entry.text.trim();
+    if (!text2) continue;
+    const source = entry.source.replace(/[^a-z0-9_.-]+/gi, "-").slice(0, 80) || "memory";
+    const bounded = text2.slice(0, remaining);
+    remaining -= bounded.length;
+    blocks.push({
+      type: "text",
+      text: `[memory_evidence source="${source}"]
+${bounded}
+[/memory_evidence]`
+    });
+  }
+  return blocks;
+}
+var EPOCH_VOLATILE_SOURCES = /* @__PURE__ */ new Set([
+  "plan",
+  "contributor",
+  "glossary",
+  "peers"
+]);
+var promptEpochPartitions = /* @__PURE__ */ new WeakMap();
+function partitionPromptEpoch(prompt) {
+  const cached2 = promptEpochPartitions.get(prompt);
+  if (cached2) return cached2;
+  const stable2 = [];
+  const tail = [];
+  for (const block of prompt) {
+    const source = SYSTEM_BLOCK_SOURCE.get(block);
+    if (source && EPOCH_VOLATILE_SOURCES.has(source)) {
+      tail.push({ type: "text", text: block.text });
+    } else {
+      stable2.push(block);
+    }
+  }
+  const partition = { stable: stable2, tail };
+  promptEpochPartitions.set(prompt, partition);
+  return partition;
+}
+var LIVE_CONTEXT_HEADER = {
+  type: "text",
+  text: "[live_context]\nThe blocks below are live session state (active plan, glossary, completed-work ledger, conversation continuity, response gates, memory evidence) re-sent with every request. They are steering context, not a new user message. Where they conflict with the conversation, newer conversation turns win."
+};
+var NEXT_STEPS_BLOCK_RE = /<nextsteps\b[^>]*>[\s\S]*?<\/nextsteps>[ \t]*\n?/gi;
+var NEXT_STEPS_STRIPPED_PLACEHOLDER = "[nextsteps suggestions were delivered to the user]";
+var strippedNextStepsCache = /* @__PURE__ */ new WeakMap();
+function stripDeliveredNextSteps(history) {
+  let out = null;
+  for (let i = 0; i < history.length; i++) {
+    const msg = history[i];
+    const replaced = stripNextStepsFromMessage(msg);
+    if (out === null && replaced !== msg) out = history.slice(0, i);
+    if (out !== null) out.push(replaced);
+  }
+  return out ?? history;
+}
+function stripNextStepsFromMessage(msg) {
+  if (msg.role !== "assistant") return msg;
+  const cached2 = strippedNextStepsCache.get(msg);
+  if (cached2) return cached2;
+  const hasTag = typeof msg.content === "string" ? msg.content.includes("<nextsteps") : msg.content.some((b) => b.type === "text" && b.text.includes("<nextsteps"));
+  if (!hasTag) {
+    strippedNextStepsCache.set(msg, msg);
+    return msg;
+  }
+  let clone;
+  if (typeof msg.content === "string") {
+    const text2 = msg.content.replace(NEXT_STEPS_BLOCK_RE, "").trimEnd();
+    clone = { ...msg, content: text2.length > 0 ? text2 : NEXT_STEPS_STRIPPED_PLACEHOLDER };
+  } else {
+    const blocks = msg.content.map(
+      (b) => b.type === "text" && b.text.includes("<nextsteps") ? { ...b, text: b.text.replace(NEXT_STEPS_BLOCK_RE, "").trimEnd() } : b
+    ).filter((b) => b.type !== "text" || b.text.trim().length > 0);
+    clone = {
+      ...msg,
+      content: blocks.length > 0 ? blocks : [{ type: "text", text: NEXT_STEPS_STRIPPED_PLACEHOLDER }]
+    };
+  }
+  strippedNextStepsCache.set(msg, clone);
+  return clone;
+}
+function composeRequestMessages(history, tail) {
+  if (history.length === 0) return null;
+  const out = history.slice();
+  const lastIdx = out.length - 1;
+  const last = out[lastIdx];
+  const blocks = typeof last.content === "string" ? [{ type: "text", text: last.content }] : last.content.slice();
+  const boundary = blocks[blocks.length - 1];
+  if (boundary && (boundary.type === "text" || boundary.type === "tool_result")) {
+    blocks[blocks.length - 1] = { ...boundary, cache_control: { type: "ephemeral" } };
+  }
+  if (tail.length === 0 || last.role !== "user") {
+    out[lastIdx] = { ...last, content: blocks };
+    if (tail.length > 0) out.push({ role: "user", content: [LIVE_CONTEXT_HEADER, ...tail] });
+    return out;
+  }
+  out[lastIdx] = { ...last, content: [...blocks, LIVE_CONTEXT_HEADER, ...tail] };
+  return out;
+}
+function createAgentResponseHandler(a) {
+  const stabilizedPromptEpochs = /* @__PURE__ */ new WeakSet();
+  function stabilizePromptEpoch() {
+    const prompt = a.ctx.systemPrompt;
+    if (stabilizedPromptEpochs.has(prompt)) return;
+    for (const block of prompt) {
+      if (block.cache_control) Object.freeze(block.cache_control);
+      Object.freeze(block);
+    }
+    Object.freeze(prompt);
+    stabilizedPromptEpochs.add(prompt);
+  }
+  async function buildAndRunRequestPipeline(opts) {
+    if (a.ctx.toolAdjacencyDirty) {
+      const repaired = repairToolUseAdjacency(a.ctx.messages);
+      a.ctx.toolAdjacencyDirty = false;
+      if (repaired.report.changed) {
+        a.ctx.state.replaceMessages(repaired.messages);
+        a.events.emit("context.repaired", {
+          sessionId: resolveEventSessionId(a.ctx),
+          ctx: a.ctx,
+          ...repaired.report
+        });
+        a.logger.warn(
+          `Repaired context tool adjacency: removed ${repaired.report.removedToolUses.length} tool_use block(s), ${repaired.report.removedToolResults.length} tool_result block(s), ${repaired.report.removedMessages} empty message(s)`
+        );
+      }
+    }
+    stabilizePromptEpoch();
+    const { stable: stableSystem, tail: epochTail } = partitionPromptEpoch(a.ctx.systemPrompt);
+    const volatileLedger = buildCompletedWorkLedgerBlock(a.ctx);
+    const continuity = buildConversationContinuityBlock(a.ctx);
+    const liveNextStepsGate = buildLiveNextStepsGateBlock(a.ctx);
+    const memoryEvidence = buildMemoryEvidenceBlocks(a.ctx);
+    const liveContextTail = [
+      ...epochTail,
+      volatileLedger,
+      continuity,
+      liveNextStepsGate,
+      ...memoryEvidence
+    ].filter((block) => block !== void 0);
+    const requestHistory = stripDeliveredNextSteps(a.ctx.messages);
+    const composedMessages = composeRequestMessages(requestHistory, liveContextTail);
+    const system = composedMessages ? stableSystem : liveContextTail.length > 0 ? [...stableSystem, ...liveContextTail] : stableSystem;
+    await a.ctx.waitForModelTransition();
+    const provider = a.ctx.provider;
+    const baseReq = {
+      model: opts.model ?? a.ctx.model,
+      system,
+      messages: composedMessages ?? requestHistory,
+      tools: a.tools.listForProvider(),
+      // `maxTokens` is deliberately NOT set here. The provider adapter
+      // resolves the ceiling from the catalog entry for the model in
+      // `req.model`, which is the only source that stays correct across a
+      // `/model` switch, a fallback hop, or a subagent on a model-matrix
+      // entry — `provider.capabilities` is resolved once, for the model the
+      // session booted on, and pinning it here would override the accurate
+      // per-request value with a stale one. Callers that genuinely want a
+      // smaller response (one-shot LLM helpers, compaction, the brain) still
+      // set `maxTokens` on their own Request and keep priority over the
+      // catalog.
+      // Provider-agnostic cache-partition key from the STABLE part of the
+      // prompt epoch. Wires that support prompt caching (OpenAI
+      // `prompt_cache_key`) read it; the config `ttl` is merged over this by
+      // the ModelRuntime middleware. Keyed off the stable partition, not the
+      // full epoch — a glossary/plan refresh must not re-route the cache
+      // partition when the actual prefix bytes did not change.
+      cache: { key: deriveCachePrefixKey(stableSystem) }
+    };
+    const request = await a.pipelines.request.run(baseReq);
+    bindRequestProvider(request, provider);
+    return { request, provider };
+  }
+  async function processResponse(raw, req, requestProvider = a.ctx.provider) {
+    let res = raw;
+    res = await a.pipelines.response.run(res);
+    res = maybeAppendPendingNextSteps(a.ctx, res);
+    a.events.emit("provider.response", {
+      sessionId: resolveEventSessionId(a.ctx),
+      ctx: a.ctx,
+      model: req.model,
+      content: res.content,
+      usage: res.usage,
+      stopReason: res.stopReason
+    });
+    a.ctx.tokenCounter.account(res.usage, req.model, requestProvider.id);
+    if (hasMeaningfulContent(res.content)) {
+      await a.ctx.session.append({
+        type: "llm_response",
+        ts: (/* @__PURE__ */ new Date()).toISOString(),
+        content: res.content,
+        stopReason: res.stopReason,
+        usage: res.usage
+      });
+      a.ctx.state.appendMessage({ role: "assistant", content: res.content });
+      if (!a.ctx.toolAdjacencyDirty) {
+        for (const block of res.content) {
+          if (block.type === "tool_use") {
+            a.ctx.toolAdjacencyDirty = true;
+            break;
+          }
+        }
+      }
+      try {
+        await a.ctx.flushConversationJournal();
+        await a.ctx.session.flush();
+      } catch (err) {
+        (a.logger.debug ?? a.logger.warn)?.(`LLM response flush failed: ${toErrorMessage(err)}`);
+      }
+    } else {
+      a.logger.warn("Empty assistant response \u2014 not appended to context or session", {
+        model: req.model,
+        stopReason: res.stopReason,
+        aborted: a.ctx.signal.aborted
+      });
+    }
+    if (a.ctx.signal.aborted) {
+      const parts2 = [];
+      for (const block of res.content) {
+        if (isTextBlock(block)) parts2.push(block.text);
+      }
+      return { finalText: parts2.join(""), aborted: true, done: false };
+    }
+    const parts = [];
+    const streamed = requestProvider.capabilities.streaming;
+    for (const block of res.content) {
+      if (isTextBlock(block)) {
+        const rendered = await a.pipelines.assistantOutput.run(block);
+        parts.push(rendered.text);
+        if (!streamed) a.renderer?.write(rendered);
+      }
+    }
+    const finalText = parts.join("");
+    markAssistantReferencedEvidence(a.ctx, finalText);
+    let directive = "none";
+    if (finalText) {
+      directive = parseContinueDirective(finalText);
+    }
+    return { finalText, aborted: false, done: false, directive };
+  }
+  return { buildAndRunRequestPipeline, processResponse };
+}
+
 // src/types/system-prompt.ts
 function flattenSystemPromptRegions(regions) {
   return [...regions.core, ...regions.session, ...regions.volatile];
@@ -10723,138 +11018,6 @@ function firstExistingDirSync(candidates) {
   return candidates[0] ?? "";
 }
 
-// src/core/instruction-template.ts
-var CANONICAL_TOOL_NAMES = new Set(
-  RUNTIME_CAPABILITY_MANIFEST.flatMap((entry) => [...entry.tools])
-);
-var DIRECTIVE_RE = /[ \t]*<!--\s*ws:(if|else|end)\b([^>]*?)-->[ \t]*(?:\r?\n)?/g;
-var PLACEHOLDER_RE = /\{\{\s*(tools:)?\s*([a-zA-Z0-9_.,\s-]+?)\s*\}\}/g;
-function renderInstructionLayer(text2, ctx) {
-  if (!text2) return text2;
-  const hasDirectives = text2.includes("<!--ws:") || text2.includes("<!-- ws:");
-  const hasPlaceholders = text2.includes("{{");
-  if (!hasDirectives && !hasPlaceholders) return text2;
-  const rendered = hasDirectives ? emit(parse2(text2), ctx) : text2;
-  const substituted = hasPlaceholders ? substitute(rendered, ctx) : rendered;
-  const guarded = ctx?.strictToolReferences ? dropLinesWithUnavailableToolReferences(
-    substituted,
-    ctx,
-    /* @__PURE__ */ new Set([...CANONICAL_TOOL_NAMES, ...declaredToolNames(text2)])
-  ) : substituted;
-  return tidy(guarded);
-}
-function declaredToolNames(text2) {
-  const names = /* @__PURE__ */ new Set();
-  for (const marker of text2.matchAll(/<!--\s*ws:if\b([^>]*?)-->/g)) {
-    for (const attr of (marker[1] ?? "").matchAll(/!?tool=([A-Za-z0-9_.,-]+)/g)) {
-      for (const name of (attr[1] ?? "").split(",")) if (name.trim()) names.add(name.trim());
-    }
-  }
-  for (const placeholder of text2.matchAll(/\{\{\s*tools:\s*([^}]+)}}/g)) {
-    for (const name of (placeholder[1] ?? "").split(",")) if (name.trim()) names.add(name.trim());
-  }
-  return names;
-}
-function dropLinesWithUnavailableToolReferences(text2, ctx, declared) {
-  const unavailable = [...declared].filter((name) => !ctx.toolNames.has(name));
-  if (unavailable.length === 0) return text2;
-  return text2.split(/(?<=\n)/).filter((line) => !unavailable.some((name) => formattedToolMention(line, name))).join("");
-}
-function formattedToolMention(line, name) {
-  const escaped = name.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
-  const token = new RegExp(`(?<![\\w-])${escaped}(?![\\w-])`);
-  if (line.split("`").some((segment, index) => {
-    if (index % 2 !== 1) return false;
-    if (segment.includes(`<${name}`) || segment.includes(`</${name}`)) return false;
-    return token.test(segment);
-  })) {
-    return true;
-  }
-  return line.split("**").some((segment, index) => index % 2 === 1 && segment.trim() === name);
-}
-function parse2(text2) {
-  const root = [];
-  const stack = [];
-  const current = () => {
-    const frame = stack[stack.length - 1];
-    if (!frame) return root;
-    return frame.branches[frame.branches.length - 1];
-  };
-  const pushText = (value) => {
-    if (value) current().push({ kind: "text", value });
-  };
-  DIRECTIVE_RE.lastIndex = 0;
-  let cursor = 0;
-  for (let m = DIRECTIVE_RE.exec(text2); m !== null; m = DIRECTIVE_RE.exec(text2)) {
-    pushText(text2.slice(cursor, m.index));
-    cursor = m.index + m[0].length;
-    const keyword = m[1];
-    if (keyword === "if") {
-      stack.push({ test: parseCondition(m[2] ?? ""), branches: [[]] });
-    } else if (keyword === "else") {
-      const frame = stack[stack.length - 1];
-      if (frame && frame.branches.length === 1) frame.branches.push([]);
-    } else {
-      const frame = stack.pop();
-      if (frame) current().push({ kind: "if", test: frame.test, body: frame.branches });
-    }
-  }
-  pushText(text2.slice(cursor));
-  while (stack.length > 0) {
-    const frame = stack.pop();
-    current().push(...frame.branches.flat());
-  }
-  return root;
-}
-function parseCondition(raw) {
-  const tokens = raw.trim().split(/\s+/).filter(Boolean);
-  if (tokens.length === 0) return null;
-  const attrs = [];
-  for (const token of tokens) {
-    const m = /^(!?)([a-zA-Z]+)=(.+)$/.exec(token);
-    if (!m) return null;
-    const key = (m[2] ?? "").toLowerCase();
-    if (key !== "tool" && key !== "tier" && key !== "role") return null;
-    const values = (m[3] ?? "").split(",").map((v) => v.trim()).filter(Boolean);
-    if (values.length === 0) return null;
-    attrs.push({ key, negated: m[1] === "!", values });
-  }
-  return attrs;
-}
-function evaluate(test, ctx) {
-  if (test === null || !ctx) return true;
-  return test.every((attr) => {
-    const matched = attr.key === "tool" ? attr.values.some((v) => ctx.toolNames.has(v)) : attr.key === "tier" ? attr.values.includes(ctx.tier) : attr.values.includes(ctx.subagent ? "subagent" : "leader");
-    return attr.negated ? !matched : matched;
-  });
-}
-function emit(nodes, ctx) {
-  let out = "";
-  for (const node of nodes) {
-    if (node.kind === "text") {
-      out += node.value;
-      continue;
-    }
-    const branch = evaluate(node.test, ctx) ? node.body[0] : node.body[1];
-    if (branch) out += emit(branch, ctx);
-  }
-  return out;
-}
-function substitute(text2, ctx) {
-  PLACEHOLDER_RE.lastIndex = 0;
-  return text2.replace(PLACEHOLDER_RE, (match, toolsPrefix, body) => {
-    if (toolsPrefix) {
-      const names = body.split(",").map((n) => n.trim()).filter(Boolean).filter((n) => !ctx || ctx.toolNames.has(n));
-      return names.map((n) => `\`${n}\``).join(", ");
-    }
-    const value = ctx?.vars?.[body.trim()];
-    return value === void 0 ? match : String(value);
-  });
-}
-function tidy(text2) {
-  return text2.replace(/(\r?\n){3,}/g, "$1$1");
-}
-
 // src/core/modes/default.ts
 import { readFileSync as readFileSync5, statSync as statSync4 } from "node:fs";
 import * as path23 from "node:path";
@@ -10889,61 +11052,6 @@ function isDirectory(candidate) {
   }
 }
 
-// src/core/system-prompt-blocks.ts
-var SYSTEM_BLOCK_SOURCE = /* @__PURE__ */ new WeakMap();
-function tagBlock(block, source) {
-  SYSTEM_BLOCK_SOURCE.set(block, source);
-  return block;
-}
-function shortSessionId(sessionId) {
-  const leaf = sessionId.split("/").pop() ?? sessionId;
-  return leaf.length > 12 ? `${leaf.slice(0, 12)}\u2026` : leaf;
-}
-function instructionSection(bundle, key, vars = {}, tplCtx) {
-  const template = bundle.sections?.[key];
-  if (!template) return "";
-  return renderInstructionLayer(
-    template,
-    tplCtx ? { ...tplCtx, vars: { ...tplCtx.vars, ...vars } } : void 0
-  ).replace(/\{\{\s*([a-zA-Z0-9_.-]+)\s*\}\}/g, (match, name) => {
-    const value = vars[name];
-    return value === void 0 ? match : String(value);
-  });
-}
-function renderToolSelectionBoundary(tool) {
-  const selection = tool.selection;
-  if (!selection?.doNotUseWhen.trim()) return "";
-  const alternatives = selection.useInstead?.filter(Boolean) ?? [];
-  const instead = alternatives.length > 0 ? ` Use ${alternatives.map((name) => `\`${name}\``).join(" or ")} instead.` : "";
-  return `Do not use when ${selection.doNotUseWhen.trim()}${instead}`;
-}
-function agentsFingerprint(agents) {
-  if (!agents || agents.length === 0) return "0";
-  let h = 2166136261;
-  for (const a of agents) {
-    const fields = [
-      a.agentId,
-      a.name,
-      a.source,
-      a.sessionId,
-      a.status,
-      a.currentTask,
-      a.currentTool,
-      a.online ? "1" : "0"
-    ];
-    for (const field of fields) {
-      const value = field ?? "";
-      for (let i = 0; i < value.length; i++) {
-        h ^= value.charCodeAt(i);
-        h = Math.imul(h, 16777619) >>> 0;
-      }
-      h ^= 255;
-      h = Math.imul(h, 16777619) >>> 0;
-    }
-  }
-  return `${agents.length}:${h.toString(36)}`;
-}
-
 // src/core/system-prompt-environment.ts
 import * as os6 from "node:os";
 import * as path25 from "node:path";
@@ -11231,24 +11339,14 @@ async function renderDomainGlossary(ctx, memory, options = {}) {
   );
   return lines.join("\n");
 }
-function makeDomainGlossaryContributor(glossary) {
-  return async (ctx) => {
-    const text2 = await renderDomainGlossary(ctx, glossary.memory);
-    if (!text2) return [];
-    return [{ type: "text", text: text2 }];
-  };
-}
 function parseTermEntry(text2) {
   const trimmed = text2.trim();
-  const separators = [" \u2014 ", " \u2013 ", " - ", " -- "];
-  for (const sep10 of separators) {
-    const idx = trimmed.indexOf(sep10);
-    if (idx > 0) {
-      return {
-        term: trimmed.slice(0, idx).trim(),
-        definition: trimmed.slice(idx + sep10.length).trim()
-      };
-    }
+  const idx = trimmed.indexOf(" \u2014 ");
+  if (idx > 0) {
+    return {
+      term: trimmed.slice(0, idx).trim(),
+      definition: trimmed.slice(idx + 3).trim()
+    };
   }
   return { term: trimmed, definition: "" };
 }
@@ -11792,7 +11890,13 @@ var DefaultSystemPromptBuilder = class {
   _lastCatalogTools;
   /** Cached rendered online agents string, keyed by content fingerprint. */
   _lastOnlineAgents;
-  /** Cached full buildToolUsage output — keyed by tools array ref + agents fingerprint + tier. */
+  /**
+   * Cached full buildToolUsage output — keyed by tools array ref + tier.
+   * Deliberately NOT keyed by the online-agents fingerprint: the live peer
+   * snapshot moved out of this layer into the `peers` volatile block, so
+   * layer2 stays byte-stable (and provider-cache-friendly) while agents
+   * join, leave, or change status.
+   */
   _toolsUsageCache;
   _instructionBundle;
   /**
@@ -11957,6 +12061,26 @@ var DefaultSystemPromptBuilder = class {
         volatile.push(tagBlock({ type: "text", text: glossary }, "glossary"));
       }
     }
+    const hasMailboxTools = ctx.tools.some(
+      (t2) => t2.name === "mailbox" || t2.name === "mail_send" || t2.name === "mail_inbox"
+    );
+    if (hasMailboxTools) {
+      const peers = this.renderOnlineAgents(ctx.onlineAgents).trim();
+      if (peers) {
+        volatile.push(
+          tagBlock(
+            {
+              type: "text",
+              text: `[online_agents]
+Live fleet peer snapshot for this request (see the Inter-agent mailbox guidance for how to coordinate):
+${peers}
+[/online_agents]`
+            },
+            "peers"
+          )
+        );
+      }
+    }
     if (!ctx.subagent) {
       session.push(
         tagBlock(
@@ -12051,9 +12175,8 @@ var DefaultSystemPromptBuilder = class {
     const instructions = await this.instructions();
     const tpl = tplCtx ?? this.templateContext(ctx);
     const section = (key, vars = {}) => instructionSection(instructions, key, vars, tpl);
-    const agentsHash = agentsFingerprint(ctx.onlineAgents);
     const tier = this.tier;
-    if (this._toolsUsageCache?.toolsRef === tools && this._toolsUsageCache?.agentsHash === agentsHash && this._toolsUsageCache?.tier === tier) {
+    if (this._toolsUsageCache?.toolsRef === tools && this._toolsUsageCache?.tier === tier) {
       return this._toolsUsageCache.text;
     }
     const byCat = /* @__PURE__ */ new Map();
@@ -12131,7 +12254,7 @@ ${hint.trim()}`);
       (t2) => t2.name === "mailbox" || t2.name === "mail_send" || t2.name === "mail_inbox"
     );
     if (hasMailbox) {
-      const onlineAgentsInfo = this.renderOnlineAgents(ctx.onlineAgents);
+      const onlineAgentsInfo = "";
       const hasMailboxPowerTool = tools.some((t2) => t2.name === "mailbox");
       const mailStatusCommand = tools.some((t2) => t2.name === "fleet_status") ? "`fleet_status`" : hasMailboxPowerTool ? "`mailbox action=status` or `mailbox action=online`" : "the online-agent list above";
       const mailInboxCommand = tools.some((t2) => t2.name === "mail_inbox") ? "`mail_inbox`" : "`mailbox action=check`";
@@ -12180,7 +12303,7 @@ ${hint.trim()}`);
       }
     }
     const text2 = lines.join("\n");
-    this._toolsUsageCache = { toolsRef: tools, agentsHash, tier, text: text2 };
+    this._toolsUsageCache = { toolsRef: tools, tier, text: text2 };
     return text2;
   }
   renderOnlineAgents(agents) {
@@ -12240,6 +12363,8 @@ var SYSTEM_BLOCK_SOURCES = [
   "leader-after-task",
   "contributor",
   "ledger",
+  "glossary",
+  "peers",
   "nextsteps",
   "other"
 ];
@@ -23658,6 +23783,7 @@ var TOOLS = {
     "glob",
     "search",
     "tree",
+    "diff",
     "write",
     "edit",
     "replace",
@@ -24689,7 +24815,7 @@ var VERIFY_AGENTS = [
       id: "e2e",
       name: "E2E",
       role: "e2e",
-      tools: [...TOOLS.build, "fetch", ...SPECIALIST_TOOLS.browser],
+      tools: [...TOOLS.build, "fetch", "e2e_plan", ...SPECIALIST_TOOLS.browser],
       prompt: agentPrompt("e2e")
     },
     budget: HEAVY_BUDGET,
@@ -25202,7 +25328,7 @@ var DOMAIN_AGENTS = [
       id: "designer",
       name: "Designer",
       role: "designer",
-      tools: [...TOOLS.docs],
+      tools: [...TOOLS.docs, "design"],
       prompt: agentPrompt("designer")
     },
     budget: MEDIUM_BUDGET,
@@ -62878,7 +63004,10 @@ var SENSITIVE_FLAG_PATTERNS = [
   // -t short flag (token): attached (-tVALUE), separated (-t VALUE), or -t=VALUE.
   // The separator group is optional so the attached form (the common one) matches.
   // (?<![-\w]) anchors to a token start so we don't match the `-t` inside `--token`.
-  /(?<![-\w])-t(?:[=\s]+)?[^\s,-]+/,
+  // The value must be token-like (>= 8 chars) so ordinary combined flags such
+  // as `tar -tf` / `ssh -tt` are not eaten. Global flag: EVERY occurrence is
+  // redacted, not just the first. Synced with packages/tools _redact-command.ts.
+  /(?<![-\w])-t(?:[=\s]+)?[^\s,-]{8,}/g,
   // -p|-password|-a (redis auth) short flags: attached + separated + =value.
   // Same token-start anchor; over-redaction is an accepted tradeoff for a
   // redaction function (false positive = cosmetic noise; false negative = leak).
@@ -62886,8 +63015,9 @@ var SENSITIVE_FLAG_PATTERNS = [
   // env var–style secrets: TOKEN=x, API_KEY=y, etc.
   /(?:TOKEN|API_KEY|API_SECRET|AUTH_TOKEN|GITHUB_TOKEN|GH_TOKEN|BEARER|JWT|OAUTH|CREDENTIAL|SECRET|PRIVATE_KEY|PASSWORD|PASSWD)\s*[=:]\s*[^\s,]+/gi,
   // Generic high-entropy look: base64 strings >32 chars or hex strings >32 digits — but only
-  // when preceded by a flag name (e.g. --github-token=EyJ...).
-  /--\w*(?:token|key|secret|password|passwd|auth|credential)\w*[=\s,][A-Za-z0-9+/=]{32,}/
+  // when preceded by a flag name (e.g. --github-token=EyJ...). Global flag so
+  // every such flag in the command line is redacted, not just the first.
+  /--\w*(?:token|key|secret|password|passwd|auth|credential)\w*[=\s,][A-Za-z0-9+/=]{32,}/g
 ];
 function redactCommand(cmd) {
   let result = cmd;
@@ -80134,7 +80264,10 @@ var DefaultPluginAPI = class {
       }
     };
     this.tools = {
-      register: (t2) => tr.register(t2, owner),
+      register: (t2) => {
+        tr.register(t2, owner);
+        tr.exposeToProvider(t2.name);
+      },
       unregister: (name) => {
         assertCanMutateTool(name, "unregister");
         return tr.unregister(name);
@@ -81212,6 +81345,15 @@ function resolveAutoReviewConfig(cfg, sessionConfig) {
     maxCascadeDepth: cfg.maxCascadeDepth ?? DEFAULT_MAX_CASCADE_DEPTH
   };
 }
+function severitiesFromFindings(findings) {
+  const severities = { critical: 0, high: 0, medium: 0 };
+  for (const finding of findings) {
+    if (finding.severity === "critical") severities.critical++;
+    else if (finding.severity === "high") severities.high++;
+    else if (finding.severity === "medium") severities.medium++;
+  }
+  return severities;
+}
 function parseReviewSeverity(text2) {
   const result = { critical: 0, high: 0, medium: 0 };
   if (!text2) return result;
@@ -81228,9 +81370,20 @@ function parseReviewSeverity(text2) {
   }
   return result;
 }
-function decideCascadeAgents(text2, severities) {
+function decideCascadeAgents(text2, severities, findings) {
   const agents = /* @__PURE__ */ new Set();
   if (severities.critical > 0 || severities.high > 0) agents.add("bug-hunter");
+  if (findings && findings.length > 0) {
+    const highPlus = findings.filter(
+      (finding) => finding.severity === "critical" || finding.severity === "high"
+    );
+    if (highPlus.some((finding) => finding.category === "security")) {
+      agents.add("security-scanner");
+    }
+    if (highPlus.every((finding) => finding.category !== void 0)) {
+      return [...agents];
+    }
+  }
   const securityKeywords = [
     "injection",
     "xss",
@@ -81603,7 +81756,9 @@ function createAutoReviewPlugin() {
               maxFiles: cfg.maxFilesPerBatch,
               autoFix: "off",
               cascadeOn: "off",
-              maxCascadeDepth: 0
+              maxCascadeDepth: 0,
+              fallbackModels: [...cfg.fallbackModels],
+              fallbackProfile: void 0
             },
             files: filesWithContent,
             activeTodos: ctxTodos,
@@ -81708,7 +81863,9 @@ function createAutoReviewPlugin() {
                 maxFiles: cfg.maxFilesPerBatch,
                 autoFix: "off",
                 cascadeOn: "off",
-                maxCascadeDepth: 0
+                maxCascadeDepth: 0,
+                fallbackModels: [...cfg.fallbackModels],
+                fallbackProfile: void 0
               },
               files: filesWithContent,
               cascadeOn: cfg.cascadeOn,
@@ -81750,23 +81907,31 @@ function createAutoReviewPlugin() {
           if (!p.reviewText) return;
           const cascadeOn = p.bundle.cascadeOn ?? "off";
           if (cascadeOn === "off") return;
-          const severities = parseReviewSeverity(p.reviewText);
+          const parsed = p.parsedReport;
+          const verifiedFindings = parsed?.findings.filter((f) => f.verification?.status === "verified") ?? [];
+          const severities = parsed ? severitiesFromFindings(verifiedFindings) : parseReviewSeverity(p.reviewText);
           const threshold = shouldCascade(cascadeOn, severities);
           if (!threshold) return;
-          const agents = decideCascadeAgents(p.reviewText, severities);
+          const agents = decideCascadeAgents(
+            p.reviewText,
+            severities,
+            parsed ? verifiedFindings : void 0
+          );
           if (agents.length === 0) {
             return;
           }
           const cascadePayload = {
             bundle: p.bundle,
+            ...p.reportId ? { reportId: p.reportId } : {},
             reviewText: p.reviewText,
             severities,
             threshold,
-            agents
+            agents,
+            ...parsed ? { verifiedFindings } : {}
           };
           api.emitCustom("chimera.cascade_needed", cascadePayload);
           api.log.info(
-            `[auto-review] cascade_needed emitted \u2014 ${severities.critical} critical, ${severities.high} high, ${severities.medium} medium; agents: ${agents.join(", ")}`
+            `[auto-review] cascade_needed emitted \u2014 ${severities.critical} critical, ${severities.high} high, ${severities.medium} medium; agents: ${agents.join(", ")}${parsed ? ` (gated on ${verifiedFindings.length} verified finding(s))` : ""}`
           );
         } catch (err) {
           api.log.warn(
@@ -81798,12 +81963,88 @@ init_review_finding_store();
 // src/plugins/review-finding-parser.ts
 init_review_finding_types();
 import { randomUUID as randomUUID40 } from "node:crypto";
+var SEVERITIES = /* @__PURE__ */ new Set(["critical", "high", "medium", "low"]);
+var CATEGORIES = /* @__PURE__ */ new Set([
+  "bug",
+  "security",
+  "performance",
+  "type",
+  "contract",
+  "test",
+  "other"
+]);
+var CONFIDENCES = /* @__PURE__ */ new Set(["high", "medium", "low"]);
+var FENCED_BLOCK = /```json[ \t]*\r?\n([\s\S]*?)\r?\n```/gi;
+function extractStructuredFindingsBlock(reportText) {
+  if (!reportText) return null;
+  let best = null;
+  for (const match of reportText.matchAll(FENCED_BLOCK)) {
+    const body = match[1];
+    if (!body?.trim()) continue;
+    let parsed;
+    try {
+      parsed = JSON.parse(body);
+    } catch {
+      continue;
+    }
+    if (typeof parsed !== "object" || parsed === null) continue;
+    const findings = parsed.findings;
+    if (!Array.isArray(findings)) continue;
+    const items = [];
+    for (const raw of findings) {
+      const item = normalizeStructuredItem(raw);
+      if (item) items.push(item);
+    }
+    if (findings.length > 0 && items.length === 0) continue;
+    const durationRaw = parsed.durationSeconds;
+    const durationSeconds = typeof durationRaw === "number" && Number.isFinite(durationRaw) && durationRaw > 0 ? Math.floor(durationRaw) : void 0;
+    best = { findings: items, ...durationSeconds !== void 0 ? { durationSeconds } : {} };
+  }
+  return best;
+}
+function normalizeStructuredItem(raw) {
+  if (typeof raw !== "object" || raw === null) return null;
+  const item = raw;
+  const severity = typeof item.severity === "string" ? item.severity.toLowerCase() : "";
+  if (!SEVERITIES.has(severity)) return null;
+  const title = typeof item.title === "string" ? item.title.trim() : "";
+  if (title.length === 0) return null;
+  const file = typeof item.file === "string" && item.file.trim().length > 0 ? item.file.trim() : void 0;
+  const line = typeof item.line === "number" && Number.isInteger(item.line) && item.line >= 1 ? item.line : void 0;
+  const categoryRaw = typeof item.category === "string" ? item.category.toLowerCase() : "";
+  const category = CATEGORIES.has(categoryRaw) ? categoryRaw : void 0;
+  const confidenceRaw = typeof item.confidence === "string" ? item.confidence.toLowerCase() : "";
+  const confidence = CONFIDENCES.has(confidenceRaw) ? confidenceRaw : void 0;
+  return {
+    severity,
+    ...file ? { file } : {},
+    ...line !== void 0 ? { line } : {},
+    ...category ? { category } : {},
+    ...confidence ? { confidence } : {},
+    title,
+    ...typeof item.description === "string" && item.description.trim().length > 0 ? { description: item.description.trim() } : {},
+    ...typeof item.suggestedFix === "string" && item.suggestedFix.trim().length > 0 ? { suggestedFix: item.suggestedFix.trim() } : {}
+  };
+}
 var SUGGEST_LINE = /^\s*(?:→|->|=>)\s*(.+)$/;
 var DURATION_LINE = /^Duration:\s*(\d+)s\s*$/im;
 function parseChimeraReviewReport(reportText, context = {}) {
   if (!reportText || reportText.trim().length === 0) {
     return { findings: [], unparseableCount: 0 };
   }
+  const structured = extractStructuredFindingsBlock(reportText);
+  if (structured) {
+    const reportId2 = context.reportId ?? randomUUID40();
+    const findings2 = structured.findings.map(
+      (item) => buildFindingFromStructuredItem(item, { ...context, reportId: reportId2 })
+    );
+    return {
+      findings: findings2,
+      unparseableCount: 0,
+      ...structured.durationSeconds !== void 0 ? { durationSeconds: structured.durationSeconds } : {},
+      structured: true
+    };
+  }
   const findings = [];
   const reportId = context.reportId ?? randomUUID40();
   let unparseableCount = 0;
@@ -81935,14 +82176,92 @@ function normalizeFindingSource(reviewType) {
       return "chimera";
   }
 }
+function buildFindingFromStructuredItem(item, context) {
+  const file = item.file;
+  const line = item.line;
+  const title = item.title;
+  const description = item.description ?? title;
+  return {
+    id: randomUUID40(),
+    fingerprint: computeFindingFingerprint(file ?? "", line ?? null, title),
+    severity: item.severity,
+    source: normalizeFindingSource(context.reviewType),
+    ...file ? { location: { file, ...line !== void 0 ? { line } : {} } } : {},
+    ...item.category ? { category: item.category } : {},
+    ...item.confidence ? { confidence: item.confidence } : {},
+    title,
+    description,
+    ...item.suggestedFix ? { suggestedFix: item.suggestedFix } : {},
+    createdAt: (/* @__PURE__ */ new Date()).toISOString(),
+    status: "active",
+    originReport: {
+      reportId: context.reportId ?? randomUUID40(),
+      sessionId: context.sessionId ?? "",
+      agentId: context.agentId ?? "",
+      reviewerModel: context.reviewerModel ?? ""
+    }
+  };
+}
 
 // src/plugins/review-report-integration.ts
 init_review_finding_store();
+
+// src/plugins/review-finding-integration.ts
+init_review_finding_store();
+function classifyChimeraReviewSource(bundle) {
+  const cascadeDepth = bundle.cascadeDepth ?? 0;
+  if (cascadeDepth > 0) return "cascade";
+  const cascadeOn = bundle.cascadeOn;
+  if (cascadeOn !== void 0 && cascadeOn !== "off") return "auto";
+  return "chimera";
+}
+async function integrateFindings(payload, projectDir, reportId) {
+  if ((!payload.reviewText || payload.reviewText.trim().length === 0) && !payload.parsedReport) {
+    return { created: 0, relinked: 0, reopened: 0, totalFindings: 0, unparseableCount: 0 };
+  }
+  const store = new JsonlFindingStore(projectDir);
+  const source = classifyChimeraReviewSource(payload.bundle);
+  const agentId = payload.bundle.fileProvenance?.find((entry) => entry.agentId)?.agentId ?? "chimera-review";
+  const sessionId = payload.sessionId ?? payload.cwd;
+  const model = payload.bundle.config.model;
+  const parsed = payload.parsedReport ?? parseChimeraReviewReport(payload.reviewText, {
+    sessionId,
+    agentId,
+    reviewerModel: model,
+    reviewType: source,
+    reportId
+  });
+  if (parsed.findings.length === 0) {
+    return {
+      created: 0,
+      relinked: 0,
+      reopened: 0,
+      totalFindings: 0,
+      unparseableCount: parsed.unparseableCount
+    };
+  }
+  const result = await store.upsert(parsed.findings, {
+    sessionId,
+    reportId,
+    agentId,
+    model
+  });
+  return {
+    created: result.created,
+    relinked: result.relinked,
+    reopened: result.reopened,
+    reportId,
+    totalFindings: parsed.findings.length,
+    unparseableCount: parsed.unparseableCount
+  };
+}
+
+// src/plugins/review-report-integration.ts
 init_review_report_store();
 async function persistReviewReport(payload, reportId, projectDir) {
   const store = new JsonlReportStore(projectDir);
   const existed = await store.get(reportId);
-  const source = classifySource(payload);
+  const source = classifyChimeraReviewSource(payload.bundle);
   const agentId = payload.bundle.fileProvenance?.find((entry) => entry.agentId)?.agentId ?? "chimera-review";
   const sessionId = payload.sessionId ?? payload.cwd;
   const model = payload.bundle.config.model;
@@ -81952,7 +82271,13 @@ async function persistReviewReport(payload, reportId, projectDir) {
     status: f.status
   }));
   const reviewStatus = payload.status === "success" ? "success" : "failed";
-  const parsed = reviewStatus === "success" ? parseChimeraReviewReport(payload.reviewText, { reportId }) : { findings: [], unparseableCount: 0, durationSeconds: void 0 };
+  const parsed = reviewStatus === "success" ? payload.parsedReport ?? parseChimeraReviewReport(payload.reviewText, {
+    sessionId,
+    agentId,
+    reviewerModel: model,
+    reviewType: source,
+    reportId
+  }) : { findings: [], unparseableCount: 0, durationSeconds: void 0 };
   const counts = { critical: 0, high: 0, medium: 0, low: 0 };
   for (const finding of parsed.findings) {
     counts[finding.severity]++;
@@ -81970,7 +82295,12 @@ async function persistReviewReport(payload, reportId, projectDir) {
     unparseableCount: parsed.unparseableCount,
     durationSeconds: parsed.durationSeconds,
     rawText: payload.reviewText,
-    ...cascadeDepth !== void 0 ? { cascadeDepth } : {}
+    ...cascadeDepth !== void 0 ? { cascadeDepth } : {},
+    // P0-3: carry the cascade evidence verification (status + per-check
+    // comparisons) so the persisted report is auditable. Absent on initial
+    // reviews — no cascade step produced evidence yet.
+    ...payload.bundle.evidenceStatus !== void 0 ? { evidenceStatus: payload.bundle.evidenceStatus } : {},
+    ...payload.bundle.evidenceChecks !== void 0 ? { evidenceChecks: payload.bundle.evidenceChecks } : {}
   };
   await store.persist(input);
   if (reviewStatus === "success" && parsed.findings.length === 0 && parsed.unparseableCount === 0 && isExplicitAllClearReview(payload.reviewText)) {
@@ -82033,13 +82363,6 @@ async function syncReportReopen(reportId, projectDir, actor, reason) {
   });
   return { reportId, reopened: true, previousLifecycle: report.lifecycle };
 }
-function classifySource(payload) {
-  const cascadeDepth = payload.bundle.cascadeDepth ?? 0;
-  if (cascadeDepth > 0) return "cascade";
-  const cascadeOn = payload.bundle.cascadeOn;
-  if (cascadeOn !== void 0 && cascadeOn !== "off") return "auto";
-  return "chimera";
-}
 
 // src/plugins/review-finding-commands.ts
 async function executeFindingCommand(args, ctx) {
@@ -82369,6 +82692,18 @@ async function showReport(id, ctx) {
     `**Review status:** ${report.reviewStatus}`,
     ...report.cascadeDepth !== void 0 ? [`**Cascade depth:** ${report.cascadeDepth}`] : [],
     ...report.durationSeconds !== void 0 ? [`**Duration:** ${report.durationSeconds}s`] : [],
+    ...report.evidenceStatus !== void 0 ? [
+      `**Evidence:** ${report.evidenceStatus === "verified" ? "\u2705 verified" : report.evidenceStatus === "failed" ? "\u274C failed" : "\u26A0\uFE0F missing"}`,
+      ...report.evidenceChecks && report.evidenceChecks.length > 0 ? [
+        "",
+        ...report.evidenceChecks.map((check) => {
+          const mark = check.ok ? "\u2713" : "\u2717";
+          const claimed = check.claimedExitCode ?? "\u2014";
+          const actual = check.actualExitCode ?? "\u2014";
+          return `  ${mark} \`${check.name}\` \u2014 \`${check.command}\` (claimed ${claimed}, observed ${actual})`;
+        })
+      ] : []
+    ] : [],
     "",
     "**Severity counts:**",
     `  \u{1F534} Critical: ${report.counts.critical}`,
@@ -82477,7 +82812,9 @@ function resolveChimeraConfig(cfg, sessionProvider, sessionModel) {
     maxFiles: cfg.maxFiles ?? DEFAULT_MAX_FILES,
     autoFix: cfg.autoFix ?? "off",
     cascadeOn: cfg.cascadeOn ?? DEFAULT_CASCADE_ON,
-    maxCascadeDepth: cfg.maxCascadeDepth ?? DEFAULT_MAX_CASCADE_DEPTH2
+    maxCascadeDepth: cfg.maxCascadeDepth ?? DEFAULT_MAX_CASCADE_DEPTH2,
+    fallbackModels: cfg.fallbackModels ? [...cfg.fallbackModels] : [],
+    fallbackProfile: cfg.fallbackProfile
   };
 }
 var CHIMERA_REVIEW_PROMPT = readBundledInstructionText("llm/chimera-review.md");
@@ -84141,49 +84478,6 @@ function dim(s) {
   return `\x1B[2m${s}\x1B[0m`;
 }
 
-// src/plugins/review-finding-integration.ts
-init_review_finding_store();
-async function integrateFindings(payload, projectDir, reportId) {
-  if (!payload.reviewText || payload.reviewText.trim().length === 0) {
-    return { created: 0, relinked: 0, reopened: 0, totalFindings: 0, unparseableCount: 0 };
-  }
-  const store = new JsonlFindingStore(projectDir);
-  const source = (payload.bundle.cascadeDepth ?? 0) > 0 ? "cascade" : payload.bundle.cascadeOn !== void 0 && payload.bundle.cascadeOn !== "off" ? "auto" : "chimera";
-  const agentId = payload.bundle.fileProvenance?.find((entry) => entry.agentId)?.agentId ?? "chimera-review";
-  const sessionId = payload.sessionId ?? payload.cwd;
-  const model = payload.bundle.config.model;
-  const parsed = parseChimeraReviewReport(payload.reviewText, {
-    sessionId,
-    agentId,
-    reviewerModel: model,
-    reviewType: source,
-    reportId
-  });
-  if (parsed.findings.length === 0) {
-    return {
-      created: 0,
-      relinked: 0,
-      reopened: 0,
-      totalFindings: 0,
-      unparseableCount: parsed.unparseableCount
-    };
-  }
-  const result = await store.upsert(parsed.findings, {
-    sessionId,
-    reportId,
-    agentId,
-    model
-  });
-  return {
-    created: result.created,
-    relinked: result.relinked,
-    reopened: result.reopened,
-    reportId,
-    totalFindings: parsed.findings.length,
-    unparseableCount: parsed.unparseableCount
-  };
-}
-
 // src/index.ts
 init_review_finding_store();
 init_review_report_store();
@@ -92778,7 +93072,7 @@ function createFallbackChainManageTool(opts) {
     name: FALLBACK_CHAIN_MANAGE_TOOL_NAME,
     description: "View or change the active rate-limit fallback chain. When the primary model is overloaded (429/5xx), the agent rotates through this chain in order. Every new entry must be a FAVORITE model \u2014 add it via favorite_manage first. Use insert to place a fallback at a specific position; use remove to delete an entry.",
     usageHint: '"list" to see the current chain. "add" with a favorite model to append. "insert" with an index (1-based) to place before that position. "remove" with index or model ref. "clear" to empty the chain (auto fallback takes over).',
-    category: "Config",
+    category: "config",
     inputSchema: FALLBACK_CHAIN_SCHEMA,
     permission: "auto",
     mutating: true,
@@ -92928,7 +93222,7 @@ function createFavoriteManageTool(opts) {
     name: FAVORITE_MANAGE_TOOL_NAME,
     description: "Manage your favorite provider/model list. Favorites are the only models that can be added to fallback chains and profiles. The LLM uses this tool to curate which models are available for fallback and role assignment.",
     usageHint: 'Start with "list" to see current favorites. Use "add <provider/model>" to add. Use "remove <index|ref>" to remove.',
-    category: "Config",
+    category: "config",
     inputSchema: FAVORITE_MANAGE_SCHEMA,
     permission: "auto",
     mutating: true,
@@ -93008,7 +93302,7 @@ async function storeProviderKey(providers, input, keyValue, opts) {
     createdAt: (/* @__PURE__ */ new Date()).toISOString()
   });
   entry.apiKeys = existingKeys;
-  entry.apiKey = void 0;
+  delete entry.apiKey;
   if (input.setActive !== false) {
     entry.activeKey = label;
   }
@@ -93051,7 +93345,7 @@ function createSystemConfigViewTool(opts) {
     name: SYSTEM_CONFIG_VIEW_TOOL_NAME,
     description: "Get a comprehensive view of all provider, model, fallback, and matrix configuration. Shows the complete state across all configurable areas so you can see what is available and make informed decisions when assigning models, creating fallback profiles, or managing providers. Use the section parameter to focus on specific areas.",
     usageHint: '"section: all" for everything. "section: providers" for configured providers and key status. "section: models" for leader model and favorites. "section: fallbacks" for chains, profiles, and toggles. "section: matrix" for per-role assignments. "section: refiner" for goal refinement config.',
-    category: "Config",
+    category: "config",
     inputSchema: SYSTEM_CONFIG_VIEW_SCHEMA,
     permission: "auto",
     mutating: false,
@@ -93355,7 +93649,7 @@ function createFallbackProfileManageTool(opts) {
     name: FALLBACK_PROFILE_MANAGE_TOOL_NAME,
     description: "Manage named fallback profiles. A profile is a reusable, ordered list of model references that can be assigned to agent roles. Every entry in a profile must be a FAVORITE model \u2014 add it via favorite_manage first. Use /setmodel or agent_model_assign to assign a profile to a role.",
     usageHint: '"list" to see all profiles. "set" with name and chain (array of model refs) to create or replace a profile. "delete" with name to remove a profile.',
-    category: "Config",
+    category: "config",
     inputSchema: FALLBACK_PROFILE_SCHEMA,
     permission: "auto",
     mutating: true,
@@ -93466,7 +93760,7 @@ function createAgentModelAssignTool(opts) {
     name: AGENT_MODEL_ASSIGN_TOOL_NAME,
     description: "Assign a provider/model or a fallback profile to a specific agent role, phase, or the fleet-wide default. This is the LLM-accessible equivalent of /setmodel set. The provider+model combination must be in your favorites list (unless only clearing). Resolution precedence: exact role \u2192 phase \u2192 * \u2192 leader model.",
     usageHint: 'Use "list" as role to see current assignments. Set with role + model, or role + provider + model, or role + profile. Set role + clear=true to remove a matrix entry. The provider/model must be a favorite.',
-    category: "Config",
+    category: "config",
     inputSchema: AGENT_MODEL_ASSIGN_SCHEMA,
     permission: "auto",
     mutating: true,
@@ -93614,7 +93908,7 @@ function createProviderManageTool(opts) {
     name: PROVIDER_MANAGE_TOOL_NAME,
     description: "View or configure provider entries. List all configured providers with their type, model lists, base URL, and key status. Add new providers, update their settings, or remove unused ones. API keys should be set via provider_key_set instead of passing them here \u2014 they are visible in the LLM output.",
     usageHint: '"list" to see all providers. "add" with provider id and type to create. "configure" to update models, baseUrl, family, or envVars. "remove" to delete a provider. Use provider_key_set for API key management.',
-    category: "Config",
+    category: "config",
     inputSchema: PROVIDER_MANAGE_SCHEMA,
     permission: "auto",
     mutating: true,
@@ -93763,9 +94057,11 @@ function createProviderKeySetTool(opts) {
     name: PROVIDER_KEY_SET_TOOL_NAME,
     description: "Set the API key for a provider. For security, prefer using envVar (reads from environment variable, value never visible to the LLM) over passing the key directly. When neither key nor envVar is provided, the tool returns a prompt for interactive key entry \u2014 the UI will present an input field and the key is stored without LLM visibility.\n\nAfter setting a key, the provider becomes usable for model assignments and fallback chains. Add its models to favorites with favorite_manage to unlock them for fallback/profile use.",
     usageHint: 'Preferred: provider_key_set({ provider: "openai", envVar: "OPENAI_API_KEY" }). For interactive input: provider_key_set({ provider: "openai" }) \u2014 the UI will prompt. Direct key: provider_key_set({ provider: "openai", key: "sk-..." }) \u2014 visible to LLM.',
-    category: "Config",
+    category: "config",
     inputSchema: PROVIDER_KEY_SET_SCHEMA,
-    permission: "auto",
+    // 'confirm', not 'auto' — this tool writes credentials to disk (and can
+    // read arbitrary env vars into the config file), so the user must see it.
+    permission: "confirm",
     mutating: true,
     riskTier: "standard",
     icon: "settings",
@@ -93861,7 +94157,7 @@ function createLeaderModelSetTool(opts) {
     name: LEADER_MODEL_SET_TOOL_NAME,
     description: 'View or change the leader provider/model and system toggles. The leader is the primary model used for the main agent interactions. "set" changes it directly. "profile" derives it from a named fallback profile (first entry becomes leader, rest become the fallback chain). "toggle" controls fallbackAuto (smart default fallback) and favoriteModelsOnly (restrict auto-fallback to favorites only).',
     usageHint: '"show" to see current state. "set" with provider+model to change. "profile" with name to derive from a profile. "toggle" with toggle name and value to change a boolean setting.',
-    category: "Config",
+    category: "config",
     inputSchema: LEADER_MODEL_SET_SCHEMA,
     permission: "auto",
     mutating: true,
@@ -93885,11 +94181,21 @@ function createLeaderModelSetTool(opts) {
         if (!input.provider || !input.model) {
           return { status: "error", message: 'Provide "provider" and "model" for the leader.' };
         }
+        if (opts.switchProviderAndModel) {
+          const switchError = await opts.switchProviderAndModel(input.provider, input.model);
+          if (switchError) {
+            return {
+              status: "error",
+              message: `Could not switch to ${input.provider}/${input.model}: ${switchError}. Config was not changed.`
+            };
+          }
+        }
         await opts.updateConfig((cfg) => {
           cfg.provider = input.provider;
           cfg.model = input.model;
         });
-        return { status: "ok", message: `\u2713 Leader \u2192 ${input.provider}/${input.model}` };
+        const liveNote = opts.switchProviderAndModel ? "" : " (config updated \u2014 the live session keeps its current model until restart or /setmodel)";
+        return { status: "ok", message: `\u2713 Leader \u2192 ${input.provider}/${input.model}${liveNote}` };
       }
       if (input.action === "profile") {
         if (!input.profile) {
@@ -93908,15 +94214,25 @@ function createLeaderModelSetTool(opts) {
           return { status: "error", message: `Cannot parse "${first}" as a valid model reference.` };
         }
         const rest = chain.slice(1);
+        if (opts.switchProviderAndModel) {
+          const switchError = await opts.switchProviderAndModel(provider, model);
+          if (switchError) {
+            return {
+              status: "error",
+              message: `Could not switch to ${provider}/${model}: ${switchError}. Config was not changed.`
+            };
+          }
+        }
         await opts.updateConfig((cfg) => {
           cfg.provider = provider;
           cfg.model = model;
           cfg.fallbackModels = rest;
         });
+        const profileLiveNote = opts.switchProviderAndModel ? "" : "\n  (config updated \u2014 the live session keeps its current model until restart or /setmodel)";
         return {
           status: "ok",
           message: `\u2713 Leader \u2192 ${provider}/${model} (profile: ${input.profile})` + (rest.length > 0 ? `
-  Fallback chain: ${rest.join(" \u2192 ")}` : "")
+  Fallback chain: ${rest.join(" \u2192 ")}` : "") + profileLiveNote
         };
       }
       if (input.action === "toggle") {
@@ -94087,20 +94403,22 @@ async function runEnable(name, deps) {
     const known = Object.keys(all).join(", ");
     return `Unknown server "${name}". Available presets: ${known}`;
   }
-  await updateJsonObjectFile(deps.configPath, (full) => {
+  const persistEnabled = () => updateJsonObjectFile(deps.configPath, (full) => {
     const current = isMcpServerRecord(full.mcpServers) ? full.mcpServers : {};
     setJsonPath(full, ["mcpServers", name], { ...current[name], ...cfg, enabled: true });
   });
   try {
     const live = deps.registry.describe().find((s) => s.name === name);
     if (live && live.state === "connected") {
-      return `${green("\u25CF")} Server "${name}" is already running (${live.toolCount} tools registered).`;
+      await persistEnabled();
+      return `Server "${name}" is already running (${live.toolCount} tools registered).`;
     }
     await deps.registry.start({ ...cfg, enabled: true });
+    await persistEnabled();
     const updated = deps.registry.describe().find((s) => s.name === name);
-    return `${green("\u2713 Enabled and started")} "${name}"${updated ? ` (${updated.toolCount} tools registered).` : "."}`;
+    return `Enabled and started "${name}"${updated ? ` (${updated.toolCount} tools registered).` : "."}`;
   } catch (err) {
-    return `${red("\u2717 Failed to start")} "${name}": ${toErrorMessage(err)}`;
+    return `Failed to start "${name}": ${toErrorMessage(err)}. Config was left unchanged (server stays disabled).`;
   }
 }
 async function runDisable(name, deps) {
@@ -94174,34 +94492,34 @@ function isMcpServerRecord(value) {
   return !!value && typeof value === "object" && !Array.isArray(value);
 }
 function bold(s) {
-  return `\x1B[1m${s}\x1B[0m`;
+  return s;
 }
 function dim2(s) {
-  return `\x1B[2m${s}\x1B[0m`;
+  return s;
 }
 function green(s) {
-  return `\x1B[32m${s}\x1B[0m`;
+  return s;
 }
 function yellow(s) {
-  return `\x1B[33m${s}\x1B[0m`;
+  return s;
 }
 function red(s) {
-  return `\x1B[31m${s}\x1B[0m`;
+  return s;
 }
 function badge(state) {
   switch (state) {
     case "connected":
-      return green("\u25CF connected");
+      return "\u25CF connected";
     case "connecting":
-      return `\x1B[36m\u25D0 connecting\x1B[0m`;
+      return "\u25D0 connecting";
     case "reconnecting":
-      return `\x1B[36m\u25D1 reconnecting\x1B[0m`;
+      return "\u25D1 reconnecting";
     case "disconnected":
-      return dim2("\u25CB disconnected");
+      return "\u25CB disconnected";
     case "failed":
-      return red("\u2717 failed");
+      return "\u2717 failed";
     default:
-      return dim2(state);
+      return state;
   }
 }
 
@@ -94243,13 +94561,19 @@ function createMcpUseTool(opts) {
       const servers = registry2.describe();
       const serverInfo = servers.find((s) => s.name === serverName);
       if (!serverInfo) {
-        return `Server "${serverName}" not found. Available: ${servers.map((s) => s.name).join(", ") || "none"}.`;
+        throw new Error(
+          `Server "${serverName}" not found. Available: ${servers.map((s) => s.name).join(", ") || "none"}.`
+        );
       }
       if (serverInfo.state !== "connected") {
-        return `Server "${serverName}" is not connected (state: ${serverInfo.state}). Use \`mcp_control({ action: "enable", server: "${serverName}" })\` first.`;
+        throw new Error(
+          `Server "${serverName}" is not connected (state: ${serverInfo.state}). Use \`mcp_control({ action: "enable", server: "${serverName}" })\` first.`
+        );
       }
-      if (registry2.activateServer) {
-        registry2.activateServer(serverName);
+      const alreadyActive = registry2.isActivated?.(serverName) === true;
+      const didActivate = !alreadyActive && Boolean(registry2.activateServer);
+      if (didActivate) {
+        registry2.activateServer?.(serverName);
       }
       try {
         const qualifiedName = mcpQualifiedToolName(serverName, toolName);
@@ -94257,7 +94581,7 @@ function createMcpUseTool(opts) {
         if (!mcpTool) {
           const allTools = toolRegistry.list().filter((t2) => t2.name.startsWith(mcpServerToolPrefix(serverName))).map((t2) => t2.name.replace(mcpServerToolPrefix(serverName), ""));
           const hint = allTools.length > 0 ? `Available tools on "${serverName}": ${allTools.join(", ")}.` : `No tools found on "${serverName}". The server may not have published any tools.`;
-          return `Tool "${toolName}" not found on server "${serverName}". ${hint}`;
+          throw new Error(`Tool "${toolName}" not found on server "${serverName}". ${hint}`);
         }
         const governedExecute = ctx.meta[GOVERNED_TOOL_EXECUTOR_META_KEY];
         if (typeof governedExecute !== "function") {
@@ -94267,7 +94591,7 @@ function createMcpUseTool(opts) {
         if (!result.success) throw new Error(result.error ?? "MCP tool execution failed");
         return result.result;
       } finally {
-        if (registry2.deactivateServer) {
+        if (didActivate && registry2.deactivateServer) {
           registry2.deactivateServer(serverName);
         }
       }
@@ -94277,6 +94601,7 @@ function createMcpUseTool(opts) {
 
 // src/tools/one-shot-llm-tool.ts
 var ONE_SHOT_LLM_TOOL_NAME = "llm";
+var MAX_TIMEOUT_MS2 = 12e4;
 var INPUT_SCHEMA2 = {
   type: "object",
   properties: {
@@ -94353,9 +94678,10 @@ var INPUT_SCHEMA2 = {
     },
     timeoutMs: {
       type: "number",
-      description: "Hard timeout in ms (default 30s)."
+      description: `Hard timeout in ms (default 30s, clamped to a maximum of ${MAX_TIMEOUT_MS2}).`
     }
-  }
+  },
+  additionalProperties: false
 };
 function createOneShotLLMTool(opts) {
   const orchestrator = new OneShotOrchestrator({
@@ -94363,6 +94689,7 @@ function createOneShotLLMTool(opts) {
     getConfig: opts.getConfig,
     fallbackProfileManager: opts.fallbackProfileManager,
     modelRouter: opts.modelRouter,
+    statusTracker: opts.statusTracker,
     logger: opts.logger,
     wrapProviderCall: opts.wrapProviderCall
   });
@@ -94371,8 +94698,23 @@ function createOneShotLLMTool(opts) {
     description: "Make a one-shot LLM call with a system prompt and user input. Supports provider selection, model routing by role, fallback chains, and timeout. Returns the response text, model info, token usage, and whether a fallback was used. Use this for summarization, classification, extraction, and any single-turn LLM task.",
     usageHint: "Provide `system` for the instruction and `userPrompt` for the input. Either set `model`+`providerId`, or have defaults configured on the tool. Set `fallbackModels` for resilience. Check `error` on the result for failure details.",
     inputSchema: INPUT_SCHEMA2,
+    // Metadata mirrors council-tool.ts — both are read-only meta tools that
+    // spend tokens but never touch the workspace.
+    category: "meta",
     permission: "auto",
     mutating: false,
+    riskTier: "safe",
+    maxOutputBytes: 262144,
+    validate(input) {
+      const hasPrompt = typeof input.userPrompt === "string" && input.userPrompt.trim().length > 0;
+      const hasMessages = Array.isArray(input.messages) && input.messages.length > 0;
+      if (!hasPrompt && !hasMessages) {
+        return [
+          "Provide `userPrompt` (a single user turn) or `messages` (a conversation array) \u2014 without either the llm tool has nothing to send to the model."
+        ];
+      }
+      return [];
+    },
     async execute(input, _ctx, { signal }) {
       if (!input.model && !input.providerId && !opts.defaultModel && !opts.defaultProvider) {
         return {
@@ -94389,7 +94731,11 @@ function createOneShotLLMTool(opts) {
         ...input,
         signal: input.signal ? AbortSignal.any([input.signal, signal]) : signal,
         model: input.model ?? opts.defaultModel,
-        providerId: input.providerId ?? opts.defaultProvider
+        providerId: input.providerId ?? opts.defaultProvider,
+        // Clamp runaway timeouts (documented on the schema). Non-positive
+        // values fall back to the orchestrator default rather than making the
+        // call instantly un-completable.
+        ...typeof input.timeoutMs === "number" && input.timeoutMs > 0 ? { timeoutMs: Math.min(input.timeoutMs, MAX_TIMEOUT_MS2) } : { timeoutMs: void 0 }
       };
       return orchestrator.call(effectiveInput);
     }
@@ -96383,7 +96729,6 @@ export {
   makeDesignStudioRequestMiddleware,
   makeDesignVerifyToolCallMiddleware,
   makeDirectorSessionFactory,
-  makeDomainGlossaryContributor,
   makeFleetEmitTool,
   makeFleetStatusTool,
   makeFleetTool,