negotium 0.1.19 → 0.1.21

package/dist/runtime/src/agents/maestro-provider.ts

@@ -63,13 +63,13 @@
 import "#platform/maestro-bootstrap-env";
 import type { HookRegistration, McpResolver } from "maestro-agent-sdk";
 import { maestroProvider as sdkMaestroProvider, setMcpResolver } from "maestro-agent-sdk";
+import { deepMapStrings } from "#agents/deep-map";
 import {
   hostedMcpServers,
   redactHostedSecrets,
   referencesHostedSecretStorage,
-  shouldRedirectHostedVaultTool,
+  substituteHostedSecrets,
 } from "#agents/execution-host";
-import { VAULT_BROKER_REDIRECT_ERROR } from "#agents/vault-tool-policy";
 import type { AgentQueryOptions, UnifiedEvent } from "#types";
 
 /**
@@ -110,28 +110,23 @@ export function buildMaestroDisallowedTools(
  *
  * Pre hook — two responsibilities in one pass:
  *   1. Block direct vault.db filesystem access (security guard).
- *   2. Block direct placeholder substitution and redirect the model to the
- *      Vault MCP credential broker, where expanded input stays outside the
- *      provider process.
+ *   2. Substitute {{KEY}} placeholders immediately before tool execution.
  *
  * Post hook — scrub raw/common encoded secret forms before tool output is sent
- * back to the model. HTTP credentials should still prefer vault_http_request,
- * whose broker never exposes the expanded request to the provider at all.
+ * back to the model.
  */
 function buildVaultHook(userId: string): HookRegistration {
   return {
     name: "vault-guard",
-    pre({ toolName, input }) {
+    pre({ input }) {
       if (referencesHostedSecretStorage(input)) {
         return { decision: "block", error: "Runtime secret storage access is not permitted" };
       }
 
-      // Detect key references from metadata only. No Vault value is decrypted
-      // or injected into the provider/tool registry on this path.
-      if (shouldRedirectHostedVaultTool(userId, toolName, input)) {
-        return { decision: "block", error: VAULT_BROKER_REDIRECT_ERROR };
-      }
-      return { decision: "allow" };
+      const substituted = deepMapStrings(input, (value) => substituteHostedSecrets(userId, value));
+      return JSON.stringify(substituted) === JSON.stringify(input)
+        ? { decision: "allow" }
+        : { decision: "modify", input: substituted as Record<string, unknown> };
     },
     post({ output }) {
       const redacted = redactHostedSecrets(userId, output);

package/dist/runtime/src/agents/mcp-tools/self-config.ts

@@ -48,7 +48,7 @@ export function createSelfConfigToolDefinitions(
     {
       name: "set_model",
       description:
-        "Set the model for THIS topic. Persists and applies from the NEXT turn. Claude: 'sonnet' / 'opus' / 'fable'. Codex: 'gpt-5.6-luna' (default) / 'gpt-5.6-terra' / 'gpt-5.6-sol'. Maestro: 'deepseek-pro' / 'deepseek-flash' / 'deepseek'. The model must match the topic's current agent's accepted list. Switch agent first via set_agent only when crossing the claude/codex/maestro runtime boundary. Fails if the user locked this setting. NEVER use 'fable' unless the user explicitly requests it.",
+        "Set the model for THIS topic; it persists and applies from the NEXT turn. Use a model from the system prompt's same-agent catalog. Cross-agent changes require an explicit user-requested set_agent first. Fails if the user locked this setting. NEVER use 'fable' unless explicitly requested.",
       schema: { model: z.string().describe("Model id valid for the topic's current agent.") },
       async handler({ model }: { model: string }) {
         return mcpResult(setSelfConfigModel(getCtx(), model));

package/dist/runtime/src/agents/mcp-tools/spawn-subagent.ts

@@ -590,7 +590,12 @@ export function createSpawnSubagentToolDefinition(ctx: SpawnSubagentToolContext)
         .enum(["claude", "codex", "maestro"])
         .optional()
         .describe("Agent backend override. Defaults to this room's agent."),
-      model: z.string().optional().describe("Model override. Defaults to the agent's default."),
+      model: z
+        .string()
+        .optional()
+        .describe(
+          `Best-fit model override from the system prompt catalog. Omit agent+model to inherit ${ctx.agent}/${ctx.model ?? "default"}; overriding agent without model uses that agent's default.`,
+        ),
     },
     async handler(input) {
       return spawnSubagent(ctx, input);

package/dist/runtime/src/agents/memory-archive-policy.ts

@@ -0,0 +1,34 @@
+export interface MemoryArchiveMessage {
+  author_id: string;
+  agent_type?: string | null;
+  kind?: string | null;
+}
+
+/**
+ * Shared lifecycle gate for reset/delete memory distillation. Keep the raw
+ * forensic archive even below this threshold, but do not spend an agent turn
+ * on a session with five or fewer completed user/assistant exchanges. This
+ * mirrors Clawgram's lifecycle-level MIN_EXCHANGE_COUNT policy.
+ */
+export const MIN_MEMORY_ARCHIVE_EXCHANGES = 6;
+
+/** Count completed conversational exchanges while ignoring system/tool/card noise. */
+export function countMemoryArchiveExchanges(rows: readonly MemoryArchiveMessage[]): number {
+  let waitingForAssistant = false;
+  let completed = 0;
+
+  for (const row of rows) {
+    const assistant = row.author_id === "ai" || Boolean(row.agent_type);
+    const conversational = row.kind == null || row.kind === "message";
+    if (!assistant && row.author_id !== "system" && conversational) {
+      waitingForAssistant = true;
+      continue;
+    }
+    if (assistant && conversational && waitingForAssistant) {
+      completed++;
+      waitingForAssistant = false;
+    }
+  }
+
+  return completed;
+}

package/dist/runtime/src/agents/model-catalog.ts

@@ -37,8 +37,40 @@ export interface SelectableModel {
   agent: AgentKind;
   /** Short comparison copy shown alongside the model in picker UIs. */
   description: string;
+  /** User-defined relative intelligence band used for routing across providers. */
+  intelligenceTier: "sonnet" | "opus" | "fable";
+  /** Compact capability/cost hint safe to inject into every turn and tool schema. */
+  routingSummary: string;
+  /** Subscription or API basis used when comparing operating cost. */
+  accessCost: string;
+  /** Marginal per-token rate after included subscription usage, or the API rate. */
+  marginalTokenCost: string;
+  /** Best available quota estimate; never presented as a provider-guaranteed token cap. */
+  estimatedUsage: string;
 }
 
+/**
+ * Pricing and quota observations were checked on 2026-07-19.
+ * Official references:
+ * - https://learn.chatgpt.com/docs/pricing
+ * - https://help.openai.com/en/articles/20001106
+ * - https://support.claude.com/en/articles/11049741-what-is-the-max-plan
+ * - https://api-docs.deepseek.com/quick_start/pricing
+ * Community token counts are deliberately labelled estimates because providers
+ * meter cached input, fresh input, output, reasoning, speed, and model choice
+ * differently and may change server-side weights without publishing a token cap.
+ */
+export const MODEL_COST_RESEARCHED_AT = "2026-07-19";
+export const MODEL_COST_ROUTING_SUMMARY =
+  "Cost basis (2026-07-19): Codex Pro 20x and Claude Max 20x are each $200/month; DeepSeek Pro is pay-per-token. Relative marginal token cost: DeepSeek Pro << Codex < Claude.";
+
+const CODEX_PRO_20X_COST = "ChatGPT Pro 20x subscription: $200/month";
+const CODEX_COMMUNITY_WEEKLY =
+  "Community plan-level observation: roughly 2–4B raw/cached tokens per week; fresh-input equivalent is much lower and unstable (low confidence)";
+const CLAUDE_MAX_20X_COST = "Claude Max 20x subscription: $200/month";
+const CLAUDE_COMMUNITY_SESSION =
+  "Community observations vary from roughly 220–250K locally displayed tokens per 5-hour session to billions of cache-heavy raw tokens per week; calibrated reports value a full weekly allowance around $680–$1,900 at API rates. Recent heavy-model reports reach the weekly cap after about 4–5 full sessions (low confidence; not a token cap)";
+
 /**
  * Canonical model picker shared by every channel. Keep this deliberately
  * finite even though the Codex backend accepts arbitrary future model ids:
@@ -48,51 +80,85 @@ export const SELECTABLE_MODELS: readonly SelectableModel[] = [
   {
     model: "gpt-5.6-sol",
     agent: "codex",
-    description: "Latest frontier agentic coding model.",
+    description: "Highest-capability Codex route for the hardest agentic coding work.",
+    intelligenceTier: "fable",
+    routingSummary: "hardest coding work; 5x Codex quota cost",
+    accessCost: CODEX_PRO_20X_COST,
+    marginalTokenCost: "Codex credits: $5/M uncached input, $0.50/M cached input, $30/M output",
+    estimatedUsage: `Official Pro 20x range: 300–1,800 local messages per 5 hours; quota weight 5x Luna. ${CODEX_COMMUNITY_WEEKLY}`,
   },
   {
     model: "gpt-5.6-terra",
     agent: "codex",
-    description: "Balanced agentic coding model for everyday work.",
+    description: "High-capability Codex route for complex coding and reasoning.",
+    intelligenceTier: "opus",
+    routingSummary: "complex coding and reasoning; 2.5x Codex quota cost",
+    accessCost: CODEX_PRO_20X_COST,
+    marginalTokenCost: "Codex credits: $2.50/M uncached input, $0.25/M cached input, $15/M output",
+    estimatedUsage: `Official Pro 20x range: 400–2,200 local messages per 5 hours; quota weight 2.5x Luna. ${CODEX_COMMUNITY_WEEKLY}`,
   },
   {
     model: "gpt-5.6-luna",
     agent: "codex",
-    description: "Fast and affordable agentic coding model.",
-  },
-  {
-    model: "gpt-5.5",
-    agent: "codex",
-    description: "Frontier model for complex coding, research, and real-world work.",
+    description: "Default Codex route with strong everyday coding intelligence.",
+    intelligenceTier: "sonnet",
+    routingSummary: "everyday coding default; lowest Codex quota cost (1x)",
+    accessCost: CODEX_PRO_20X_COST,
+    marginalTokenCost: "Codex credits: $1/M uncached input, $0.10/M cached input, $6/M output",
+    estimatedUsage: `Official Pro 20x range: 1,000–5,600 local messages per 5 hours; lowest Codex quota weight (1x). ${CODEX_COMMUNITY_WEEKLY}`,
   },
   {
     model: "fable",
     agent: "claude",
-    description:
-      "Fable 5 · Most capable for your hardest and longest-running tasks · $10/$50 per Mtok",
+    description: "Highest-capability Claude route for the hardest and longest-running tasks.",
+    intelligenceTier: "fable",
+    routingSummary: "hardest long-running work; highest Claude cost; explicit request only",
+    accessCost: CLAUDE_MAX_20X_COST,
+    marginalTokenCost:
+      "Claude API/extra usage: $10/M input, $12.50/M cache write, $1/M cache read, $50/M output",
+    estimatedUsage: `${CLAUDE_COMMUNITY_SESSION}; Fable drains weighted quota fastest, so use only on explicit user request. No stable per-model token cap is published.`,
   },
   {
     model: "opus",
     agent: "claude",
-    description: "Opus 4.8 with 1M context · Best for everyday, complex tasks · $5/$25 per Mtok",
+    description: "High-capability Claude route for complex reasoning and tool-heavy work.",
+    intelligenceTier: "opus",
+    routingSummary: "complex reasoning and tool-heavy work; about 2.5x Sonnet marginal cost",
+    accessCost: CLAUDE_MAX_20X_COST,
+    marginalTokenCost:
+      "Claude API/extra usage: $5/M input, $6.25/M cache write, $0.50/M cache read, $25/M output",
+    estimatedUsage: `${CLAUDE_COMMUNITY_SESSION}; Opus uses the shared all-model weekly pool more quickly than Sonnet. No stable per-model token cap is published.`,
   },
   {
     model: "sonnet",
     agent: "claude",
-    description: "Sonnet 5 · Efficient for routine tasks · $2/$10 per Mtok · promo through Aug 31",
+    description: "Default Claude route for capable, efficient everyday work.",
+    intelligenceTier: "sonnet",
+    routingSummary: "capable everyday default; lowest Claude model cost",
+    accessCost: CLAUDE_MAX_20X_COST,
+    marginalTokenCost:
+      "Claude API/extra usage introductory rate: $2/M input, $2.50/M cache write, $0.20/M cache read, $10/M output through 2026-08-31; then $3/M input and $15/M output",
+    estimatedUsage: `${CLAUDE_COMMUNITY_SESSION}; Sonnet also has a separate weekly allowance and normally provides the highest Claude throughput. No stable weekly token cap is published.`,
   },
   {
     model: "deepseek-pro",
     agent: "maestro",
-    description: "Sonnet-level performance at lower cost · Best for everyday tasks over coding",
-  },
-  {
-    model: "deepseek-flash",
-    agent: "maestro",
-    description: "DeepSeek V4 Flash · Fast and low-cost for lightweight everyday tasks",
+    description: "API-priced Sonnet-level route for cost-efficient everyday work.",
+    intelligenceTier: "sonnet",
+    routingSummary: "cost-efficient everyday work; pay-per-token and cheapest route",
+    accessCost: "DeepSeek V4 Pro pay-as-you-go API; no monthly subscription required",
+    marginalTokenCost:
+      "DeepSeek API: $0.435/M uncached input, $0.003625/M cached input, $0.87/M output",
+    estimatedUsage:
+      "No subscription token cap; pay per token. Official account concurrency limit is 500 requests.",
   },
 ];
 
+export function formatSelectableModel(candidate: SelectableModel): string {
+  const tier = `${candidate.intelligenceTier[0].toUpperCase()}${candidate.intelligenceTier.slice(1)}`;
+  return `${candidate.agent} / \`${candidate.model}\` [${tier}-level]: ${candidate.routingSummary}`;
+}
+
 export function selectableModel(value: string): SelectableModel | undefined {
   const normalized = value.trim().toLowerCase();
   return SELECTABLE_MODELS.find((candidate) => candidate.model === normalized);

package/dist/runtime/src/mcp/factories/vault.ts

@@ -7,6 +7,7 @@ import type { VaultCredentialHost } from "./vault-host";
 
 export interface VaultMcpContext {
   userId?: string;
+  listOnly?: boolean;
   httpOnly?: boolean;
   cwd?: string;
 }
@@ -50,7 +51,7 @@ export function createVaultMcpServer(
     },
   );
 
-  if (!context.httpOnly) {
+  if (!context.listOnly && !context.httpOnly) {
     server.tool(
       "vault_run",
       "Run a shell command containing {{KEY}} references inside Otium's credential broker. Expanded command input never reaches the model/provider, and stdout/stderr are redacted before return. Prefer vault_http_request for HTTP APIs.",
@@ -86,39 +87,40 @@ export function createVaultMcpServer(
     );
   }
 
-  server.tool(
-    "vault_http_request",
-    "Make an HTTPS request with {{KEY}} references resolved inside Otium. Put secrets in headers or body, never in the URL. The expanded request is not returned; the response is redacted before the model sees it.",
-    {
-      method: z.enum(["GET", "POST", "PUT", "PATCH", "DELETE"]).default("GET"),
-      url: z.string().url().describe("Absolute HTTPS URL without Vault placeholders"),
-      headers: z.record(z.string(), z.string()).optional(),
-      body: z.string().optional(),
-      timeout_ms: z.number().int().min(1_000).max(120_000).optional(),
-      max_response_bytes: z
-        .number()
-        .int()
-        .min(1_024)
-        .max(1024 * 1024)
-        .optional(),
-    },
-    async ({ method, url, headers, body, timeout_ms, max_response_bytes }) => {
-      if (!context.userId) return mcpError("Vault unavailable: no user context");
-      const result = await http(
-        context.userId,
-        {
-          method,
-          url,
-          headers,
-          body,
-          timeoutMs: timeout_ms,
-          maxResponseBytes: max_response_bytes,
-        },
-        host,
-      );
-      return result.error ? mcpError(result.error) : mcpOk(JSON.stringify(result, null, 2));
-    },
-  );
+  if (!context.listOnly)
+    server.tool(
+      "vault_http_request",
+      "Make an HTTPS request with {{KEY}} references resolved inside Otium. Put secrets in headers or body, never in the URL. The expanded request is not returned; the response is redacted before the model sees it.",
+      {
+        method: z.enum(["GET", "POST", "PUT", "PATCH", "DELETE"]).default("GET"),
+        url: z.string().url().describe("Absolute HTTPS URL without Vault placeholders"),
+        headers: z.record(z.string(), z.string()).optional(),
+        body: z.string().optional(),
+        timeout_ms: z.number().int().min(1_000).max(120_000).optional(),
+        max_response_bytes: z
+          .number()
+          .int()
+          .min(1_024)
+          .max(1024 * 1024)
+          .optional(),
+      },
+      async ({ method, url, headers, body, timeout_ms, max_response_bytes }) => {
+        if (!context.userId) return mcpError("Vault unavailable: no user context");
+        const result = await http(
+          context.userId,
+          {
+            method,
+            url,
+            headers,
+            body,
+            timeoutMs: timeout_ms,
+            maxResponseBytes: max_response_bytes,
+          },
+          host,
+        );
+        return result.error ? mcpError(result.error) : mcpOk(JSON.stringify(result, null, 2));
+      },
+    );
 
   return server;
 }

package/dist/runtime/src/mcp/vault-server.ts

@@ -19,6 +19,7 @@ await connectStdio(
   createVaultMcpServer(
     {
       userId: parseUserIdArg(args),
+      listOnly: args.includes("--list-only=true"),
       httpOnly: args.includes("--http-only=true"),
       cwd: process.cwd(),
     },

package/dist/runtime/src/platform/mcp-config.ts

@@ -455,15 +455,11 @@ const MCP_CATALOG: Record<string, RuntimeMcpCatalogEntry> = {
   },
   vault: {
     ...commonRuntimeMcpPolicy("vault"),
-    // Credential references must never dead-end because a topic whitelist
-    // omitted the broker while provider hooks correctly block raw expansion.
     build({ userId, agent }) {
-      const args = [`--user-id=${userId}`];
-      // Codex has no host-side PostToolUse redaction. Keep its Vault surface
-      // broker-only and non-persistent: HTTPS responses are scrubbed before the
-      // MCP result reaches Codex, while arbitrary shell could write a secret to
-      // disk and reveal it through a later Read.
-      if (agent === "codex") args.push("--http-only=true");
+      // Normal turns use {{KEY}} directly in tool inputs. Keep the Vault MCP
+      // surface focused on key discovery; broker tools remain available from
+      // the public factory for compatibility but are no longer the default UX.
+      const args = [`--user-id=${userId}`, "--list-only=true"];
       return buildStdioMcpServer(agent, VAULT_SERVER, args);
     },
   },

package/dist/runtime/src/platform/playwright/manager.ts

@@ -726,6 +726,10 @@ async function spawnPlaywright(
     // browser tools would silently disappear from maestro turns.
     "--host",
     "127.0.0.1",
+    // Keep the browser visible. Do not rely on the launcher's implicit
+    // default: an explicit flag prevents wrapper/upstream default changes
+    // from silently switching topic browsers back to headless mode.
+    "--headed",
     "--user-data-dir",
     userDataDir,
     // NOTE: mcp-patchright throws on unknown CLI args, so the old
@@ -736,8 +740,6 @@ async function spawnPlaywright(
     //     (requires real Google Chrome at /opt/google/chrome/chrome).
     //   --init-script <stealth>  → unneeded; Patchright is stealth by default.
   ];
-  // Headed by default, matching clawgram's production behavior.
-
   // Pass the egress proxy to the child through the environment rather than
   // argv so the credentials never surface in `ps`/`/proc` command lines. The
   // launcher (scripts/mcp-patchright-http.mjs) reads these NEGOTIUM_BROWSER_PROXY_*
@@ -747,6 +749,7 @@ async function spawnPlaywright(
   const childEnv = {
     ...process.env,
     NEGOTIUM_BROWSER_CAPABILITY: capability,
+    NEGOTIUM_BROWSER_VAULT_USER_ID: userId,
     ...(proxy
       ? {
           NEGOTIUM_BROWSER_PROXY_SERVER: proxy.server,

package/dist/runtime/src/prompts/builders.ts

@@ -1,8 +1,13 @@
 import { readFileSync } from "node:fs";
 import { resolve } from "node:path";
+import {
+  formatSelectableModel,
+  MODEL_COST_ROUTING_SUMMARY,
+  SELECTABLE_MODELS,
+} from "#agents/model-catalog";
 import { AGENTS_PROMPTS_DIR, PROJECT_ROOT, RESOURCES_DIR } from "#platform/config";
 import { logger } from "#platform/logger";
-import type { AgentKind } from "#types";
+import type { AgentKind, EffortLevel } from "#types";
 
 const PROMPTS_DIR = resolve(PROJECT_ROOT, "src/prompts");
 const SESSIONS_DIR = resolve(PROMPTS_DIR, "sessions");
@@ -151,6 +156,10 @@ export interface SessionSystemPromptOpts {
   topicTitle: string;
   workspaceCwd: string;
   agentKind: AgentKind;
+  /** Resolved model actually used for this turn. */
+  currentModel?: string;
+  /** Resolved effort actually used for this turn; absent means provider default/off. */
+  currentEffort?: EffortLevel;
   description?: string | null;
   /** True only for top-level agent rooms — the runtime spawn_subagent tool is registered there. */
   canSpawnSubagents?: boolean;
@@ -165,6 +174,8 @@ function buildRuntimeToolSection(
   canSpawnSubagents = false,
   visualTools = false,
   fileDeliveryTools = false,
+  currentModel?: string,
+  currentEffort?: EffortLevel,
 ): string {
   const runtimeNamespace = "mcp__runtime";
   const taskNamespace = "mcp__task";
@@ -262,20 +273,28 @@ function buildRuntimeToolSection(
     ...spawnSubagentSection,
   ];
 
+  const modelCatalog = SELECTABLE_MODELS.map(
+    (candidate) => `- ${formatSelectableModel(candidate)}`,
+  );
   const topicConfig = [
     "",
     "## Topic Configuration (model / agent / effort)",
+    `Current execution: agent=\`${agentKind}\`, model=\`${currentModel ?? "unknown (call get_model)"}\`, effort=\`${currentEffort ?? "provider default/off"}\`.`,
     "The user's configured agent/model/effort is intentional. Preserve it by default.",
     `When the user explicitly asks to change the model, agent backend, or reasoning effort for THIS topic, call "${runtimeNamespace}__set_model", "${runtimeNamespace}__set_agent", or "${runtimeNamespace}__set_effort". The change applies from your NEXT turn. After calling, briefly confirm and the system will continue with the new setting.`,
     "`set_effort` is available but discouraged; use it only when the user explicitly requests an effort change.",
-    "`set_model` may be called autonomously only when the current model is clearly below the task's required capability, such as complex algorithm design, proof-level math, or broad multi-file refactoring. In that case, move up one step within the same agent and end the turn. Do not use vague task complexity as a trigger.",
+    "`set_model` may be called autonomously only when the current model is clearly below the task's required capability, such as complex algorithm design, proof-level math, or broad multi-file refactoring. Choose the best-fit model directly from the same-agent catalog; model selection is not a mandatory one-step ladder. End the turn after changing it. Do not use vague task complexity as a trigger.",
     "`set_agent` autonomous calls are forbidden. Only switch agent when the user explicitly asks to switch runtime, e.g. “codex로 가”, “claude로 바꿔”.",
     "Never use `fable` unless the user explicitly requests it; it is expensive.",
     "",
-    "Accepted values:",
-    "- claude models: `sonnet` (default), `opus`, `fable`; efforts: `low`, `medium`, `high`, `xhigh`, `max`.",
-    "- codex models: `gpt-5.6-luna` (default), `gpt-5.6-terra`, `gpt-5.6-sol`; efforts: `low`, `medium`, `high`, `xhigh`, `max`.",
-    "- maestro models: `deepseek-pro` (default), `deepseek-flash`, `deepseek`; efforts: `low`, `medium`, `high`, `xhigh`, `max`.",
+    "Model catalog (capability/cost routing guidance):",
+    MODEL_COST_ROUTING_SUMMARY,
+    ...modelCatalog,
+    "",
+    "Accepted effort values:",
+    "- claude: `low`, `medium`, `high`, `xhigh`, `max`.",
+    "- codex: `low`, `medium`, `high`, `xhigh`, `max`.",
+    "- maestro: `low`, `medium`, `high`, `xhigh`, `max`.",
     "Agent guidance when the user explicitly asks to switch: `codex` for deepest reasoning and complex code/math; `claude` for tool-heavy MCP/file automation; `maestro` for inexpensive fast drafts and lighter experiments.",
   ];
 
@@ -306,6 +325,8 @@ export function buildTopicSystemPrompt(opts: SessionSystemPromptOpts): string {
       opts.canSpawnSubagents,
       opts.visualTools,
       opts.fileDeliveryTools,
+      opts.currentModel,
+      opts.currentEffort,
     );
   if (opts.description?.trim()) {
     prompt += `\n\n## Topic-Specific Instructions\n${opts.description.trim()}`;
@@ -321,7 +342,15 @@ export function buildChannelSystemPrompt(opts: SessionSystemPromptOpts): string
       TOPIC_TITLE: opts.topicTitle,
       WORKSPACE_CWD: opts.workspaceCwd,
       UPLOADS_DIR: uploadsDir,
-    }) + buildRuntimeToolSection(opts.agentKind, false, opts.visualTools, opts.fileDeliveryTools)
+    }) +
+    buildRuntimeToolSection(
+      opts.agentKind,
+      false,
+      opts.visualTools,
+      opts.fileDeliveryTools,
+      opts.currentModel,
+      opts.currentEffort,
+    )
   );
 }
 

package/dist/runtime/src/runtime/turn-runner.ts

@@ -1897,6 +1897,8 @@ export function startAiTurn(params: StartAiTurnParams): string | null {
     topicTitle: topic.title,
     workspaceCwd,
     agentKind,
+    currentModel: resolvedModel,
+    currentEffort: resolvedEffort,
     description: topic.description,
     canSpawnSubagents:
       peerBridge?.canSpawnSubagents ?? (topicRecord?.kind === "agent" && !topicRecord.isSubagent),

package/dist/runtime/src/storage/topic-archive.ts

@@ -1,5 +1,6 @@
 import { mkdirSync, writeFileSync } from "node:fs";
 import { join } from "node:path";
+import { countMemoryArchiveExchanges } from "#agents/memory-archive-policy";
 import { logger } from "#platform/logger";
 import { sanitizeTopicName } from "#security/sanitize";
 import { getAllMessagesForTopic, getMessagesForTopicAfterRowid } from "#storage/api-messages";
@@ -9,6 +10,7 @@ import { getSharedWikiDir } from "#storage/wiki";
 export interface TopicArchiveResult {
   path: string;
   messageCount: number;
+  exchangeCount: number;
   lastRowid: number;
 }
 
@@ -71,9 +73,10 @@ export function archiveTopicMessages(
     }
   }
 
+  const exchangeCount = countMemoryArchiveExchanges(rows);
   logger.info(
-    { topicId, topicTitle, archive: path, messageCount: rows.length, lastRowid },
+    { topicId, topicTitle, archive: path, messageCount: rows.length, exchangeCount, lastRowid },
     "archiveTopicMessages: archived topic messages",
   );
-  return { path, messageCount: rows.length, lastRowid };
+  return { path, messageCount: rows.length, exchangeCount, lastRowid };
 }

package/dist/runtime/src/topics/lifecycle.ts

@@ -9,6 +9,7 @@ import { rmSync } from "node:fs";
 import { runArchiverTurn } from "#agents/archiver";
 import { cancelIdleArchiveForTopic } from "#agents/idle-archiver";
 import { cancelSubagentWatchForDeletedTopic } from "#agents/mcp-tools/spawn-subagent";
+import { MIN_MEMORY_ARCHIVE_EXCHANGES } from "#agents/memory-archive-policy";
 import { type PurgeSessionRef, purgeTopicLogs } from "#agents/topic-cleanup";
 import { WsHub } from "#bus";
 import { killBgBash } from "#platform/background-bash/manager";
@@ -210,7 +211,7 @@ async function deleteTopicCascadeImpl(
       }
     }
 
-    if (archived) {
+    if (archived && archived.exchangeCount >= MIN_MEMORY_ARCHIVE_EXCHANGES) {
       try {
         const memoryTopic = getTopicMemoryOrigin(topicId) ?? topic;
         runArchiverTurn({

package/dist/runtime/src/topics/session.ts

@@ -3,6 +3,7 @@
 import { randomUUID } from "node:crypto";
 import { archiveActiveTopicForMemory, cancelIdleArchiveForTopic } from "#agents/idle-archiver";
 import { runAgent } from "#agents/index";
+import { MIN_MEMORY_ARCHIVE_EXCHANGES } from "#agents/memory-archive-policy";
 import { resolveModelForAgent } from "#agents/model-catalog";
 import { getRegistry } from "#agents/registry";
 import { cleanupTopicRollouts, purgeTopicLogs } from "#agents/topic-cleanup";
@@ -119,6 +120,7 @@ export async function restartTopicSession(
     (options.archiveMemory ?? archiveActiveTopicForMemory)(topicId, userId, {
       reason: "reset",
       minMessages: 1,
+      minExchanges: MIN_MEMORY_ARCHIVE_EXCHANGES,
       allowMentionOnly: true,
       skipBusyCheck: true,
     });

package/dist/runtime/src/version.ts

@@ -1 +1 @@
-export const NEGOTIUM_VERSION = "0.1.19";
+export const NEGOTIUM_VERSION = "0.1.21";

package/dist/storage.js

@@ -2823,6 +2823,25 @@ __export(exports_topic_archive, {
 import { mkdirSync as mkdirSync9, writeFileSync as writeFileSync7 } from "fs";
 import { join as join9 } from "path";
 
+// ../../packages/core/src/agents/memory-archive-policy.ts
+function countMemoryArchiveExchanges(rows) {
+  let waitingForAssistant = false;
+  let completed = 0;
+  for (const row of rows) {
+    const assistant = row.author_id === "ai" || Boolean(row.agent_type);
+    const conversational = row.kind == null || row.kind === "message";
+    if (!assistant && row.author_id !== "system" && conversational) {
+      waitingForAssistant = true;
+      continue;
+    }
+    if (assistant && conversational && waitingForAssistant) {
+      completed++;
+      waitingForAssistant = false;
+    }
+  }
+  return completed;
+}
+
 // ../../packages/core/src/storage/topic-transcript.ts
 var exports_topic_transcript = {};
 __export(exports_topic_transcript, {
@@ -2966,8 +2985,9 @@ function archiveTopicMessages(topicId, topicTitle, options = {}) {
       counter++;
     }
   }
-  logger.info({ topicId, topicTitle, archive: path, messageCount: rows.length, lastRowid }, "archiveTopicMessages: archived topic messages");
-  return { path, messageCount: rows.length, lastRowid };
+  const exchangeCount = countMemoryArchiveExchanges(rows);
+  logger.info({ topicId, topicTitle, archive: path, messageCount: rows.length, exchangeCount, lastRowid }, "archiveTopicMessages: archived topic messages");
+  return { path, messageCount: rows.length, exchangeCount, lastRowid };
 }
 // ../../packages/core/src/storage/topic-archive-state.ts
 var exports_topic_archive_state = {};
@@ -3340,4 +3360,4 @@ export {
   DEFAULT_AI_NAME
 };
 
-//# debugId=AE15B7A45D14E36264756E2164756E21
+//# debugId=957BDB73C3C8AF8A64756E2164756E21