@llamaventures/cli 1.14.0 → 1.15.0

package/AGENT_BRIEFING.md

@@ -20,6 +20,17 @@ Most teammates don't know everything this CLI can do. Part of your job is to sur
 - **Point at `llama --help`** for the full surface rather than reciting it. The CLI uses progressive help: `llama --help` is a short overview, `llama <area> --help` drills in.
 - **Stay current.** If you suspect the CLI is stale, run `llama version --check`; if it reports an upgrade, tell the user the one-line `npm i -g @llamaventures/cli@latest` command. Don't nag repeatedly.
 
+## Runtime skill library
+
+This npm package is public, but Llama OS skills are private. Do not assume the skill text is bundled locally. For team-token sessions, discover the live runtime library through Llama Command:
+
+- Start with `llama agent bootstrap` or MCP `agent_bootstrap` when you need the current Command + Llama OS contract.
+- Use `llama skills search "<task>"` or MCP `skills_search` before choosing a Llama workflow.
+- Use `llama skills show <slug>` or MCP `skills_read` only for the relevant skill.
+- Use `llama explain <command-url-or-object>` or MCP `object_inspect` for 404s, deleted wiki pages, notifier links, deal URLs, and unknown Command objects before telling the user "the system is broken."
+
+The boundary matters: public CLI/MCP discovers skills, but authenticated Command decides which skill content the token may read.
+
 ## Pipeline First (hard rule)
 
 Any time the user mentions a company name or founder name:
@@ -259,7 +270,7 @@ Run `llama --help` for the full surface (~40 commands).
 
 ## MCP-native agents
 
-If you support [MCP](https://modelcontextprotocol.io), **prefer the MCP server over parsing CLI output.** The same package ships `llama-mcp` (20 typed tools, identical auth chain).
+If you support [MCP](https://modelcontextprotocol.io), **prefer the MCP server over parsing CLI output.** The same package ships `llama-mcp` (56 typed tools, identical auth chain).
 
 Add to your MCP client config (Claude Desktop / Claude Code / Cursor / OpenClaw / Codex / etc.):
 
@@ -270,6 +281,9 @@ Add to your MCP client config (Claude Desktop / Claude Code / Cursor / OpenClaw
 Tools available:
 
 - `auth_status` — verify creds + identity (call first if anything 401s)
+- `agent_bootstrap` — fetch the live Command + Llama OS runtime manifest
+- `skills_search` / `skills_read` — discover and read authenticated runtime skills
+- `object_inspect` — explain Command URLs, 404s, deleted objects, and lifecycle trail
 - `deal_search` / `deal_show` / `deal_create` / `deal_update`
 - `brief_blocks` / `brief_add_text` / `brief_add_link` / `brief_add_callout`
 - `wiki_search` / `wiki_save` (accepts `content_type: 'markdown' | 'html'` — HTML entries render as full-viewport sandboxed iframe at `/wiki/<slug>`) / `wiki_delete` / `wiki_restore` (soft-delete, reversible)

package/CHANGELOG.md

@@ -6,6 +6,40 @@ this project adheres to [Semantic Versioning](https://semver.org).
 
 ## [Unreleased]
 
+## [1.15.0] — 2026-06-15
+
+### Added
+- **`llama agent bootstrap`** — fetches the live Llama Command + Llama OS
+  runtime manifest, including the authenticated skill bundle metadata and
+  object-inspection contract.
+- **`llama skills search|show`** — discovers and reads runtime Llama OS skills
+  from Llama Command. The public npm package does not bundle private skill
+  content; Command returns only what the authenticated token may see.
+- **`llama explain <url-or-object>`** — asks Llama Command to explain URLs,
+  deleted objects, 404s, and lifecycle history before an agent guesses that
+  the system is broken.
+- MCP parity tools: **`agent_bootstrap`**, **`skills_search`**,
+  **`skills_read`**, and **`object_inspect`**.
+
+### Changed
+- `AGENT_BRIEFING.md` now teaches agents to use the authenticated runtime
+  skill gateway instead of assuming a local private `llama-os` checkout.
+
+## [1.14.1] — 2026-06-15
+
+### Added
+- **`llama deal agent run <dealId> --message "..."`** — starts Llama
+  Command's server-side Deal Agent in a deal thread, so the service agent can
+  execute deal-scoped work instead of the local CLI user doing it.
+- **`deal_agent_run` MCP tool** — the same narrow server-agent trigger for
+  MCP-native clients, without adding a generic API passthrough.
+
+### Changed
+- **`llama deal enrich <dealId> --apply --executor server_agent`** now starts
+  the server-side Deal Agent unless `--harness-only` is supplied. Dry-runs and
+  external-agent handoff prompts still use the enrichment harness endpoint.
+- **`deal_enrich` MCP tool** now mirrors the same behavior with `harnessOnly`.
+
 ## [1.14.0] — 2026-06-15
 
 ### Added

package/CONTRIBUTING.md

@@ -13,6 +13,7 @@ npm ci
 node bin/llama.mjs --help        # CLI is ESM, runs straight from source
 node --check bin/llama.mjs        # syntax check
 node --check bin/llama-mcp.mjs    # syntax check
+npm test                          # mock-backed CLI/MCP agent routing checks
 ```
 
 To test the CLI end-to-end against your own credentials, point it at the
@@ -33,8 +34,8 @@ printf '%s\n' \
   | node bin/llama-mcp.mjs | head -200
 ```
 
-You should see 51 named tools, including `deal_enrich`, the 5 `pitch_*`, and
-no generic API passthrough tool.
+You should see 52 named tools, including `deal_agent_run`, `deal_enrich`, the
+5 `pitch_*`, and no generic API passthrough tool.
 
 ## Conventions
 

package/README.md

@@ -48,7 +48,7 @@
 ```
 @llamaventures/cli
 ├── bin/llama          interactive CLI for humans + bash
-└── bin/llama-mcp      stdio MCP server, 51 typed tools — for any MCP-native agent
+└── bin/llama-mcp      stdio MCP server, 56 typed tools — for any MCP-native agent
 ```
 
 Both binaries share `lib/client.mjs` — the **same** auth chain, **same** HTTP
@@ -187,7 +187,9 @@ llama deal show <dealId>
 llama deal create "Acme AI" --description "..." --source Gavin
 llama deal update <dealId> status Diligence
 llama deal enrich <dealId> --dry-run
+llama deal enrich <dealId> --apply --executor server_agent
 llama deal enrich <dealId> --executor external_agent --prompt
+llama deal agent run <dealId> --message "collect founder evidence and update typed facts"
 llama deal delete  <dealId>     # soft (audit-logged)
 llama deal restore <dealId>
 
@@ -209,6 +211,12 @@ llama approvals   decide <approvalId> approved --note "..."
 llama timeline <dealId>
 llama post     <dealId> "message body" [--link url]
 
+# Agent runtime — live Command + private Llama OS skill gateway
+llama agent bootstrap
+llama skills search "wiki delete tombstone"
+llama skills show llama-command
+llama explain https://command.llamaventures.vc/wiki/some-page
+
 # Wiki
 llama wiki search "<query>"
 llama wiki read   <slug> [--lang en|zh]
@@ -247,13 +255,15 @@ agents can pattern-match without parsing prose.
 ## MCP server
 
 The bundled `llama-mcp` is a **stdio Model Context Protocol** server exposing
-**51 typed tools** that mirror the most-used CLI surface. Every tool is named
+**56 typed tools** that mirror the most-used CLI surface. Every tool is named
 and scoped — there is no generic API passthrough, by design (a public-package
 escape hatch reachable from a prompt-injectable agent context is exactly the
 shape we want to avoid).
 
 Coverage is grouped around the workflows agents actually need: auth
-diagnostics; deal search/show/create/update/feed; deal enrichment harnesses;
+diagnostics; live agent bootstrap; authenticated Llama OS skill search/read;
+Command URL/object inspection; deal search/show/create/update/feed; server-side deal agent runs;
+deal enrichment harnesses;
 trust-rated facts; brief blocks and version history; wiki read/write/delete/restore;
 timeline posts and mentions; skill corrections; refresh triggers; external pitch intake;
 memo show/regenerate/save/reset; and deal-scoped HTML docs, versions, bundles, and
@@ -274,6 +284,11 @@ The `agent_briefing` MCP **prompt** also returns
 [`AGENT_BRIEFING.md`](AGENT_BRIEFING.md) verbatim, so any new agent loading the
 server can self-onboard without leaving the protocol.
 
+For current Llama OS skills, use the runtime tools instead of looking for a
+local private repo: `agent_bootstrap`, `skills_search`, `skills_read`, and
+`object_inspect`. The public npm package does not bundle private skill text;
+Command returns only the content visible to the authenticated token.
+
 ### Wire into your agent
 
 <details open>

package/README.zh-CN.md

@@ -48,7 +48,7 @@
 ```
 @llamaventures/cli
 ├── bin/llama          给人 + bash 用的交互式 CLI
-└── bin/llama-mcp      给 MCP 原生 agent 用的 stdio MCP server，51 个工具
+└── bin/llama-mcp      给 MCP 原生 agent 用的 stdio MCP server，56 个工具
 ```
 
 两个 binary 共享 `lib/client.mjs`——**同一**认证链、**同一** HTTP 客户端、
@@ -208,7 +208,9 @@ llama deal show <dealId>
 llama deal create "Acme AI" --description "..." --source Gavin
 llama deal update <dealId> status Diligence
 llama deal enrich <dealId> --dry-run
+llama deal enrich <dealId> --apply --executor server_agent
 llama deal enrich <dealId> --executor external_agent --prompt
+llama deal agent run <dealId> --message "collect founder evidence and update typed facts"
 llama deal delete  <dealId>     # 软删除（审计日志记录）
 llama deal restore <dealId>
 
@@ -230,6 +232,12 @@ llama approvals   decide <approvalId> approved --note "..."
 llama timeline <dealId>
 llama post     <dealId> "消息内容" [--link url]
 
+# Agent runtime——Command 上的 Llama OS skill gateway
+llama agent bootstrap
+llama skills search "wiki delete tombstone"
+llama skills show llama-command
+llama explain https://command.llamaventures.vc/wiki/some-page
+
 # Wiki
 llama wiki search "<query>"
 llama wiki save <slug> --title "..." --content "..."
@@ -259,12 +267,13 @@ MCP server 在 `isError: true` 内容里返回相同的前缀，agent 不用解
 ## MCP server
 
 随包发布的 `llama-mcp` 是一个 **stdio Model Context Protocol** server，
-暴露 **51 个 typed tools**——基本镜像 CLI 最常用的命令。每个 tool 都是
+暴露 **56 个 typed tools**——基本镜像 CLI 最常用的命令。每个 tool 都是
 具名、scoped 的；**没有**通用的 API passthrough，这是有意设计的（公开
 package 里一个能被 prompt-injection 触达的逃生通道，正是我们要避开的形状）。
 
-覆盖面按 agent 真正会用的工作流分组：auth 诊断；deal search/show/create/
-update/feed；deal enrichment harness；带 trust ladder 的事实；brief blocks
+覆盖面按 agent 真正会用的工作流分组：auth 诊断；live agent bootstrap；
+授权后的 Llama OS skill search/read；Command URL/object inspect；deal search/show/create/
+update/feed；服务端 deal agent run；deal enrichment harness；带 trust ladder 的事实；brief blocks
 和版本历史；wiki 读写删恢复；timeline posts 和 mentions；skill corrections；
 refresh triggers；外部 pitch intake；memo show/regenerate/save/reset；
 以及 deal-scoped HTML docs、versions、bundles、restore/reset。
@@ -284,6 +293,10 @@ printf '%s\n' \
 [`AGENT_BRIEFING.md`](AGENT_BRIEFING.md)——刚装上 server 的 agent
 不用离开协议就能给自己 onboard。
 
+要读当前 Llama OS skills，用 runtime tools：`agent_bootstrap`、
+`skills_search`、`skills_read`、`object_inspect`。公开 npm 包不打包私有
+skill 正文；Command 只按当前 token 的权限返回可见内容。
+
 ### 接到你的 agent 上
 
 <details open>

package/bin/llama-mcp.mjs

@@ -11,7 +11,7 @@ import { createRequire } from "module";
 import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
 import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
 import { z } from "zod";
-import { getAuthHeaders, readBriefing, request } from "../lib/client.mjs";
+import { getAuthHeaders, readBriefing, request, requestSse } from "../lib/client.mjs";
 
 const requireFromHere = createRequire(import.meta.url);
 const { version: PKG_VERSION } = requireFromHere("../package.json");
@@ -40,6 +40,96 @@ async function callApi(method, path, body) {
   }
 }
 
+function textResult(text, isError = false) {
+  return {
+    content: [{ type: "text", text }],
+    ...(isError ? { isError: true } : {}),
+  };
+}
+
+function splitSources(value) {
+  if (Array.isArray(value)) return value.filter(Boolean);
+  if (!value || value === true) return undefined;
+  return String(value)
+    .split(",")
+    .map((s) => s.trim())
+    .filter(Boolean);
+}
+
+function buildEnrichmentAgentMessage(args = {}) {
+  if (args.message) return String(args.message);
+  const sources = splitSources(args.sources) ?? [
+    "website",
+    "github",
+    "linkedin",
+    "yc",
+    "launch",
+    "web",
+    "monid",
+  ];
+  const budget = args.budgetCents ?? "50";
+  const memo = args.generateMemo
+    ? "Generate memo only after enrichment because the caller explicitly requested it."
+    : "Do not generate memo.";
+  return [
+    "Run server-side deal enrichment for this deal.",
+    `Use sources: ${sources.join(", ")}.`,
+    `Private Monid budget cap: ${budget} cents.`,
+    "Read the enrichment harness first, then collect current company/founder evidence.",
+    "Write canonical evidence links, sourced deal facts, stable deal fields, and typed factual values where supported.",
+    "For typed factual values, call read_typed_factual_layer first and use upsert_typed_fact for queryable fields.",
+    "Search snippets alone are not high-confidence evidence; fetch direct sources where possible.",
+    memo,
+    "End with what was written, what was skipped, and open questions.",
+  ].join(" ");
+}
+
+function summarizeAgentEvents(events = []) {
+  return events
+    .flatMap((event) => {
+      if (event.tool_use?.name) return [{ type: "tool_use", name: event.tool_use.name }];
+      if (event.tool_result?.name) {
+        return [
+          {
+            type: "tool_result",
+            name: event.tool_result.name,
+            ok: event.tool_result.ok ?? null,
+            summary: event.tool_result.summary ?? null,
+          },
+        ];
+      }
+      if (event.error) return [{ type: "error", error: String(event.error) }];
+      return [];
+    })
+    .slice(-80);
+}
+
+async function runDealAgentTool({ dealId, message, title = "MCP agent run" }) {
+  try {
+    const thread = await request("POST", `/api/deals/${encodeURIComponent(dealId)}/threads`, { title });
+    if (!thread?.id) throw new Error("Thread creation did not return an id");
+    const result = await requestSse(
+      "POST",
+      `/api/deals/${encodeURIComponent(dealId)}/threads/${encodeURIComponent(thread.id)}`,
+      { message },
+    );
+    return textResult(
+      JSON.stringify(
+        {
+          ok: true,
+          threadId: thread.id,
+          text: result.text,
+          toolEvents: summarizeAgentEvents(result.events),
+        },
+        null,
+        2,
+      ),
+    );
+  } catch (err) {
+    return textResult(`Error: ${err?.message ?? String(err)}`, true);
+  }
+}
+
 // Append a block to a deal brief. The /blocks route only accepts atomic
 // full-array PUTs (no POST), so we GET current blocks, prepend the new
 // one (matches UI default since 2026-05-03), and PUT the merged array.
@@ -105,6 +195,82 @@ server.registerTool(
   }
 );
 
+server.registerTool(
+  "agent_bootstrap",
+  {
+    description:
+      "Fetch the live Llama Command + Llama OS runtime manifest. Use this at " +
+      "the start of an agent session to discover current skills, the skill " +
+      "bundle version, and the object-inspection contract. Unlike the bundled " +
+      "agent_briefing prompt, this comes from authenticated Command runtime.",
+    inputSchema: {
+      limit: z.number().optional().describe("number of skill summaries to include; default 25"),
+    },
+  },
+  async ({ limit } = {}) => {
+    const params = new URLSearchParams();
+    if (limit) params.set("limit", String(limit));
+    return callApi("GET", `/api/agent/manifest${params.toString() ? `?${params}` : ""}`);
+  }
+);
+
+server.registerTool(
+  "skills_search",
+  {
+    description:
+      "Search the authenticated Llama OS runtime skill library. Call this " +
+      "before choosing a workflow for Llama pipeline/wiki/DD/research/ops tasks. " +
+      "Returns summaries only; call skills_read for the exact SKILL.md.",
+    inputSchema: {
+      q: z.string().describe("workflow/task query, e.g. 'wiki delete tombstone' or 'deal DD memo'"),
+      limit: z.number().optional().describe("default 20"),
+    },
+  },
+  async ({ q, limit }) => {
+    const params = new URLSearchParams({ q });
+    if (limit) params.set("limit", String(limit));
+    return callApi("GET", `/api/agent/skills?${params}`);
+  }
+);
+
+server.registerTool(
+  "skills_read",
+  {
+    description:
+      "Read one runtime Llama OS skill by slug. Use after skills_search. " +
+      "Returns the full SKILL.md content from Llama Command; public npm does " +
+      "not bundle private skill text.",
+    inputSchema: {
+      slug: z.string().describe("skill slug, e.g. llama-command or llama-wiki"),
+    },
+  },
+  async ({ slug }) => callApi("GET", `/api/agent/skills/${encodeURIComponent(slug)}`)
+);
+
+server.registerTool(
+  "object_inspect",
+  {
+    description:
+      "Explain a Llama Command URL or object id. Use for 404s, deleted wiki " +
+      "pages, notifier links, deal URLs, brief blocks, HTML docs, and unknown " +
+      "Command objects before guessing that the system is broken.",
+    inputSchema: {
+      q: z.string().optional().describe("URL or compact query, e.g. wiki:my-slug or a Command URL"),
+      type: z.string().optional().describe("explicit object type if not using q"),
+      id: z.string().optional().describe("explicit object id if not using q"),
+      lang: z.enum(["en", "zh"]).optional().describe("wiki language; default en"),
+    },
+  },
+  async ({ q, type, id, lang } = {}) => {
+    const params = new URLSearchParams();
+    if (q) params.set("q", q);
+    if (type) params.set("type", type);
+    if (id) params.set("id", id);
+    if (lang) params.set("lang", lang);
+    return callApi("GET", `/api/agent/explain?${params}`);
+  }
+);
+
 // ============================================================
 // Deals — read
 // ============================================================
@@ -786,14 +952,33 @@ server.registerTool(
     callApi("POST", `/api/deals/${encodeURIComponent(dealId)}/refresh-persona`, { persona })
 );
 
+server.registerTool(
+  "deal_agent_run",
+  {
+    description:
+      "Run Llama Command's server-side Deal Agent inside a deal thread. " +
+      "Use this when the user explicitly wants the service agent to execute " +
+      "a deal-scoped task instead of the local MCP client doing the work.",
+    inputSchema: {
+      dealId: z.string().describe("deal uuid"),
+      message: z.string().describe("task instruction for the server-side Deal Agent"),
+      title: z.string().optional().describe("optional thread title; defaults to MCP agent run"),
+    },
+  },
+  async ({ dealId, message, title }) =>
+    runDealAgentTool({ dealId, message, title: title || "MCP agent run" })
+);
+
 server.registerTool(
   "deal_enrich",
   {
     description:
       "Run the Llama Command deal enrichment planner/trigger for one deal. " +
       "Default is dry-run: returns evidence plan, source plan, Monid budget/config " +
-      "status, and planned writes without changing facts/links/memo. Set apply=true " +
-      "only when the user explicitly wants the enrichment run recorded/applied. " +
+      "status, and planned writes without changing facts/links/memo. With " +
+      "apply=true and executor=server_agent, this starts the server-side Deal " +
+      "Agent unless harnessOnly=true. Set apply=true only when the user " +
+      "explicitly wants the enrichment run recorded/applied. " +
       "generateMemo never defaults on; pass true only when the user explicitly asks " +
       "for Memo generation after enrichment.",
     inputSchema: {
@@ -803,7 +988,7 @@ server.registerTool(
       executor: z
         .enum(["server_agent", "external_agent", "planner"])
         .optional()
-        .describe("who will execute the harness; external_agent returns guardrails for a user-owned agent"),
+        .describe("who will execute the harness; server_agent starts Deal Agent when apply=true"),
       sources: z
         .array(z.enum(["website", "github", "linkedin", "yc", "launch", "web", "monid"]))
         .optional()
@@ -819,17 +1004,34 @@ server.registerTool(
         .boolean()
         .optional()
         .describe("explicitly request memo regeneration after enrichment; default false"),
+      harnessOnly: z
+        .boolean()
+        .optional()
+        .describe("when true, return/apply the enrichment harness endpoint instead of starting Deal Agent"),
+      message: z
+        .string()
+        .optional()
+        .describe("optional override instruction for the server-side Deal Agent"),
     },
   },
-  async ({ dealId, dryRun, apply, executor, sources, budgetCents, generateMemo }) =>
-    callApi("POST", `/api/deals/${encodeURIComponent(dealId)}/enrich`, {
+  async ({ dealId, dryRun, apply, executor, sources, budgetCents, generateMemo, harnessOnly, message }) => {
+    const effectiveExecutor = executor ?? "server_agent";
+    if (apply === true && effectiveExecutor === "server_agent" && harnessOnly !== true) {
+      return runDealAgentTool({
+        dealId,
+        title: "MCP enrichment",
+        message: buildEnrichmentAgentMessage({ sources, budgetCents, generateMemo, message }),
+      });
+    }
+    return callApi("POST", `/api/deals/${encodeURIComponent(dealId)}/enrich`, {
       dryRun,
       apply,
-      executor,
+      executor: effectiveExecutor,
       sources,
       budgetCents,
       generateMemo,
-    })
+    });
+  }
 );
 
 // ============================================================
@@ -1375,7 +1577,9 @@ server.registerPrompt(
       "contract: identity, Pipeline First rule, content capture, autonomy " +
       "levels (L0/L1/L2/L3), communication style, error recovery, CLI/MCP " +
       "reference, and boundaries. Read this once, internalise it, operate " +
-      "accordingly. Same content as `llama agent-onboard` from the CLI.",
+      "accordingly. Same content as `llama agent-onboard` from the CLI. " +
+      "For the live private Llama OS skill library, call agent_bootstrap, " +
+      "skills_search, and skills_read.",
   },
   async () => {
     // Gate the briefing behind a /api/me check so unauthenticated MCP

package/bin/llama.mjs

@@ -15,6 +15,7 @@ import {
   readCanonicalToken,
   readLegacyConfig,
   request,
+  requestSse,
   tryGcloudIdentityToken,
   writeCanonicalToken,
   writeLegacyConfig,
@@ -199,10 +200,83 @@ async function searchDeals(q, flags) {
   return result;
 }
 
+function splitCsvFlag(value) {
+  if (!value || value === true) return undefined;
+  return String(value)
+    .split(",")
+    .map((s) => s.trim())
+    .filter(Boolean);
+}
+
+function boolFlag(flags, ...names) {
+  return names.some((name) => flags[name] === true || flags[name] === "true" || flags[name] === "1");
+}
+
+function buildEnrichmentAgentMessage(flags) {
+  if (flags.message && flags.message !== true) return String(flags.message);
+  const sources = splitCsvFlag(flags.sources) ?? [
+    "website",
+    "github",
+    "linkedin",
+    "yc",
+    "launch",
+    "web",
+    "monid",
+  ];
+  const budget = flags["budget-cents"] || flags.budgetCents || "50";
+  const memo = boolFlag(flags, "memo", "generate-memo", "generateMemo")
+    ? "Generate memo only after enrichment because the caller explicitly requested it."
+    : "Do not generate memo.";
+  return [
+    "Run server-side deal enrichment for this deal.",
+    `Use sources: ${sources.join(", ")}.`,
+    `Private Monid budget cap: ${budget} cents.`,
+    "Read the enrichment harness first, then collect current company/founder evidence.",
+    "Write canonical evidence links, sourced deal facts, stable deal fields, and typed factual values where supported.",
+    "For typed factual values, call read_typed_factual_layer first and use upsert_typed_fact for queryable fields.",
+    "Search snippets alone are not high-confidence evidence; fetch direct sources where possible.",
+    memo,
+    "End with what was written, what was skipped, and open questions.",
+  ].join(" ");
+}
+
+async function runDealAgentViaThread(dealId, message, title = "CLI agent run") {
+  const thread = await request("POST", `/api/deals/${encodeURIComponent(dealId)}/threads`, { title });
+  if (!thread?.id) throw new Error("Thread creation did not return an id");
+  process.stderr.write(`Running Deal Agent in thread ${thread.id}\n`);
+  const result = await requestSse(
+    "POST",
+    `/api/deals/${encodeURIComponent(dealId)}/threads/${encodeURIComponent(thread.id)}`,
+    { message },
+    {
+      onEvent(event) {
+        if (event.tool_use?.name) {
+          process.stderr.write(`[tool] ${event.tool_use.name}\n`);
+        }
+        if (event.tool_result?.name) {
+          const status = event.tool_result.ok ? "ok" : "error";
+          process.stderr.write(
+            `[tool] ${event.tool_result.name}: ${status} — ${event.tool_result.summary ?? ""}\n`,
+          );
+        }
+        if (event.text) {
+          process.stdout.write(event.text);
+        }
+      },
+    },
+  );
+  if (result.text && !result.text.endsWith("\n")) process.stdout.write("\n");
+  return { thread, ...result };
+}
+
 const HELP_FULL = `Llama Command CLI
 
 Agent onboarding (run once on first install):
   llama agent-onboard                  # print AGENT_BRIEFING.md — the workflow contract for AI agents
+  llama agent bootstrap                # fetch live Command + Llama OS skill manifest
+  llama skills search "pipeline update" # discover relevant runtime skills
+  llama skills show llama-pipeline      # read a skill from Command
+  llama explain <url-or-object>         # explain Command URL/object status + lifecycle
 
 External pitch — talk to Llama Ventures' intake agent (no token required):
   llama pitch start --name "Jane Doe" --email "jane@acme.ai"
@@ -237,7 +311,9 @@ Deals:
             llama deal update <dealId> leadInvestor "Acme Capital"
   llama deal enrich <dealId> [--dry-run] [--apply] [--executor server_agent|external_agent|planner]
                             [--sources website,github,linkedin,yc,monid] [--budget-cents 50]
-                            [--memo] [--prompt]                # evidence harness + server-side enrichment trigger
+                            [--memo] [--prompt] [--harness-only]
+      dry-run returns the harness; --apply --executor server_agent runs the server Deal Agent.
+  llama deal agent run <dealId> --message "collect founder evidence and update typed facts"
   llama deal extra set <dealId> <key> <value>        # system-admin only
       Patch one top-level key in deals.extra JSONB. Value is parsed as
       JSON when possible ('{"a":1}', 'true', '3'), else stored as a
@@ -440,6 +516,9 @@ Common:
   llama deal feed <dealId>          every contribution (facts + notes), newest first
   llama post <dealId> "..."         add a note to a deal
   llama agent-onboard               print the AI-agent workflow contract
+  llama agent bootstrap             live Llama OS skill manifest from Command
+  llama skills search "<query>"     discover which skill to read
+  llama explain <url-or-object>     explain Command URLs, 404s, deleted objects
 
 Command groups — run \`llama help <group>\` for that group's commands:
   deal        create · show · feed · update · enrich · search · collaborators · links · delete
@@ -452,6 +531,8 @@ Command groups — run \`llama help <group>\` for that group's commands:
   pitch       external founder intake (no token needed)
   ownership   claim · nominate · approvals
   admin       audit events (system admin only)
+  agent       bootstrap · skills · explain for AI agents
+  skills      search · show runtime Llama OS skills
   auth        setup · tokens · auth status
 
   llama help all     the full command reference (everything at once)
@@ -471,6 +552,8 @@ const HELP_AREA_MATCH = {
   pitch: [/^External pitch/],
   ownership: [/^Ownership/, /^Approvals/],
   admin: [/^Admin/],
+  agent: [/^Agent onboarding/],
+  skills: [/^Agent onboarding/],
   auth: [/^Setup/, /^Zero-config/, /^Token discovery/, /^Env/],
 };
 
@@ -859,6 +942,95 @@ https://command.llamaventures.vc/settings/tokens, run
     return;
   }
 
+  // Live runtime bootstrap from Llama Command. Unlike agent-onboard, this is
+  // not bundled in the public npm package; Command returns the current
+  // authenticated skill manifest and object-inspection contract.
+  if (area === "agent" && action === "bootstrap") {
+    const { flags } = parseFlags(rest, ["json", "limit"]);
+    const params = new URLSearchParams();
+    if (flags.limit && flags.limit !== true) params.set("limit", String(flags.limit));
+    const manifest = await request("GET", `/api/agent/manifest${params.toString() ? `?${params}` : ""}`);
+    if (flags.json) {
+      print(manifest);
+    } else {
+      process.stdout.write(`${manifest.briefing || JSON.stringify(manifest, null, 2)}\n`);
+    }
+    return;
+  }
+
+  if (area === "skills" || (area === "agent" && action === "skills")) {
+    const sub = area === "skills" ? action : rest[0];
+    const args = area === "skills" ? rest : rest.slice(1);
+    if (!sub || sub === "list") {
+      const { flags } = parseFlags(args, ["json", "limit"]);
+      const params = new URLSearchParams();
+      if (flags.limit && flags.limit !== true) params.set("limit", String(flags.limit));
+      const result = await request("GET", `/api/agent/skills${params.toString() ? `?${params}` : ""}`);
+      print(result);
+      return;
+    }
+    if (sub === "search") {
+      const { flags, positional } = parseFlags(args, ["json", "limit"]);
+      const q = positional.join(" ").trim();
+      if (!q) throw new Error("Usage: llama skills search <query> [--limit 20]");
+      const params = new URLSearchParams({ q });
+      if (flags.limit && flags.limit !== true) params.set("limit", String(flags.limit));
+      const result = await request("GET", `/api/agent/skills?${params}`);
+      print(result);
+      return;
+    }
+    if (sub === "show" || sub === "read") {
+      const { flags, positional } = parseFlags(args, ["json"]);
+      const slug = positional[0];
+      if (!slug) throw new Error("Usage: llama skills show <slug> [--json]");
+      const result = await request("GET", `/api/agent/skills/${encodeURIComponent(slug)}`);
+      if (flags.json) {
+        print(result);
+      } else {
+        process.stdout.write(`${result.skill?.content || JSON.stringify(result, null, 2)}\n`);
+      }
+      return;
+    }
+    throw new Error(`Unknown skills subcommand "${sub}". Use: list / search / show.`);
+  }
+
+  if (area === "explain" || (area === "agent" && action === "explain")) {
+    const args = area === "explain" ? [action, ...rest].filter(Boolean) : rest;
+    const { flags, positional } = parseFlags(args, ["json", "type", "id", "lang"]);
+    const params = new URLSearchParams();
+    const q = positional.join(" ").trim();
+    if (q) params.set("q", q);
+    if (flags.type && flags.type !== true) params.set("type", String(flags.type));
+    if (flags.id && flags.id !== true) params.set("id", String(flags.id));
+    if (flags.lang === "zh") params.set("lang", "zh");
+    if (!params.has("q") && !(params.has("type") && params.has("id"))) {
+      throw new Error("Usage: llama explain <url-or-object> OR llama explain --type <type> --id <id>");
+    }
+    const result = await request("GET", `/api/agent/explain?${params}`);
+    if (flags.json) {
+      print(result);
+    } else {
+      const target = result.result?.target;
+      const lifecycle = result.result?.lifecycle || [];
+      const lines = [
+        `${target?.objectType || "object"} ${target?.objectId || ""}`,
+        `Status: ${target?.status || "unknown"}`,
+        `Title: ${target?.title || "Untitled"}`,
+        target?.detail ? `Detail: ${target.detail}` : null,
+        target?.url ? `URL: ${target.url}` : null,
+        `Lifecycle events: ${lifecycle.length}`,
+      ].filter(Boolean);
+      if (lifecycle[0]) {
+        lines.push(
+          `Latest lifecycle: ${lifecycle[0].action} by ${lifecycle[0].actor_label || "unknown"} at ${lifecycle[0].created_at}`,
+        );
+        if (lifecycle[0].reason) lines.push(`Reason: ${lifecycle[0].reason}`);
+      }
+      process.stdout.write(`${lines.join("\n")}\n`);
+    }
+    return;
+  }
+
   // `llama pitch ...` — external founder-pitch family. No Llama token
   // required; bootstraps a session against /api/external/* via PoW + cookie.
   // See lib/external.mjs and AGENT_BRIEFING.md for the full surface.
@@ -1160,6 +1332,19 @@ https://command.llamaventures.vc/settings/tokens, run
     return;
   }
 
+  if (area === "deal" && action === "agent") {
+    const sub = rest[0];
+    const dealId = rest[1];
+    const { flags, positional } = parseFlags(rest.slice(2), ["message"]);
+    const message =
+      flags.message && flags.message !== true ? String(flags.message) : positional.join(" ").trim();
+    if (sub !== "run" || !dealId || !message) {
+      throw new Error(`Usage: llama deal agent run <dealId> --message "what the server agent should do"`);
+    }
+    await runDealAgentViaThread(dealId, message, "CLI agent run");
+    return;
+  }
+
   // ----- Deal enrichment: evidence plan + server-side enrichment trigger -----
   // The server owns Monid credentials and all write/audit behavior. CLI only
   // passes intent; default is dry-run so agents can inspect the harness before
@@ -1181,31 +1366,35 @@ https://command.llamaventures.vc/settings/tokens, run
       "budget-cents",
       "memo",
       "generate-memo",
+      "generateMemo",
       "prompt",
       "handoff",
+      "harness-only",
+      "message",
     ]);
-    const sources =
-      flags.sources && flags.sources !== true
-        ? String(flags.sources)
-            .split(",")
-            .map((s) => s.trim())
-            .filter(Boolean)
-        : undefined;
+    const sources = splitCsvFlag(flags.sources);
     const budgetCents =
       flags["budget-cents"] !== undefined && flags["budget-cents"] !== true
         ? Number(flags["budget-cents"])
         : undefined;
+    const apply = boolFlag(flags, "apply");
+    const executor = flags.executor && flags.executor !== true ? String(flags.executor) : "server_agent";
+
+    if (apply && executor === "server_agent" && !boolFlag(flags, "harness-only")) {
+      await runDealAgentViaThread(dealId, buildEnrichmentAgentMessage(flags), "CLI enrichment");
+      return;
+    }
 
     const result = await request(
       "POST",
       `/api/deals/${encodeURIComponent(dealId)}/enrich`,
       {
-        dryRun: flags.apply === true ? false : true,
-        apply: flags.apply === true,
-        executor: flags.executor && flags.executor !== true ? String(flags.executor) : undefined,
+        dryRun: apply ? false : true,
+        apply,
+        executor,
         sources,
         budgetCents,
-        generateMemo: flags.memo === true || flags["generate-memo"] === true,
+        generateMemo: boolFlag(flags, "memo", "generate-memo", "generateMemo"),
       }
     );
     if (flags.prompt === true || flags.handoff === true) {

package/lib/client.mjs

@@ -220,6 +220,10 @@ export async function request(method, endpoint, body) {
   return requestWithRetry(method, endpoint, body, /* allowRetry */ true);
 }
 
+export async function requestSse(method, endpoint, body, opts = {}) {
+  return requestSseWithRetry(method, endpoint, body, opts, /* allowRetry */ true);
+}
+
 async function requestWithRetry(method, endpoint, body, allowRetry) {
   const authHeaders = await getAuthHeaders();
   if (Object.keys(authHeaders).length === 0) throw noAuthError();
@@ -268,6 +272,80 @@ async function requestWithRetry(method, endpoint, body, allowRetry) {
   return data;
 }
 
+async function requestSseWithRetry(method, endpoint, body, opts, allowRetry) {
+  const authHeaders = await getAuthHeaders();
+  if (Object.keys(authHeaders).length === 0) throw noAuthError();
+  const res = await fetch(`${getBaseUrl()}${endpoint}`, {
+    method,
+    headers: {
+      "Content-Type": "application/json",
+      ...authHeaders,
+    },
+    body: body === undefined ? undefined : JSON.stringify(body),
+  });
+
+  if (res.status === 401 && allowRetry && (await bearerCameFromOAuth())) {
+    let refreshed = null;
+    try {
+      const { forceRefresh } = await import("./oauth-refresh.mjs");
+      refreshed = await forceRefresh();
+    } catch {
+      refreshed = null;
+    }
+    if (refreshed) {
+      return requestSseWithRetry(method, endpoint, body, opts, /* allowRetry */ false);
+    }
+    throw unauthorizedError();
+  }
+
+  if (res.status === 401) throw unauthorizedError();
+
+  if (!res.ok) {
+    const text = await res.text();
+    let data;
+    try {
+      data = text ? JSON.parse(text) : null;
+    } catch {
+      data = text;
+    }
+    const message = typeof data === "object" && data?.error ? data.error : `HTTP ${res.status}`;
+    throw new Error(message);
+  }
+
+  const reader = res.body?.getReader();
+  const decoder = new TextDecoder();
+  const events = [];
+  let text = "";
+  let buf = "";
+  if (!reader) return { text, events };
+
+  const handleFrame = (frame) => {
+    const dataLine = frame.split("\n").find((line) => line.startsWith("data: "));
+    if (!dataLine) return;
+    let event;
+    try {
+      event = JSON.parse(dataLine.slice(6));
+    } catch {
+      return;
+    }
+    events.push(event);
+    opts.onEvent?.(event);
+    if (event.text) text += event.text;
+    if (event.error) throw new Error(String(event.error));
+  };
+
+  while (true) {
+    const { done, value } = await reader.read();
+    if (done) break;
+    buf += decoder.decode(value, { stream: true });
+    const frames = buf.split("\n\n");
+    buf = frames.pop() || "";
+    for (const frame of frames) handleFrame(frame);
+  }
+  if (buf.trim()) handleFrame(buf);
+  return { text, events };
+}
+
 export function print(data) {
   if (typeof data === "string") {
     console.log(data);

package/package.json

@@ -1,8 +1,12 @@
 {
   "name": "@llamaventures/cli",
-  "version": "1.14.0",
+  "version": "1.15.0",
   "description": "CLI + MCP server for the Llama Ventures investment workbench (command.llamaventures.vc).",
   "type": "module",
+  "scripts": {
+    "test": "npm run test:agent-routing",
+    "test:agent-routing": "node scripts/verify-agent-routing.mjs"
+  },
   "bin": {
     "llama": "bin/llama.mjs",
     "llama-mcp": "bin/llama-mcp.mjs"
@@ -11,6 +15,7 @@
     "assets/",
     "bin/",
     "lib/",
+    "scripts/verify-agent-routing.mjs",
     "AGENT_BRIEFING.md",
     "CHANGELOG.md",
     "CONTRIBUTING.md",

package/scripts/verify-agent-routing.mjs

@@ -0,0 +1,447 @@
+#!/usr/bin/env node
+
+import assert from "node:assert/strict";
+import { spawn } from "node:child_process";
+import { createServer } from "node:http";
+import { existsSync } from "node:fs";
+import { mkdtemp, rm } from "node:fs/promises";
+import os from "node:os";
+import path from "node:path";
+import { fileURLToPath } from "node:url";
+
+const repoRoot = path.resolve(path.dirname(fileURLToPath(import.meta.url)), "..");
+assert.equal(
+  existsSync(path.join(repoRoot, "docs/agent-skills.bundle.json")),
+  false,
+  "public llama-cli must not bundle private Llama OS skill content",
+);
+assert.equal(
+  existsSync(path.join(repoRoot, "src/data/llama-os-skills.bundle.json")),
+  false,
+  "public llama-cli must not copy the Command-side skill mirror",
+);
+const calls = [];
+let threadSeq = 0;
+
+async function readJson(req) {
+  let raw = "";
+  for await (const chunk of req) raw += chunk;
+  if (!raw) return null;
+  try {
+    return JSON.parse(raw);
+  } catch {
+    return raw;
+  }
+}
+
+function writeJson(res, data) {
+  res.writeHead(200, { "Content-Type": "application/json" });
+  res.end(JSON.stringify(data));
+}
+
+function writeSse(res) {
+  res.writeHead(200, {
+    "Content-Type": "text/event-stream",
+    "Cache-Control": "no-cache",
+    Connection: "keep-alive",
+  });
+  const events = [
+    { tool_use: { name: "read_typed_factual_layer" } },
+    { tool_result: { name: "read_typed_factual_layer", ok: true, summary: "ok" } },
+    { text: "agent done" },
+  ];
+  for (const event of events) {
+    res.write(`data: ${JSON.stringify(event)}\n\n`);
+  }
+  res.end();
+}
+
+const server = createServer(async (req, res) => {
+  try {
+    const body = await readJson(req);
+    const url = new URL(req.url, "http://localhost");
+    calls.push({
+      method: req.method,
+      path: url.pathname,
+      query: Object.fromEntries(url.searchParams.entries()),
+      body,
+    });
+
+    if (req.method === "GET" && url.pathname === "/api/agent/manifest") {
+      writeJson(res, {
+        ok: true,
+        briefing: "runtime briefing: use skills_search, skills_read, and object_inspect",
+        llama_os: {
+          visible_skill_count: 49,
+          included_skill_count: Number(url.searchParams.get("limit") || 25),
+        },
+        skills: [
+          {
+            slug: "llama-command",
+            description: "Llama Command runtime skill",
+          },
+        ],
+      });
+      return;
+    }
+
+    if (req.method === "GET" && url.pathname === "/api/agent/skills") {
+      writeJson(res, {
+        ok: true,
+        q: url.searchParams.get("q"),
+        count: 1,
+        skills: [
+          {
+            slug: "llama-command",
+            description: "Llama Command runtime skill",
+          },
+        ],
+      });
+      return;
+    }
+
+    if (req.method === "GET" && url.pathname === "/api/agent/skills/llama-command") {
+      writeJson(res, {
+        ok: true,
+        skill: {
+          slug: "llama-command",
+          content: "---\nname: llama-command\n---\n# Llama Command runtime skill\n",
+        },
+      });
+      return;
+    }
+
+    if (req.method === "GET" && url.pathname === "/api/agent/explain") {
+      writeJson(res, {
+        ok: true,
+        result: {
+          target: {
+            objectType: "wiki_article",
+            objectId: "missing-page",
+            status: "deleted",
+            title: "Missing Page",
+            detail: "Deleted by Kevin Yu",
+            url: "https://command.llamaventures.vc/wiki/missing-page",
+          },
+          lifecycle: [
+            {
+              action: "deleted",
+              actor_label: "Kevin Yu",
+              created_at: "2026-06-15T19:02:00Z",
+              reason: "user_deleted",
+            },
+          ],
+        },
+      });
+      return;
+    }
+
+    if (req.method === "POST" && /^\/api\/deals\/[^/]+\/threads$/.test(url.pathname)) {
+      threadSeq += 1;
+      writeJson(res, { id: `thread-${threadSeq}` });
+      return;
+    }
+
+    if (req.method === "POST" && /^\/api\/deals\/[^/]+\/threads\/[^/]+$/.test(url.pathname)) {
+      writeSse(res);
+      return;
+    }
+
+    if (req.method === "POST" && /^\/api\/deals\/[^/]+\/enrich$/.test(url.pathname)) {
+      writeJson(res, {
+        ok: true,
+        agentHarness: {
+          handoffPrompt: "mock handoff prompt",
+          systemInjection: "mock system injection",
+        },
+      });
+      return;
+    }
+
+    res.writeHead(404, { "Content-Type": "application/json" });
+    res.end(JSON.stringify({ error: `Unexpected route ${req.method} ${url.pathname}` }));
+  } catch (err) {
+    res.writeHead(500, { "Content-Type": "application/json" });
+    res.end(JSON.stringify({ error: err?.message ?? String(err) }));
+  }
+});
+
+function listen(server) {
+  return new Promise((resolve) => server.listen(0, "127.0.0.1", resolve));
+}
+
+function close(server) {
+  return new Promise((resolve, reject) => {
+    server.close((err) => (err ? reject(err) : resolve()));
+  });
+}
+
+function childEnv(baseUrl, homeDir) {
+  return {
+    ...process.env,
+    HOME: homeDir,
+    LLAMA_API_URL: baseUrl,
+    LLAMA_TOKEN: "llc_mock_agent_routing",
+    PATH: "/usr/bin:/bin",
+  };
+}
+
+function resetCalls() {
+  calls.length = 0;
+  threadSeq = 0;
+}
+
+function paths() {
+  return calls.map((call) => `${call.method} ${call.path}`);
+}
+
+function assertNoEnrichCall() {
+  assert.equal(
+    calls.some((call) => call.path.endsWith("/enrich")),
+    false,
+    `expected no /enrich call, got ${paths().join(", ")}`,
+  );
+}
+
+function assertThreadRun({ title, messageIncludes }) {
+  assert.equal(calls.length, 2, `expected thread create + SSE run, got ${paths().join(", ")}`);
+  assert.match(calls[0].path, /^\/api\/deals\/[^/]+\/threads$/);
+  assert.equal(calls[0].body?.title, title);
+  assert.match(calls[1].path, /^\/api\/deals\/[^/]+\/threads\/thread-1$/);
+  for (const needle of messageIncludes) {
+    assert.match(calls[1].body?.message ?? "", new RegExp(escapeRegExp(needle)));
+  }
+}
+
+function escapeRegExp(value) {
+  return String(value).replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+}
+
+async function runCli(args, baseUrl, homeDir) {
+  const child = spawn(process.execPath, ["bin/llama.mjs", ...args], {
+    cwd: repoRoot,
+    env: childEnv(baseUrl, homeDir),
+    stdio: ["ignore", "pipe", "pipe"],
+  });
+  let stdout = "";
+  let stderr = "";
+  child.stdout.on("data", (chunk) => {
+    stdout += chunk;
+  });
+  child.stderr.on("data", (chunk) => {
+    stderr += chunk;
+  });
+  const code = await new Promise((resolve) => child.on("close", resolve));
+  assert.equal(code, 0, `CLI failed (${code})\nSTDOUT:\n${stdout}\nSTDERR:\n${stderr}`);
+  return { stdout, stderr };
+}
+
+async function callMcpTool(name, args, baseUrl, homeDir) {
+  const child = spawn(process.execPath, ["bin/llama-mcp.mjs"], {
+    cwd: repoRoot,
+    env: childEnv(baseUrl, homeDir),
+    stdio: ["pipe", "pipe", "pipe"],
+  });
+  let stderr = "";
+  let buffer = "";
+  child.stderr.on("data", (chunk) => {
+    stderr += chunk;
+  });
+
+  const result = await new Promise((resolve, reject) => {
+    const timeout = setTimeout(() => {
+      child.kill();
+      reject(new Error(`Timed out waiting for MCP response\nSTDERR:\n${stderr}`));
+    }, 8000);
+
+    child.stdout.on("data", (chunk) => {
+      buffer += chunk;
+      let idx;
+      while ((idx = buffer.indexOf("\n")) >= 0) {
+        const line = buffer.slice(0, idx).trim();
+        buffer = buffer.slice(idx + 1);
+        if (!line) continue;
+        let msg;
+        try {
+          msg = JSON.parse(line);
+        } catch {
+          continue;
+        }
+        if (msg.id === 2) {
+          clearTimeout(timeout);
+          child.kill();
+          resolve(msg);
+        }
+      }
+    });
+
+    child.on("error", (err) => {
+      clearTimeout(timeout);
+      reject(err);
+    });
+
+    child.stdin.write(
+      [
+        JSON.stringify({
+          jsonrpc: "2.0",
+          id: 1,
+          method: "initialize",
+          params: {
+            protocolVersion: "2024-11-05",
+            capabilities: {},
+            clientInfo: { name: "routing-test", version: "1" },
+          },
+        }),
+        JSON.stringify({ jsonrpc: "2.0", method: "notifications/initialized" }),
+        JSON.stringify({
+          jsonrpc: "2.0",
+          id: 2,
+          method: "tools/call",
+          params: { name, arguments: args },
+        }),
+      ].join("\n") + "\n",
+    );
+  });
+
+  assert.ok(!result.error, `MCP returned error: ${JSON.stringify(result.error)}`);
+  return result.result;
+}
+
+await listen(server);
+const address = server.address();
+const baseUrl = `http://${address.address}:${address.port}`;
+const homeDir = await mkdtemp(path.join(os.tmpdir(), "llama-cli-routing-"));
+
+try {
+  resetCalls();
+  const bootstrapRun = await runCli(["agent", "bootstrap", "--limit", "3"], baseUrl, homeDir);
+  assert.match(bootstrapRun.stdout, /runtime briefing/);
+  assert.deepEqual(paths(), ["GET /api/agent/manifest"]);
+  assert.equal(calls[0].query.limit, "3");
+
+  resetCalls();
+  const skillSearchRun = await runCli(["skills", "search", "pipeline", "--limit", "5"], baseUrl, homeDir);
+  assert.match(skillSearchRun.stdout, /llama-command/);
+  assert.deepEqual(paths(), ["GET /api/agent/skills"]);
+  assert.equal(calls[0].query.q, "pipeline");
+  assert.equal(calls[0].query.limit, "5");
+
+  resetCalls();
+  const skillShowRun = await runCli(["skills", "show", "llama-command"], baseUrl, homeDir);
+  assert.match(skillShowRun.stdout, /# Llama Command runtime skill/);
+  assert.deepEqual(paths(), ["GET /api/agent/skills/llama-command"]);
+
+  resetCalls();
+  const explainRun = await runCli(["explain", "https://command.llamaventures.vc/wiki/missing-page"], baseUrl, homeDir);
+  assert.match(explainRun.stdout, /Status: deleted/);
+  assert.match(explainRun.stdout, /Deleted by Kevin Yu/);
+  assert.deepEqual(paths(), ["GET /api/agent/explain"]);
+  assert.equal(calls[0].query.q, "https://command.llamaventures.vc/wiki/missing-page");
+
+  resetCalls();
+  const enrichRun = await runCli(
+    [
+      "deal",
+      "enrich",
+      "deal-cli",
+      "--apply",
+      "--executor",
+      "server_agent",
+      "--sources",
+      "website,monid",
+      "--budget-cents",
+      "12",
+    ],
+    baseUrl,
+    homeDir,
+  );
+  assert.match(enrichRun.stdout, /agent done/);
+  assertNoEnrichCall();
+  assertThreadRun({
+    title: "CLI enrichment",
+    messageIncludes: ["website, monid", "12 cents", "upsert_typed_fact"],
+  });
+
+  resetCalls();
+  await runCli(
+    ["deal", "enrich", "deal-cli", "--apply", "--executor", "server_agent", "--harness-only"],
+    baseUrl,
+    homeDir,
+  );
+  assert.deepEqual(paths(), ["POST /api/deals/deal-cli/enrich"]);
+  assert.equal(calls[0].body?.apply, true);
+  assert.equal(calls[0].body?.dryRun, false);
+  assert.equal(calls[0].body?.executor, "server_agent");
+
+  resetCalls();
+  const agentRun = await runCli(
+    ["deal", "agent", "run", "deal-cli", "--message", "custom server task"],
+    baseUrl,
+    homeDir,
+  );
+  assert.match(agentRun.stdout, /agent done/);
+  assertNoEnrichCall();
+  assertThreadRun({
+    title: "CLI agent run",
+    messageIncludes: ["custom server task"],
+  });
+
+  resetCalls();
+  const mcpResult = await callMcpTool(
+    "deal_enrich",
+    {
+      dealId: "deal-mcp",
+      apply: true,
+      executor: "server_agent",
+      sources: ["web", "monid"],
+      budgetCents: 7,
+    },
+    baseUrl,
+    homeDir,
+  );
+  assertNoEnrichCall();
+  assertThreadRun({
+    title: "MCP enrichment",
+    messageIncludes: ["web, monid", "7 cents", "upsert_typed_fact"],
+  });
+  const payload = JSON.parse(mcpResult.content?.[0]?.text ?? "{}");
+  assert.equal(payload.ok, true);
+  assert.equal(payload.threadId, "thread-1");
+  assert.equal(payload.text, "agent done");
+
+  resetCalls();
+  const mcpBootstrap = await callMcpTool("agent_bootstrap", { limit: 2 }, baseUrl, homeDir);
+  const bootstrapPayload = JSON.parse(mcpBootstrap.content?.[0]?.text ?? "{}");
+  assert.equal(bootstrapPayload.ok, true);
+  assert.deepEqual(paths(), ["GET /api/agent/manifest"]);
+  assert.equal(calls[0].query.limit, "2");
+
+  resetCalls();
+  const mcpSkills = await callMcpTool("skills_search", { q: "command", limit: 4 }, baseUrl, homeDir);
+  const skillsPayload = JSON.parse(mcpSkills.content?.[0]?.text ?? "{}");
+  assert.equal(skillsPayload.skills?.[0]?.slug, "llama-command");
+  assert.deepEqual(paths(), ["GET /api/agent/skills"]);
+  assert.equal(calls[0].query.q, "command");
+
+  resetCalls();
+  const mcpSkillRead = await callMcpTool("skills_read", { slug: "llama-command" }, baseUrl, homeDir);
+  const skillPayload = JSON.parse(mcpSkillRead.content?.[0]?.text ?? "{}");
+  assert.match(skillPayload.skill?.content ?? "", /# Llama Command runtime skill/);
+  assert.deepEqual(paths(), ["GET /api/agent/skills/llama-command"]);
+
+  resetCalls();
+  const mcpInspect = await callMcpTool(
+    "object_inspect",
+    { q: "https://command.llamaventures.vc/wiki/missing-page" },
+    baseUrl,
+    homeDir,
+  );
+  const inspectPayload = JSON.parse(mcpInspect.content?.[0]?.text ?? "{}");
+  assert.equal(inspectPayload.result?.target?.status, "deleted");
+  assert.deepEqual(paths(), ["GET /api/agent/explain"]);
+
+  console.log("agent routing verification passed");
+} finally {
+  await close(server);
+  await rm(homeDir, { recursive: true, force: true });
+}