@socialneuron/mcp-server 1.7.14 → 1.7.15

package/CHANGELOG.md

@@ -2,6 +2,21 @@
 
 All notable changes to `@socialneuron/mcp-server` will be documented in this file.
 
+## [1.7.15] - 2026-07-06
+
+### Changed
+
+- **Public tool surface tightened.** Internal service tools used by Social Neuron's own automation are still registered and scope-gated at runtime, but are no longer advertised on the hosted server card, HTTP discovery, `search_tools`, knowledge documents, or the CLI tool listing. The hosted endpoint now advertises 85 tools; the npm stdio server exposes 87 discoverable tools (including 2 local screen-capture tools).
+- **Tool descriptions cleaned up.** Internal project references and implementation jargon removed from tool descriptions across the catalog.
+- **`run_content_pipeline` scheduling guard restored.** Scheduling again requires `schedule_confirmed=true` after explicit user approval, cannot run with the quality stage skipped, caps generated posts to the requested plan size, drops posts targeting unrequested platforms, and counts scheduled posts against the credit budget. (This hardening was unintentionally dropped in the 1.7.14 source sync.)
+- **Metadata contract unified + CI-guarded.** `server.json` version/pricing/tool-count now match `package.json` and the canonical pricing (MCP requires Pro $49/mo or higher; free tier is 50 credits/mo with no MCP access). Added `mcpName` for MCP Registry ownership verification. New `npm run verify:metadata` gate (wired into CI and release) blocks stale counts, stale pricing, internal codenames, and dead endpoint links from re-entering the public surface.
+- **Docs corrected.** REST API docs (tool count, response `version` example, plan limits), troubleshooting boot log line (95/95), integration methods (stdio vs hosted tool split), and auth docs (plan-scope matrix; removed internal implementation notes). Removed links to the not-yet-deployed `/v1/openapi.json` endpoint.
+- **SDK release path hardened.** `packages/sdk` now ships a `package-lock.json`, the SDK release workflow uses `npm ci --ignore-scripts`, and Dependabot watches `/packages/sdk` and `/apps/content-calendar`.
+
+## [1.7.12–1.7.14] - 2026-06/07
+
+Released without changelog entries (see git history): OAuth connector flow + DCR, MCP Apps content calendar, trial-key scopes and post-trial degrade, funnel instrumentation, 96-tool catalog sync, dependency cooldown pins.
+
 ## [1.7.11] - 2026-05-15
 
 ### Changed

package/README.md

@@ -1,6 +1,6 @@
 # @socialneuron/mcp-server
 
-> 96 MCP tools for AI-powered social media management. Create content, schedule posts, track analytics, and optimize performance — all from Claude Code or any MCP client.
+> 85+ MCP tools for AI-powered social media management. Create content, schedule posts, track analytics, and optimize performance — all from Claude Code or any MCP client.
 
 [![npm version](https://img.shields.io/npm/v/@socialneuron/mcp-server)](https://www.npmjs.com/package/@socialneuron/mcp-server)
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
@@ -92,7 +92,9 @@ Ask Claude things like:
 - "Check my analytics and suggest improvements"
 - "Set up autopilot to post 3 times per week"
 
-## Tool Categories (96 tools)
+## Tool Categories
+
+The hosted endpoint advertises **85 tools** (live count: [server card](https://mcp.socialneuron.com/.well-known/mcp/server-card.json)); the npm stdio server adds 2 local screen-capture tools (87 discoverable). A small set of internal service tools used by Social Neuron's own automation are registered but not part of the public surface.
 
 These tools are available to AI agents (Claude, Cursor, etc.) via the MCP protocol.
 
@@ -217,7 +219,7 @@ socialneuron-mcp sn loop
 socialneuron-mcp sn credits
 
 # Agent-native CLI v2
-socialneuron-mcp sn tools [--module ideation] [--scope mcp:write]  # List/filter all 96 tools
+socialneuron-mcp sn tools [--module ideation] [--scope mcp:write]  # List/filter available tools
 socialneuron-mcp sn info                                            # Version, auth, credits, tool count
 socialneuron-mcp sn plan list|view|approve                          # Content plan management
 socialneuron-mcp sn preset list|show|save|delete                    # Platform presets (6 built-in)

package/dist/http.js

@@ -707,10 +707,10 @@ var TOOL_SCOPES = {
   record_campaign_spend: "mcp:write",
   // mcp:read
   get_active_campaigns: "mcp:read",
-  // mcp:read / mcp:write (Skills — PR #4.4)
+  // mcp:read / mcp:write (Skills)
   list_skills: "mcp:read",
   run_skill: "mcp:write",
-  // mcp:read (Loop observability — growth-loop KPIs + bandit posteriors)
+  // mcp:read (Loop observability — growth-loop KPIs + content learning state)
   get_loop_pulse: "mcp:read",
   get_bandit_state: "mcp:read"
 };
@@ -1213,7 +1213,7 @@ function checkRateLimit(category, key) {
 init_supabase();
 
 // src/lib/version.ts
-var MCP_VERSION = "1.7.14";
+var MCP_VERSION = "1.7.15";
 
 // src/tools/ideation.ts
 function asEnvelope(data) {
@@ -9555,7 +9555,7 @@ var TOOL_CATALOG = [
   // niche research
   {
     name: "find_winning_content",
-    description: "Find QA-gated high-performing short-form videos in the project's niche. Returns extracted hook patterns, content structures, and pre-compiled replication prompts (backed by niche_winners view, qa_score >= 0.5).",
+    description: "Find QA-gated high-performing short-form videos in the project's niche. Returns extracted hook patterns, content structures, and pre-compiled replication prompts.",
     module: "research",
     scope: "mcp:read"
   },
@@ -9706,59 +9706,68 @@ var TOOL_CATALOG = [
   // agentic-harness — learning loop write-back
   {
     name: "write_agent_reflection",
-    description: "Persist a verbal reflection for an agent loop. Provenance keys are restricted (Anti-Goodhart safety): only content_history_id, outcome_event_id, prm_score_ids, and handoff_ids are accepted.",
+    description: "Persist a verbal reflection for an agent loop. Provenance keys are restricted to an allowlist: only content_history_id, outcome_event_id, prm_score_ids, and handoff_ids are accepted.",
     module: "harness",
-    scope: "mcp:write"
+    scope: "mcp:write",
+    internal: true
   },
   {
     name: "record_outcome",
-    description: "Record an outcome for a published decision event. Idempotent on (decision_event_id, horizon). Only horizon=24h triggers a content_bandits posterior update.",
+    description: "Record an outcome for a published decision event. Idempotent on (decision_event_id, horizon). Only horizon=24h triggers a learning-loop update.",
     module: "harness",
-    scope: "mcp:write"
+    scope: "mcp:write",
+    internal: true
   },
   // agentic-harness — learning loop read-back
   {
     name: "read_agent_reflection",
     description: "Read past agent reflections for a brand. Ordered by created_at DESC, id ASC (deterministic tiebreak). Only active reflections returned (superseded_by IS NULL). Optional generated_by_agent filter.",
     module: "harness",
-    scope: "mcp:read"
+    scope: "mcp:read",
+    internal: true
   },
   // hermes — autonomous agent integration (closed-loop content)
   {
     name: "save_draft_to_library",
-    description: "Save a draft post to the SN content library with origin=hermes. Used by Hermes to persist drafts before the founder approves them.",
+    description: "Save a draft post to the SN content library for review before publishing. Drafts land in the content library pending approval.",
     module: "hermes",
-    scope: "mcp:write"
+    scope: "mcp:write",
+    internal: true
   },
   {
     name: "record_voice_lesson",
-    description: "Persist a learned voice lesson to brand_profiles.brand_context.voiceProfile.voice_lessons via atomic RPC. Used by Hermes reflection cron.",
+    description: "Persist a learned voice lesson to the brand voice profile.",
     module: "hermes",
-    scope: "mcp:write"
+    scope: "mcp:write",
+    internal: true
   },
   {
     name: "record_observation",
-    description: 'Record an agent observation (e.g. "topic X engagement up 23%") for the UnifiedAnalytics > Playbook surface.',
+    description: 'Record an agent observation (e.g. "topic X engagement up 23%") for the analytics playbook.',
     module: "hermes",
-    scope: "mcp:write"
+    scope: "mcp:write",
+    internal: true
   },
   {
     name: "record_intel_signal",
-    description: "Record a research/trend signal from Hermes watchers (news, HN, competitor, etc.) for Niche Intelligence. Dedupes by URL.",
+    description: "Record a research/trend signal (news, competitor, community sources) for niche intelligence. Dedupes by URL.",
     module: "hermes",
-    scope: "mcp:write"
+    scope: "mcp:write",
+    internal: true
   },
   {
     name: "record_campaign_spend",
-    description: "Log a campaign cost line (hermes_drafts, carousel_renders, analytics_pulls, paid_amplification, other). Ownership-checked.",
+    description: "Log a campaign cost line item. Ownership-checked.",
     module: "hermes",
-    scope: "mcp:write"
+    scope: "mcp:write",
+    internal: true
   },
   {
     name: "get_active_campaigns",
-    description: "List currently-running campaigns with thesis, budget, hero format, and current spend. Used by Hermes pitch skill to bias drafts.",
+    description: "List currently-running campaigns with thesis, budget, hero format, and current spend.",
     module: "hermes",
-    scope: "mcp:read"
+    scope: "mcp:read",
+    internal: true
   },
   // skills (workflow skills — multi-step brand-locked content pipelines)
   {
@@ -9769,20 +9778,20 @@ var TOOL_CATALOG = [
   },
   {
     name: "run_skill",
-    description: "Run a Social Neuron workflow skill end-to-end (brand-locked content production). PR #4.4 v1 returns a structured run preview with the step plan, credit cost, and a deep-link to launch in the SN dashboard.",
+    description: "Run a Social Neuron workflow skill end-to-end (brand-locked content production). Returns a structured run preview with the step plan, credit cost, and a deep-link to launch in the SN dashboard.",
     module: "skills",
     scope: "mcp:write"
   },
-  // loop observability (growth-loop KPIs + Thompson Sampling bandit posteriors)
+  // loop observability (growth-loop KPIs + content learning state)
   {
     name: "get_loop_pulse",
-    description: "Read dynamic loop-health KPIs for the growth loop over the last 7 days (reflection/decision coverage, visual gate pass rate, bandit-update application rate, per-platform uptake, autopilot lag) \u2014 each with an ok/warn/bad status. Use to decide whether the loop is closing or where it is stuck.",
+    description: "Read dynamic loop-health KPIs for the growth loop over the last 7 days (reflection/decision coverage, visual gate pass rate, learning-update application rate, per-platform uptake, autopilot lag) \u2014 each with an ok/warn/bad status. Use to decide whether the loop is closing or where it is stuck.",
     module: "loop",
     scope: "mcp:read"
   },
   {
     name: "get_bandit_state",
-    description: "Read the current Thompson Sampling bandit posteriors for a project \u2014 top-K arms per (arm_type, platform) with Beta(alpha,beta) posterior mean and uncertainty. Use to reason about which hook family / format / timing slot the bandit currently prefers per platform.",
+    description: "Read the current content learning state for a project \u2014 top-K arms per (arm_type, platform) with expected performance and uncertainty. Use to reason about which hook family / format / timing slot currently performs best per platform.",
     module: "loop",
     scope: "mcp:read"
   }
@@ -9892,7 +9901,10 @@ function toolKnowledgeDocument(tool) {
   };
 }
 function getKnowledgeDocuments() {
-  return [...STATIC_KNOWLEDGE_DOCUMENTS, ...TOOL_CATALOG.map(toolKnowledgeDocument)];
+  return [
+    ...STATIC_KNOWLEDGE_DOCUMENTS,
+    ...TOOL_CATALOG.filter((t) => !t.internal).map(toolKnowledgeDocument)
+  ];
 }
 function tokenize(input) {
   return input.toLowerCase().split(/[^a-z0-9:_-]+/).map((token) => token.trim()).filter(Boolean);
@@ -10008,6 +10020,7 @@ function registerDiscoveryTools(server) {
       if (query) {
         results = searchTools(query);
       }
+      results = results.filter((t) => !t.internal);
       if (module) {
         const moduleTools = getToolsByModule(module);
         results = results.filter((t) => moduleTools.some((mt) => mt.name === t.name));
@@ -10084,6 +10097,7 @@ var PLATFORM_ENUM2 = z24.enum([
 ]);
 var BASE_PLAN_CREDITS = 15;
 var SOURCE_EXTRACTION_CREDITS = 5;
+var SCHEDULE_POST_CREDITS = 1;
 function registerPipelineTools(server) {
   server.tool(
     "check_pipeline_readiness",
@@ -10200,7 +10214,7 @@ function registerPipelineTools(server) {
   );
   server.tool(
     "run_content_pipeline",
-    "Run the full content pipeline: research trends \u2192 generate plan \u2192 quality check \u2192 auto-approve \u2192 schedule posts. Chains all stages in one call for maximum efficiency. Set dry_run=true to preview the plan without publishing. Check check_pipeline_readiness first to verify credits, OAuth, and brand profile are ready.",
+    "Run the full content pipeline: research trends \u2192 generate plan \u2192 quality check \u2192 auto-approve \u2192 schedule posts. Chains all stages in one call for maximum efficiency. Set dry_run=true to preview the plan without publishing. To schedule posts, set schedule_confirmed=true after the user explicitly approves publishing. Check check_pipeline_readiness first to verify credits, OAuth, and brand profile are ready.",
     {
       project_id: z24.string().uuid().optional().describe("Project ID (auto-detected if omitted)"),
       topic: z24.string().optional().describe("Content topic (required if no source_url)"),
@@ -10216,6 +10230,9 @@ function registerPipelineTools(server) {
       ),
       max_credits: z24.number().optional().describe("Credit budget cap"),
       dry_run: z24.boolean().default(false).describe("If true, skip scheduling and return plan only"),
+      schedule_confirmed: z24.boolean().default(false).describe(
+        "Required to schedule posts. Set true only after explicit user confirmation to publish/schedule."
+      ),
       skip_stages: z24.array(z24.enum(["research", "quality", "schedule"])).optional().describe("Stages to skip"),
       response_format: z24.enum(["text", "json"]).default("json")
     },
@@ -10230,6 +10247,7 @@ function registerPipelineTools(server) {
       auto_approve_threshold,
       max_credits,
       dry_run,
+      schedule_confirmed,
       skip_stages,
       response_format
     }) => {
@@ -10245,6 +10263,29 @@ function registerPipelineTools(server) {
         };
       }
       const skipSet = new Set(skip_stages ?? []);
+      const schedulingRequested = !dry_run && !skipSet.has("schedule");
+      if (schedulingRequested && !schedule_confirmed) {
+        return {
+          content: [
+            {
+              type: "text",
+              text: 'Scheduling requires explicit confirmation. Re-run with schedule_confirmed=true after the user approves publishing, or set dry_run=true / skip_stages=["schedule"].'
+            }
+          ],
+          isError: true
+        };
+      }
+      if (schedulingRequested && skipSet.has("quality")) {
+        return {
+          content: [
+            {
+              type: "text",
+              text: "Scheduling cannot run when the quality stage is skipped."
+            }
+          ],
+          isError: true
+        };
+      }
       try {
         const resolvedProjectId = project_id ?? await getDefaultProjectId() ?? void 0;
         const estimatedCost = BASE_PLAN_CREDITS + (source_url ? SOURCE_EXTRACTION_CREDITS : 0);
@@ -10287,6 +10328,7 @@ function registerPipelineTools(server) {
               approval_mode,
               auto_approve_threshold,
               dry_run,
+              schedule_confirmed,
               skip_stages: skip_stages ?? []
             },
             current_stage: "planning",
@@ -10352,7 +10394,9 @@ function registerPipelineTools(server) {
         stagesCompleted.push("planning");
         const rawText = String(planData.text ?? planData.content ?? "");
         const postsArray = extractJsonArray(rawText);
-        const posts = (postsArray ?? []).map((p) => ({
+        const requestedPlatformSet = new Set(platforms);
+        const maxPosts = platforms.length * days * posts_per_day;
+        const parsedPosts = (postsArray ?? []).map((p) => ({
           id: String(p.id ?? randomUUID3().slice(0, 8)),
           day: Number(p.day ?? 1),
           date: String(p.date ?? ""),
@@ -10366,6 +10410,23 @@ function registerPipelineTools(server) {
           visual_direction: p.visual_direction ? String(p.visual_direction) : void 0,
           media_type: p.media_type ? String(p.media_type) : void 0
         }));
+        const platformFilteredPosts = parsedPosts.filter(
+          (post) => requestedPlatformSet.has(post.platform)
+        );
+        const posts = platformFilteredPosts.slice(0, maxPosts);
+        if (parsedPosts.length > maxPosts) {
+          errors.push({
+            stage: "planning",
+            message: `AI returned ${parsedPosts.length} posts; truncated to ${maxPosts}.`
+          });
+        }
+        const invalidPlatformCount = parsedPosts.length - platformFilteredPosts.length;
+        if (invalidPlatformCount > 0) {
+          errors.push({
+            stage: "planning",
+            message: `Dropped ${invalidPlatformCount} post(s) with unrequested or invalid platform.`
+          });
+        }
         let postsApproved = 0;
         let postsFlagged = 0;
         if (!skipSet.has("quality")) {
@@ -10518,6 +10579,7 @@ function registerPipelineTools(server) {
                 });
               } else {
                 postsScheduled++;
+                creditsUsed += SCHEDULE_POST_CREDITS;
               }
             } catch (schedErr) {
               errors.push({
@@ -12577,13 +12639,13 @@ function asEnvelope24(data) {
 function registerNicheResearchTools(server) {
   server.tool(
     "find_winning_content",
-    "Find QA-gated high-performing short-form videos in the project's niche. Returns extracted hook patterns, content structures, and pre-compiled replication prompts you can use to generate new content on a different topic. Backed by niche_winners view (qa_score >= 0.5 + replication_prompt populated).",
+    "Find QA-gated high-performing short-form videos in the project's niche. Returns extracted hook patterns, content structures, and pre-compiled replication prompts you can use to generate new content on a different topic. Only QA-gated winners with a populated replication prompt are returned.",
     {
       project_id: z29.string().uuid().optional().describe("Project ID (auto-detected if omitted)"),
       platform: z29.enum(["tiktok", "instagram", "youtube", "reddit", "twitter"]).optional().describe("Filter to one platform. Omit for all platforms."),
       days: z29.number().int().min(1).max(365).default(30).describe("Window: only return winners scanned within this many days."),
       limit: z29.number().int().min(1).max(50).default(10).describe("Number of winners to return (1-50)."),
-      min_qa_score: z29.number().min(0).max(1).default(0.5).describe("Minimum QA score (0..1). Default 0.5 matches the niche_winners view floor."),
+      min_qa_score: z29.number().min(0).max(1).default(0.5).describe("Minimum QA score (0..1). Default 0.5."),
       response_format: z29.enum(["text", "json"]).optional()
     },
     async ({ project_id, platform: platform2, days, limit, min_qa_score, response_format }) => {
@@ -13241,7 +13303,7 @@ var writeReflectionSchema = {
     prm_score_ids: z33.array(z33.string()).optional().describe("PRM score IDs that informed this reflection."),
     handoff_ids: z33.array(z33.string()).optional().describe("Handoff event IDs that triggered this reflection.")
   }).strict().describe(
-    "Source evidence for this reflection. Only these four keys are accepted (Anti-Goodhart guard \u2014 unknown keys are rejected to prevent spurious provenance claims)."
+    "Source evidence for this reflection. Only these four keys are accepted \u2014 unknown keys are rejected to prevent spurious provenance claims."
   ),
   brand_id: z33.string().describe("Brand profile UUID this reflection belongs to."),
   pipeline_id: z33.string().optional().describe("Optional: pipeline run UUID."),
@@ -13259,7 +13321,7 @@ var readReflectionSchema = {
 var recordOutcomeSchema = {
   decision_event_id: z33.string().describe("UUID of the decision_events row to record an outcome for."),
   horizon: z33.enum(["1h", "6h", "24h"]).describe(
-    "Observation horizon. Only horizon=24h with a non-null reward triggers a content_bandits posterior update; 1h/6h are stored but inert for learning."
+    "Observation horizon. Only horizon=24h with a non-null reward triggers a learning-loop update; 1h/6h are stored but inert for learning."
   ),
   reward: z33.number().min(0).max(1).describe("Normalised reward signal in [0, 1]. Higher = better outcome."),
   outcome_metrics: z33.record(z33.string(), z33.number()).optional().describe(
@@ -13269,7 +13331,7 @@ var recordOutcomeSchema = {
 function registerHarnessTools(server, _ctx) {
   server.tool(
     "write_agent_reflection",
-    "Persist a verbal reflection for an agent loop. Provenance keys are restricted (Anti-Goodhart safety): only content_history_id, outcome_event_id, prm_score_ids, and handoff_ids are accepted \u2014 unknown keys are rejected at the input layer. Requires mcp:write scope. Returns the created reflection UUID on success.",
+    "Persist a verbal reflection for an agent loop. Provenance keys are restricted to an allowlist: only content_history_id, outcome_event_id, prm_score_ids, and handoff_ids are accepted \u2014 unknown keys are rejected at the input layer. Requires mcp:write scope. Returns the created reflection UUID on success.",
     writeReflectionSchema,
     // eslint-disable-next-line @typescript-eslint/no-explicit-any
     async (args) => {
@@ -13297,7 +13359,7 @@ function registerHarnessTools(server, _ctx) {
   );
   server.tool(
     "record_outcome",
-    "Record an outcome for a published decision event. Idempotent on (decision_event_id, horizon) \u2014 safe to call multiple times. Returns idempotent:true when the row already existed (UPDATE), idempotent:false on fresh INSERT. Note: only horizon=24h with reward != null triggers a content_bandits posterior update; 1h/6h outcomes are stored but are inert for the learning loop. Requires mcp:write scope.",
+    "Record an outcome for a published decision event. Idempotent on (decision_event_id, horizon) \u2014 safe to call multiple times. Returns idempotent:true when the row already existed (UPDATE), idempotent:false on fresh INSERT. Note: only horizon=24h with reward != null triggers a learning-loop update; 1h/6h outcomes are stored but are inert for the learning loop. Requires mcp:write scope.",
     recordOutcomeSchema,
     // eslint-disable-next-line @typescript-eslint/no-explicit-any
     async (args) => {
@@ -13381,13 +13443,13 @@ var PLATFORM = z34.enum([
 function registerHermesTools(server) {
   server.tool(
     "save_draft_to_library",
-    "Save a draft post to the SN content library. Use when an autonomous agent (Hermes) wants to persist a draft before the founder approves it. Lands in ContentLibrary with status='draft', origin='hermes'. The draft can then be approved/edited in the SN UI.",
+    "Save a draft post to the SN content library. Use when an autonomous agent wants to persist a draft for review before publishing. Lands in the content library with status='draft'. The draft can then be approved/edited in the SN UI.",
     {
       platform: PLATFORM.describe("Target platform for the draft."),
       copy: z34.string().min(1).max(8e3).describe("The draft post body."),
       project_id: z34.string().optional().describe("SN project UUID. Optional but recommended."),
       media_url: z34.string().url().optional().describe("Optional cover/media URL (R2 signed URL)."),
-      hermes_run_id: z34.string().optional().describe("Hermes cron run id, for traceability."),
+      hermes_run_id: z34.string().optional().describe("Agent run id, for traceability."),
       source_intel_ids: z34.array(z34.string()).optional().describe("Optional list of intel_signals.id rows that informed this draft."),
       response_format: z34.enum(["text", "json"]).optional()
     },
@@ -13469,11 +13531,11 @@ function registerHermesTools(server) {
   );
   server.tool(
     "record_observation",
-    'Record an agent observation ("Hermes noticed X this week"). Surfaces in UnifiedAnalytics > Playbook tab. Use for weekly reflection digests and mid-campaign pulse summaries.',
+    'Record an agent observation (e.g. "topic X engagement up 23% this week"). Surfaces in the analytics Playbook. Use for weekly reflection digests and mid-campaign pulse summaries.',
     {
-      summary: z34.string().min(1).max(2e3).describe("One-paragraph summary, founder-facing."),
+      summary: z34.string().min(1).max(2e3).describe("One-paragraph summary, shown to the account owner."),
       deltas: z34.record(z34.string(), z34.union([z34.number(), z34.string(), z34.boolean()])).optional().describe("Optional structured key/value payload (e.g. {topic_x_er_pct: 23})."),
-      run_id: z34.string().optional().describe("Hermes cron run id for traceability."),
+      run_id: z34.string().optional().describe("Agent run id for traceability."),
       response_format: z34.enum(["text", "json"]).optional()
     },
     async ({ summary, deltas, run_id, response_format }) => {
@@ -13543,7 +13605,7 @@ function registerHermesTools(server) {
   );
   server.tool(
     "record_campaign_spend",
-    "Log a campaign cost line. Use when Hermes incurs spend that should be attributed to a campaign \u2014 drafts, renders, paid amp. Categories: hermes_drafts / carousel_renders / analytics_pulls / paid_amplification / other. Read aggregate via get_active_campaigns.",
+    "Log a campaign cost line. Use when an autonomous agent incurs spend that should be attributed to a campaign \u2014 drafts, renders, paid amp. Read aggregate via get_active_campaigns.",
     {
       campaign_id: z34.string().min(1).max(200).describe("Campaign identifier (slug)."),
       category: z34.enum([
@@ -13587,7 +13649,7 @@ function registerHermesTools(server) {
   );
   server.tool(
     "get_active_campaigns",
-    "List currently-running campaigns with thesis, budget, hero format, and current spend. Used by Hermes pitch skill to bias drafts toward active campaign themes.",
+    "List currently-running campaigns with thesis, budget, hero format, and current spend. Use to bias drafts toward active campaign themes.",
     {
       response_format: z34.enum(["text", "json"]).optional()
     },
@@ -13730,7 +13792,7 @@ ${blocks}` }]
   );
   server.tool(
     "run_skill",
-    "Run a Social Neuron workflow skill end-to-end (brand-locked content production). PR #4.4 v1: returns a structured run preview with the exact step plan, credit cost, and a deep-link to launch the run in the SN dashboard. PR #4.5 (next release) executes in-process so you can stream step-by-step progress back to the user. Call list_skills first to discover available skill ids.",
+    "Run a Social Neuron workflow skill end-to-end (brand-locked content production). Returns a structured run preview with the exact step plan, credit cost, and a deep-link to launch the run in the SN dashboard. A future release executes in-process so you can stream step-by-step progress back to the user. Call list_skills first to discover available skill ids.",
     {
       skill_id: z35.string().describe(
         'The id of the skill to run (e.g. "skill-brand-locked-viral-hook-reel"). Use list_skills to discover available ids.'
@@ -13775,7 +13837,7 @@ ${blocks}` }]
                   inputs,
                   runUrl,
                   previewedAt,
-                  note: "PR #4.4 v1 returns a preview. PR #4.5 will execute in-process via the run-skill EF."
+                  note: "This release returns a preview. A future release will execute in-process."
                 }),
                 null,
                 2
@@ -13804,7 +13866,7 @@ ${blocks}` }]
         "",
         `Launch in dashboard: ${runUrl}`,
         "",
-        "\u26A0\uFE0F  PR #4.4 v1: preview only. End-to-end MCP execution arrives in the next release."
+        "\u26A0\uFE0F  Preview only. End-to-end MCP execution arrives in a future release."
       ];
       return {
         content: [{ type: "text", text: lines.join("\n") }]
@@ -13824,7 +13886,7 @@ function asEnvelope27(data) {
 function registerLoopPulseTools(server) {
   server.tool(
     "get_loop_pulse",
-    'Read the dynamic loop-health KPIs for the Social Neuron growth loop over the last 7 days. Returns reflection coverage, decision coverage, visual gate pass rate, bandit-update application rate, per-platform bandit uptake, autopilot lag, and pattern aggregation counts \u2014 each with a status ("ok" / "warn" / "bad") and a why-line explaining what the metric measures. Use this to decide whether the loop is closing or where it is stuck before recommending next moves.',
+    'Read the dynamic loop-health KPIs for the Social Neuron growth loop over the last 7 days. Returns reflection coverage, decision coverage, visual gate pass rate, learning-update application rate, per-platform learning uptake, autopilot lag, and pattern aggregation counts \u2014 each with a status ("ok" / "warn" / "bad") and a why-line explaining what the metric measures. Use this to decide whether the loop is closing or where it is stuck before recommending next moves.',
     {
       response_format: z36.enum(["text", "json"]).optional().describe("Output format. Defaults to text.")
     },
@@ -13880,7 +13942,7 @@ function asEnvelope28(data) {
 function registerBanditStateTools(server) {
   server.tool(
     "get_bandit_state",
-    "Read the current Thompson Sampling bandit posteriors for a project. Returns top-K arms per (arm_type, platform) with Beta(alpha, beta) posterior mean and uncertainty. Use this to reason about which hook family / format / timing slot the bandit currently prefers on each platform before recommending next moves. SN real arm types: hook_family (6 fallback families), length_bucket (xs/s/m/l/xl by platform), posting_time_bucket (morning/midday/evening/late), content_format (video/carousel/image/caption/text/avatar/storyboard). Legacy/dead-taxonomy types also present: hook_type, format, timing_slot, topic_cluster, caption_style, platform, story_type, emoji_type.",
+    "Read the current content learning state for a project. Returns top-K arms per (arm_type, platform) with expected performance and uncertainty. Use this to reason about which hook family / format / timing slot currently performs best on each platform before recommending next moves. Arm types: hook_family, length_bucket (xs/s/m/l/xl by platform), posting_time_bucket (morning/midday/evening/late), content_format (video/carousel/image/caption/text/avatar/storyboard).",
     {
       project_id: z37.string().uuid().optional().describe("Project UUID. Defaults to the authenticated user default project."),
       platform: z37.enum([
@@ -13931,9 +13993,7 @@ function registerBanditStateTools(server) {
       });
       if (error || !data) {
         return {
-          content: [
-            { type: "text", text: `Failed to read bandit state: ${error || "no data"}` }
-          ],
+          content: [{ type: "text", text: `Failed to read bandit state: ${error || "no data"}` }],
           isError: true
         };
       }
@@ -15280,7 +15340,7 @@ async function computeDiscoveryCatalog() {
       err?.message
     );
   }
-  return TOOL_CATALOG.filter((t) => !t.localOnly).map((t) => ({
+  return TOOL_CATALOG.filter((t) => !t.localOnly && !t.internal).map((t) => ({
     name: t.name,
     description: t.description,
     inputSchema: schemaByName.get(t.name) ?? { type: "object", properties: {} }
@@ -15633,8 +15693,8 @@ app.get("/.well-known/mcp/server-card.json", (_req, res) => {
       required: true,
       schemes: ["oauth2"]
     },
-    toolCount: TOOL_CATALOG.filter((t) => !t.localOnly).length,
-    tools: TOOL_CATALOG.filter((t) => !t.localOnly).map((t) => ({
+    toolCount: TOOL_CATALOG.filter((t) => !t.localOnly && !t.internal).length,
+    tools: TOOL_CATALOG.filter((t) => !t.localOnly && !t.internal).map((t) => ({
       name: t.name,
       description: t.description,
       module: t.module,

package/dist/index.js

@@ -19,7 +19,7 @@ var MCP_VERSION;
 var init_version = __esm({
   "src/lib/version.ts"() {
     "use strict";
-    MCP_VERSION = "1.7.14";
+    MCP_VERSION = "1.7.15";
   }
 });
 
@@ -1235,7 +1235,7 @@ var init_tool_catalog = __esm({
       // niche research
       {
         name: "find_winning_content",
-        description: "Find QA-gated high-performing short-form videos in the project's niche. Returns extracted hook patterns, content structures, and pre-compiled replication prompts (backed by niche_winners view, qa_score >= 0.5).",
+        description: "Find QA-gated high-performing short-form videos in the project's niche. Returns extracted hook patterns, content structures, and pre-compiled replication prompts.",
         module: "research",
         scope: "mcp:read"
       },
@@ -1386,59 +1386,68 @@ var init_tool_catalog = __esm({
       // agentic-harness — learning loop write-back
       {
         name: "write_agent_reflection",
-        description: "Persist a verbal reflection for an agent loop. Provenance keys are restricted (Anti-Goodhart safety): only content_history_id, outcome_event_id, prm_score_ids, and handoff_ids are accepted.",
+        description: "Persist a verbal reflection for an agent loop. Provenance keys are restricted to an allowlist: only content_history_id, outcome_event_id, prm_score_ids, and handoff_ids are accepted.",
         module: "harness",
-        scope: "mcp:write"
+        scope: "mcp:write",
+        internal: true
       },
       {
         name: "record_outcome",
-        description: "Record an outcome for a published decision event. Idempotent on (decision_event_id, horizon). Only horizon=24h triggers a content_bandits posterior update.",
+        description: "Record an outcome for a published decision event. Idempotent on (decision_event_id, horizon). Only horizon=24h triggers a learning-loop update.",
         module: "harness",
-        scope: "mcp:write"
+        scope: "mcp:write",
+        internal: true
       },
       // agentic-harness — learning loop read-back
       {
         name: "read_agent_reflection",
         description: "Read past agent reflections for a brand. Ordered by created_at DESC, id ASC (deterministic tiebreak). Only active reflections returned (superseded_by IS NULL). Optional generated_by_agent filter.",
         module: "harness",
-        scope: "mcp:read"
+        scope: "mcp:read",
+        internal: true
       },
       // hermes — autonomous agent integration (closed-loop content)
       {
         name: "save_draft_to_library",
-        description: "Save a draft post to the SN content library with origin=hermes. Used by Hermes to persist drafts before the founder approves them.",
+        description: "Save a draft post to the SN content library for review before publishing. Drafts land in the content library pending approval.",
         module: "hermes",
-        scope: "mcp:write"
+        scope: "mcp:write",
+        internal: true
       },
       {
         name: "record_voice_lesson",
-        description: "Persist a learned voice lesson to brand_profiles.brand_context.voiceProfile.voice_lessons via atomic RPC. Used by Hermes reflection cron.",
+        description: "Persist a learned voice lesson to the brand voice profile.",
         module: "hermes",
-        scope: "mcp:write"
+        scope: "mcp:write",
+        internal: true
       },
       {
         name: "record_observation",
-        description: 'Record an agent observation (e.g. "topic X engagement up 23%") for the UnifiedAnalytics > Playbook surface.',
+        description: 'Record an agent observation (e.g. "topic X engagement up 23%") for the analytics playbook.',
         module: "hermes",
-        scope: "mcp:write"
+        scope: "mcp:write",
+        internal: true
       },
       {
         name: "record_intel_signal",
-        description: "Record a research/trend signal from Hermes watchers (news, HN, competitor, etc.) for Niche Intelligence. Dedupes by URL.",
+        description: "Record a research/trend signal (news, competitor, community sources) for niche intelligence. Dedupes by URL.",
         module: "hermes",
-        scope: "mcp:write"
+        scope: "mcp:write",
+        internal: true
       },
       {
         name: "record_campaign_spend",
-        description: "Log a campaign cost line (hermes_drafts, carousel_renders, analytics_pulls, paid_amplification, other). Ownership-checked.",
+        description: "Log a campaign cost line item. Ownership-checked.",
         module: "hermes",
-        scope: "mcp:write"
+        scope: "mcp:write",
+        internal: true
       },
       {
         name: "get_active_campaigns",
-        description: "List currently-running campaigns with thesis, budget, hero format, and current spend. Used by Hermes pitch skill to bias drafts.",
+        description: "List currently-running campaigns with thesis, budget, hero format, and current spend.",
         module: "hermes",
-        scope: "mcp:read"
+        scope: "mcp:read",
+        internal: true
       },
       // skills (workflow skills — multi-step brand-locked content pipelines)
       {
@@ -1449,20 +1458,20 @@ var init_tool_catalog = __esm({
       },
       {
         name: "run_skill",
-        description: "Run a Social Neuron workflow skill end-to-end (brand-locked content production). PR #4.4 v1 returns a structured run preview with the step plan, credit cost, and a deep-link to launch in the SN dashboard.",
+        description: "Run a Social Neuron workflow skill end-to-end (brand-locked content production). Returns a structured run preview with the step plan, credit cost, and a deep-link to launch in the SN dashboard.",
         module: "skills",
         scope: "mcp:write"
       },
-      // loop observability (growth-loop KPIs + Thompson Sampling bandit posteriors)
+      // loop observability (growth-loop KPIs + content learning state)
       {
         name: "get_loop_pulse",
-        description: "Read dynamic loop-health KPIs for the growth loop over the last 7 days (reflection/decision coverage, visual gate pass rate, bandit-update application rate, per-platform uptake, autopilot lag) \u2014 each with an ok/warn/bad status. Use to decide whether the loop is closing or where it is stuck.",
+        description: "Read dynamic loop-health KPIs for the growth loop over the last 7 days (reflection/decision coverage, visual gate pass rate, learning-update application rate, per-platform uptake, autopilot lag) \u2014 each with an ok/warn/bad status. Use to decide whether the loop is closing or where it is stuck.",
         module: "loop",
         scope: "mcp:read"
       },
       {
         name: "get_bandit_state",
-        description: "Read the current Thompson Sampling bandit posteriors for a project \u2014 top-K arms per (arm_type, platform) with Beta(alpha,beta) posterior mean and uncertainty. Use to reason about which hook family / format / timing slot the bandit currently prefers per platform.",
+        description: "Read the current content learning state for a project \u2014 top-K arms per (arm_type, platform) with expected performance and uncertainty. Use to reason about which hook family / format / timing slot currently performs best per platform.",
         module: "loop",
         scope: "mcp:read"
       }
@@ -2637,6 +2646,7 @@ async function handleTools(args, asJson) {
   if (typeof module === "string") {
     tools = getToolsByModule(module);
   }
+  tools = tools.filter((t) => !t.internal);
   if (asJson) {
     emitSnResult({ ok: true, command: "tools", toolCount: tools.length, tools }, true);
     return;
@@ -2667,7 +2677,7 @@ Module: ${moduleName} (${moduleTools.length} tools)`);
 async function handleInfo(args, asJson) {
   const info = {
     version: MCP_VERSION,
-    toolCount: TOOL_CATALOG.length,
+    toolCount: TOOL_CATALOG.filter((t) => !t.internal).length,
     modules: getModules()
   };
   try {
@@ -4270,10 +4280,10 @@ var TOOL_SCOPES = {
   record_campaign_spend: "mcp:write",
   // mcp:read
   get_active_campaigns: "mcp:read",
-  // mcp:read / mcp:write (Skills — PR #4.4)
+  // mcp:read / mcp:write (Skills)
   list_skills: "mcp:read",
   run_skill: "mcp:write",
-  // mcp:read (Loop observability — growth-loop KPIs + bandit posteriors)
+  // mcp:read (Loop observability — growth-loop KPIs + content learning state)
   get_loop_pulse: "mcp:read",
   get_bandit_state: "mcp:read"
 };
@@ -12615,7 +12625,10 @@ function toolKnowledgeDocument(tool) {
   };
 }
 function getKnowledgeDocuments() {
-  return [...STATIC_KNOWLEDGE_DOCUMENTS, ...TOOL_CATALOG.map(toolKnowledgeDocument)];
+  return [
+    ...STATIC_KNOWLEDGE_DOCUMENTS,
+    ...TOOL_CATALOG.filter((t) => !t.internal).map(toolKnowledgeDocument)
+  ];
 }
 function tokenize(input) {
   return input.toLowerCase().split(/[^a-z0-9:_-]+/).map((token) => token.trim()).filter(Boolean);
@@ -12731,6 +12744,7 @@ function registerDiscoveryTools(server2) {
       if (query) {
         results = searchTools(query);
       }
+      results = results.filter((t) => !t.internal);
       if (module) {
         const moduleTools = getToolsByModule(module);
         results = results.filter((t) => moduleTools.some((mt) => mt.name === t.name));
@@ -12810,6 +12824,7 @@ var PLATFORM_ENUM2 = z24.enum([
 ]);
 var BASE_PLAN_CREDITS = 15;
 var SOURCE_EXTRACTION_CREDITS = 5;
+var SCHEDULE_POST_CREDITS = 1;
 function registerPipelineTools(server2) {
   server2.tool(
     "check_pipeline_readiness",
@@ -12926,7 +12941,7 @@ function registerPipelineTools(server2) {
   );
   server2.tool(
     "run_content_pipeline",
-    "Run the full content pipeline: research trends \u2192 generate plan \u2192 quality check \u2192 auto-approve \u2192 schedule posts. Chains all stages in one call for maximum efficiency. Set dry_run=true to preview the plan without publishing. Check check_pipeline_readiness first to verify credits, OAuth, and brand profile are ready.",
+    "Run the full content pipeline: research trends \u2192 generate plan \u2192 quality check \u2192 auto-approve \u2192 schedule posts. Chains all stages in one call for maximum efficiency. Set dry_run=true to preview the plan without publishing. To schedule posts, set schedule_confirmed=true after the user explicitly approves publishing. Check check_pipeline_readiness first to verify credits, OAuth, and brand profile are ready.",
     {
       project_id: z24.string().uuid().optional().describe("Project ID (auto-detected if omitted)"),
       topic: z24.string().optional().describe("Content topic (required if no source_url)"),
@@ -12942,6 +12957,9 @@ function registerPipelineTools(server2) {
       ),
       max_credits: z24.number().optional().describe("Credit budget cap"),
       dry_run: z24.boolean().default(false).describe("If true, skip scheduling and return plan only"),
+      schedule_confirmed: z24.boolean().default(false).describe(
+        "Required to schedule posts. Set true only after explicit user confirmation to publish/schedule."
+      ),
       skip_stages: z24.array(z24.enum(["research", "quality", "schedule"])).optional().describe("Stages to skip"),
       response_format: z24.enum(["text", "json"]).default("json")
     },
@@ -12956,6 +12974,7 @@ function registerPipelineTools(server2) {
       auto_approve_threshold,
       max_credits,
       dry_run,
+      schedule_confirmed,
       skip_stages,
       response_format
     }) => {
@@ -12971,6 +12990,29 @@ function registerPipelineTools(server2) {
         };
       }
       const skipSet = new Set(skip_stages ?? []);
+      const schedulingRequested = !dry_run && !skipSet.has("schedule");
+      if (schedulingRequested && !schedule_confirmed) {
+        return {
+          content: [
+            {
+              type: "text",
+              text: 'Scheduling requires explicit confirmation. Re-run with schedule_confirmed=true after the user approves publishing, or set dry_run=true / skip_stages=["schedule"].'
+            }
+          ],
+          isError: true
+        };
+      }
+      if (schedulingRequested && skipSet.has("quality")) {
+        return {
+          content: [
+            {
+              type: "text",
+              text: "Scheduling cannot run when the quality stage is skipped."
+            }
+          ],
+          isError: true
+        };
+      }
       try {
         const resolvedProjectId = project_id ?? await getDefaultProjectId() ?? void 0;
         const estimatedCost = BASE_PLAN_CREDITS + (source_url ? SOURCE_EXTRACTION_CREDITS : 0);
@@ -13013,6 +13055,7 @@ function registerPipelineTools(server2) {
               approval_mode,
               auto_approve_threshold,
               dry_run,
+              schedule_confirmed,
               skip_stages: skip_stages ?? []
             },
             current_stage: "planning",
@@ -13078,7 +13121,9 @@ function registerPipelineTools(server2) {
         stagesCompleted.push("planning");
         const rawText = String(planData.text ?? planData.content ?? "");
         const postsArray = extractJsonArray(rawText);
-        const posts = (postsArray ?? []).map((p) => ({
+        const requestedPlatformSet = new Set(platforms);
+        const maxPosts = platforms.length * days * posts_per_day;
+        const parsedPosts = (postsArray ?? []).map((p) => ({
           id: String(p.id ?? randomUUID3().slice(0, 8)),
           day: Number(p.day ?? 1),
           date: String(p.date ?? ""),
@@ -13092,6 +13137,23 @@ function registerPipelineTools(server2) {
           visual_direction: p.visual_direction ? String(p.visual_direction) : void 0,
           media_type: p.media_type ? String(p.media_type) : void 0
         }));
+        const platformFilteredPosts = parsedPosts.filter(
+          (post) => requestedPlatformSet.has(post.platform)
+        );
+        const posts = platformFilteredPosts.slice(0, maxPosts);
+        if (parsedPosts.length > maxPosts) {
+          errors.push({
+            stage: "planning",
+            message: `AI returned ${parsedPosts.length} posts; truncated to ${maxPosts}.`
+          });
+        }
+        const invalidPlatformCount = parsedPosts.length - platformFilteredPosts.length;
+        if (invalidPlatformCount > 0) {
+          errors.push({
+            stage: "planning",
+            message: `Dropped ${invalidPlatformCount} post(s) with unrequested or invalid platform.`
+          });
+        }
         let postsApproved = 0;
         let postsFlagged = 0;
         if (!skipSet.has("quality")) {
@@ -13244,6 +13306,7 @@ function registerPipelineTools(server2) {
                 });
               } else {
                 postsScheduled++;
+                creditsUsed += SCHEDULE_POST_CREDITS;
               }
             } catch (schedErr) {
               errors.push({
@@ -15313,13 +15376,13 @@ function asEnvelope24(data) {
 function registerNicheResearchTools(server2) {
   server2.tool(
     "find_winning_content",
-    "Find QA-gated high-performing short-form videos in the project's niche. Returns extracted hook patterns, content structures, and pre-compiled replication prompts you can use to generate new content on a different topic. Backed by niche_winners view (qa_score >= 0.5 + replication_prompt populated).",
+    "Find QA-gated high-performing short-form videos in the project's niche. Returns extracted hook patterns, content structures, and pre-compiled replication prompts you can use to generate new content on a different topic. Only QA-gated winners with a populated replication prompt are returned.",
     {
       project_id: z29.string().uuid().optional().describe("Project ID (auto-detected if omitted)"),
       platform: z29.enum(["tiktok", "instagram", "youtube", "reddit", "twitter"]).optional().describe("Filter to one platform. Omit for all platforms."),
       days: z29.number().int().min(1).max(365).default(30).describe("Window: only return winners scanned within this many days."),
       limit: z29.number().int().min(1).max(50).default(10).describe("Number of winners to return (1-50)."),
-      min_qa_score: z29.number().min(0).max(1).default(0.5).describe("Minimum QA score (0..1). Default 0.5 matches the niche_winners view floor."),
+      min_qa_score: z29.number().min(0).max(1).default(0.5).describe("Minimum QA score (0..1). Default 0.5."),
       response_format: z29.enum(["text", "json"]).optional()
     },
     async ({ project_id, platform: platform3, days, limit, min_qa_score, response_format }) => {
@@ -15981,7 +16044,7 @@ var writeReflectionSchema = {
     prm_score_ids: z33.array(z33.string()).optional().describe("PRM score IDs that informed this reflection."),
     handoff_ids: z33.array(z33.string()).optional().describe("Handoff event IDs that triggered this reflection.")
   }).strict().describe(
-    "Source evidence for this reflection. Only these four keys are accepted (Anti-Goodhart guard \u2014 unknown keys are rejected to prevent spurious provenance claims)."
+    "Source evidence for this reflection. Only these four keys are accepted \u2014 unknown keys are rejected to prevent spurious provenance claims."
   ),
   brand_id: z33.string().describe("Brand profile UUID this reflection belongs to."),
   pipeline_id: z33.string().optional().describe("Optional: pipeline run UUID."),
@@ -15999,7 +16062,7 @@ var readReflectionSchema = {
 var recordOutcomeSchema = {
   decision_event_id: z33.string().describe("UUID of the decision_events row to record an outcome for."),
   horizon: z33.enum(["1h", "6h", "24h"]).describe(
-    "Observation horizon. Only horizon=24h with a non-null reward triggers a content_bandits posterior update; 1h/6h are stored but inert for learning."
+    "Observation horizon. Only horizon=24h with a non-null reward triggers a learning-loop update; 1h/6h are stored but inert for learning."
   ),
   reward: z33.number().min(0).max(1).describe("Normalised reward signal in [0, 1]. Higher = better outcome."),
   outcome_metrics: z33.record(z33.string(), z33.number()).optional().describe(
@@ -16009,7 +16072,7 @@ var recordOutcomeSchema = {
 function registerHarnessTools(server2, _ctx) {
   server2.tool(
     "write_agent_reflection",
-    "Persist a verbal reflection for an agent loop. Provenance keys are restricted (Anti-Goodhart safety): only content_history_id, outcome_event_id, prm_score_ids, and handoff_ids are accepted \u2014 unknown keys are rejected at the input layer. Requires mcp:write scope. Returns the created reflection UUID on success.",
+    "Persist a verbal reflection for an agent loop. Provenance keys are restricted to an allowlist: only content_history_id, outcome_event_id, prm_score_ids, and handoff_ids are accepted \u2014 unknown keys are rejected at the input layer. Requires mcp:write scope. Returns the created reflection UUID on success.",
     writeReflectionSchema,
     // eslint-disable-next-line @typescript-eslint/no-explicit-any
     async (args) => {
@@ -16037,7 +16100,7 @@ function registerHarnessTools(server2, _ctx) {
   );
   server2.tool(
     "record_outcome",
-    "Record an outcome for a published decision event. Idempotent on (decision_event_id, horizon) \u2014 safe to call multiple times. Returns idempotent:true when the row already existed (UPDATE), idempotent:false on fresh INSERT. Note: only horizon=24h with reward != null triggers a content_bandits posterior update; 1h/6h outcomes are stored but are inert for the learning loop. Requires mcp:write scope.",
+    "Record an outcome for a published decision event. Idempotent on (decision_event_id, horizon) \u2014 safe to call multiple times. Returns idempotent:true when the row already existed (UPDATE), idempotent:false on fresh INSERT. Note: only horizon=24h with reward != null triggers a learning-loop update; 1h/6h outcomes are stored but are inert for the learning loop. Requires mcp:write scope.",
     recordOutcomeSchema,
     // eslint-disable-next-line @typescript-eslint/no-explicit-any
     async (args) => {
@@ -16123,13 +16186,13 @@ var PLATFORM = z34.enum([
 function registerHermesTools(server2) {
   server2.tool(
     "save_draft_to_library",
-    "Save a draft post to the SN content library. Use when an autonomous agent (Hermes) wants to persist a draft before the founder approves it. Lands in ContentLibrary with status='draft', origin='hermes'. The draft can then be approved/edited in the SN UI.",
+    "Save a draft post to the SN content library. Use when an autonomous agent wants to persist a draft for review before publishing. Lands in the content library with status='draft'. The draft can then be approved/edited in the SN UI.",
     {
       platform: PLATFORM.describe("Target platform for the draft."),
       copy: z34.string().min(1).max(8e3).describe("The draft post body."),
       project_id: z34.string().optional().describe("SN project UUID. Optional but recommended."),
       media_url: z34.string().url().optional().describe("Optional cover/media URL (R2 signed URL)."),
-      hermes_run_id: z34.string().optional().describe("Hermes cron run id, for traceability."),
+      hermes_run_id: z34.string().optional().describe("Agent run id, for traceability."),
       source_intel_ids: z34.array(z34.string()).optional().describe("Optional list of intel_signals.id rows that informed this draft."),
       response_format: z34.enum(["text", "json"]).optional()
     },
@@ -16211,11 +16274,11 @@ function registerHermesTools(server2) {
   );
   server2.tool(
     "record_observation",
-    'Record an agent observation ("Hermes noticed X this week"). Surfaces in UnifiedAnalytics > Playbook tab. Use for weekly reflection digests and mid-campaign pulse summaries.',
+    'Record an agent observation (e.g. "topic X engagement up 23% this week"). Surfaces in the analytics Playbook. Use for weekly reflection digests and mid-campaign pulse summaries.',
     {
-      summary: z34.string().min(1).max(2e3).describe("One-paragraph summary, founder-facing."),
+      summary: z34.string().min(1).max(2e3).describe("One-paragraph summary, shown to the account owner."),
       deltas: z34.record(z34.string(), z34.union([z34.number(), z34.string(), z34.boolean()])).optional().describe("Optional structured key/value payload (e.g. {topic_x_er_pct: 23})."),
-      run_id: z34.string().optional().describe("Hermes cron run id for traceability."),
+      run_id: z34.string().optional().describe("Agent run id for traceability."),
       response_format: z34.enum(["text", "json"]).optional()
     },
     async ({ summary, deltas, run_id, response_format }) => {
@@ -16285,7 +16348,7 @@ function registerHermesTools(server2) {
   );
   server2.tool(
     "record_campaign_spend",
-    "Log a campaign cost line. Use when Hermes incurs spend that should be attributed to a campaign \u2014 drafts, renders, paid amp. Categories: hermes_drafts / carousel_renders / analytics_pulls / paid_amplification / other. Read aggregate via get_active_campaigns.",
+    "Log a campaign cost line. Use when an autonomous agent incurs spend that should be attributed to a campaign \u2014 drafts, renders, paid amp. Read aggregate via get_active_campaigns.",
     {
       campaign_id: z34.string().min(1).max(200).describe("Campaign identifier (slug)."),
       category: z34.enum([
@@ -16329,7 +16392,7 @@ function registerHermesTools(server2) {
   );
   server2.tool(
     "get_active_campaigns",
-    "List currently-running campaigns with thesis, budget, hero format, and current spend. Used by Hermes pitch skill to bias drafts toward active campaign themes.",
+    "List currently-running campaigns with thesis, budget, hero format, and current spend. Use to bias drafts toward active campaign themes.",
     {
       response_format: z34.enum(["text", "json"]).optional()
     },
@@ -16473,7 +16536,7 @@ ${blocks}` }]
   );
   server2.tool(
     "run_skill",
-    "Run a Social Neuron workflow skill end-to-end (brand-locked content production). PR #4.4 v1: returns a structured run preview with the exact step plan, credit cost, and a deep-link to launch the run in the SN dashboard. PR #4.5 (next release) executes in-process so you can stream step-by-step progress back to the user. Call list_skills first to discover available skill ids.",
+    "Run a Social Neuron workflow skill end-to-end (brand-locked content production). Returns a structured run preview with the exact step plan, credit cost, and a deep-link to launch the run in the SN dashboard. A future release executes in-process so you can stream step-by-step progress back to the user. Call list_skills first to discover available skill ids.",
     {
       skill_id: z35.string().describe(
         'The id of the skill to run (e.g. "skill-brand-locked-viral-hook-reel"). Use list_skills to discover available ids.'
@@ -16518,7 +16581,7 @@ ${blocks}` }]
                   inputs,
                   runUrl,
                   previewedAt,
-                  note: "PR #4.4 v1 returns a preview. PR #4.5 will execute in-process via the run-skill EF."
+                  note: "This release returns a preview. A future release will execute in-process."
                 }),
                 null,
                 2
@@ -16547,7 +16610,7 @@ ${blocks}` }]
         "",
         `Launch in dashboard: ${runUrl}`,
         "",
-        "\u26A0\uFE0F  PR #4.4 v1: preview only. End-to-end MCP execution arrives in the next release."
+        "\u26A0\uFE0F  Preview only. End-to-end MCP execution arrives in a future release."
       ];
       return {
         content: [{ type: "text", text: lines.join("\n") }]
@@ -16569,7 +16632,7 @@ function asEnvelope27(data) {
 function registerLoopPulseTools(server2) {
   server2.tool(
     "get_loop_pulse",
-    'Read the dynamic loop-health KPIs for the Social Neuron growth loop over the last 7 days. Returns reflection coverage, decision coverage, visual gate pass rate, bandit-update application rate, per-platform bandit uptake, autopilot lag, and pattern aggregation counts \u2014 each with a status ("ok" / "warn" / "bad") and a why-line explaining what the metric measures. Use this to decide whether the loop is closing or where it is stuck before recommending next moves.',
+    'Read the dynamic loop-health KPIs for the Social Neuron growth loop over the last 7 days. Returns reflection coverage, decision coverage, visual gate pass rate, learning-update application rate, per-platform learning uptake, autopilot lag, and pattern aggregation counts \u2014 each with a status ("ok" / "warn" / "bad") and a why-line explaining what the metric measures. Use this to decide whether the loop is closing or where it is stuck before recommending next moves.',
     {
       response_format: z36.enum(["text", "json"]).optional().describe("Output format. Defaults to text.")
     },
@@ -16627,7 +16690,7 @@ function asEnvelope28(data) {
 function registerBanditStateTools(server2) {
   server2.tool(
     "get_bandit_state",
-    "Read the current Thompson Sampling bandit posteriors for a project. Returns top-K arms per (arm_type, platform) with Beta(alpha, beta) posterior mean and uncertainty. Use this to reason about which hook family / format / timing slot the bandit currently prefers on each platform before recommending next moves. SN real arm types: hook_family (6 fallback families), length_bucket (xs/s/m/l/xl by platform), posting_time_bucket (morning/midday/evening/late), content_format (video/carousel/image/caption/text/avatar/storyboard). Legacy/dead-taxonomy types also present: hook_type, format, timing_slot, topic_cluster, caption_style, platform, story_type, emoji_type.",
+    "Read the current content learning state for a project. Returns top-K arms per (arm_type, platform) with expected performance and uncertainty. Use this to reason about which hook family / format / timing slot currently performs best on each platform before recommending next moves. Arm types: hook_family, length_bucket (xs/s/m/l/xl by platform), posting_time_bucket (morning/midday/evening/late), content_format (video/carousel/image/caption/text/avatar/storyboard).",
     {
       project_id: z37.string().uuid().optional().describe("Project UUID. Defaults to the authenticated user default project."),
       platform: z37.enum([
@@ -16678,9 +16741,7 @@ function registerBanditStateTools(server2) {
       });
       if (error || !data) {
         return {
-          content: [
-            { type: "text", text: `Failed to read bandit state: ${error || "no data"}` }
-          ],
+          content: [{ type: "text", text: `Failed to read bandit state: ${error || "no data"}` }],
           isError: true
         };
       }

package/dist/sn.js

@@ -19,7 +19,7 @@ var MCP_VERSION;
 var init_version = __esm({
   "src/lib/version.ts"() {
     "use strict";
-    MCP_VERSION = "1.7.14";
+    MCP_VERSION = "1.7.15";
   }
 });
 
@@ -2377,7 +2377,7 @@ var init_tool_catalog = __esm({
       // niche research
       {
         name: "find_winning_content",
-        description: "Find QA-gated high-performing short-form videos in the project's niche. Returns extracted hook patterns, content structures, and pre-compiled replication prompts (backed by niche_winners view, qa_score >= 0.5).",
+        description: "Find QA-gated high-performing short-form videos in the project's niche. Returns extracted hook patterns, content structures, and pre-compiled replication prompts.",
         module: "research",
         scope: "mcp:read"
       },
@@ -2528,59 +2528,68 @@ var init_tool_catalog = __esm({
       // agentic-harness — learning loop write-back
       {
         name: "write_agent_reflection",
-        description: "Persist a verbal reflection for an agent loop. Provenance keys are restricted (Anti-Goodhart safety): only content_history_id, outcome_event_id, prm_score_ids, and handoff_ids are accepted.",
+        description: "Persist a verbal reflection for an agent loop. Provenance keys are restricted to an allowlist: only content_history_id, outcome_event_id, prm_score_ids, and handoff_ids are accepted.",
         module: "harness",
-        scope: "mcp:write"
+        scope: "mcp:write",
+        internal: true
       },
       {
         name: "record_outcome",
-        description: "Record an outcome for a published decision event. Idempotent on (decision_event_id, horizon). Only horizon=24h triggers a content_bandits posterior update.",
+        description: "Record an outcome for a published decision event. Idempotent on (decision_event_id, horizon). Only horizon=24h triggers a learning-loop update.",
         module: "harness",
-        scope: "mcp:write"
+        scope: "mcp:write",
+        internal: true
       },
       // agentic-harness — learning loop read-back
       {
         name: "read_agent_reflection",
         description: "Read past agent reflections for a brand. Ordered by created_at DESC, id ASC (deterministic tiebreak). Only active reflections returned (superseded_by IS NULL). Optional generated_by_agent filter.",
         module: "harness",
-        scope: "mcp:read"
+        scope: "mcp:read",
+        internal: true
       },
       // hermes — autonomous agent integration (closed-loop content)
       {
         name: "save_draft_to_library",
-        description: "Save a draft post to the SN content library with origin=hermes. Used by Hermes to persist drafts before the founder approves them.",
+        description: "Save a draft post to the SN content library for review before publishing. Drafts land in the content library pending approval.",
         module: "hermes",
-        scope: "mcp:write"
+        scope: "mcp:write",
+        internal: true
       },
       {
         name: "record_voice_lesson",
-        description: "Persist a learned voice lesson to brand_profiles.brand_context.voiceProfile.voice_lessons via atomic RPC. Used by Hermes reflection cron.",
+        description: "Persist a learned voice lesson to the brand voice profile.",
         module: "hermes",
-        scope: "mcp:write"
+        scope: "mcp:write",
+        internal: true
       },
       {
         name: "record_observation",
-        description: 'Record an agent observation (e.g. "topic X engagement up 23%") for the UnifiedAnalytics > Playbook surface.',
+        description: 'Record an agent observation (e.g. "topic X engagement up 23%") for the analytics playbook.',
         module: "hermes",
-        scope: "mcp:write"
+        scope: "mcp:write",
+        internal: true
       },
       {
         name: "record_intel_signal",
-        description: "Record a research/trend signal from Hermes watchers (news, HN, competitor, etc.) for Niche Intelligence. Dedupes by URL.",
+        description: "Record a research/trend signal (news, competitor, community sources) for niche intelligence. Dedupes by URL.",
         module: "hermes",
-        scope: "mcp:write"
+        scope: "mcp:write",
+        internal: true
       },
       {
         name: "record_campaign_spend",
-        description: "Log a campaign cost line (hermes_drafts, carousel_renders, analytics_pulls, paid_amplification, other). Ownership-checked.",
+        description: "Log a campaign cost line item. Ownership-checked.",
         module: "hermes",
-        scope: "mcp:write"
+        scope: "mcp:write",
+        internal: true
       },
       {
         name: "get_active_campaigns",
-        description: "List currently-running campaigns with thesis, budget, hero format, and current spend. Used by Hermes pitch skill to bias drafts.",
+        description: "List currently-running campaigns with thesis, budget, hero format, and current spend.",
         module: "hermes",
-        scope: "mcp:read"
+        scope: "mcp:read",
+        internal: true
       },
       // skills (workflow skills — multi-step brand-locked content pipelines)
       {
@@ -2591,20 +2600,20 @@ var init_tool_catalog = __esm({
       },
       {
         name: "run_skill",
-        description: "Run a Social Neuron workflow skill end-to-end (brand-locked content production). PR #4.4 v1 returns a structured run preview with the step plan, credit cost, and a deep-link to launch in the SN dashboard.",
+        description: "Run a Social Neuron workflow skill end-to-end (brand-locked content production). Returns a structured run preview with the step plan, credit cost, and a deep-link to launch in the SN dashboard.",
         module: "skills",
         scope: "mcp:write"
       },
-      // loop observability (growth-loop KPIs + Thompson Sampling bandit posteriors)
+      // loop observability (growth-loop KPIs + content learning state)
       {
         name: "get_loop_pulse",
-        description: "Read dynamic loop-health KPIs for the growth loop over the last 7 days (reflection/decision coverage, visual gate pass rate, bandit-update application rate, per-platform uptake, autopilot lag) \u2014 each with an ok/warn/bad status. Use to decide whether the loop is closing or where it is stuck.",
+        description: "Read dynamic loop-health KPIs for the growth loop over the last 7 days (reflection/decision coverage, visual gate pass rate, learning-update application rate, per-platform uptake, autopilot lag) \u2014 each with an ok/warn/bad status. Use to decide whether the loop is closing or where it is stuck.",
         module: "loop",
         scope: "mcp:read"
       },
       {
         name: "get_bandit_state",
-        description: "Read the current Thompson Sampling bandit posteriors for a project \u2014 top-K arms per (arm_type, platform) with Beta(alpha,beta) posterior mean and uncertainty. Use to reason about which hook family / format / timing slot the bandit currently prefers per platform.",
+        description: "Read the current content learning state for a project \u2014 top-K arms per (arm_type, platform) with expected performance and uncertainty. Use to reason about which hook family / format / timing slot currently performs best per platform.",
         module: "loop",
         scope: "mcp:read"
       }
@@ -2628,6 +2637,7 @@ async function handleTools(args, asJson) {
   if (typeof module === "string") {
     tools = getToolsByModule(module);
   }
+  tools = tools.filter((t) => !t.internal);
   if (asJson) {
     emitSnResult({ ok: true, command: "tools", toolCount: tools.length, tools }, true);
     return;
@@ -2658,7 +2668,7 @@ Module: ${moduleName} (${moduleTools.length} tools)`);
 async function handleInfo(args, asJson) {
   const info = {
     version: MCP_VERSION,
-    toolCount: TOOL_CATALOG.length,
+    toolCount: TOOL_CATALOG.filter((t) => !t.internal).length,
     modules: getModules()
   };
   try {

package/package.json

@@ -1,6 +1,7 @@
 {
   "name": "@socialneuron/mcp-server",
-  "version": "1.7.14",
+  "mcpName": "com.socialneuron/mcp-server",
+  "version": "1.7.15",
   "description": "Official MCP server, REST API & CLI for Social Neuron \u2014 create, schedule, and optimize social content with 35+ AI models across YouTube, TikTok, and more.",
   "type": "module",
   "main": "dist/index.js",
@@ -78,7 +79,8 @@
     "test:watch": "vitest",
     "clean": "rm -rf dist/",
     "prepublishOnly": "npm run clean && npm run build:all && npm test",
-    "build:docs": "node scripts/build-tools-reference.mjs"
+    "build:docs": "node scripts/build-tools-reference.mjs",
+    "verify:metadata": "node scripts/verify-metadata.mjs"
   },
   "dependencies": {
     "@modelcontextprotocol/ext-apps": "1.7.4",

package/tools.lock.json

@@ -43,16 +43,16 @@
     "fetch_trends": "9efd2a24c9f9d9a3a6fa707f2355c8ef0d12dbc25f416e9237c0b9b73d15007b",
     "fetch_youtube_analytics": "c1d6cf67a5d0ad1696b4320ee91de7ddf58af6cf4a3be537636be746a4822af7",
     "find_next_slots": "fccf3bc7a8ac54b7ee243c59272252dbfeb307ca8d2987b1527b3b904408c909",
-    "find_winning_content": "d5112618e5c386ff7cde1a18ec694284365888794b8b0d44154b4e2e527f4ff3",
+    "find_winning_content": "f124ae937ea5c3bda8ff9eae82c0ad5b6b095c9a8b1b1812931e10a5b6b35766",
     "generate_carousel": "e6420e3ccc766f5f0fa2e783ecada527dc4ca23b2bb34954963c912eacba84a9",
     "generate_content": "1425c9bc1a8ce800448344a5dea61dd1359eb30872a563b7014d461a696fc6bb",
     "generate_image": "0aa3f14e2b4475ec5909f3b5cc141556b3315989af886a817d9fa7efa1be40e0",
     "generate_performance_digest": "8bbb53edc46698927a10372444239ba9973c5c84e2d0a8160f74192495dae74b",
     "generate_video": "371eda383af033b4f99a1f4447e8a555d16d10688995f6fd55b9c6eb0a619313",
     "generate_voiceover": "ced5ecaf3d936bb7ba0f17d2efe3c15594a061226fbcec1fb572ec87555eff20",
-    "get_active_campaigns": "3a28a6eb17131d48df8ec6938fba7af807a9c4148cb570dc4a82ae891730c631",
+    "get_active_campaigns": "2e593f72a0a4e1a7744049556444c64eda760a40f5c08d603c32b3b34b1d760a",
     "get_autopilot_status": "c6e689867a32a491a7afbbae46326d223bba412f0c66113ddf4be831c73fd769",
-    "get_bandit_state": "bcb21aa0c2ca055fd7093d05c1d78a23691627920de9dc1dd7087472c57b8a34",
+    "get_bandit_state": "1e58ef7836e24d6c9d474bea52916c15397611200e779ded692cc34fe882352f",
     "get_best_posting_times": "78932988a8a2238d81c7b2c065e95349fda3e226702998fb08262b74baff6a53",
     "get_brand_profile": "a0ebd63ebd7032de497751d2b35c0ec78c847869d64e54f7019baa5e532fed0e",
     "get_brand_runtime": "5ac3149d38eac100aca5d769ddb0f5d900832fb29eef78c18dc97678902e3a41",
@@ -60,7 +60,7 @@
     "get_content_plan": "24a3845da81533d8c22affd76ce28dc179d79f2556c50317eafafbff6b1c520a",
     "get_credit_balance": "e5a086ab75278640ec589893a7575aa561fe0cbc30bbbd67a328fb4032d3164d",
     "get_ideation_context": "fa07379c69b5045de910d8f70445b4bab0bded9f04a1cb41b157cab5a5385f9d",
-    "get_loop_pulse": "58fdca648d6da7b92f36ae08f94791697d5aebdd8e8084788833b04b7307143c",
+    "get_loop_pulse": "cf3c386feaaf6c07d0d1c831451da8f7043ab43ef1e685ae7f90c824f58dd4cd",
     "get_loop_summary": "e37bbc18048c039876dd4f4ba640dbb5e1efaf968fb887cb5479683e2840c1cc",
     "get_mcp_usage": "3e75743113a8e232c831f915edec38143528e92c504219ac3921d524f738df10",
     "get_media_url": "2132e420dc3df1a3c6b8d421b3530be23a7282b8f03af53d26707b337de73370",
@@ -84,22 +84,22 @@
     "quality_check": "a627658f804100276c6335cda8c2359edea163c155fa48aba6adb8749a80416a",
     "quality_check_plan": "9ce43929418427b4b256afc621283c7e887b97e208da849f4ae02418c2103954",
     "read_agent_reflection": "bdf834b25577eb7d8b992b7e7b2462949b2ffd6da664a38eba8b53d1ec8fa5a6",
-    "record_campaign_spend": "7cade3b29ccfb9bdc1634ec8199eea14abd79594da1bb1ad59f6f481db7ca2e6",
-    "record_intel_signal": "9f3db98ec4343e53b488367273383e15f1d89ef27a7e96b3b2db1aa3d6175a91",
-    "record_observation": "64f26ca189c6b4ebf6c02436fc3392ac12fdc86361ba5a50feb7d8c9dab1a849",
-    "record_outcome": "8d5be6f47e9d925bb8f13d6a94041e72c097337f34deecb246e576c7d651e122",
-    "record_voice_lesson": "47707d359ccdae2953e63aa8443ea981784c4e3f0a0e986ecc84ea6b564814db",
+    "record_campaign_spend": "a7d94f00b4f8875b21f851d5f608a590c958e0acf75cfc45f16c1688f8cd2d2e",
+    "record_intel_signal": "35e0841dd77b8d388692fdc15de13f745b8629a57444ca41fc431a4bb2bff3fd",
+    "record_observation": "6b560716f8e594016e7a00cca06c1c7988cf72e9890132c680510713e03f1e70",
+    "record_outcome": "4b89861c57ea54cbefe48e372c4f0e7a1dbedd1aa3588729c359e5dcf8029f6a",
+    "record_voice_lesson": "c296119b1985179f9196e5c220847f0e696dc6b1eaa0017f21f817948bb15914",
     "refresh_platform_analytics": "2cb4e3c060aa26c6616c09fd5c8831f324fb232a2cdb17c69413db26ee732503",
     "render_demo_video": "cd7c3e1c0aef1f4c785f3dda46134cf37717911cfccb8c0513cd1768213005f2",
     "render_hyperframes": "517ccab72014c626074c6da94867b0e81cda51b20b6609287dec9b8dd22f8386",
     "render_template_video": "e696bc84a44e797807f247d5ceb85c66e65657edb82d8ed3e23ef82760fadace",
     "reply_to_comment": "25756f7c2a8c2cd133d7b9623b395c76a452be11f1e9e795a7b03111fd2c1858",
     "respond_plan_approval": "60ef3827a8e199b140a1008746947d6753e766505b2dc69291dccddb276216e2",
-    "run_content_pipeline": "bea1d7d9fab5248af8cc5668527606b8299fa8f884cea0e8f226f6d32d91f98e",
-    "run_skill": "b7e6fd660ba7763be53536ac72af56bbf601ba23c2d2a8a55dcc7b678ba4190e",
+    "run_content_pipeline": "f34d8e01125aa430bde4b3ac06d8d2bdcee64d45e7227ba8e08100e300ad2a69",
+    "run_skill": "c9180db96375a64952e292d06f4e39e33241906ba855e77443a528d20cbf474d",
     "save_brand_profile": "70c40bbd3f80b0fb7251050cbd75e5dab79e03b7b6cacc4fa702301e6c7cda67",
     "save_content_plan": "e876db2b255acd39bdced9d5ff9ea0f1b6472343cb6833f9b9d402c7edb27e67",
-    "save_draft_to_library": "87ed0dcd83f92c76ebd2f080119c92415b7bdffefd404caa32f91be09d3bd9e5",
+    "save_draft_to_library": "cb7c13b408e959ce7b78ed4f9c87872558366a9b9bd7cec8a32f0d9c59541f75",
     "schedule_content_plan": "c61002f034b198e8f405cf9a3a05032d9639ddee75ea6649f08a432fca46542d",
     "schedule_post": "4bd2efdcf9584c579e8adacdfca8d75bdb2c4783013432b047bf3cc521bdd5c1",
     "search": "a985b8fb8cdfc6260bab53bd7aa0c2f02b1f0107c2d3cc4e6c845f7b5f5b0018",
@@ -114,6 +114,6 @@
     "visual_gate_constraints": "77d8c1eeba11d9aa4d3af08609c4c1f1ee365c2d14410efd16a2b9911dd76a8e",
     "visual_quality_check": "63a52d8f38864162b901bdb38cc69032e9f530a8a53f11db617f87b62d70f36c",
     "wait_for_connection": "3f3a587e9a6d3a32e6dd23600665eee1a428721ea352daf6eefafd746bfff8e7",
-    "write_agent_reflection": "04384e55b02a2fbefa6d1d6777dbb161cac91f7d6f080a09094ff2d12f9c2538"
+    "write_agent_reflection": "eb8faeb1dc7e1e6484a475a82b801108e9c7bbcca20918f73a07f7f4eae724b0"
   }
 }