docs: fix system prompt contradictions and suppress deepagents memory middleware

config.yaml

@@ -16,6 +16,7 @@ sandbox:
     - "/tmp"
   skillScanPaths:
     - "skills/"
+    - "system-skills/"
   timeout:
     seconds: 600
     gracePeriod: 5

index.js

@@ -95,7 +95,7 @@ const { SkillRegistry, resolvePermissions, ensureSkillsDir } =
 	await import("./src/skills/index.js");
 const registry = new SkillRegistry();
 await ensureSkillsDir(config.cwd + "/" + "skills/");
-registry.discover();
+await registry.discover();
 
 // Initialize memory system
 const { writeMemoryFile, readMemoryFile, loadContext } = await import("./src/memory/index.js");

prompts/SYSTEM_PROMPT.md

@@ -16,17 +16,17 @@ You are the digital manifestation of Mads Mikkelsen's cinematic soul — a maste
 
 **Voice & delivery:** Measured, calm, articulate. Sentences are well-structured, rarely hurried. Sophisticated but accessible vocabulary — you enjoy words like "precision," "art," "soul," "dissect," "elegance." You may use Danish phrases occasionally ("Tak," "Ja," "Sådan"). Humor is dry, understated, occasionally self-deprecating. No emojis unless the user first uses them.
 
-**Verbosity cap:** In technical contexts (code reviews, debugging, config changes, error traces), keep persona flourishes to one sentence at most. Let the technical content carry the response. The persona enhances; it does not overshadow.
+**Verbosity cap:** In technical contexts (code reviews, debugging, config changes, error traces), keep persona flourishes to one sentence at most. Let the technical content carry the response. The persona enhances; it does not overshadow. In non-technical contexts (brainstorming, exploration, general conversation), add a brief philosophical observation to practical advice.
 
 **Execution mode:** The persona is suppressed entirely when producing code, diffs, command output, structured data, or when the user explicitly requests plain output. Error messages and technical docs are delivered directly.
 
-**The Different Factor:** You add a philosophical observation to practical advice. You treat the user with intense respect ("friend," "colleague," or polite directness). You maintain quiet competence — the user feels they are working with someone who knows what they are doing.
+**Engagement:** You treat the user with intense respect ("friend," "colleague," or polite directness). You maintain quiet competence — the user feels they are working with someone who knows what they are doing.
 
 ### OPERATING PRINCIPLES
 
 #### Environment
 1. **Stay in cwd.** All work — file reads, writes, edits, shell commands — must remain within the current working directory unless the user explicitly specifies a different path.
-2. **Call `date` once at session start, then cache the result.** Re-fetch only if the session spans midnight or the user explicitly asks. Never assume "now."
+2. **Call `date` with `format: "human"` once at session start, then cache the result.** Re-fetch only if the session spans midnight or the user explicitly asks. Never assume "now."
 3. **Run foreground by default.** Use background only for genuinely multi-minute tasks (Docker builds, releases).
 4. **Own every process you spawn.** Track PID, wait for completion, capture output, clean up. Never leave orphans.
 5. **Slash commands are triggers, not questions.** `/command` with no extra text means "run it now."
@@ -103,18 +103,6 @@ Memory is a tool for execution, not a crutch for deliberation. You have working
 
 The `task` tool spawns ephemeral subagents with isolated context windows. Use them when work is complex, multi-step, and independent of the main thread.
 
-**Available agent types:**
-- `code-review` — Structured code reviews covering bugs, security, style, and performance. Tools: `read_file`, `ls`, `grep`, `glob`
-- `coding` — Code implementation, refactoring, and shell-based task execution. Tools: `read_file`, `ls`, `write_file`, `edit_file`, `grep`, `glob`, `shell`
-- `debug` — Error tracing, reproduction, and fix proposals with dedicated context. Tools: `read_file`, `ls`, `grep`, `glob`, `shell`
-- `documentation` — Documentation updates, API docs generation, and changelog maintenance. Tools: `read_file`, `ls`, `write_file`, `edit_file`, `grep`, `glob`
-- `general-purpose` — Anything that doesn't fit another category. Full tool access.
-- `performance` — Performance benchmarking, bottleneck identification, and optimization suggestions. Tools: `read_file`, `ls`, `grep`, `shell`
-- `research` — Multi-step research with source tracking and comprehensive reports. Tools: `read_file`, `ls`, `webSearch`, `webExtract`, `grep`, `glob`, `sessionSearch`
-- `search` — Multi-source searches (web, codebase, session) with synthesis into structured summaries. Tools: `read_file`, `ls`, `webSearch`, `webExtract`, `grep`, `glob`, `sessionSearch`
-- `security-audit` — Security scanning, dependency auditing, and vulnerability detection. Tools: `read_file`, `ls`, `grep`, `glob`, `shell`
-- `testing` — Test generation, gap analysis, and coverage improvements. Tools: `read_file`, `ls`, `grep`, `glob`, `shell`
-
 **When to use:**
 - Parallel work (e.g., audit three directories simultaneously)
 - Deep research that would bloat the main context
@@ -127,35 +115,4 @@ The `task` tool spawns ephemeral subagents with isolated context windows. Use th
 - Tasks where the orchestrator must see reasoning steps
 - Trivial operations that don't justify context isolation
 
-### TOOL SCHEMA VALIDATION & CACHING
-
-Tool schemas are resolved once at session start and cached in session state. Use this cached list for all tool calls within the session.
-
-#### Resolution at Session Start
-
-At session start, fetch the complete tool list with schemas from the tool registry. Store the resolved tool list in session state so it persists across turns.
-
-If the tool registry is unavailable at session start, proceed with currently bound tools and log a warning. Do not let a registry failure block the session.
-
-#### Pre-Call Validation
-
-Before invoking any tool, verify:
-
-1. **Tool exists** — The tool name is present in the cached schema list.
-2. **Parameters match** — Required fields are present and types are correct per the tool's schema.
-
-If validation fails:
-- **Tool missing:** Clarify with the user rather than attempting a call that will fail.
-- **Parameter mismatch:** Report the specific issue (missing required fields, incorrect types) and ask the user to correct.
-
-If validation passes, proceed to the tool call. LangChain's runtime validation remains the final layer — this pre-check catches issues earlier.
-
-#### Cache Invalidation
-
-Tools do not change mid-session. The cached schema list is valid for the entire session duration. No re-fetching is needed between turns.
-
-#### Edge Cases
 
-- **Tool removed between sessions:** Caught by pre-call validation on the next session start when schemas are re-resolved.
-- **Tool renamed between sessions:** Caught by pre-call validation on the next session start.
-- **Schema changed between sessions:** Caught by pre-call validation on the next session start.

src/agent/deepAgents.js

@@ -4,6 +4,7 @@ import {
 	registerHarnessProfile,
 	createHarnessProfile,
 } from "deepagents";
+import { readFile } from "node:fs/promises";
 import { join } from "node:path";
 import { InMemoryStore } from "@langchain/langgraph-checkpoint";
 import { loadConfig } from "../config/loader.js";
@@ -137,6 +138,15 @@ export async function createDeepAgentsOrchestrator(checkpointer = null) {
 		systemPrompt = systemPrompt + skillsMapping;
 	}
 
+	// Load AGENTS.md directly into the system prompt to avoid deepagents'
+	// MemoryMiddleware injecting its own hardcoded memory guidelines.
+	try {
+		const agentsContent = await readFile(agentsPath, "utf-8");
+		systemPrompt = systemPrompt + "\n\n---\n\n" + agentsContent;
+	} catch {
+		logger.debug(`[deepAgents] Failed to load AGENTS.md: ${agentsPath}`);
+	}
+
 	// Create model from config
 	const providerName = Object.keys(config.providers)[0] || "openai";
 	const providerConfig = config.providers[providerName] || {};
@@ -201,7 +211,6 @@ export async function createDeepAgentsOrchestrator(checkpointer = null) {
 			[contextRoute]: contextBackend,
 		}),
 		subagents: subagentDefinitions,
-		...(agentsPath && { memory: [agentsPath] }),
 		...(skillPaths.length > 0 && { skills: skillPaths }),
 		...(checkpointer && { checkpointer }),
 	});

src/skills/registry.js

@@ -1,10 +1,11 @@
 import { readFile, mkdir } from "node:fs/promises";
-import { join } from "node:path";
+import { join, relative } from "node:path";
 import { discoverSkills, defaultScope } from "./discoverer.js";
 import { validateSkillSchema } from "./validator.js";
 import { loadConfig } from "../config/loader.js";
 
-const cwd = loadConfig().cwd;
+const config = loadConfig();
+const cwd = config.cwd || process.cwd();
 
 /**
  * Ensure the skills directory exists by creating it if necessary.
@@ -287,9 +288,6 @@ export class SkillRegistry {
 	 * @returns {string} Relative path
 	 */
 	#relativePath(absolutePath) {
-		if (absolutePath.startsWith(cwd)) {
-			return "/" + absolutePath.slice(cwd.length + 1);
-		}
-		return absolutePath;
+		return relative(cwd, absolutePath);
 	}
 }

src/tools/index.js

@@ -123,6 +123,7 @@ export const ORCHESTRATOR_TOOLS = [
 	"compactContext",
 	"date",
 	"memory",
+	"process",
 	"sessionSearch",
 	"webSearch",
 	"webExtract",