fix(prompt): reduce mechanical chat behavior + optimize system prompt

- Add Tool Call Style section with narration minimalism + non-disclosure
  rule (from TS reference): agents must never expose tool names to users
- Consolidate 3 redundant memory recall reminders into 1 dedicated section
- Remove "tell the user you checked but found nothing" instruction that
  caused agents to describe internal tool mechanics in responses
- Remove 11 tool aliases from system prompt listing (~300 tokens saved);
  aliases still work via provider definitions
- Filter alias tool names out of system prompt ToolNames in loop_history
- Update AGENTS.md: remove tool name references from Memory section,
  add group chat framing from V1 ("participant, not their proxy")

internal/agent/loop_history.go

@@ -202,6 +202,21 @@ func (l *Loop) buildMessages(ctx context.Context, history []providers.Message, s
 		}
 		toolNames = filtered
 	}
+	// Exclude tool aliases from the system prompt tool list.
+	// Aliases are sent as separate provider definitions (LLM can still call them),
+	// but listing them in the prompt adds ~300 tokens of noise that dilutes persona.
+	if l.tools != nil {
+		aliasSet := l.tools.Aliases()
+		if len(aliasSet) > 0 {
+			noAlias := toolNames[:0:0]
+			for _, n := range toolNames {
+				if _, isAlias := aliasSet[n]; !isAlias {
+					noAlias = append(noAlias, n)
+				}
+			}
+			toolNames = noAlias
+		}
+	}
 	// Always build MCP tool descriptions for inline tools — in hybrid search
 	// mode the kept inline tools still need descriptions in the system prompt.
 	mcpToolDescs := l.buildMCPToolDescs(toolNames)

internal/agent/systemprompt.go

@@ -130,20 +130,9 @@ var coreToolSummaries = map[string]string{
 	"list_group_members":      "List all members of the current group chat (Feishu/Lark only)",
 	"create_forum_topic":      "Create a forum topic in a Telegram supergroup",
 
-	// Legacy tool aliases — kept for backward compatibility with older clients
-	"edit_file":      "Alias for edit — Edit a file by replacing exact text matches",
-	"sessions_spawn": "Alias for spawn — Spawn a self-clone subagent to handle a task in the background",
-
-	// Claude Code tool aliases — enable Claude Code skills without modification
-	"Read":       "Alias for read_file — Read file contents",
-	"Write":      "Alias for write_file — Create or overwrite files",
-	"Edit":       "Alias for edit — Edit a file by replacing exact text matches",
-	"Bash":       "Alias for exec — Run shell commands",
-	"WebFetch":   "Alias for web_fetch — Fetch and extract content from a URL",
-	"WebSearch":  "Alias for web_search — Search the web",
-	"Agent":      "Alias for spawn — Spawn a subagent or delegate to another agent",
-	"Skill":      "Alias for use_skill — Invoke a skill by name",
-	"ToolSearch": "Alias for mcp_tool_search — Search for available MCP tools",
+	// Tool aliases (edit_file, sessions_spawn, Read, Write, Edit, Bash, etc.)
+	// are registered in the tool registry but excluded from the system prompt
+	// to reduce prompt size (~300 tokens). They work without being listed here.
 }
 
 // BuildSystemPrompt constructs the full system prompt with all sections.
@@ -223,6 +212,11 @@ func BuildSystemPrompt(cfg SystemPromptConfig) string {
 	// 2. ## Tooling
 	lines = append(lines, buildToolingSection(cfg.ToolNames, cfg.SandboxEnabled, cfg.ShellDenyGroups)...)
 
+	// 2.3. ## Tool Call Style — narration minimalism + non-disclosure of tool internals
+	if !cfg.IsBootstrap {
+		lines = append(lines, buildToolCallStyleSection()...)
+	}
+
 	// 2.5. Credentialed CLI context (appended after tooling, before safety) — skip during bootstrap
 	if !cfg.IsBootstrap && cfg.CredentialCLIContext != "" {
 		lines = append(lines, cfg.CredentialCLIContext, "")
@@ -328,19 +322,14 @@ func BuildSystemPrompt(cfg SystemPromptConfig) string {
 	lines = append(lines, buildRuntimeSection(cfg)...)
 
 	// 16. Recency reinforcements — skip during bootstrap (short prompt, no drift risk)
+	// Consolidated: persona reminder + slim AGENTS.md reminder (no memory duplication).
+	// Memory recall is covered by the dedicated ## Memory Recall section above.
 	if !cfg.IsBootstrap {
 		if len(personaFiles) > 0 {
 			lines = append(lines, buildPersonaReminder(personaFiles, cfg.AgentType, cfg.ProviderType)...)
 		}
 		if !isMinimal {
-			lines = append(lines, "Reminder: Follow AGENTS.md rules — memory recall before answering, NO_REPLY when silent, match the user's language.", "")
-		}
-		if !isMinimal && cfg.HasMemory {
-			memReminder := "Reminder: Before answering questions about prior work, decisions, or preferences, always run memory_search first. If nothing relevant is found, say you checked but found nothing."
-			if cfg.HasKnowledgeGraph {
-				memReminder += " Also run knowledge_graph_search when the question involves people, teams, projects, or connections."
-			}
-			lines = append(lines, memReminder, "")
+			lines = append(lines, "Reminder: Follow AGENTS.md rules — NO_REPLY when silent, match the user's language.", "")
 		}
 	}
 

internal/agent/systemprompt_sections.go

@@ -89,8 +89,28 @@ func buildSandboxSection(cfg SystemPromptConfig) []string {
 	return lines
 }
 
+// buildToolCallStyleSection generates the ## Tool Call Style section.
+// Matches TS system-prompt.ts "Tool Call Style" — narration minimalism + non-disclosure.
+// Prevents the agent from exposing internal tool names to users.
+func buildToolCallStyleSection() []string {
+	return []string{
+		"## Tool Call Style",
+		"",
+		"Default: do not narrate routine, low-risk tool calls (just call the tool).",
+		"Narrate only when it helps: multi-step work, complex problems, sensitive actions, or when the user explicitly asks.",
+		"Keep narration brief and value-dense; avoid repeating obvious steps.",
+		"Use plain human language — never mention tool names, function names, or internal mechanics to users.",
+		"",
+		"WRONG: \"I searched memory_search and knowledge_graph_search but results were sparse...\"",
+		"RIGHT: \"Hmm, I remember you mentioned that before...\" or \"I don't recall that specifically.\"",
+		"",
+		"When runtime-generated events (subagent completions, cron results) provide metadata, rewrite in your natural voice before relaying to the user.",
+		"When a first-class tool exists for an action, use the tool directly instead of asking the user to run equivalent CLI commands.",
+		"",
+	}
+}
+
 // buildMemoryRecallSection generates the ## Memory Recall section for the system prompt.
-// Supplements (does NOT replace) the recency reminder at the end of the prompt.
 func buildMemoryRecallSection(hasMemoryGet, hasKG bool) []string {
 	lines := []string{"## Memory Recall", ""}
 
@@ -98,12 +118,12 @@ func buildMemoryRecallSection(hasMemoryGet, hasKG bool) []string {
 		lines = append(lines,
 			"Before answering questions about prior work, decisions, people, preferences, or todos: "+
 				"call memory_search with a relevant query; then use memory_get to pull only the needed lines. "+
-				"If no relevant results found, tell the user you checked but found nothing.")
+				"If no relevant results found, say so naturally without mentioning tool names.")
 	} else {
 		lines = append(lines,
 			"Before answering questions about prior work, decisions, people, preferences, or todos: "+
 				"call memory_search with a relevant query and answer from the matching results. "+
-				"If no relevant results found, tell the user you checked but found nothing.")
+				"If no relevant results found, say so naturally without mentioning tool names.")
 	}
 
 	if hasKG {

internal/bootstrap/templates/AGENTS.md

@@ -21,26 +21,23 @@ Talk like a person, not a customer service bot.
 
 ## Memory
 
-You start fresh each session. Use tools to maintain continuity:
+You start fresh each session. Your tools handle recall automatically.
 
-- **Recall:** Use `memory_search` before answering about prior work, decisions, or preferences
-- **Save:** Use `write_file` to persist important information:
-  - Daily notes → `memory/YYYY-MM-DD.md` (raw logs, what happened today)
-  - Long-term → `MEMORY.md` (curated: key decisions, lessons, significant events)
-- **No "mental notes"** — if you want to remember something, write it to a file NOW with a tool call
+- Before answering about past events, check your memory first — then answer naturally
+- Save important info to files NOW — "mental notes" don't survive sessions
+- Daily notes → `memory/YYYY-MM-DD.md` | Long-term → `MEMORY.md`
 - When asked to "remember this" → write immediately, don't just acknowledge
-- **Recall details:** Use `memory_search` first, then `memory_get` to pull only the needed lines.
-  If `knowledge_graph_search` is available, also run it for questions about people, teams, projects, or connections — it finds multi-hop relationships that `memory_search` misses.
-- When asked to save or remember something, you MUST call a write tool (`write_file` or `edit`) in THIS turn. Never claim "already saved" without a tool call.
+- When asked to save or remember something, you MUST write in THIS turn. Never claim "already saved" without actually saving.
 
-### Memory Privacy
+### Privacy
 
-- In group chats: do NOT surface personal memory content (private preferences, personal history)
-- Memory details (file paths, raw notes) should only be shared in **private/direct chats**
-- In groups, use memory to inform your answers, but don't quote or reference it directly
+- In group chats: use memory to inform your answers, but don't quote or reference it directly
+- Memory details should only be shared in private/direct chats
 
 ## Group Chats
 
+You have access to your human's stuff. That doesn't mean you _share_ their stuff. In groups, you're a participant — not their voice, not their proxy.
+
 ### Know When to Speak
 
 **Respond when:**
@@ -63,6 +60,8 @@ You start fresh each session. Use tools to maintain continuity:
 
 **Avoid the triple-tap:** Don't respond multiple times to the same message. One thoughtful response beats three fragments.
 
+Participate, don't dominate.
+
 ### NO_REPLY Format
 
 When you have nothing to say, respond with ONLY: NO_REPLY