System Audit: Context & Prompt Fragmentation

Date: 2026-03-12 Scope: All LLM-calling code paths in orchestrator/ Purpose: Map every location that builds prompts, loads tools, injects memory, or constructs LLM context — to inform PRD-80 (Unified Context Service)

---

1. Codebase Overview

Metric	Value
Total .py files in orchestrator/	651
Total lines of Python	~202,831
Largest file	modules/tools/discovery/platform_executor.py (3,806 lines)
Files containing class definitions	~60
Code paths that call LLMs	9 (confirmed)

---

2. Discovery: The 9 Code Paths

Path 1 — Chatbot (Auto)

File: consumers/chatbot/smart_orchestrator.py (569 lines) Entry point: SmartChatOrchestrator.prepare_request()

Capability	How it works	Lines
A. Prompt Building	Calls get_happy_system_prompt() from personality.py which assembles 6 sections: base identity, platform skill, memory context, tool guidance, action response style, self-learning instruction. Then appends session context, daily logs, platform action summary, datetime as separate string concatenations.	263-378
B. Tool Loading	Receives available_tools from caller (service.py:598 calls get_tools_for_agent()). Then SmartToolRouter.route() filters by intent classification, with semantic ranking (PRD-64) or keyword fallback. Platform tools always included.	223-261
C. Memory Injection	Two paths: (1) Context Router (unified_memory.retrieve_context()) returning a ContextBundle with L1/L2/L3 data, or (2) fallback SmartMemoryManager.retrieve_memories() doing two-tier Mem0 search. Memory formatted as bullet list via _format_memories_for_llm().	159-221
D. Daily Logs	smart_memory.get_daily_logs() fetches from Mem0, max 2000 chars. Appended to system prompt as ## Recent Activity. Only in fallback path; Context Router path gets daily logs from bundle.	316-358
E. Personality	AutomatosPersonality class in personality.py (401 lines). Workspace-level settings (personality_mode, communication_style) loaded with 2-min TTL cache from DB. PromptRegistry integration (PRD-58).	personality.py
F. Platform Actions	get_action_registry().build_prompt_summary() injected into system prompt. Markdown catalog of all platform_execute actions grouped by category.	361-367
G. Datetime	Injected as a system message at position 1 in the message array (not in system prompt). Format: Current date/time (UTC): YYYY-MM-DD HH:MM.	373-379
H. Message Conversion	_convert_messages() strips system messages, converts parts format to plain text, handles file attachments.	431-465
I. Token Budget	NONE. No token awareness. Prompt grows unbounded.
J. Composio Hints	Not in chatbot — handled by tool_router's get_tools_for_agent() which enriches composio_execute description with available actions.
K. Skill Content	Not in chatbot path. Auto uses personality.py platform skill, not agent skills.

Dependencies: intent_classifier.py, smart_memory.py, smart_tool_router.py, personality.py, unified_memory_service.py, context_router.py

---

Path 2 — Agent Task Execution (AgentFactory)

File: modules/agents/factory/agent_factory.py (1,424 lines) Entry point: AgentFactory.execute_with_prompt() (line 714)

Capability	How it works	Lines
A. Prompt Building	_build_agent_system_prompt() assembles: identity (name, ID, type, description), persona (custom or from persona table), task context, skills (SKILL.md via skill_loader), plugins (tier1+tier2 from PluginContextService), Composio apps section, dependency context instructions, skill tool usage instructions, response formatting rules. Returns (prompt_text, skill_tool_schemas).	1136-1308
B. Tool Loading	get_tools_for_agent() from tool_router.py — loads from ToolRegistry, validates per-agent access, enriches composio_execute with app actions, adds platform_execute dispatcher. Skill tool schemas appended separately.	811-824
C. Memory Injection	Short-term memory from agent_runtime.memory[-3:] injected as user/assistant message pairs. No long-term Mem0 retrieval (controlled by use_memory param, defaults True but only uses runtime memory).	780-785
D. Daily Logs	NONE.
E. Personality	NONE. No personality/workspace settings applied. Raw agent identity only.
F. Platform Actions	NOT in system prompt. But platform_execute dispatcher schema IS in the tool list via get_tools_for_agent(). The LLM sees the tool but has no markdown catalog unless the agent's skill mentions it.
G. Datetime	NONE. No datetime injection.
H. Message Conversion	Direct message construction: system prompt, optional memory messages, optional recipe context, then user prompt.	759-845
I. Token Budget	NONE.
J. Composio Hints	_inject_composio_hints() and _inject_composio_recipe_hints() — adds action enum + hints as system messages.	826-839
K. Skill Content	Full SKILL.md loaded via skill_loader.load_skill_core() for each active skill. Plus tool schemas extracted from skill.tools_schema.	1172-1219

Key finding: build_prompt_summary() for platform actions is NOT called in agent_factory. Agents executing tasks cannot see the platform action catalog in their prompt, only the raw platform_execute tool schema. This means task-executing agents don't know what actions are available without the skill mentioning them.

---

Path 3 — Heartbeat Service (Orchestrator Tick)

File: services/heartbeat_service.py (1,446 lines) Entry point: _orchestrator_tick_llm() (line 382)

Capability	How it works	Lines
A. Prompt Building	Inline f-string: personality mode, level instruction (silent/notify/act_notify/autonomous), communication style suffix, checklist block, platform action summary. Hardcoded "You are the Automatos orchestrator performing a scheduled health check."	426-433
B. Tool Loading	registry.to_dispatcher_schema() — single platform_execute dispatcher only. No core tools, no composio.	399-400
C. Memory Injection	NONE.
D. Daily Logs	NONE.
E. Personality	Partial: loads personality_mode and communication_style from workspace settings. No full AutomatosPersonality template.	393-396
F. Platform Actions	registry.build_prompt_summary() injected into system prompt.	425
G. Datetime	Injected in user message: Time: {datetime.utcnow().strftime('%Y-%m-%d %H:%M UTC')}	437
H. Message Conversion	Direct construction: system + user messages. Context trimming at exchange boundaries after 10 messages.	435-524
I. Token Budget	NONE. Max 5 tool loop iterations as implicit limit. Context trimming keeps last 2 exchanges.	447, 510-524
J. Composio Hints	NONE.
K. Skill Content	NONE.

---

Path 4 — Heartbeat Service (Agent Tick)

File: services/heartbeat_service.py (1,446 lines) Entry point: _agent_tick() (line 579)

Capability	How it works	Lines
A. Prompt Building	Delegates to AgentFactory.execute_with_prompt() — the prompt is just the heartbeat instruction + board task context. System prompt comes from agent_factory's _build_agent_system_prompt().	676-698
B. Tool Loading	Via AgentFactory which calls get_tools_for_agent().	688-698
C. Memory Injection	use_memory=False passed to factory — no memory.	698
D. Daily Logs	NONE.
E. Personality	NONE. (Via agent_factory which has no personality.)
F. Platform Actions	Via agent_factory tool list (dispatcher schema). No markdown catalog in prompt.
G. Datetime	Injected in user prompt text: Time: {datetime.utcnow().strftime('%Y-%m-%d %H:%M UTC')}	678
H. Message Conversion	Via agent_factory.
I. Token Budget	NONE.
J. Composio Hints	Via agent_factory.
K. Skill Content	Via agent_factory's _build_agent_system_prompt().

---

Path 5 — Recipe Executor

File: api/recipe_executor.py (1,453 lines) Entry point: _execute_step() (line 66)

Capability	How it works	Lines
A. Prompt Building	_build_system_prompt() (line 443) — uses AgentRuntime's pre-built system_prompt from factory activation. Falls back to minimal identity-only prompt. Step-level context injected via Scratchpad and prior step outputs.	125-126, 443-466
B. Tool Loading	get_chat_tools() (alias for get_agent_tools()) which wraps get_tools_for_agent(). Scratchpad write tool injected separately.	228-239
C. Memory Injection	Recipe memory via RecipeMemoryService — retrieves recipe-scoped and agent-scoped memories from Mem0. Injected as system message.	Not in _execute_step directly
D. Daily Logs	NONE.
E. Personality	NONE.
F. Platform Actions	Via tool list (dispatcher schema). No markdown catalog.
G. Datetime	NONE.
H. Message Conversion	Direct message construction with scratchpad context, prior step outputs, and user instruction.
I. Token Budget	NONE.
J. Composio Hints	Pre-resolved action names passed to factory via composio_action_names param.
K. Skill Content	Via agent_factory's pre-built prompt (from activation).

---

Path 6 — Execution Manager

File: modules/agents/execution/execution_manager.py (1,344 lines) Entry point: _execute_subtask() (around line 810)

Capability	How it works	Lines
A. Prompt Building	Hardcoded "professional_system_prompt" (40 lines of instructions), then prepends _build_agent_system_prompt() from factory if agent has skills. Appends workspace guidance and conversion tool instructions.	821-904
B. Tool Loading	Via AgentFactory.execute_with_prompt() which calls get_tools_for_agent().	906
C. Memory Injection	use_memory=True passed to factory — uses runtime memory.	910
D. Daily Logs	NONE.
E. Personality	NONE. Custom "professional" prompt replaces personality.	821-847
F. Platform Actions	Via factory tool list. No markdown catalog.
G. Datetime	NONE.
H. Message Conversion	Via factory.
I. Token Budget	NONE.
J. Composio Hints	Via factory.
K. Skill Content	Via factory's _build_agent_system_prompt() — DUPLICATED. The execution manager calls _build_agent_system_prompt() directly AND then passes a system_prompt to execute_with_prompt(), which means the factory won't rebuild it. But the execution_manager concatenation with professional_system_prompt creates a different prompt than what the factory alone would produce.	863-868

Key finding: The execution_manager and agent_factory produce DIFFERENT system prompts for the same agent. The execution_manager prepends skill content and appends a hardcoded professional prompt, workspace guidance, and conversion tool instructions.

---

Path 7 — Universal Router (Tier 3 — LLM Classification)

File: core/routing/engine.py (887 lines) Entry point: _classify_with_llm() (line 452)

Capability	How it works	Lines
A. Prompt Building	_build_classification_prompt() — hardcoded routing prompt with agent descriptions, semantic hints. PRD-58 PromptRegistry integration attempted.	662-726
B. Tool Loading	NONE. Text-only LLM call.
C-K	NONE. Router is special-purpose — no memory, no tools, no personality, no datetime.

---

Path 8 — Orchestrator Stages

Files: modules/orchestrator/stages/*.py (11 files) Key files: task_decomposer.py, complexity_analyzer.py, quality_assessor.py, agent_negotiation.py, context_engineering.py

Capability	How it works
A. Prompt Building	Each stage builds its own prompt. task_decomposer.py has a hardcoded system prompt ("You are an expert task decomposition system") with PromptRegistry fallback. complexity_analyzer.py, quality_assessor.py, agent_negotiation.py all have custom _build_*_prompt() methods. context_engineering.py has _build_enhanced_prompt() for RAG-augmented subtask prompts.
B. Tool Loading	NONE. All stages are text-only LLM calls.
C-K	NONE. Stages are internal orchestration — no memory, tools, or personality.

---

Path 9 — NL2SQL

File: modules/nl2sql/query/nl2sql_service.py Entry point: NaturalLanguageToSQLService.generate_sql()

Capability	How it works
A. Prompt Building	_build_prompt() — schema-specific prompt with SQL rules, table schema, semantic layer, few-shot examples, error context for self-correction. Fully self-contained.
B. Tool Loading	NONE. Text-only.
C-K	NONE.

---

3. Capability Matrix

Capability	Path 1 (Chatbot)	Path 2 (Factory)	Path 3 (HB Orch)	Path 4 (HB Agent)	Path 5 (Recipe)	Path 6 (ExecMgr)	Path 7 (Router)	Path 8 (Stages)	Path 9 (NL2SQL)
A. Prompt Building	personality.py	_build_agent_system_prompt	inline f-string	via factory	via factory	hardcoded + factory	_build_classification_prompt	per-stage	_build_prompt
B. Tool Loading	SmartToolRouter	get_tools_for_agent	dispatcher only	via factory	get_chat_tools	via factory	none	none	none
C. Memory	Context Router / SmartMemory	runtime memory only	none	none	recipe memory	runtime memory	none	none	none
D. Daily Logs	smart_memory.get_daily_logs	none	none	none	none	none	none	none	none
E. Personality	full AutomatosPersonality	none	partial (mode+style)	none	none	hardcoded professional	none	none	none
F. Platform Actions (prompt)	build_prompt_summary	MISSING	build_prompt_summary	MISSING	MISSING	MISSING	none	none	none
G. Datetime	system message	none	in user msg	in user msg	none	none	none	none	none
H. Message Convert	_convert_messages	direct	direct	via factory	direct	via factory	direct	direct	direct
I. Token Budget	NONE	NONE	implicit (5 iter)	NONE	NONE	NONE	NONE	NONE	NONE
J. Composio Hints	via tool_router	_inject_composio_hints	none	via factory	pre-resolved	via factory	none	none	none
K. Skill Content	platform skill only	full SKILL.md	none	via factory	via factory	via factory	none	none	none

---

4. Classification of Findings

4.1 Critical Fragmentation Issues

F1: Platform action catalog missing from 4 of 6 agent-facing paths.

• Chatbot and heartbeat orchestrator inject build_prompt_summary() into the system prompt.

• Agent factory, agent heartbeat, recipe executor, and execution manager do NOT inject it. The LLM gets the platform_execute tool schema but has no markdown catalog explaining what actions exist or what parameters they take.

• Impact: Task-executing agents cannot effectively use platform actions unless their SKILL.md happens to list them.

F2: Memory injection inconsistency across 4 paths.

• Chatbot: full Context Router or SmartMemory with two-tier Mem0 search.

• Factory: only runtime memory (last 3 exchanges, no Mem0).

• Heartbeat (agent): explicitly use_memory=False.

• Recipe: recipe-scoped Mem0, separate path.

• Impact: Agents performing tasks have no long-term user context. Heartbeat agents have no memory at all.

F3: Personality applied to only 1.5 of 6 agent-facing paths.

• Chatbot: full AutomatosPersonality with workspace settings.

• Heartbeat orchestrator: personality_mode and communication_style only (no template).

• All other paths: no personality at all.

• Impact: Agent responses have inconsistent tone depending on how they're invoked.

F4: No token budget management anywhere.

• Zero paths implement token budget awareness.

• Prompt sections grow unbounded.

• Adding platform action summary (~1200 tokens) + skill content (~1800 tokens) + memory (~400 tokens) + daily logs (~500 tokens) is fine now (~7K total for task execution), but adding more sections will silently overflow model context limits.

• Only the heartbeat orchestrator has implicit limits (5 tool iterations, context trimming at 10 messages).

4.2 Duplication

D1: _build_agent_system_prompt() called from 3 different places with different wrapping.

1. agent_factory.activate_agent() — stored on AgentRuntime, used as-is.

2. agent_factory.execute_with_prompt() — called as fallback when no cached prompt exists.

3. execution_manager._execute_subtask() — called directly, then prepended to a separate hardcoded prompt.

• The execution_manager produces a fundamentally different prompt than the factory alone.

D2: Tool loading has 3 entry points that all resolve to get_tools_for_agent().

• get_tools_for_agent() in tool_router.py (primary)

• get_agent_tools() in tool_router.py (convenience wrapper)

• get_chat_tools() re-exported via consumers/chatbot/tool_router.py

• All resolve to the same function. This is fine structurally but confusing.

D3: Platform action summary injected via string concatenation in 2 places.

• smart_orchestrator.py:363 — chatbot path

• heartbeat_service.py:425 — orchestrator heartbeat path

• Both call get_action_registry().build_prompt_summary() and append to system prompt. Any change to injection format requires updating both.

D4: Datetime context injected 3 different ways.

• Chatbot: system message at position 1 in message array.

• Heartbeat orchestrator: in the user message text.

• Heartbeat agent: in the user prompt text.

• Factory, recipe, execution_manager, router, stages, NL2SQL: not injected at all.

4.3 Dead / Redundant Code

R1: SmartToolRouter in consumers/chatbot/smart_tool_router.py (442 lines).

• Used ONLY by smart_orchestrator.py for intent-based tool filtering.

• Contains its own IntentClassifier, TOOL_CATEGORIES, INTENT_TO_TOOLS mappings, and semantic embedding ranking.

• Could be replaced by a tool loading strategy in ContextService.

R2: AutomatosPersonality.get_platform_skill() in personality.py.

• Returns a hardcoded platform skill block. Only used by the chatbot.

• Should be a section in ContextService.

R3: professional_system_prompt in execution_manager.py (lines 821-847).

• 40 lines of hardcoded instructions. Completely different from any other path.

• Duplicates some instructions that should be in the agent's skill or a shared section.

4.4 Missing Capabilities

Missing	Affected Paths	Severity
Platform action catalog in prompt	Factory, Recipe, ExecMgr, HB Agent	HIGH
Long-term memory (Mem0)	Factory, ExecMgr, HB Agent	MEDIUM
Daily logs	All except chatbot	MEDIUM
Personality/tone	Factory, Recipe, ExecMgr, HB Agent	MEDIUM
Datetime context	Factory, Recipe, ExecMgr	LOW
Token budget	ALL paths	HIGH
Session context (L1)	All except chatbot	LOW

---

5. Dependency Map

smart_orchestrator.py
  -> personality.py (get_happy_system_prompt, load_orchestrator_settings)
  -> smart_memory.py (retrieve_memories, get_daily_logs, store_daily_summary)
  -> smart_tool_router.py (route)
  -> intent_classifier.py (classify)
  -> unified_memory_service.py (retrieve_context, store_exchange, update_session)
  -> action_registry.py (build_prompt_summary)

agent_factory.py
  -> tool_router.py (get_tools_for_agent)
  -> skill_loader.py (load_skill_core)
  -> plugin_context_service.py (get_assigned_plugins, build_tier1_summary, build_tier2_content_sync)
  -> composio_cache models (AgentAppAssignment, ComposioAppCache, ComposioActionCache)

heartbeat_service.py
  -> personality.py (load_orchestrator_settings)
  -> action_registry.py (to_dispatcher_schema, build_prompt_summary)
  -> agent_factory.py (execute_with_prompt)
  -> platform_executor.py (PlatformActionExecutor)
  -> LLMManager (direct)

recipe_executor.py
  -> agent_factory.py (activate_agent)
  -> tool_router.py (get_agent_tools)
  -> LLMManager (direct)

execution_manager.py
  -> agent_factory.py (_build_agent_system_prompt, execute_with_prompt)
  -> MemoryPromptInjector (imported but usage unclear in current code)

engine.py (UniversalRouter)
  -> LLMManager (create_llm_manager)
  -> IntentClassifier
  -> PromptRegistry

stages/*.py
  -> LLMManager variants
  -> ContextOptimizer (from search module)

---

6. File Inventory (All Files Inspected)

File	Lines	Role in Context Building
consumers/chatbot/smart_orchestrator.py	569	Central chatbot coordinator — prompt + tools + memory
consumers/chatbot/personality.py	401	AutomatosPersonality prompt template
consumers/chatbot/smart_tool_router.py	442	Intent-based tool filtering
consumers/chatbot/smart_memory.py	788	Two-tier Mem0 retrieval + daily logs
consumers/chatbot/intent_classifier.py	~400	Intent classification (9 intents)
consumers/chatbot/service.py	1,976	StreamingChatService, tool loop, SSE
modules/agents/factory/agent_factory.py	1,424	Agent activation, prompt building, tool loop
modules/agents/execution/execution_manager.py	1,344	Subtask execution with custom prompts
services/heartbeat_service.py	1,446	Scheduled ticks, LLM tool loop
api/recipe_executor.py	1,453	Step-by-step recipe execution
core/routing/engine.py	887	Universal router with LLM tier
modules/orchestrator/stages/*.py	~6,000	9-stage pipeline (11 files)
modules/nl2sql/query/nl2sql_service.py	~400	Text-to-SQL prompt builder
modules/tools/tool_router.py	762	Tool loading + execution
modules/tools/discovery/action_registry.py	192	Platform action registry
modules/memory/unified_memory_service.py	2,068	5-layer memory stack
modules/memory/context_router.py	~500	Context assembly (PRD-79)
channels/base.py	~150	Channel adapter base (delegates to router + factory)

---

7. Key Metrics for Migration Planning

Metric	Count
Distinct prompt-building functions	8
Distinct tool-loading entry points	3 (all resolve to 1)
Distinct memory retrieval paths	4
Places build_prompt_summary() is called	2
Places to_dispatcher_schema() is called	2
Places get_happy_system_prompt() is called	2
Places _build_agent_system_prompt() is called	3
Lines of prompt-building code (estimated)	~600
Lines deletable after ContextService migration	~450