refactor: skills-based prompt architecture

client/src/app/chat/components/ChatClient.tsx

@@ -374,7 +374,7 @@ export default function ChatClient({ initialSessionId, shouldStartNewChat, initi
     // eslint-disable-next-line react-hooks/exhaustive-deps
   }, [handleChatSessionSelect, handleNewChat, currentSessionId]);
 
-  // Auto-send initial message from URL
+  // Auto-send initial message from URL or sessionStorage (e.g., Next Steps execution)
   useEffect(() => {
     if (initialMessage && chatWebSocket.isReady && userId && !isLoadingSessionMessages && !isSending && !initialMessageSentRef.current) {
       initialMessageSentRef.current = true;
@@ -384,6 +384,9 @@ export default function ChatClient({ initialSessionId, shouldStartNewChat, initi
             sessionStorage.removeItem('pendingChatMessage');
             const sessionIdToUse = currentSessionId;
             window.history.replaceState({}, '', `/chat${sessionIdToUse ? `?sessionId=${sessionIdToUse}` : ''}`);
+          } else {
+            // Reset so the effect can retry when conditions are met again
+            initialMessageSentRef.current = false;
           }
         });
       }, 500);

client/src/app/incidents/components/SuggestionModal.tsx

@@ -89,6 +89,7 @@ export default function SuggestionModal({
     if (!suggestion.command || isExecuting) return;
     setIsExecuting(true);
 
+    // Mark suggestion as executed (non-blocking — don't let this prevent navigation)
     try {
       const res = await fetch(`/api/incidents/suggestions/${suggestion.id}/mark-executed`, {
         method: 'POST',
@@ -97,13 +98,9 @@ export default function SuggestionModal({
       });
       if (!res.ok) {
         console.error('Failed to mark suggestion as executed:', res.status);
-        setIsExecuting(false);
-        return;
       }
     } catch (err) {
       console.error('Failed to mark suggestion as executed:', err);
-      setIsExecuting(false);
-      return;
     }
 
     const message = `Execute this command and report the output:\n\n\`\`\`\n${suggestion.command}\n\`\`\`\n\nRun ONLY this command. Report the output, then stop. Do not run follow-up commands or investigate further.`;

server/chat/backend/agent/agent.py

@@ -205,16 +205,15 @@ async def agentic_tool_flow(self, state: State) -> State:
                     logging.error(f"Invalid user_id format: '{state.user_id}' - must be a non-empty string")
                     # Don't fail completely, but log the error
 
-                # Get connected providers from database (always fetch from DB, no preferences stored)
+                # Get verified providers (cloud + SkillRegistry-validated integrations)
                 provider_preference = getattr(state, 'provider_preference', None)
                 if provider_preference is None:
                     try:
-                        from utils.auth.stateless_auth import get_connected_providers
-                        provider_preference = get_connected_providers(state.user_id)
+                        from chat.background.rca_prompt_builder import get_user_providers
+                        provider_preference = get_user_providers(state.user_id)
                     except Exception as e:
                         logging.error(f"Error getting connected providers: {e}")
                         provider_preference = []
-                    # Update state so downstream code can access it
                     state.provider_preference = provider_preference
 
                 selected_project_id = getattr(state, 'selected_project_id', None)

server/chat/backend/agent/prompt/README.md

@@ -0,0 +1,263 @@
+# Prompt + Skills Architecture (Modular)
+
+This document explains the new modular prompt architecture for the Aurora agent.
+
+It covers:
+- what moved out of the old monolithic prompt builder
+- where each responsibility now lives
+- how interactive chat and background RCA prompts are assembled
+- how to extend the system safely
+
+## 1) High-level goals
+
+The refactor separates:
+- prompt content (Markdown files)
+- prompt orchestration (Python modules)
+- skill metadata and connection logic (YAML frontmatter + registry)
+- runtime loading policy (connected integrations, RCA budget, on-demand load)
+
+This gives:
+- faster prompt iteration (edit `.md`, not Python string blocks)
+- lower risk changes (small focused modules)
+- better context discipline (load only relevant skills)
+- clearer ownership boundaries
+
+## 2) Directory layout
+
+### Prompt package
+
+`server/chat/backend/agent/prompt/`
+
+- `prompt_builder.py`
+  - Backward-compatible facade for existing imports.
+  - Re-exports key functions/constants from split modules.
+- `schema.py`
+  - `PromptSegments` dataclass.
+- `composer.py`
+  - Top-level orchestration:
+    - `build_system_invariant()`
+    - `build_prompt_segments()`
+    - `assemble_system_prompt()`
+- `provider_rules.py`
+  - Provider/mode/rule segments and validation helpers:
+    - `CLOUD_EXEC_PROVIDERS`
+    - provider constraints/context/prerequisites
+    - Terraform validation
+    - model overlay
+    - failure recovery
+    - regional rules
+    - mode rules (`ask` vs `agent`)
+- `context_fetchers.py`
+  - DB-backed dynamic context:
+    - managed VM access hints
+    - knowledge base memory
+- `background.py`
+  - Background RCA system-prompt assembly for autonomous investigations.
+  - Loads source-specific markdown segments (Slack/Google Chat/general).
+- `cache_registration.py`
+  - Prefix-cache segment registration (`register_prompt_cache_breakpoints`).
+
+### Skills package
+
+`server/chat/backend/agent/skills/`
+
+- `loader.py`
+  - parses YAML frontmatter + markdown body
+  - discovers skills and segment files
+  - template substitution and token estimation
+- `registry.py`
+  - singleton registry for discovered skills
+  - connection checks
+  - connected skill index
+  - on-demand skill loading
+  - RCA preloading with token budget
+- `load_skill_tool.py`
+  - LangChain tool used by the model to load integration skill details on demand
+
+### Markdown prompt content
+
+`server/chat/backend/agent/skills/core/`
+- always-loaded core system prompt segments
+
+`server/chat/backend/agent/skills/integrations/*/SKILL.md`
+- one skill per integration (YAML frontmatter + markdown instructions)
+
+`server/chat/backend/agent/skills/rca/*.md`
+- RCA provider/general investigation guides
+
+`server/chat/backend/agent/skills/rca/segments/*.md`
+- shared RCA requirements/output blocks (used by `server/chat/background/rca_prompt_builder.py`)
+
+`server/chat/backend/agent/skills/rca/background/*.md`
+- background RCA source-mode blocks (used by `prompt/background.py`)
+
+## 3) Runtime flow
+
+## 3.1 Interactive/normal chat flow
+
+1. `agent.py` calls `build_prompt_segments(...)`.
+2. `composer.py` builds `PromptSegments` by combining:
+   - provider rules (`provider_rules.py`)
+   - background segment if in background mode (`background.py`)
+   - DB context (`context_fetchers.py`)
+   - skill index (`SkillRegistry.build_index`)
+   - core prompt markdown (`skills/core/*.md`)
+3. `assemble_system_prompt(...)` concatenates segments in stable order.
+4. `cache_registration.py` registers cache boundaries for stable segments.
+5. Model can call `load_skill(...)` tool to fetch full integration guidance only when needed.
+
+## 3.2 Background RCA flow
+
+There are two related pieces:
+
+1. Background system prompt assembly (`prompt/background.py`)
+   - builds channel/source operating instructions
+   - injects connected skill content from `SkillRegistry.load_skills_for_rca(...)`
+   - uses `skills/rca/background/*.md` for source-specific behavior
+
+2. RCA investigation prompt assembly (`chat/background/rca_prompt_builder.py`)
+   - builds alert details prompt
+   - appends RCA skills/guides via registry
+   - appends RCA shared requirement segments from `skills/rca/segments/*.md`
+
+## 4) Connection checks and loading policy
+
+`SkillRegistry` evaluates connectivity using `connection_check` frontmatter.
+
+Supported methods:
+- `get_credentials_from_db`
+- `get_token_data`
+- `is_connected_function`
+- `provider_in_preference`
+- `always`
+
+Supported field requirements:
+- `required_field`
+- `required_any_fields`
+- optional `feature_flag` function name
+
+RCA preloading:
+- connected skills are sorted by `rca_priority`
+- loaded until token budget reached
+- defaults to `4000` tokens
+- overridable by env:
+  - `RCA_SKILLS_TOKEN_BUDGET`
+  - `RCA_TOKEN_BUDGET` (fallback)
+
+## 5) Prompt segment categories
+
+You can reason about all prompt content in these categories:
+
+- Core invariants (`skills/core/*.md`)
+- Provider/mode rules (`provider_rules.py`)
+- Dynamic org/user context (`context_fetchers.py`)
+- Connected integration index (`SkillRegistry.build_index`)
+- On-demand integration guidance (`load_skill_tool.py` + `integrations/*/SKILL.md`)
+- RCA provider/general playbooks (`skills/rca/*.md`)
+- RCA shared requirement segments (`skills/rca/segments/*.md`)
+- Background source/channel behavior (`skills/rca/background/*.md`)
+
+## 6) How to extend
+
+## 6.1 Add a new integration skill
+
+1. Create `server/chat/backend/agent/skills/integrations/<id>/SKILL.md`.
+2. Add YAML frontmatter:
+   - `id`, `name`, `tools`, `connection_check`, `index`, `rca_priority`
+3. Add markdown body with usage workflow and constraints.
+4. Ensure the listed tool names exist in `get_cloud_tools()`.
+5. If using `is_connected_function`, expose a safe import path.
+
+## 6.2 Add a new background source behavior
+
+1. Add a new markdown segment under `skills/rca/background/`.
+2. Update branch selection in `prompt/background.py` to load the segment.
+3. Keep branch behavior read-only unless intentionally changing policy.
+
+## 6.3 Tune RCA requirements
+
+1. Edit `skills/rca/segments/*.md` files.
+2. Keep section titles stable if downstream logic expects them.
+3. Validate prompt output in a background RCA session.
+
+## 7) Backward compatibility
+
+`prompt_builder.py` remains as a facade so existing imports still work:
+
+- `build_prompt_segments`
+- `assemble_system_prompt`
+- `register_prompt_cache_breakpoints`
+- `CLOUD_EXEC_PROVIDERS`
+- other previously exported helpers
+
+No call sites need to change immediately.
+
+## 8) Caching behavior
+
+Prefix-cache boundaries are registered by `cache_registration.py`.
+Stable sections are cached with long TTL behavior.
+Dynamic/user-sensitive sections use ephemeral TTL (`300s`).
+
+Segments typically cached:
+- system invariant
+- provider constraints
+- regional rules
+- provider context
+- integration index
+- prerequisite checks
+- terraform validation
+- model overlay
+- failure recovery
+- tool manifest
+- ephemeral mode rules
+
+## 9) Guardrails and safety assumptions
+
+- The model should not blindly execute provider tools outside selected providers.
+- `provider_in_preference` ensures provider-bound skills are only shown when connected.
+- Feature-flag-gated skills are hidden when disabled.
+- Unknown/invalid connection methods fail closed (not connected).
+- Untrusted module paths for `is_connected_function` are blocked.
+
+## 10) Validation checklist after edits
+
+Run:
+
+```bash
+python -m compileall -q server/chat/backend/agent/prompt
+python -m compileall -q server/chat/backend/agent/skills
+python -m compileall -q server/chat/background/rca_prompt_builder.py
+```
+
+Then smoke-check:
+- interactive chat prompt creation
+- background RCA prompt creation
+- `load_skill` tool availability
+- one integration connected/disconnected scenario
+
+## 11) Common navigation shortcuts
+
+Start here:
+- `prompt/prompt_builder.py` (facade entrypoint)
+
+Then jump to:
+- orchestration: `prompt/composer.py`
+- background mode: `prompt/background.py`
+- provider logic: `prompt/provider_rules.py`
+- DB context: `prompt/context_fetchers.py`
+- skills registry/loader: `skills/registry.py`, `skills/loader.py`
+
+If editing prompt content only:
+- `skills/core/*.md`
+- `skills/rca/segments/*.md`
+- `skills/rca/background/*.md`
+- `skills/integrations/*/SKILL.md`
+
+## 12) Current tradeoff
+
+Total lines across modules may be similar or slightly higher than the old single file.
+The benefit is not line-count reduction alone, but local reasoning:
+- each file has one concern
+- content changes happen in markdown
+- orchestration changes happen in focused Python modules
+- debugging has clear entrypoints

server/chat/backend/agent/prompt/__init__.py

@@ -1 +1,56 @@
-
+
+"""Prompt assembly package.
+
+Module layout:
+- prompt_builder.py: backward-compatible facade
+- composer.py: top-level segment composition
+- provider_rules.py: provider/mode/rule text segments
+- context_fetchers.py: DB-backed prompt context helpers
+- background.py: background RCA prompt assembly
+- cache_registration.py: prefix-cache segment registration
+- schema.py: PromptSegments dataclass
+"""
+
+from .prompt_builder import (
+    CLOUD_EXEC_PROVIDERS,
+    PREFIX_CACHE_EPHEMERAL_TTL,
+    PromptSegments,
+    assemble_system_prompt,
+    build_background_mode_segment,
+    build_ephemeral_rules,
+    build_failure_recovery_segment,
+    build_knowledge_base_memory_segment,
+    build_long_documents_note,
+    build_manual_vm_access_segment,
+    build_model_overlay_segment,
+    build_prerequisite_segment,
+    build_prompt_segments,
+    build_provider_constraints,
+    build_provider_context_segment,
+    build_regional_rules,
+    build_system_invariant,
+    build_terraform_validation_segment,
+    register_prompt_cache_breakpoints,
+)
+
+__all__ = [
+    "CLOUD_EXEC_PROVIDERS",
+    "PREFIX_CACHE_EPHEMERAL_TTL",
+    "PromptSegments",
+    "assemble_system_prompt",
+    "build_background_mode_segment",
+    "build_ephemeral_rules",
+    "build_failure_recovery_segment",
+    "build_knowledge_base_memory_segment",
+    "build_long_documents_note",
+    "build_manual_vm_access_segment",
+    "build_model_overlay_segment",
+    "build_prerequisite_segment",
+    "build_prompt_segments",
+    "build_provider_constraints",
+    "build_provider_context_segment",
+    "build_regional_rules",
+    "build_system_invariant",
+    "build_terraform_validation_segment",
+    "register_prompt_cache_breakpoints",
+]