* feat(observe): ExecutionDetail with tree view, headers, and resizable drawer
## Summary
Ports the legacy `ExecutionDetails.jsx` (~984 lines) into the `@flowiseai/observe` SDK as a composable `ExecutionDetail` orchestrator.
- **`ExecutionDetail`**: split-pane (sidebar tree + node detail) with sidebar header that holds the agentflow chip, Copy-ID chip, formatted updated date, and refresh button. Auto-selects the first STOPPED node (or first top-level otherwise) on initial tree build.
- **`ExecutionTreeSidebar`**: `@mui/x-tree-view` `RichTreeView` with a custom `TreeItem2` slot — per-branch border colored from `AGENTFLOW_ICONS`, status-icon variants on each row, virtual `Iteration #N` containers, controlled expand/collapse defaulting to all-expanded.
- **`useExecutionTree`** rewritten to mirror the legacy three-pass `buildTreeData`: `previousNodeIds` parent→child with most-recent disambiguation, iteration grouping under the latest parent instance, sub-tree linking inside virtual containers, iterations-first sibling sort, and recursive `FLOWISE_CREDENTIAL_ID` scrub. Iteration container status is the legacy `ERROR > INPROGRESS > all-FINISHED > UNKNOWN` rollup.
- **`useDrawerWidths`** + `useResizableSidebar({ inverted })` extension drive a resizable `ExecutionsViewer` Drawer (parity with the legacy left-edge drag handle). Adds a public `ExecutionsViewerProps.drawer?: { defaultWidth?, minWidth?, maxWidth? }` override.
- **`AGENTFLOW_ICONS`** lifted from `atoms/NodeIcon.tsx` to `core/primitives/agentflowIcons.ts` so the tree slot can consume it without re-importing through `atoms`.
- **`ExecutionDetailProps.agentflow?: AgentflowRef`** seeds the header chip while the detail endpoint loads (server-side `getExecutionById` doesn't perform the agentflow join — only `getAllExecutions` does).
- Adds `@mui/x-tree-view` as a runtime dep.
## Test plan
- [x] `pnpm -C packages/observe test:coverage` — 331 tests pass, all `coverageThreshold` entries met (85.37% statements / 82.98% branches overall; `useExecutionTree` 96.5% / 85.18%, new `useDrawerWidths` 100% / 90.9%, `ExecutionDetail` 89.18% / 92.5%).
- [x] Manually verify in `packages/observe/examples` (`pnpm dev` after copying `.env.example` to `.env`):
- List → row click opens the resizable Drawer; left-edge handle widens/narrows.
- Tree shows status icons for each row plus rolled-up status on virtual iteration containers; `IconLoader` spins for INPROGRESS rows.
- Agentflow chip is clickable and opens the canvas in a new tab when `VITE_AGENTFLOW_CANVAS_URL` is set; non-clickable otherwise.
- Copy-ID chip flips inline to `Copied!` for 2s. **Behavior change from legacy**: legacy dispatched a Redux/notistack snackbar toast ("ID copied to clipboard") via the host app; the SDK can't assume a snackbar provider exists, so the confirmation is rendered on the chip itself.
- Refresh button re-fetches; date formatted as `MMM d, yyyy h:mm a`.
- Standalone `<ExecutionDetail>` demo by execution id renders the same surface without the list.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* address gemini review feedback
* fix flaky test and console warning in test
* fix drawer min/max edge case bug
* fix Copy ID chip falsely showing 'Copied!' on clipboard reject
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
* fix(observe): iteration tree parity with legacy
Iteration metadata (parentNodeId/iterationIndex) is nested under `data`
in the runtime payload per IAgentflowExecutedData; the SDK was reading
from the top level so virtual Iteration #N containers never materialized
and iteration children dangled at root. Same shape fix applies to
humanInputAgentflow detection in useNodeData. Iteration parent input
now shows "No data" matching legacy's `data.input.question || '*No
data*'` fallback, and virtual Iteration #N containers are selectable
with empty Input/Output panels.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <[email protected]>
20 KiB
@flowiseai/observe - Architecture
This document describes the internal architecture of the @flowiseai/observe package.
Overview
The package follows a Domain-Driven Modular Architecture with clear separation of concerns — the same four-layer structure as @flowiseai/agentflow. The goal is to keep UI primitives, business logic, API communication, and domain features cleanly separated, so individual features (executions, evaluations, chat history) can be added, tested, and removed without disturbing each other.
src/
├── index.ts # Public Package API ("Public Face")
│
├── atoms/ # ⚛️ UI Primitives ("What it looks like" (Dumb))
├── features/ # 🧩 Domain Features ("What it does" (Smart))
├── core/ # 🧠 Business Logic ("Business Rules" (Types/Logic))
└── infrastructure/ # 🌐 External Services ("Outside World" (API/Storage))
Directory Structure
atoms/ - UI Primitives
"What it looks like." Tiny, irreducible UI components with no business logic. Shared building blocks used across features.
Example layout (illustrative — not exhaustive; check the directory for the current set):
atoms/
├── StatusIndicator.tsx # Execution / node status icon (INPROGRESS, FINISHED, ERROR…)
├── MetricsDisplay.tsx # Token / cost / time metrics row
├── NodeIcon.tsx # AGENTFLOW_ICONS-driven avatar for a node type
├── … # Other primitives as the SDK grows
└── index.ts # Central export
Rules:
- Must be "dumb" and stateless (or minimal local state for animations)
- No business logic
- No API calls
- Imported by features, never the reverse
- Forbidden: Importing from
features/orinfrastructure/— onlycore/types,core/theme, andcore/primitives
Goal: 100% visual consistency across all observe features.
features/ - Domain Features
"What it does." Self-contained domain modules. Each feature owns its components, hooks, and feature-local utilities. Adding a new observability feature (evaluations, chat history, MALT) means adding a new directory here — it does not touch existing feature exports.
Example layout (illustrative — not exhaustive; the executions feature currently has more sub-components and hooks than shown here):
features/
├── executions/ # Execution viewer
│ ├── components/
│ │ ├── ExecutionsViewer.tsx # Filterable list + pagination + resizable detail drawer
│ │ ├── ExecutionDetail.tsx # Resizable split-pane: tree sidebar + node detail
│ │ ├── ExecutionTreeSidebar.tsx # @mui/x-tree-view RichTreeView with custom slot
│ │ ├── NodeExecutionDetail.tsx # Step viewer: markdown/JSON/code/HTML renderers, HITL
│ │ ├── ExecutionsListTable.tsx # Sortable/selectable MUI table
│ │ └── … # Other components (chat bubbles, panels) as the feature grows
│ ├── hooks/
│ │ ├── useExecutionPoll.ts # Auto-poll while INPROGRESS, stop on terminal state
│ │ ├── useExecutionTree.ts # Build ExecutionTreeNode[] from flat NodeExecutionData[]
│ │ ├── useResizableSidebar.ts # Drag-to-resize width state (left- or right-anchored panels)
│ │ └── … # Other feature-local hooks
│ └── index.ts # Public API for this feature
│
└── (evaluations/) # Future — GA
└── (chat-history/) # Future — GA
Rules:
- Each feature has an
index.tsgatekeeper — only re-export what consumers need - Features never import from other features directly — if two features share logic, move it to
core/ - Feature-specific utilities stay in the feature folder
- Hooks are co-located with their feature
Goal: High cohesion. Deleting a feature folder should not break other features.
core/ - Business Logic
"The Brain." Framework-agnostic types, utilities, and the MUI theme factory. No React components, no API calls.
Example layout (illustrative — not exhaustive; new sub-files land under types/, primitives/, and utils/ as the SDK grows):
core/
├── types/
│ ├── execution.ts # Execution, ExecutionState, NodeExecutionData,
│ │ # ExecutionTreeNode, HumanInputParams
│ ├── observe.ts # ObserveBaseProps, ExecutionsViewerProps,
│ │ # ExecutionDetailProps, ExecutionFilters
│ ├── nodeDetail.ts # ChatMessage, ConditionEntry, AvailableToolEntry, …
│ └── index.ts
│
├── primitives/ # Domain-free, safe to import from atoms/
│ ├── agentflowIcons.ts # AGENTFLOW_ICONS registry (icon + color per node type)
│ ├── json.ts # JSON tokenizer / safe parser
│ └── index.ts
│
├── utils/ # Domain-aware helpers — not importable from atoms/
│ ├── guards.ts # Type predicates over core/types/* shapes
│ ├── tools.ts # Tool-call extraction / resolution
│ └── index.ts
│
└── theme/
├── tokens.ts # Color palette, spacing, shadows
├── createObserveTheme.ts # MUI theme factory (isDarkMode toggle)
└── index.ts
Rules:
- No React components — pure TypeScript only
- No browser-specific APIs unless unavoidable
- No side effects
- Can be tested in isolation
core/primitives/ vs core/utils/
primitives/— Domain-free utilities with no knowledge of executions or node data. Safe to import fromatoms/.utils/— Domain-aware utilities (e.g., execution state helpers). Only importable byfeatures/andinfrastructure/.
When adding a utility, ask: "Does this function need to know what an Execution or NodeExecutionData is?" If no → primitives/. If yes → utils/.
Goal: Framework-agnostic source of truth.
infrastructure/ - External Services
"The Outside World." API client factory and React context providers.
Example layout (illustrative — not exhaustive; new API modules land under api/ as features grow):
infrastructure/
├── api/
│ ├── client.ts # bindApiClient(apiBaseUrl, token) — axios with Bearer header
│ ├── executions.ts # createExecutionsApi(client): getAllExecutions, getExecutionById,
│ │ # deleteExecutions, updateExecution
│ └── index.ts
│
└── store/
└── ObserveContext.tsx # ObserveProvider, useObserveApi(), useObserveConfig()
Rules:
- All external communication goes through
infrastructure/api/ - Contexts are internal — composed in
ObserveProvider, not exported directly to consumers useObserveApi()throws if called outsideObserveProvider(fast failure)
Goal: Wrap external dependencies so they can be mocked in tests.
Dependency Flow
Dependencies must only flow downwards.
┌─────────────────────────────────────────────────────────┐
│ Root Files (Public API) │
│ (index.ts) │
│ "The Public Face" │
└─────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────┐
│ features/ │
│ (executions, evaluations, chat-history, …) │
│ "What it does" │
└─────────────────────────────────────────────────────────┘
│ │
▼ ▼
┌──────────────────┐ ┌────────────────────────┐
│ atoms/ │ │ infrastructure/ │
│ (UI primitives) │ │ (api, store) │
│ "What it looks │ │ "The Outside World" │
│ like" │ │ │
└──────────────────┘ └────────────────────────┘
│ │
└──────────────┬─────────────────────┘
▼
┌─────────────────────┐
│ core/ │
│ (types, theme, │
│ primitives) │
│ "The Brain" │
└─────────────────────┘
Import Rules:
features→atoms,infrastructure,core✅infrastructure→core✅atoms→core/types,core/theme,core/primitivesonly ✅core→ nothing (leaf node) ✅- No reverse imports — enforced by ESLint
import/no-restricted-paths
Observe-Specific Patterns
ObserveProvider — Root Provider
ObserveProvider composes three providers:
<ThemeProvider theme={createObserveTheme(isDarkMode)}>
<ObserveApiContext.Provider value={api}>
<ObserveConfigContext.Provider value={config}>{children}</ObserveConfigContext.Provider>
</ObserveApiContext.Provider>
</ThemeProvider>
ObserveApiContext— holds the bound axios client (created once fromapiBaseUrl+token+requestInterceptor)ObserveConfigContext— holdsisDarkModeand other display config- Features call
useObserveApi()to get the client; they never construct it themselves
Request Interceptor — Auth Bridging
The SDK sends Authorization: Bearer <token> by default. Different consumers authenticate differently:
| Consumer | Auth mechanism | How to configure |
|---|---|---|
| DevSite | AgentForge proxy headers | requestInterceptor to inject proxy token |
| OSS (embedded in Flowise UI) | Session cookie + x-request-from: internal |
requestInterceptor to set withCredentials + header |
| OSS (standalone / external) | Flowise API key | token prop (default) |
requestInterceptor is a callback on the Axios request config, identical in shape to the one in @flowiseai/agentflow:
interface ObserveBaseProps {
requestInterceptor?: (config: InternalAxiosRequestConfig) => InternalAxiosRequestConfig
}
It is passed to bindApiClient and registered via axios.interceptors.request.use. If the callback throws, the error is caught, logged, and the original unmodified config is used so the request still proceeds safely.
OSS session auth example:
<ObserveProvider
apiBaseUrl='http://localhost:3000'
requestInterceptor={(config) => {
config.withCredentials = true
config.headers['x-request-from'] = 'internal'
return config
}}
>
<ExecutionsViewer agentflowId='...' />
</ObserveProvider>
requestInterceptor should only be provided by trusted, developer-authored code — never from user input or dynamically evaluated strings.
Auto-Poll (useExecutionPoll)
useExecutionPoll sets up a setInterval while execution.state === 'INPROGRESS':
INPROGRESS → polls every pollInterval ms
FINISHED | ERROR | TERMINATED | TIMEOUT | STOPPED → interval cleared immediately
pollInterval={0}disables auto-poll entirelyrefresh()is returned for manual refresh buttons- The hook clears the interval on unmount
Execution Tree (useExecutionTree)
useExecutionTree converts the flat NodeExecutionData[] stored as JSON in execution.executionData into a hierarchical ExecutionTreeNode[]. Two parenting mechanisms run in sequence (mirrors legacy ExecutionDetails.jsx buildTreeData):
previousNodeIdsparent→child — each non-iteration node attaches to the most-recent prior instance of any node listed in itspreviousNodeIds. Nodes with empty / unmatchedpreviousNodeIdsbecome roots. The "most recent" rule disambiguates when the samenodeIdruns more than once (e.g. inside a loop).- Iteration grouping — children with
parentNodeId+iterationIndexare bundled into virtualIteration #Ncontainer nodes (isVirtualNode: true) under the iteration agent's most-recent instance, then linked to one another inside the iteration via the samepreviousNodeIdsrule (restricted to same-iteration siblings).
Tree-node ids are ${nodeId}_${arrayIndex} so duplicate nodeIds remain addressable. The tree is consumed by ExecutionTreeSidebar (a RichTreeView slot) inside ExecutionDetail.
HITL Callback — Opaque Consumer Contract
The onHumanInput prop is intentionally opaque:
interface ExecutionsViewerProps {
onHumanInput?: (agentflowId: string, params: HumanInputParams) => Promise<void>
}
- In OSS: consumer calls the Flowise prediction endpoint directly
- In DevSite: consumer routes through the AgentForge proxy
The SDK never calls the prediction API itself. Approve/Reject buttons are only rendered when onHumanInput is provided AND the node is a human input node in INPROGRESS state. This keeps auth and routing logic entirely out of the SDK.
Tenant Isolation
Tenant isolation is handled server-side by Flowise's ExtendRequestContextMiddleware via request headers. The SDK does not accept or pass a tenantId prop — the API returns only what the authenticated token is permitted to see.
Gatekeeper Pattern
Each module exposes only what's needed via its index.ts:
// features/executions/index.ts (illustrative — re-export only what consumers need)
// ✅ Public API
export { ExecutionsViewer } from './components/ExecutionsViewer'
export { ExecutionDetail } from './components/ExecutionDetail'
export { NodeExecutionDetail } from './components/NodeExecutionDetail'
export { useExecutionPoll } from './hooks/useExecutionPoll'
export { useExecutionTree } from './hooks/useExecutionTree'
// ❌ Internal sub-components (e.g. ExecutionTreeSidebar, ChatMessageBubble) and helpers stay private
Adding a New Feature
Adding evaluations, chat history, or MALT follows the same pattern:
-
Create a folder under
features/:features/evaluations/ ├── components/ │ └── EvaluationsViewer.tsx ├── hooks/ │ └── useEvaluationPoll.ts └── index.ts -
Add the API module under
infrastructure/api/:infrastructure/api/ └── evaluations.ts # createEvaluationsApi(client) -
Add any new types to
core/types/(e.g.,evaluation.ts) -
Export only from the feature's
index.ts, then re-export from the packagesrc/index.ts -
Import from infrastructure/core, never from other features:
// ✅ Good import { useObserveApi } from '../../../infrastructure/store' import type { Execution } from '../../../core/types' // ❌ Bad — cross-feature import import { ExecutionDetail } from '../executions/components/ExecutionDetail'
Root Files (src/index.ts)
The barrel export exposes only the public surface. Example shape (illustrative — refer to src/index.ts for the current set):
// Components
export { ExecutionsViewer, ExecutionDetail, NodeExecutionDetail } from './features/executions'
export { ObserveProvider } from './infrastructure/store'
// Hooks (for consumers building custom UIs on top of observe infrastructure)
export { useExecutionPoll, useExecutionTree } from './features/executions'
export { useObserveApi, useObserveConfig } from './infrastructure/store'
// Types
export type { ExecutionsViewerProps, ExecutionDetailProps, ExecutionFilters } from './core/types'
export type { Execution, ExecutionState, NodeExecutionData, HumanInputParams, AgentflowRef } from './core/types'
Everything else is internal implementation detail.
Development Principles
Barrel Exports
Every directory has an index.ts gatekeeper.
✅ Good:
import { StatusIcon } from '../../atoms'
import { useObserveApi } from '../../infrastructure/store'
❌ Bad:
// Never deep-link into files
import { StatusIcon } from '../../atoms/StatusIcon'
import { useObserveApi } from '../../infrastructure/store/ObserveContext'
Naming Conventions
| Type | Convention | Example |
|---|---|---|
| Component | PascalCase.tsx | ExecutionDetail.tsx |
| Hook | camelCase.ts (use prefix) |
useExecutionPoll.ts |
| Logic/Types | camelCase.ts | execution.ts, observe.ts |
| API module | camelCase.ts | executions.ts |
| Styles | kebab-case (co-located) | executions.css |
Theme Token Ordering
Inside core/theme/tokens.ts, all maps are ordered alphanumerically: baseColors, top-level tokens groups, subgroups, and keys (e.g. gray50 < gray75 < gray100 < gray200). Each { dark, light } / { dark, light, main } pair follows the same rule. Keep new entries in order so diffs stay small.
Type Location Rule
Domain types live in core/types/, not alongside their primary component, when they cross any of these boundaries:
- Imported by a hook (
features/*/hooks/*) - Imported by another component in a different file
- Used in a public API exported via
src/index.ts
Component-internal types (props interfaces, file-private shapes used only inside the file that defines the component) stay co-located. The rule prevents the inversion where a hook reaches "upward" into a sibling component file just to read a type.
Props interfaces (e.g. ChatMessageBubbleProps) are always file-local — never lift them to core/types/ even if another component would re-use one. Re-using a props interface across components is a code smell on its own; the right fix is a shared underlying type, not a shared props interface.
Predicates that narrow into a core/types/ shape (e.g. isChatMessageArray(value): value is ChatMessage[]) live in core/utils/ next to other domain-aware helpers — never in atoms/ (atoms cannot import from core/utils, only from core/primitives).
Example layout:
core/types/nodeDetail.ts—ChatMessage,ConditionEntry,AvailableToolEntry,UsedToolEntry,UsedToolRef,ToolNodeRef,NormalizedToolCallcore/utils/guards.ts—isChatMessageArray,isConditionArray,isAvailableToolArray,isUsedToolArraycore/utils/tools.ts—resolveTool,extractToolCalls(consumed byChatMessageBubble+ToolAccordionList)