kandev/AGENTS.md

Kandev Engineering Guide

Purpose: Architecture notes, key patterns, and conventions for LLM agents working on Kandev.

Repo Layout

apps/
├── backend/          # Go backend (orchestrator, lifecycle, agentctl, WS gateway)
├── web/              # Next.js frontend (SSR + WS + Zustand)
├── cli/              # CLI tool (TypeScript)
├── landing/          # Landing page
└── packages/         # Shared packages/types

Tooling

• Package manager: pnpm workspace (run from apps/, not repo root)
• Backend: Go with Make (make -C apps/backend test|lint|build)
• Frontend: Next.js (cd apps && pnpm --filter @kandev/web dev|lint|typecheck)
• UI: Shadcn components via @kandev/ui
• GitHub repo: https://github.com/kdlbs/kandev
• Container image: ghcr.io/kdlbs/kandev (GitHub Container Registry)

---

Backend Architecture

Package Structure

apps/backend/
├── cmd/
│   ├── kandev/           # Main backend binary entry point
│   ├── agentctl/         # Agentctl binary (runs inside containers or standalone)
│   └── mock-agent/       # Mock agent for testing
├── internal/
│   ├── agent/
│   │   ├── lifecycle/    # Agent instance management (see below)
│   │   ├── agents/       # Agent type implementations
│   │   ├── controller/   # Agent control operations
│   │   ├── credentials/  # Agent credential management
│   │   ├── discovery/    # Agent discovery
│   │   ├── docker/       # Docker-specific agent logic
│   │   ├── dto/          # Agent data transfer objects
│   │   ├── executor/     # Executor types, checks, and service
│   │   ├── handlers/     # Agent event handlers
│   │   ├── registry/     # Agent type registry and defaults
│   │   ├── settings/     # Agent settings
│   │   ├── mcpconfig/    # MCP server configuration
│   │   └── remoteauth/   # Remote auth catalog and method IDs for remote executors/UI
│   ├── agentctl/
│   │   ├── client/       # HTTP client for talking to agentctl
│   │   └── server/       # agentctl HTTP server
│   │       ├── acp/      # ACP protocol implementation
│   │       ├── adapter/  # Protocol adapters + transport/ (ACP, Codex, OpenCode, Copilot, Amp)
│   │       ├── api/      # HTTP endpoints
│   │       ├── config/   # agentctl configuration
│   │       ├── instance/ # Multi-instance management
│   │       ├── mcp/      # MCP server integration
│   │       ├── process/  # Agent subprocess management
│   │       ├── shell/    # Shell session management
│   │       └── utility/  # agentctl utilities
│   ├── orchestrator/     # Task execution coordination
│   │   ├── dto/          # Orchestrator data transfer objects
│   │   ├── executor/     # Launches agents via lifecycle manager
│   │   ├── handlers/     # Orchestrator event handlers
│   │   ├── messagequeue/ # Message queue for agent prompts
│   │   ├── queue/        # Task queue
│   │   ├── scheduler/    # Task scheduling
│   │   └── watcher/      # Event handlers
│   ├── task/
│   │   ├── controller/   # Task HTTP/WS controllers
│   │   ├── dto/          # Task data transfer objects
│   │   ├── events/       # Task event types
│   │   ├── handlers/     # Task event handlers
│   │   ├── models/       # Task, Session, Executor, Message models
│   │   ├── repository/   # Database access (SQLite)
│   │   └── service/      # Task business logic
│   ├── analytics/        # Usage analytics
│   ├── clarification/    # Agent clarification handling
│   ├── common/           # Shared utilities
│   ├── db/               # Database initialization
│   ├── debug/            # Debug tooling
│   ├── editors/          # Editor integration
│   ├── events/           # Event bus for internal pub/sub
│   ├── gateway/          # WebSocket gateway
│   ├── github/           # GitHub API integration (PRs, reviews, webhooks)
│   ├── integration/      # External integrations
│   ├── integrations/     # Shared shapes for third-party integrations
│   │   ├── healthpoll/   # Reusable 90s auth-health Poller (used by jira, linear)
│   │   └── secretadapter/ # Upsert-style adapter over secrets.SecretStore
│   ├── jira/             # Jira/Atlassian Cloud integration (config, REST client, poller)
│   ├── linear/           # Linear integration (config, GraphQL client, poller)
│   ├── lsp/              # LSP server
│   ├── mcp/              # MCP protocol support
│   ├── health/           # Health check endpoints
│   ├── notifications/    # Notification system
│   ├── persistence/      # Persistence layer
│   ├── prompts/          # Prompt management
│   ├── repoclone/        # Repository cloning for remote executors
│   ├── scriptengine/     # Script placeholder resolution and interpolation
│   ├── secrets/          # Secret management
│   ├── sprites/          # Sprites AI integration
│   ├── sysprompt/        # System prompt injection
│   ├── task/
│   │   └── ...
│   ├── tools/            # Tool integrations
│   ├── user/             # User management
│   ├── utility/          # Shared utility functions
│   ├── workflow/         # Workflow engine
│   │   ├── engine/       # Typed state-machine engine (trigger evaluation, action callbacks, transition store)
│   │   ├── models/       # Workflow step, template, and history models
│   │   ├── repository/   # Workflow persistence (SQLite)
│   │   └── service/      # Workflow CRUD and step resolution
│   └── worktree/         # Git worktree management for workspace isolation

Key Concepts

Orchestrator coordinates task execution:

• Receives task start/stop/resume requests via WebSocket
• Delegates to lifecycle manager for agent operations
• Handles event-driven state transitions via workflow engine
• Located in internal/orchestrator/

Workflow Engine (internal/workflow/engine/) provides typed state-machine evaluation:

• Engine.HandleTrigger() evaluates step actions for triggers (on_enter, on_turn_start, on_turn_complete, on_exit)
• TransitionStore interface abstracts persistence (implemented by orchestrator.workflowStore)
• CallbackRegistry maps action kinds to callbacks (plan mode, auto-start, context reset)
• First-transition-wins: multiple transition actions in one trigger, first eligible wins
• EvaluateOnly mode: engine evaluates without persisting, caller orchestrates on_exit → DB → on_enter
• RequiresApproval on actions: transitions requiring review gating are skipped
• Idempotent by OperationID; session-scoped data bag via MachineState.Data

Lifecycle Manager (internal/agent/lifecycle/) manages agent instances:

• Manager (manager.go, manager_*.go) - central coordinator for agent lifecycle
• ExecutorBackend interface (executor_backend.go) - abstracts execution environment (Docker, Standalone, Sprites, Remote Docker)
• ExecutionStore (execution_store.go) - thread-safe in-memory execution tracking
• session.go - ACP session initialization and resume
• streams.go - WebSocket stream connections to agentctl
• process_runner.go - agent process launch and management
• profile_resolver.go - resolves agent profiles/settings

agentctl is an HTTP server that:

• Runs inside Docker containers or as standalone process
• Manages agent subprocess via stdin/stdout (ACP protocol)
• Exposes workspace operations (shell, git, files)
• Supports multiple concurrent instances on different ports

Executor Types (database model):

• local_pc - Standalone process on host ✅
• local_docker - Docker container on host ✅
• sprites - Sprites cloud environment ✅
• remote_docker, remote_vps, k8s - Planned

Execution Flow

Client (WS) → Orchestrator → Lifecycle Manager → ExecutorBackend (container/process) → agentctl
                                                                                          ↓
Client (WS) ← Orchestrator ← Lifecycle Manager ←──── stream updates (WS) ──────── agent subprocess

1. Orchestrator receives task.start via WS
2. Lifecycle Manager creates executor instance (container or process)
3. agentctl starts inside the instance, agent subprocess is configured and started
4. Agent events stream back via WS through the chain

Session Resume: TaskSession.ACPSessionID stored for resume; ExecutorRunning tracks active state; on restart RecoverInstances() reconnects.

Provider Pattern: Packages expose Provide(cfg, log) (*impl, cleanup, error) for DI. Returns implementation, cleanup function, and error. Cleanup called during graceful shutdown.

Worktrees: internal/worktree/Manager provides workspace isolation. Each session can have its own worktree (branch) to prevent conflicts between concurrent agents.

Executor default scripts: Default prepare scripts are in internal/agent/lifecycle/default_scripts.go; internal/scriptengine/ handles placeholder resolution.

---

agentctl Server

API Groups

agentctl exposes these route groups (see internal/agentctl/server/api/):

• /health, /info, /status - Health and status
• /instances/* - Multi-instance management
• /processes/* - Agent subprocess management (start/stop)
• /agent/configure, /agent/stream - Agent configuration and event streaming
• /git/* - Git operations (status, commit, push, pull, rebase, stage, etc.)
• /shell/* - Shell session management
• /workspace/* - File operations, search, tree
• /vscode/* - VS Code integration proxy

Adapter Model

Protocol adapters in adapter/transport/ normalize different agent CLIs:

• AgentAdapter interface defines Start(), Stop(), Prompt(), Cancel()
• Transports: acp (Claude Code), codex (OpenAI Codex), opencode, shared, streamjson
• Top-level adapters: CopilotAdapter (GitHub Copilot SDK), AmpAdapter (Sourcegraph Amp)
• process.Manager owns subprocess, wires stdio to adapter
• Factory pattern in adapter/factory.go selects adapter by agent type

ACP Protocol

JSON-RPC 2.0 over stdin/stdout between agentctl and agent process. Requests: initialize, session/new, session/load, session/prompt, session/cancel. Notifications: session/update with types message_chunk, tool_call, tool_update, complete, error, permission_request, context_window.

---

Frontend Architecture

UI Components

Shadcn Components: Import from @kandev/ui package:

import { Badge } from '@kandev/ui/badge';
import { Button } from '@kandev/ui/button';
import { Dialog } from '@kandev/ui/dialog';
// etc...

Do NOT import from @/components/ui/* - always use @kandev/ui package.

Data Flow Pattern (Critical)

SSR Fetch -> Hydrate Store -> Components Read Store -> Hooks Subscribe

Never fetch data directly in components.

Store Structure (Domain Slices)

lib/state/
├── store.ts                        # Root composition
├── default-state.ts                # Default state + initial state merge
├── slices/                         # Domain slices
│   ├── kanban/                    # boards, tasks, columns
│   ├── session/                   # sessions, messages, turns, worktrees
│   ├── session-runtime/           # shell, processes, git, context
│   ├── workspace/                 # workspaces, repos, branches
│   ├── settings/                  # executors, agents, editors, prompts (incl. userSettings)
│   ├── comments/                  # code review diff comments
│   ├── github/                    # GitHub PRs, reviews
│   └── ui/                        # preview, connection, active state, sidebar views
├── hydration/                     # SSR merge strategies

hooks/domains/{kanban,session,workspace,settings,comments,github}/  # Domain-organized hooks
lib/api/domains/                    # API clients
├── kanban-api, session-api, workspace-api, settings-api, process-api
├── plan-api, queue-api, workflow-api, stats-api, github-api
├── user-shell-api, debug-api, secrets-api, sprites-api, vscode-api
├── health-api, utility-api

Key State Paths:

• messages.bySession[sessionId], shell.outputs[sessionId], gitStatus.bySessionId[sessionId]
• tasks.activeTaskId, tasks.activeSessionId, workspaces.activeId
• repositories.byWorkspace, repositoryBranches.byRepository

Hydration: lib/state/hydration/merge-strategies.ts has deepMerge(), mergeSessionMap(), mergeLoadingState() to avoid overwriting live client state. Pass activeSessionId to protect active sessions.

Hooks Pattern: Hooks in hooks/domains/ encapsulate WS subscription + store selection. WS client deduplicates subscriptions automatically.

WS

Format: {id, type, action, payload, timestamp}.

---

Best Practices

Commit Conventions (enforced by CI)

Commits to main must follow Conventional Commits (type: description). PRs are squash-merged — the PR title becomes the commit, validated by CI. Changelog is auto-generated from these via git-cliff (cliff.toml). See .agents/skills/commit/SKILL.md for allowed types and examples.

Release & Versioning Convention

Kandev uses a single SemVer X.Y.Z shared across all distribution channels:

• apps/cli/package.json version → X.Y.Z
• npm main package: kandev@X.Y.Z
• npm runtime packages: @kdlbs/runtime-{platform}@X.Y.Z (5 platforms; declared as optionalDependencies in main package)
• Git tag: vX.Y.Z (three-part; legacy vM.m tags normalize to M.m.0)
• Homebrew formula: kdlbs/homebrew-kandev Formula/kandev.rb version "X.Y.Z"
• GitHub release: vX.Y.Z with platform tarballs kandev-{platform}.tar.gz + .sha256

npm and Homebrew are sibling channels, not chained. Both consume the same GitHub release artifacts; neither depends on the other.

Release flow (entirely in CI via .github/workflows/release.yml, triggered by maintainer from GitHub Actions UI):

1. Maintainer clicks "Run workflow" → picks bump (patch/minor/major) → optional dry_run.
2. prepare job bumps version + regenerates CHANGELOG, opens release PR, squash-merges, tags vX.Y.Z.
3. build-web + build-cli + build-bundles (5 platforms) build the release artifacts.
4. publish-release creates the GitHub release with platform tarballs + sha256 + auto-generated notes.
5. publish-npm publishes 5 @kdlbs/runtime-* packages + main kandev package to npmjs.
6. update-homebrew-tap pushes updated Formula/kandev.rb to kdlbs/homebrew-kandev via SSH deploy key.

There is no local release script — the entire flow runs in GHA.

Runtime resolution (in apps/cli/src/runtime.ts):

1. KANDEV_BUNDLE_DIR env var (set by Homebrew wrapper, used by tests).
2. Installed @kdlbs/runtime-{platform} npm package via require.resolve().
3. --runtime-version <tag> cache fallback (debug only — downloads from GitHub).

Code Quality (enforced by linters)

Static analysis runs in CI and pre-commit. New code must stay within these limits:

Go (apps/backend/.golangci.yml - errors on new code only):

• Functions: ≤80 lines, ≤50 statements
• Cyclomatic complexity: ≤15 · Cognitive complexity: ≤30
• Nesting depth: ≤5 · Naked returns only in functions ≤30 lines
• No duplicated blocks (≥150 tokens) · Repeated strings → constants (≥3 occurrences)

TypeScript (apps/web/eslint.config.mjs - warnings, will become errors):

• Files: ≤600 lines · Functions: ≤100 lines
• Cyclomatic complexity: ≤15 · Cognitive complexity: ≤20
• Nesting depth: ≤4 · Parameters: ≤5
• No duplicated strings (≥4 occurrences) · No identical functions · No unused imports
• No nested ternaries

When you hit a limit: extract a helper function, custom hook, or sub-component. Prefer composition over growing a single function.

Testing

Every code change must include tests for new or changed logic. Backend: *_test.go files alongside the source. Frontend: *.test.ts files for utility functions, hooks, API clients, and store slices. Exceptions: config files, generated code, React component markup. Use /tdd for test-driven development.

Knowledge

• Specs: Feature specs live in docs/specs/<slug>/spec.md — the "what & why" of a feature, written before coding. Optional siblings: plan.md (how), notes.md (post-ship). Use /spec to write or update a spec. See docs/specs/INDEX.md.
• Decisions: Architecture decisions are recorded in docs/decisions/. Read docs/decisions/INDEX.md for an overview. When making significant architectural choices, create a new ADR via /record decision.
• Plans: Retrospective implementation plans live in docs/plans/ (created via /record plan). A forward-looking plan tied to a spec goes at docs/specs/<slug>/plan.md.

Backend

• Provider pattern for DI; stderr for logs, stdout for ACP only
• Pass context through chains; event bus for cross-component comm
• Execution access: Workspace-oriented handlers (files, shell, inference, ports, vscode, LSP) MUST use GetOrEnsureExecution(ctx, sessionID) — it recovers from backend restarts by creating executions on-demand. Only use GetExecutionBySessionID for operations that require a running agent process (prompt, cancel, mode).
• Testing: Prefer testing/synctest (Go 1.24+) over time.Sleep for time-dependent tests. Use synctest.Test to wrap tests with tickers or timeouts — it advances fake time instantly when all goroutines are idle. When synctest is not feasible (e.g., tests spawning external processes like git), use channel-based synchronization (<-started, non-blocking select) instead of sleep-based waits. Reserve time.Sleep only for integration tests that need real subprocess execution time.

Frontend

• Data: SSR fetch → hydrate → read store. Never fetch in components
• UI Components:
  • Import shadcn components from @kandev/ui, NOT @/components/ui/*
  • Always prefer native shadcn components over custom implementations
  • Check apps/packages/ui/src/ for available components (pagination, table, dialog, etc.)
  • For data tables, use @kandev/ui/table with TanStack Table; use shadcn Pagination components
  • Only create custom components when shadcn doesn't provide what's needed
• Components: <200 lines, extract to domain components, composition over props
• Hooks: Domain-organized in hooks/domains/, encapsulate subscription + selection
• WS: Use subscription hooks only; client auto-deduplicates
• Interactivity: All buttons and links with actions must have cursor-pointer class

Plan Implementation

• After implementing a plan, run make fmt first to format code, then run make typecheck test lint to verify the changes. Formatting must come first because formatters may split lines, which can trigger complexity linter warnings.

GitHub Operations

Skills use gh CLI by default. If a gh command fails (not installed, not authenticated, etc.), use whatever GitHub tools are available in the environment (MCP GitHub tools, API tools, etc.) to accomplish the same operation. The goal is the same — the tool may differ.

Adding a new third-party integration

Jira and Linear are the model: per-workspace credentials, a 90s auth-health poller, a settings page with status banner + reconnect CTA, link/import buttons that gate on availability. New integrations should reuse the shared shapes rather than copying either.

Backend (apps/backend/internal/<name>/):

• Mirror the package layout: service.go, store.go, client.go, provider.go, handlers.go, models.go, poller.go. Expose Provide(writer, reader *sqlx.DB, secrets SecretStore, log *logger.Logger) (*Service, func() error, error) (add eventBus bus.EventBus only if the integration actually publishes events — Linear does not).
• Use internal/integrations/secretadapter instead of writing your own upsert wrapper around secrets.SecretStore. The adapter satisfies any per-integration SecretStore interface shaped as {Reveal, Set, Delete, Exists}.
• Use internal/integrations/healthpoll for the auth-health loop. Implement the Prober interface (ListConfiguredWorkspaces + RecordAuthHealth) on a small adapter and let healthpoll.New("name", prober, log) own Start/Stop/ticker. Keep integration-specific loops (JQL polling, webhook reconciliation, etc.) separate, like jira's issue-watch loop.
• Wire the service via a per-domain init<Name>Service(...) helper in cmd/kandev/services.go, not inline in provideServices.
• Ship a mock_client.go + mock_controller.go next to the real client. Provide branches on KANDEV_MOCK_<NAME>=true and returns the in-memory client; RegisterMockRoutes(router, svc, log) mounts /api/v1/<name>/mock/* only when the service was built with the mock. The e2e backend fixture sets the env var so Playwright tests drive the mock via apiClient.mock<Name>*() helpers — see jira/linear for the layout.

Frontend:

• Hooks live under hooks/domains/<name>/, not components/<name>/.
• Use hooks/domains/integrations/use-integration-availability.ts and use-integration-enabled.ts — each integration's useXAvailable / useXEnabled should be a one-line wrapper passing the storage key + sync event + config-fetch function.
• Settings page reuses <IntegrationAuthStatusBanner> (components/integrations/auth-status-banner.tsx).
• "Auth required / reconnect" UI reuses <IntegrationAuthErrorMessage> (components/integrations/auth-error-message.tsx) — supply the integration's display name, regex check, and reconnect href.
• Link / import popovers reuse <ValidatedPopover> (components/integrations/validated-popover.tsx) — supply the icon, label, key regex, fetch function, and success callback.

Where Jira and Linear deliberately diverge:

• Issue model: Jira uses transitions + JQL; Linear uses state IDs + structured filters. Don't merge these schemas — the upstream APIs are genuinely different.
• Event publishing: only Jira accepts eventBus and emits NewJiraIssueEvent from its issue-watch loop. Linear has no equivalent feature today.
• Health column extras: Linear's linear_configs row carries an org_slug captured from successful probes; Jira's row does not.

---

Maintaining This File

This file is read by AI coding agents (Claude Code via CLAUDE.md symlink, Codex via AGENTS.md). If your changes make any section of this file outdated or inaccurate - e.g., you add/remove/rename packages, change architectural patterns, add new adapters, modify store slices, or change conventions - update the relevant sections of this file as part of the same PR. Keep descriptions concise and factual. Do not add speculative or aspirational content.

---

Last Updated: 2026-05-02