Skip to content
/

Ptah (Coding Orchestra)

ptah-orchestra

Primitive shape
No installable primitives
00

Summary

Ptah (Coding Orchestra) — Summary

Ptah is a massively ambitious VS Code extension (plus Electron desktop app and headless CLI) that delivers multi-agent orchestration, workspace intelligence, and a built-in MCP server in a single product. It ships 13 named agent personas (Backend Developer, Frontend Developer, Software Architect, Project Manager, Team Leader, Code Logic Reviewer, Code Style Reviewer, Senior Tester, Research Expert, DevOps Engineer, Modernization Detector, UI/UX Designer, Technical Content Writer), 7 workflow types (Feature, Bugfix, Refactoring, Documentation, Research, DevOps, Creative), a 14-namespace MCP server with AST parsing + LSP diagnostics + dependency graphs, a Letta-style memory curator backed by SQLite, Telegram/Discord/Slack messaging gateway, SQLite-backed cron scheduler, and a 6-step setup wizard that auto-generates agent configurations for 13+ project types. Built on a Nx monorepo with Angular 21 webview, NestJS 11 license server, and @anthropic-ai/claude-agent-sdk. Licensed under FSL-1.1-MIT (converts to full MIT 2 years after release). Freemium model with ptah.live Pro trial. This is by far the most architecturally complex framework in this batch — comparable only to claude-flow in scope. Compared to the seeds, Ptah is closest to claude-flow (MCP-anchored, multi-agent, custom agent loop) but ships a VS Code extension + Electron desktop app + headless CLI instead of a pure CLI tool.

01

Overview

Ptah (Coding Orchestra) — Overview

Origin

Built by Hive-Academy (github.com/Hive-Academy). 3 contributors on the extension repo. 26 stars. Last commit 2026-05-25. Published on VS Code Marketplace as ptah-extensions.ptah-coding-orchestra.

Philosophy

"Provider-agnostic AI orchestration with intelligent workspace analysis and multi-agent workflows."

"Spawn multiple AI agents as headless background workers. The primary agent delegates subtasks and checks back for results — enabling true parallel multi-agent workflows."

"Drive Ptah from any MCP-compliant host over stdio. External orchestrators can spawn agents, monitor and steer them, and delegate whole tasks to the Team Leader harness via the session_submit tool — no bespoke integration required."

Licensing Philosophy

FSL-1.1-MIT (Functional Source License v1.1): each release becomes full MIT two years after publication. This is a "time-delay open source" model that enables commercial development while guaranteeing eventual full openness.

AI-Native Dogfood

From CLAUDE.md: 50+ module-scoped CLAUDE.md files across the monorepo (one per lib/app). This is the most thorough CLAUDE.md usage in the entire catalog — every library has its own agent guidance. The .ptah/specs/ directory contains TASK-prefixed spec files (e.g., TASK_2025_294) similar to Kiro's task decomposition.

VS Code Marketplace Compliance Note

The CLAUDE.md includes a notable constraint: VS Code Scanner rejects extensions containing trademarked AI product names (copilot, codex, claude, openai, anthropic) in non-JS files. Ptah works around this by downloading plugin/template content at runtime from GitHub rather than bundling it in the VSIX.

02

Architecture

Ptah (Coding Orchestra) — Architecture

Distribution

  • VS Code Marketplace: ext install ptah-extensions.ptah-coding-orchestra
  • Electron desktop app (ptah-electron)
  • Headless CLI: @hive-academy/ptah-cli (JSON-RPC stdio)

Technology Stack

  • Monorepo: Nx 22.6 + hexagonal architecture (ports & adapters)
  • AI SDKs: @anthropic-ai/claude-agent-sdk, @github/copilot-sdk, @openai/codex-sdk, Tavily, Exa
  • Frontend: Angular 21 (signals, OnPush, zoneless libs) + Tailwind 3 + daisyui 4 + Monaco + xterm.js + gridstack
  • Backend libs: TypeScript 5.9, tsyringe DI, NestJS 11 (license server), Electron 40
  • Persistence: better-sqlite3 + sqlite-vec (extension/desktop), Prisma 7 + PostgreSQL (license server only)
  • Validation: Zod 4 at all external boundaries
  • Build: esbuild (extension), ng-packagr (libs), electron-builder (desktop), Astro 6 (docs)

10 Runtime Targets (apps/)

App Purpose
ptah-extension-vscode VS Code extension host
ptah-extension-webview Angular 21 webview shell
ptah-electron Electron 40 desktop app
ptah-electron-e2e Playwright electron tests
ptah-cli Headless JSON-RPC stdio CLI
ptah-license-server NestJS 11 + Prisma + Paddle + WorkOS + Resend
ptah-license-server-e2e License server e2e
ptah-landing-page Angular marketing site
ptah-docs Astro Starlight documentation
infra-test MCP/trial-cron probes

16 Backend Libs (libs/backend/)

platform-core, platform-cli, platform-electron, platform-vscode, agent-sdk, agent-generation, workspace-intelligence, rpc-handlers, vscode-core, vscode-lm-tools, persistence-sqlite, memory-contracts, memory-curator, messaging-gateway, cron-scheduler, skill-synthesis

21 Frontend Libs (libs/frontend/)

core, ui, markdown, editor, chat, chat-state, chat-streaming, chat-routing, chat-ui, chat-types, chat-execution-tree, canvas, dashboard, setup-wizard, harness-builder, thoth-shell, memory-curator-ui, cron-scheduler-ui, messaging-gateway-ui, skill-synthesis-ui, webview-e2e-harness

Key Config Files

  • ~/.ptah/ptah.db — SQLite database (persistence-sqlite lib)
  • ~/.ptah/settings.json — Provider settings (API keys, model preferences)
  • agent-pack-manifest.json — Agent pack registry
  • content-manifest.json — Plugin/template manifest
  • .ptah/specs/ — TASK-prefixed spec files

Required Runtime

  • VS Code (extension) / Electron 40 (desktop) / Node.js (CLI)
  • Node.js (postinstall triggers electron native rebuild)
  • Optional: Docker (for license server dev)
03

Components

Ptah (Coding Orchestra) — Components

13 Named Agent Personas (agent-pack-manifest.json)

Agent Category Description
Team Leader planning Task Decomposition and Batch Orchestration Specialist
Software Architect planning Elite Software Architect for system design
Project Manager planning Technical Lead for task orchestration
Backend Developer development Server-side architecture and API design
Frontend Developer development Modern UI frameworks and component design
Code Logic Reviewer qa Ensures business logic correctness
Code Style Reviewer qa Coding standards, patterns, best practices
Senior Tester qa Comprehensive quality assurance
Research Expert specialist Deep technical analysis and strategic insights
DevOps Engineer specialist CI/CD, containerization, deployment automation
Modernization Detector specialist Identifies technology modernization opportunities
UI/UX Designer creative Visual design systems and production-ready specs
Technical Content Writer creative Technical documentation and content

7 Orchestration Workflow Types

Feature, Bugfix, Refactoring, Documentation, Research, DevOps, Creative — each with specialized agent delegation and user approval checkpoints.

MCP Server (14 API Namespaces)

Workspace analysis, LSP diagnostics, AST parsing, file search, dependency graphs, editor state. Exposed via ptah mcp-serve as stdio MCP server.

Key Backend Libs (Features)

Lib Feature
workspace-intelligence AST + symbol indexer + analysis (13+ project types)
memory-curator Letta-style memory + IndexingControl (sqlite-vec)
messaging-gateway Telegram/Discord/Slack + voice integration
cron-scheduler SQLite-backed slot-claim cron jobs
skill-synthesis Trajectory extraction + judge (learn skills from runs)
agent-sdk Claude/Codex SDK wrapper + compaction
agent-generation Setup-wizard generation pipeline
vscode-lm-tools Code-exec MCP + browser/web capabilities

UI Components (Notable)

  • Canvas: Multi-tile orchestra (gridstack, 9-tile cap) for parallel agent visualization
  • Chat execution tree: Real-time visualization of agent task decomposition
  • Setup wizard: 7-step premium-gated workspace analysis onboarding
  • Harness builder: Streamed agent harness configuration
  • Thoth shell: 4-tab inner chrome for Electron (Memory/Skills/Cron/Gateway)
  • Performance dashboard: Cost + tokens + sessions + agent analytics

Agent Communication

External orchestrators reach Ptah via session_submit MCP tool — delegates whole tasks to Team Leader harness without bespoke integration.

Plugin System

Skill plugins downloadable at runtime from GitHub (via ContentDownloadService). Available for Angular, React, NestJS, NX workspaces.

05

Prompts

Ptah (Coding Orchestra) — Prompts & Agent Instructions

Agent Pack Manifest Structure

Agent configurations live in agent-pack-manifest.json (runtime registry) and in per-library CLAUDE.md files scattered across the Nx monorepo. The Setup Wizard (agent-generation lib) generates project-specific agent configs at workspace onboarding time.

System Prompt Pattern

Each of the 13 agent personas has a role description that serves as its system prompt seed:

Agent Role Description (from manifest)
Team Leader "Task Decomposition and Batch Orchestration Specialist"
Software Architect "Elite Software Architect for system design"
Project Manager "Technical Lead for task orchestration"
Backend Developer "Server-side architecture and API design"
Frontend Developer "Modern UI frameworks and component design"
Code Logic Reviewer "Ensures business logic correctness"
Code Style Reviewer "Coding standards, patterns, best practices"
Senior Tester "Comprehensive quality assurance"
Research Expert "Deep technical analysis and strategic insights"
DevOps Engineer "CI/CD, containerization, deployment automation"
Modernization Detector "Identifies technology modernization opportunities"
UI/UX Designer "Visual design systems and production-ready specs"
Technical Content Writer "Technical documentation and content"

Module-Scoped CLAUDE.md Files

The monorepo contains 50+ CLAUDE.md files — one per lib/app — each providing agent guidance scoped to that module. This is Ptah's primary mechanism for injecting codebase-specific context into agent prompts. Each CLAUDE.md covers:

  • Module purpose and hexagonal architecture role (port vs adapter)
  • Key abstractions and naming conventions
  • Which other libs to import from (dependency rules)
  • Forbidden patterns (e.g., no business logic in adapters)
  • Testing conventions for that module

Task Spec Files (.ptah/specs/)

TASK-prefixed spec files (e.g., TASK_2025_294, TASK_2026_127) serve as structured prompts for the orchestration pipeline. Each spec contains:

  • Objective and motivation
  • Implementation requirements
  • Acceptance criteria
  • Status tracking (active/completed/planned)

These function as the persistent task memory that survives between agent sessions.

Setup Wizard Agent Generation

The agent-generation lib runs a pipeline during Setup Wizard that:

  1. Scans workspace (workspace-intelligence lib) to detect project type
  2. Feeds detected frameworks/deps into a generation prompt
  3. Produces project-specific agent configuration (system prompt variants)
  4. Writes generated configs to .ptah/ directory

VS Code Marketplace Constraint

A notable prompt-engineering constraint: VS Code Scanner rejects extensions containing trademarked AI product names (copilot, codex, claude, openai, anthropic) in non-JS files. Plugin/template content containing these terms is downloaded at runtime from GitHub via ContentDownloadService rather than bundled in the VSIX.

09

Uniqueness

Ptah (Coding Orchestra) — Uniqueness & Differentiation

Compared to Seeds

vs. claude-flow (closest architectural peer)

claude-flow is a pure CLI/MCP tool with a SPARC methodology. Ptah ships the same MCP-anchored multi-agent concept but packages it as a VS Code extension + Electron desktop app + headless CLI, adds a visual Canvas for parallel agent monitoring, and includes commercial infrastructure (license server, freemium model, Paddle payments, WorkOS auth).

vs. taskmaster-ai

taskmaster-ai focuses on task decomposition as a standalone service. Ptah embeds task decomposition (Team Leader agent) as one component of a larger orchestration runtime with 13 agent personas, workspace intelligence, and memory persistence.

vs. kiro

Kiro uses Spec/Hook/Steering artifacts as its core primitive. Ptah uses TASK-prefixed spec files similarly but builds on top of them a live execution runtime with parallel agents, not just a planning scaffold.

vs. superpowers

superpowers is a skills-only behavioral framework with no runtime. Ptah ships a complete runtime with its own agent loop, memory, cron, messaging gateway, and visual UI.

vs. bmad-method

bmad-method defines agent personas in markdown. Ptah takes the same persona-definition idea but backs it with a live multi-agent executor, tool infrastructure, and a learning system (skill-synthesis).

Distinctive Features

50+ Module-Scoped CLAUDE.md Files

The densest CLAUDE.md usage in the entire catalog: every lib and app in the Nx monorepo has its own agent guidance file. This creates a self-describing codebase where agents working on any module receive highly relevant context without needing to scan the entire repo.

FSL-1.1-MIT Time-Delay Open Source

A novel licensing model: each release becomes full MIT exactly two years after publication date. This enables commercial development and monetization while guaranteeing the community eventually gets full open-source access. No other framework in this batch uses this license.

Skill Synthesis with Quality Judge

Unlike simple session memory, Ptah's skill-synthesis lib extracts trajectories from agent runs and evaluates them with a quality judge before storing. Only high-quality patterns become retrievable skills, preventing the memory store from degrading with low-quality examples.

Workspace Intelligence for 13+ Project Types

The workspace-intelligence lib performs deep static analysis (AST + symbol indexer) across Angular, React, NestJS, NX, and 10+ other project types. This feeds the Setup Wizard's agent configuration generation — the generated prompts are workspace-aware from day one.

AI-Native Dogfood at Scale

Ptah is built using Ptah. The .ptah/specs/ registry contains hundreds of TASK entries documenting the tool's own development. This is not just a stated philosophy but demonstrably practiced across the entire Nx monorepo.

VS Code Marketplace Compliance Workaround

The constraint that VS Code Scanner rejects trademarked AI names in non-JS files forced an interesting architecture: all AI provider-related plugin/template content is downloaded at runtime from GitHub (ContentDownloadService) rather than bundled. This decouples the VSIX release cycle from AI provider content updates.

Letta-Style Memory in a VS Code Extension

Bringing pgvector-class memory (sqlite-vec) and Letta-inspired IndexingControl into a VS Code extension context is architecturally unusual. Most extensions treat memory as ephemeral; Ptah treats it as a first-class durable store.

Access Topology (Batch Context)

Unlike other frameworks in this batch where the agent runs elsewhere (CCBot: agent runs in tmux on developer's machine; OnePilot: agent runs on SSH remote server; IM.codes: agent runs on cloud daemon), Ptah runs the agent loop inside the VS Code extension host process or Electron app process on the developer's local machine. The MCP server exposes local workspace tools outward to external orchestrators rather than the agent being remotely operated.

04

Workflow

Ptah (Coding Orchestra) — Workflow

Setup Workflow

  1. Install extension from VS Code Marketplace or Electron app
  2. Sign up at ptah.live (free 100-day Pro trial)
  3. Configure AI provider in sidebar settings
  4. Run Setup Wizard (6-step automated workspace scanning)
  5. Wizard detects: project type, frameworks, dependencies across 13+ project types
  6. Wizard generates: project-specific agent configurations and rules

Task Orchestration Workflow (Normal)

  1. Open Ptah chat → type request or use /orchestrate
  2. Select workflow type (Feature/Bugfix/Refactoring/Documentation/Research/DevOps/Creative)
  3. Team Leader agent decomposes task → assigns to specialized agents
  4. Specialized agents (Backend Developer, Frontend Developer, etc.) execute subtasks in parallel
  5. Approval checkpoints: user reviews at each stage
  6. Code Logic Reviewer + Code Style Reviewer + Senior Tester run QA
  7. Results aggregated by Team Leader
  8. Final implementation delivered

External Orchestration via MCP

External MCP orchestrator
    ↓ session_submit tool
Team Leader harness (inside Ptah)
    ↓ delegates
Backend Developer / Frontend Developer / etc.
    ↓ results
Team Leader → back to external orchestrator

Memory-Augmented Workflow

  1. Each agent run extracts trajectory (skill-synthesis lib)
  2. Trajectory judged for quality
  3. High-quality patterns stored in memory (sqlite-vec)
  4. Future runs retrieve relevant past patterns via vector search

Phase → Artifact Map

Phase Artifact
Setup Wizard Project-specific agent configs + rules
Task decomposition Execution tree (displayed in UI)
Agent execution Code changes + tool calls
QA review Logic + style review comments
Session ~/.ptah/ptah.db rows
Skill synthesis Extracted skill pattern in SQLite

Approval Gates

Gate Type Description
Workflow stage completion file-review User reviews before next stage
Agent spawn authorization yes-no Confirm spawning background agents
06

Memory Context

Ptah (Coding Orchestra) — Memory & Context

Memory Architecture

Ptah implements a Letta-style memory system backed by SQLite with the sqlite-vec extension for vector similarity search.

Storage Layer

Store Location Purpose
Session DB ~/.ptah/ptah.db Sessions, messages, agent runs, costs
Vector index sqlite-vec extension (same DB) Embedding storage for skill retrieval
Settings ~/.ptah/settings.json API keys, model preferences, provider config
Spec files .ptah/specs/ TASK-prefixed persistent task specifications

Memory Curator Lib

The memory-curator lib (backend) implements IndexingControl — a Letta-inspired approach where memory is explicitly managed:

  • Ingestion: agent trajectories fed into memory pipeline
  • Indexing: content chunked and embedded
  • Retrieval: vector similarity search via sqlite-vec at query time
  • UI: memory-curator-ui (frontend lib) exposes Memory tab in Thoth shell

Skill Synthesis

The skill-synthesis lib provides trajectory-based learning:

  1. After each agent run, trajectory (sequence of tool calls + decisions) is extracted
  2. A judge evaluates trajectory quality
  3. High-quality patterns are stored as skill records in SQLite
  4. Future runs retrieve relevant past patterns via vector similarity
  5. UI exposed in skill-synthesis-ui (frontend lib) — Skills tab in Thoth shell

Context Window Management

The agent-sdk lib wraps @anthropic-ai/claude-agent-sdk and implements compaction:

  • Monitors token usage across agent turns
  • Triggers compaction when approaching context limits
  • Preserves task state and key decisions across compaction boundaries

TASK Spec Persistence

The .ptah/specs/ directory acts as long-running task memory:

  • Registry tracks hundreds of TASK entries (TASK_2025_282 through TASK_2026_127+)
  • Each entry records: objective, status (active/completed/planned), completion date
  • Agents read spec files as structured context at task start
  • Prevents re-solving previously solved problems

Multi-Session Continuity

Session 1: Agent runs → trajectory extracted → skill stored (sqlite-vec)
Session 2: New task → vector search retrieves relevant past skills → injected into context

The memory-curator ensures continuity is content-addressable (semantic similarity) rather than session-order-dependent.

Cron-Scheduled Context Refresh

The cron-scheduler lib (SQLite-backed slot-claim system) can schedule periodic memory consolidation or background indexing jobs, enabling the memory store to stay current without blocking interactive sessions.

07

Orchestration

Ptah (Coding Orchestra) — Orchestration

Orchestration Model

Ptah uses a hierarchical multi-agent model: one Team Leader agent decomposes tasks and delegates to specialized agents, which execute in parallel. Results flow back to Team Leader for aggregation.

Task Decomposition

  1. User submits request (chat or /orchestrate command)
  2. User selects workflow type: Feature / Bugfix / Refactoring / Documentation / Research / DevOps / Creative
  3. Team Leader receives task + workflow type → builds execution tree
  4. Execution tree displayed in real-time via chat-execution-tree UI lib
  5. Team Leader assigns subtasks to appropriate specialized agents

Parallel Agent Execution

The Canvas (multi-tile orchestra, gridstack, 9-tile cap) visualizes up to 9 agents running concurrently. Each tile shows:

  • Agent name/persona
  • Current tool call
  • Status (running/waiting/complete)

Agents share no direct state — all communication routes through Team Leader.

Approval Gates

Two gate types enforce user control:

Gate Trigger User Action
Workflow stage completion Before next stage begins File-review checkpoint
Agent spawn authorization Before spawning background agents Yes/no confirmation

QA Pipeline

After development agents complete, three QA agents run in sequence:

  1. Code Logic Reviewer — business logic correctness
  2. Code Style Reviewer — coding standards and patterns
  3. Senior Tester — comprehensive quality assurance

External Orchestration via MCP

Ptah exposes a 14-namespace MCP server (ptah mcp-serve, stdio transport):

External MCP orchestrator
    ↓ session_submit tool
Team Leader harness (inside Ptah)
    ↓ delegates
Backend Developer / Frontend Developer / etc.
    ↓ results
Team Leader → back to external orchestrator

The session_submit tool allows any MCP-compliant host to delegate a whole task to Ptah's Team Leader without bespoke integration. This is the primary cross-tool integration point.

MCP Namespaces (14)

Workspace analysis, LSP diagnostics, AST parsing, file search, dependency graphs, editor state — all exposed via the MCP server. The workspace-intelligence lib backs the workspace/AST/dependency namespaces.

Agent SDK

The agent-sdk lib wraps three AI SDKs:

  • @anthropic-ai/claude-agent-sdk (primary)
  • @github/copilot-sdk
  • @openai/codex-sdk

Provider selection is configurable per-agent via ~/.ptah/settings.json. The SDK wrapper handles compaction, retry, and token accounting uniformly across providers.

Skill Plugins

Downloadable at runtime from GitHub via ContentDownloadService. Available for Angular, React, NestJS, NX workspaces. Plugins extend agent capabilities without requiring a VSIX update.

08

Ui Cli Surface

Ptah (Coding Orchestra) — UI & CLI Surface

Distribution Targets

Target Package Entry
VS Code Extension ptah-extensions.ptah-coding-orchestra VS Code sidebar + chat
Electron Desktop ptah-electron (Electron 40) Native app window
Headless CLI @hive-academy/ptah-cli ptah binary, JSON-RPC stdio

VS Code Extension UI

The extension (ptah-extension-vscode) hosts an Angular 21 webview (ptah-extension-webview). Key panels:

  • Sidebar: Provider settings, model selection, session history
  • Chat: Primary interaction surface with task input and /orchestrate command
  • Execution tree: Real-time task decomposition visualization (chat-execution-tree lib)
  • Canvas: Multi-tile orchestra view (gridstack, 9-tile cap) for parallel agent monitoring

Electron Desktop (Thoth Shell)

The Electron app adds the Thoth Shell — a 4-tab inner chrome exposing backend capabilities:

Tab Lib Content
Memory memory-curator-ui Letta-style memory browser
Skills skill-synthesis-ui Extracted skill patterns
Cron cron-scheduler-ui Scheduled jobs management
Gateway messaging-gateway-ui Telegram/Discord/Slack config

Setup Wizard

7-step premium-gated workspace analysis onboarding:

  1. Project type detection (13+ types)
  2. Framework/dependency scanning
  3. Agent configuration generation
  4. Rules generation
  5. Workspace validation
  6. Configuration review
  7. Activation

Headless CLI (ptah binary)

JSON-RPC stdio interface from @hive-academy/ptah-cli:

  • Suitable for CI/CD pipelines and scripted automation
  • Same agent capabilities as VS Code/Electron without the UI
  • Accepts JSON-RPC requests, returns structured JSON responses

MCP Server

ptah mcp-serve launches a stdio MCP server exposing 14 namespaces. Any MCP-compliant host can connect and use workspace analysis, LSP diagnostics, AST parsing, file search, dependency graphs, and editor state tools.

Messaging Gateway

Configured in Thoth Shell's Gateway tab. Supports:

  • Telegram
  • Discord
  • Slack
  • Voice integration

Allows users to interact with agents and receive notifications through external chat platforms.

Performance Dashboard

Built-in analytics showing:

  • Cost tracking per session/agent
  • Token usage
  • Session history
  • Agent performance metrics

Notable UI Tech

  • Angular 21: Signals, OnPush, zoneless libs throughout
  • Monaco Editor: Code display/editing in webview
  • xterm.js: Terminal emulation
  • gridstack: Tile-based canvas layout
  • Tailwind 3 + daisyui 4: Styling

Related frameworks

same archetype · same primary tool · same memory type

Claude-Flow / Ruflo ★ 55k

Eliminates single-agent context limits and sequential bottlenecks by orchestrating fault-tolerant swarms of specialized AI agents…

Hermes Agent (NousResearch) ★ 168k

Self-improving personal AI agent with closed learning loop, 7 terminal backends, and messaging gateway — not tied to any AI…

OpenCode ★ 165k

Terminal-first AI coding agent with multi-model routing, native desktop app, and a typed .opencode/ configuration system for…

OpenHands ★ 75k

Open-source AI software development platform (open-source Devin alternative) with Docker sandbox isolation, 77.6% SWE-bench…

DeerFlow ★ 70k

Long-horizon superagent that researches, codes, and creates by orchestrating parallel sub-agents with isolated contexts in Docker…

oh-my-openagent (omo) ★ 60k

Multi-provider AI agent orchestration for OpenCode: escape vendor lock-in by routing Sisyphus (Claude/Kimi/GLM) and Hephaestus…