Skip to content

README.md

Latest commit

 

History

History
44 lines (32 loc) · 3.56 KB

README.md

File metadata and controls

44 lines (32 loc) · 3.56 KB

forever-claude-template

A self-contained template for running a persistent Claude agent that communicates via Telegram, delegates work to sub-agents, and can manage its own background services.

Usage

mngr create my-workspace main -t local \
    --host-env MINDS_WORKSPACE_NAME=my-workspace \
    --project ~/project/forever-claude-template \
    --pass-env TELEGRAM_BOT_TOKEN \
    --pass-env TELEGRAM_USER_NAME

Structure

  • CLAUDE.md - Agent instructions
  • parent.toml - Upstream repo for pulling updates
  • .mngr/settings.toml - Agent types, create templates, command defaults
  • skills/ - Agent skills (telegram, task delegation, services, self-update)
  • scripts/ - Utility scripts (reviewer settings)
  • event-processor/ - Pre-configured directory for creating persistent sub-agents
  • supervisord.conf - Supervisord config defining the background services
  • libs/telegram_bot/ - Telegram bot, send CLI, and history viewer
  • libs/bootstrap/ - First-boot setup, then launches supervisord to supervise the services
  • vendor/mngr/ - A vendored, mutable copy of mngr. Note that making changes here will affect the behavior of the mngr command
  • vendor/tk/ - A vendored copy of the tk ticket tracker. The ticket script (also callable as tk) manages tickets stored as markdown. We point TICKETS_DIR at runtime/tickets/ (set in .mngr/settings.toml's host_env) so tickets are backed up alongside the rest of runtime/ on the mindsbackup/$MNGR_AGENT_ID branch.

Create templates

  • worker - For sub-agents created via the launch-task skill (includes code review)
  • subskill-worker - Sub-agent for any flow that hands its worker the generic harden worker (the crystallize / update / heal artifact lifecycle, including the update-system-interface flow). Inherits from worker and pre-installs the single generic worker from .agents/shared/worker/ into its own .agents/skills/ as harden-worker.

Artifact harden lifecycle

The main agent can promote ad-hoc work into reusable artifacts, fix artifacts that fail, and extend artifacts that came up short -- across skills, web services, and the system interface. The user-invokable surface is three generic operation leads (main agent side), each parameterized by the artifact:

  • crystallize-artifact - Create a new artifact (default: a skill reconstructed from the just-finished turn). Invoked directly post-turn, or by the live-half wrappers (build-web-service, fetch-process-show) once a prototype is confirmed.
  • heal-artifact - Fix a skill or service that errored or produced wrong results.
  • update-artifact - Extend / refactor / verify a skill, service, or shared reference; one flow with a committed-vs-emergent design-gate toggle.

Each lead spawns a subskill-worker sub-agent that runs the single generic harden-worker sub-skill. The worker reads the operation and artifact from its task file and composes the universal harden-artifact.md contract with one op-*.md and one artifact-*.md reference under .agents/shared/worker/references/. Workers commit to mngr/<task-name> branches; main merges on user approval. (The same template also backs the update-system-interface flow, which wraps update-artifact with artifact=system-interface and adds its preview / safe-reveal go-live.)

Crystallized skills are marked with metadata.crystallized: true in their SKILL.md frontmatter and follow the agentskills.io layout (scripts/run.py as a PEP 723 script, companion SKILL.md, optional references/ and assets/).