atlasbridge 0.4.0

Universal human-in-the-loop control plane for AI developer agents — forward prompts to your phone, reply from anywhere

AtlasBridge (AtlasBridge)

Universal human-in-the-loop control plane for AI developer agents.

---

AtlasBridge sits between you and your AI coding agent. Whenever your agent pauses and requires human input — approval, confirmation, a choice, or clarification — AtlasBridge forwards that prompt to your phone.

You respond from your phone, from a channel such as Telegram or WhatsApp, Slack or others. AtlasBridge relays your decision back to the CLI. Execution resumes.

No walking back to your desk. No missed prompts. You stay in control.

┌──────────────┐        ┌───────────────┐        ┌─────────────────┐
│  AI Agent    │──────► │     AtlasBridge     │──────► │   Your Phone    │
│ (Claude CLI) │        │  Prompt Relay │        │   (Telegram)    │
│              │◄────── │               │◄────── │                 │
└──────────────┘        └───────────────┘        └─────────────────┘
   paused waiting           detects &                you reply
   for input                forwards prompt          from anywhere

---

How it works

1. atlasbridge run claude — wraps your AI CLI in a PTY supervisor
2. The tri-signal prompt detector watches the output stream
3. When a prompt is detected, AtlasBridge sends it to your Telegram (or Slack)
4. You tap a button or send a reply on your phone
5. AtlasBridge injects your answer into the CLI's stdin
6. The agent continues

That's it. AtlasBridge is a relay, not a firewall. It does not interpret commands, score risks, or block actions. It asks you — and only you — at the exact moment the agent needs human input.

---

Status

Version	Status	Description
v0.1.0	Design release	Architecture, docs, and code stubs
v0.2.0	In progress	macOS MVP — working Telegram relay
v0.3.0	Planned	Linux support
v0.4.0	Planned	Slack channel
v0.5.0	Planned	Windows (ConPTY, experimental)

---

Quick start (v0.2.0 target)

pip install atlasbridge

# First-time setup (creates ~/.atlasbridge/config.toml)
aegis setup

# Wrap Claude Code
aegis run claude

# In another terminal, check status
aegis status

When Claude Code asks you a question, your Telegram bot will send you a message with buttons. Tap Yes or No. Claude Code answers and continues.

---

Design

See the docs/ directory:

Document	What it covers
architecture.md	System diagram, component overview, sequence diagrams
reliability.md	PTY supervisor, tri-signal detector, Prompt Lab
adapters.md	BaseAdapter interface, Claude Code adapter
channels.md	BaseChannel interface, Telegram implementation
cli-ux.md	All CLI commands, output formats, exit codes
roadmap-90-days.md	6-phase roadmap, v0.2.0–v0.5.0
qa-top-20-failure-scenarios.md	20 mandatory QA scenarios
dev-workflow-multi-agent.md	Branch model, agent roles, CI pipeline

---

Repository structure

src/aegis/
  core/
    prompt/     — detector, state machine, models
    session/    — session manager and lifecycle
    routing/    — prompt router (events → channel, replies → PTY)
    store/      — SQLite database
    audit/      — append-only audit log with hash chaining
    daemon/     — daemon manager (orchestrates all subsystems)
    scheduler/  — TTL sweeper and periodic tasks
  os/tty/       — PTY supervisors (macOS, Linux, Windows stub)
  adapters/     — CLI tool adapters (Claude Code, OpenAI CLI, custom)
  channels/     — notification channels (Telegram, Slack stub)
  cli/          — Click CLI entry point and subcommands
tests/
  unit/         — pure unit tests (no I/O)
  integration/  — SQLite + mocked HTTP
  e2e/          — real PTY + mocked Telegram
  prompt_lab/   — deterministic QA scenario runner
    scenarios/  — QA-001 through QA-020 scenario implementations
docs/           — design documents (see table above)

---

Core invariants

AtlasBridge guarantees the following regardless of channel, adapter, or concurrency:

1. No duplicate injection — nonce idempotency via atomic SQL guard
2. No expired injection — TTL enforced in the database WHERE clause
3. No cross-session injection — prompt_id + session_id binding checked
4. No unauthorised injection — allowlisted identities only
5. No echo loops — 500ms suppression window after every injection
6. No lost prompts — daemon restart reloads pending prompts from SQLite
7. Bounded memory — rolling 4096-byte buffer, never unbounded growth

---

Development

# Install in editable mode with dev dependencies
pip install -e ".[dev]"

# Run tests
pytest tests/

# Run a Prompt Lab scenario
aegis lab run partial-line-prompt

# Lint and format
ruff check . && ruff format --check .

# Type check
mypy src/aegis/

# Full CI equivalent (local)
ruff check . && ruff format --check . && mypy src/aegis/ && pytest tests/ --cov=aegis

---

Contributing

See CONTRIBUTING.md. All contributions require:

• Existing tests to remain green
• New code to have unit tests
• Prompt Lab scenarios for any PTY/detection changes

---

License

MIT — see LICENSE.