ReidX

Terminal-native personal intelligence and coding CLI with an agent-first runtime.

A real runtime — not a chat wrapper. Sessions, tasks, tools, policy gates, providers, and persistence are first-class. A genuine full-screen TUI (not an inline redraw hack) with a locked-to-bottom footer, full-width transcript, input history, token streaming, drag-to-copy, collapsible reasoning, a live subagent panel, and a live / command menu.

Status: Phase 5 complete, plus TUI rewrite, multi-provider HTTP clients (Anthropic / OpenAI / OpenAI-compatible / Ollama / OpenCode Go), SSE streaming, auto context windows, compact + cost, soft provider errors, DeepReid, Nyx, web search, and workflows. See docs/ for architecture notes.

Target stack

• Python 3.12+
• Typer — CLI command surface
• Pydantic v2 — schemas and validation
• Rich — markdown, tables, panels (width tracks the live terminal)
• prompt_toolkit — full-screen TUI (layout, input, completion, mouse)
• stdlib urllib only for HTTP/SSE (no extra HTTP client package)

Quick start

Option A: install via npm

Requires Python 3.12+ on PATH. The package pip-installs on npm install and seeds ~/.reidcli/settings.json.

npm install -g reidx

Option B: install from source

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -e ".[dev]"

macOS/Linux: source .venv/bin/activate

Verify and run

reid doctor
reid

• Offline: stub provider works without keys; use /connect for a real model.
• Real chat: env keys and/or /connect, then /use <provider>.

The interactive TUI

Full-screen prompt_toolkit app (alternate screen; host scrollback restored on exit). Output fills the live terminal width; the footer stays one compact line.

| Feature | Details | |---|---| | Footer | Mode · model · effort · tokens/context · workspace · cost · Copied! | | Token streaming | OpenAI-compatible SSE; /stream auto (default) / on / off | | Input history | ↑ / ↓ (and Ctrl+P / Ctrl+N); saved under ~/.reidcli/input_history | | Scroll | Mouse wheel; PageUp/PageDown; Shift+↑/↓ line scroll | | Copy | Drag (red highlight), double/triple-click; /copy or Ctrl+Y = last AI reply | | Reasoning | Collapsible CoT (Ctrl+O); splits <think>, GLM <parameter name="reasoning">, fences | | / menu | Commands, enums, /use aliases, /model list, /resume session pick list | | Pastes | Large/multi-line → [Pasted text #N +… lines] | | Escape | Cancel in-flight turn at next safe point | | DeepReid | Type deepread / deep reid at start of input (border pulses green) | | Nyx | /nyx on redteam persona (prompt only) |

Keyboard shortcuts

| Key | Action | |---|---| | ↑ / ↓ | Input history (when not in / completion menu) | | Ctrl+P / Ctrl+N | Same as ↑ / ↓ history | | ← / → | Cycle effort when the box is empty; else move cursor | | Shift+↑ / ↓ | Scroll transcript by line | | PageUp / PageDown | Scroll transcript by page | | Tab | Accept / completion | | Ctrl+O | Expand/collapse reasoning + tool calls | | Ctrl+Y | Copy last AI reply | | Ctrl+C | Copy selection if any, else clear line | | Esc | Stop in-flight response | | Ctrl+D | Exit |

Command surface

Top-level CLI

| Command | Purpose | |---|---| | reid | Interactive TUI (default) | | reid interactive "<prompt>" | TUI + first turn | | reid --file <path> / -f | Initial prompt from file | | reid --nyx | Start with Nyx on | | <cmd> \| reid | Stdin as first turn | | reid exec "<prompt>" | Headless one-shot | | reid deepreid "<task>" | Headless plan pipeline | | reid resume <session-id> | Resume then TUI | | reid sessions | List sessions | | reid config-show / tools / doctor / version | Utility |

Slash commands (highlights)

Session

| Command | Purpose | |---|---| | /status | Session summary | | /sessions | List sessions (with message counts) | | /resume <id> | Resume; type /resume for a pick list; re-prints conversation | | /transcript [n] | Show last n messages | | /rewind | Drop last turn | | /compact [n\|--force] | Summarize older turns | | /cost [reset] | Session spend estimate | | /copy | Last AI reply → clipboard | | /recap / /review <pr> | Recap / PR review helpers |

Config

| Command | Purpose | |---|---| | /model [list\|id] | List or set model; context meter updates for that model | | /effort auto\|low\|medium\|high\|xhigh | Reasoning effort (auto classifies the prompt) | | /stream auto\|on\|off | Token streaming (default auto) | | /mode / /nyx / /permissions / /tools | Policy and tools |

Providers

| Command | Purpose | |---|---| | /providers | List connected + built-in | | /connect | Palette or CLI connect | | /use <name> | Switch provider (aliases: nvidia, opencode, …) | | /models | Models across providers |

Providers

| Kind | Notes | |---|---| | anthropic | Messages API | | openai | Chat Completions + SSE streaming | | openai-compatible | NIM, Groq, xAI, OpenCode Go, vLLM, LM Studio, … + streaming | | ollama | Local Ollama |

OpenCode Go

OpenCode Go — Zen subscription for open coding models.

$env:OPENCODE_API_KEY = "your-zen-key"   # or OPENCODE_GO_API_KEY
# optional: $env:OPENCODE_GO_MODEL = "glm-5.2"
reid
/use opencode
/model list

Or /connect → OpenCode Go → paste key.

| Catalog entry | Kind | Base URL | Example models | |---|---|---|---| | OpenCode Go | openai-compatible | https://opencode.ai/zen/go/v1 | glm-5.2, kimi-k2.7-code, deepseek-v4-flash, mimo-v2.5 | | OpenCode Go (Anthropic) | anthropic | https://opencode.ai/zen/go | qwen3.7-plus, minimax-m2.7 |

Aliases: opencode, opencode-go, zen-go, go.

HTTP clients send a normal User-Agent so Cloudflare does not block stdlib urllib (403 HTML / error 1010).

Context windows

The footer meter (used/max) updates when you change model or provider:

1. Provider /models metadata when present
2. Known model-id table (e.g. GLM-5.2 → 1M, Claude → 200k)
3. Session cache / id size tags / 128k default

/model glm-5.2 should show something like 1.4k/1.0M, not a stale 200k seed.

Streaming

| Mode | Behavior | |---|---| | auto (default) | Stream when the provider supports SSE | | on | Prefer streaming | | off | Wait for full reply |

Tools

| Tool | Risk | Purpose | |---|---|---| | read_file | low | Read file text | | write_file | medium | Create/overwrite | | patch_file | medium | Unique substring replace | | list_dir | low | Directory listing as Name / Type columns | | find_files | low | Glob | | grep_files | low | Regex search | | run_command | high | Shell + policy (PowerShell on Windows; rewrites head/tail) | | set_context_window | low | Agent can fix footer max context (1M, 200k, …) | | list_provider_catalog | low | Search known providers (OpenCode Go, NIM, …) | | list_connected_providers | low | What’s registered in this session | | connect_provider | high | Save API key + register (user approves; like /connect) | | use_provider | low | Switch active provider (like /use) | | set_model | low | Set session model id (like /model) | | disconnect_provider | high | Remove a saved provider | | web_search | high | DuckDuckGo (no API key) | | spawn_agent | medium | Scoped child agent |

Configuration

Precedence (low → high): defaults → ~/.reidcli/config.json → project .reidx/config.json → settings.json reidx block → environment.

Storage (~/.reidcli)

~/.reidcli/
  settings.json          # auto-seeded; empty env keys ignored
  input_history          # ↑/↓ prompt history
  providers.db
  workflows.json
  sessions/<id>/
    meta.json
    transcript.jsonl
    tasks.json
    goals.json
    events.jsonl
    costs.jsonl
  deepreid/<run-id>.md

Override root with REIDX_STORAGE. Legacy .reidx / AppData Reid migrate once.

Environment variables

| Variable | Effect | |---|---| | REIDX_STORAGE | Storage root (default ~/.reidcli) | | REIDX_PROVIDER / REIDX_WORKSPACE / REIDX_PERMISSION_MODE | Defaults | | REIDX_LOG_LEVEL / REIDX_INSECURE / REIDX_COLOR / NO_COLOR | Runtime | | REIDCHAT_SETTINGS | Explicit settings path | | ANTHROPIC_* / OPENAI_* | Auto-register those providers | | OPENCODE_API_KEY or OPENCODE_GO_API_KEY | OpenCode Go auto-register | | OPENCODE_GO_MODEL / OPENCODE_GO_BASE_URL | Go model / base override |

See settings.example.json.

Permission modes

| Mode | Behavior | |---|---| | strict | Approve nearly everything | | balanced | (Default) Medium/high prompt | | autonomous | Low/medium free; high prompts | | custom | Explicit allowlists only |

Path confinement + shell denylist apply in all modes.

Sessions

• Resume: /resume <id> or /resume (completion list). Restores transcript into model context and re-prints the conversation in the TUI.
• Empty sessions (0 messages) are starts that never finished a turn.
• Compact: /compact or auto near the model window limit.
• Cost: per-turn estimates + /cost (from public price table).

Nyx / DeepReid / subagents

• Nyx — redteam persona (--nyx or /nyx on); tools/policy unchanged
• DeepReid — Researcher → Planner → Critic; reid deepreid or deepread in TUI
• spawn_agent — scoped child; live panel under the input box

Headless

reid exec "list the current dir"
reid exec --nyx "recon plan for <authorized target>"

Exit 1 on provider errors. Approver auto-allows in exec mode.

Development

pytest
ruff check src tests

Coverage includes policy, tools, session, reasoning tags (incl. GLM parameter blocks), soft provider errors, aliases, context windows / bind-on-model, compact, cost, effort auto, settings seed, and SSE streaming.

src/reidx/
  app/  config/  diagnostics/
  session/  tasks/  goals/  workflows/
  policy/  provider/  provider_manager/  tools/
  runtime/   # agent, orchestrator, compact, cost, effort_auto, reasoning
  deepreid/  nyx/
  ui/        # TUI, slash commands, terminal_host

What works now

• Full-width TUI: live terminal width, compact footer, input history (↑/↓)
• Streaming (/stream auto), drag-copy, /copy / Ctrl+Y
• OpenCode Go + NIM + Anthropic/OpenAI/Ollama; Cloudflare-safe User-Agent
• Auto context windows on /model / /use (GLM-5.2 → 1M, etc.)
• Soft HTTP errors; provider aliases; settings seed under ~/.reidcli
• /resume pick list + visible transcript restore
• list_dir Name/Type listings; compact/cost/effort auto
• DeepReid, Nyx, workflows, goals, headless exec

What is stubbed

• MCP full stdio/JSON-RPC
• Structured patch / diff preview
• Anthropic native event-stream (OpenAI-compatible stream is done)
• DeepReid Builder role
• Escape does not cancel DeepReid stages

License

MIT