privateer-agent 0.2.1 → 0.3.0

package/README.md

@@ -5,20 +5,18 @@
 <h1 align="center">⚓ Privateer</h1>
 
 <p align="center">
-  <strong>A provider-agnostic terminal coding agent — bring your own model.</strong>
+  <strong>A provider-agnostic terminal coding agent — bring your own model, keep your own privacy.</strong>
 </p>
 
 <p align="center">
-  <a href="https://github.com/privateer-agent/privateer-agent/actions/workflows/ci.yml">
-    <img src="https://github.com/privateer-agent/privateer-agent/actions/workflows/ci.yml/badge.svg" alt="CI" />
-  </a>
   <a href="https://www.npmjs.com/package/privateer-agent">
     <img src="https://img.shields.io/npm/v/privateer-agent" alt="npm" />
   </a>
-  <img src="https://img.shields.io/badge/node-%E2%89%A520-brightgreen" alt="Node >= 20" />
+  <img src="https://img.shields.io/badge/node-%E2%89%A522.19-brightgreen" alt="Node >= 22.19" />
   <img src="https://img.shields.io/badge/license-MIT-blue" alt="MIT License" />
-  <img src="https://img.shields.io/badge/providers-OpenRouter%20·%20Anthropic%20·%20OpenAI%20·%20Ollama%20·%20NEAR%20AI-5b8def" alt="Providers" />
-  <img src="https://img.shields.io/badge/built%20on-Vercel%20AI%20SDK-black" alt="Vercel AI SDK" />
+  <img src="https://img.shields.io/badge/permissions-safe%20by%20default-2ea44f" alt="Safe by default" />
+  <img src="https://img.shields.io/badge/inference-TEE%20attested%20·%20on--device%20PII%20gate-5b8def" alt="Private inference" />
+  <img src="https://img.shields.io/badge/providers-OpenRouter%20·%20Anthropic%20·%20OpenAI%20·%20Google%20·%20xAI%20·%20Groq%20·%20Mistral%20·%20Z.ai%20·%20DeepSeek%20·%20Qwen%20·%20Ollama%20·%20NEAR%20AI%20·%20Tinfoil%20·%20Venice-5b8def" alt="Providers" />
 </p>
 
 ```bash
@@ -26,75 +24,69 @@ curl -fsSL https://privateer.pro/install.sh | sh    # installs the `privateer` c
 npx privateer-agent                                 # or run it instantly, nothing installed
 ```
 
-Switch between **OpenRouter**, **Anthropic**, **OpenAI**, local **Ollama**, and **NEAR AI**
-(private, attestable inference) with one command. Built on the Vercel AI SDK, so tool-calling
-and streaming work identically across every provider — no model lock-in, no separate code paths.
-MCP servers, Claude Code-compatible skills, scheduled routines, and approval from your phone
-included.
-
-<p align="center">
-  <img src="docs/screenshot.png" alt="Privateer running in the terminal" width="820" />
-</p>
+Point it at a frontier model today and a local Ollama model tomorrow — **OpenRouter**,
+**Anthropic**, **OpenAI**, **Google**, **xAI**, **Groq**, **Mistral**, **Z.ai** (GLM),
+**DeepSeek**, **Qwen**, local **Ollama**, **NEAR AI** or **Tinfoil** (verifiable TEE
+inference), **Venice** / **Fireworks** (no-retention inference), and any **custom
+OpenAI-compatible endpoint** (LM Studio, vLLM, llama.cpp…) are interchangeable at
+`/model` time, including mid-session. No model lock-in, no separate code paths. MCP
+servers, sub-agents, scheduled routines, and one-tap approval from your phone are
+included — and every one of the agent's actions runs through a **safe-by-default
+permission gate**.
 
 ## Why Privateer?
 
-- **No lock-in.** Point it at a frontier model today and a local Ollama model tomorrow —
-  `/model` swaps mid-session. Your config, commands, and agents come along for the ride.
-- **No API key required.** Bring your own key (BYOK) from any supported provider, run
-  keyless against a local Ollama — or `/login` to bill a Privateer account instead.
-- **The agent UX you already know.** Plan mode, checkpoint/rewind, a modal prompt, slash
-  commands, sub-agents, and project memory — but vendor-neutral.
-- **Genuinely extensible.** MCP servers, lifecycle hooks, custom commands, output styles,
-  sub-agents, and skills are all just files under `.privateer/`. No plugins to compile.
-- **Zero binary deps.** The file/search/shell tools are pure Node — nothing to install
-  beyond `node`.
+- **No lock-in.** One agent, every provider. `/model` swaps mid-session and your config,
+  commands, and agents come along for the ride. No vendor's models are privileged.
+- **No API key required.** Bring your own key from any supported provider, run keyless
+  against a local Ollama — or `/signin` to bill a Privateer account instead.
+- **Safe by default.** Every edit, shell command, and network call is classified and gated
+  before it runs; destructive commands are blocked even in unattended runs. You stay in
+  control, whether you're watching or not.
+- **Privacy you can verify, not just trust.** Confidential-enclave (TEE) inference is
+  cryptographically **attested** — not a policy promise — and an **on-device PII gate**
+  warns before structured personal data ever leaves your machine for an unverified model.
+- **Genuinely extensible.** MCP servers, sub-agents, custom commands, and skills — all just
+  files. Nothing to compile.
 
 ## Highlights
 
-- **MCP servers** (local stdio + remote HTTP/SSE, with interactive OAuth), lifecycle
-  **hooks**, and **custom sub-agents**
-- **Claude Code-compatible skills** — published Agent Skills drop in unchanged; install
-  from GitHub with `/skills install owner/repo`
-- **Scheduled routines** — a daemon runs approved tasks unattended, cron or one-off
-- **Approve it from your phone** — link the terminal to the Privateer app with
-  `/remote-access` (off by default) and Allow/Deny every action remotely
-- **Zero-Data-Retention surfacing** for OpenRouter: a status-bar shield colors the selected
-  model's retention posture, and `/zdr` pins routing to zero-retention endpoints
-- **Private, verifiable inference** via NEAR AI: every model runs in a TEE, a `⛉ TEE` status
-  shield reflects the live attestation, and `/verify` fetches the attestation report (validate
-  the raw quote chains with the NEAR Cloud Verifier for full cryptographic proof)
-- **Plan mode** (read-only → present a plan → approve), **checkpoint/rewind** of
-  conversation and files
-- A modal prompt with `/` command and `@` file autocomplete, `!` shell passthrough,
-  `#` memory append, input history, optional **vim** mode, and **ctrl-r** history search
-- Layered `settings.json` (user → project → local → managed), **custom slash commands**
-  and **output styles** as markdown files
-- Background shells, bounded parallel sub-agents, thinking display, structured compaction,
-  and image attachment for vision-capable models
+- **Private, verifiable inference** via **NEAR AI** and **Tinfoil**: every model runs inside
+  a Trusted Execution Environment, a live status shield reflects the attestation, and
+  `/verify` fetches and checks the cryptographic report on demand — genuine proof the
+  inference ran on real confidential hardware, not a terms-of-service page.
+- **On-device PII gate.** Before a prompt goes to an *unverified* channel, Privateer scans it
+  locally for structured personal data (emails, phone numbers, SSNs, cards, IBANs, IPs…) and
+  offers to redact or hold it — detection never leaves your machine, and an attested TEE
+  channel skips the check because it provably can't read your data anyway.
+- **Honest privacy posture, graded.** A verified TEE and a "we promise not to retain"
+  policy are **never rendered the same** — the badge tells you exactly how strong the
+  guarantee is (cryptographically verified → observable → policy → none).
+- **Approve it from your phone.** Link the terminal to the Privateer app with
+  `/remote-access` (off by default) and Allow/Deny every action remotely while execution
+  stays on your machine — supervise long agent runs from anywhere.
+- **Scheduled routines.** A background daemon runs approved tasks unattended — cron or
+  one-off — and the agent can schedule its own follow-up work. Results deliver to a file,
+  the next session, your phone, email, or a webhook.
+- **MCP servers, sub-agents & skills.** Connect Model Context Protocol servers (local stdio
+  or remote HTTP with OAuth), delegate work to bounded parallel sub-agents, and drop in
+  skills — all gated like everything else.
+- **Zero-Data-Retention surfacing** for OpenRouter — see the selected model's retention
+  posture before you send, and pin routing to zero-retention endpoints.
+- **Plan mode**, checkpoint/rewind, session branching, a modal prompt with `/` command and
+  `@` file autocomplete, `!` shell passthrough, background shells, and image attachment for
+  vision-capable models.
 
 ## Quickstart
 
 ```bash
 curl -fsSL https://privateer.pro/install.sh | sh    # or: npm install -g privateer-agent
-export OPENROUTER_API_KEY=sk-or-...                 # one provider is enough — or skip and run /login
-privateer                                           # launches the interactive TUI
+export OPENROUTER_API_KEY=sk-or-...                 # one provider is enough — or skip and /signin
+privateer                                           # launches the interactive agent
 ```
 
 First run walks you through picking a provider and default model. From there, just type.
-(No install at all: `npx privateer-agent`.)
-
-## Contents
-
-- [Requirements](#requirements) · [Install](#install) · [Configure a provider](#configure-a-provider) · [Model routing](#model-routing) · [Data retention (ZDR)](#data-retention-zdr) · [Private inference (NEAR AI)](#private-inference-near-ai) · [Privateer account](#privateer-account-billed-inference--what-it-sees) · [Usage](#usage)
-- [The prompt](#the-prompt) · [Slash commands](#slash-commands) · [Tools](#tools)
-- [Customize & extend](#customize--extend) · [Permission modes](#permission-modes) · [Project context](#project-context)
-- [How it compares](#how-it-compares) · [Develop](#develop) · [Caveats](#caveats) · [Docs](#docs) · [License](#license)
-
-## Requirements
-
-- macOS or Linux
-- Node.js ≥ 20 (pure Node, zero binary dependencies)
-- An API key for at least one provider — or a local Ollama install, or a Privateer account (`/login`)
+No install at all: `npx privateer-agent`.
 
 ## Install
 
@@ -102,348 +94,95 @@ First run walks you through picking a provider and default model. From there, ju
 npm install -g privateer-agent          # installs the `privateer` command
 # or run it without installing:
 npx privateer-agent
-# or the one-liner installer (verifies Node >= 20, then installs globally):
+# or the one-liner installer (verifies Node, then installs globally):
 curl -fsSL https://privateer.pro/install.sh | sh
 ```
 
-**From source** (for hacking on Privateer):
+**Requirements:** macOS or Linux, Node.js ≥ 22.19.0.
+
+**From source:**
 
 ```bash
 git clone https://github.com/privateer-agent/privateer-agent.git
 cd privateer-agent
-npm install        # install dependencies
-npm start          # launches the interactive TUI (or: node bin/privateer.mjs)
-npm link           # optional: put your local `privateer` on PATH
+npm install
+npm start            # launches the interactive agent
 ```
 
 ## Configure a provider
 
-Privateer reads credentials from environment variables or a config file.
-
-**Env vars (quickest):**
+Privateer reads credentials from environment variables (or sign in to an account and skip
+keys entirely). One provider is enough to start:
 
 ```bash
 export OPENROUTER_API_KEY=sk-or-...      # gateway to ~everything
 export ANTHROPIC_API_KEY=sk-ant-...
 export OPENAI_API_KEY=sk-...
-export OLLAMA_BASE_URL=http://localhost:11434/api   # optional; defaults to this
-export NEAR_AI_API_KEY=...               # private TEE inference (cloud.near.ai)
+export GEMINI_API_KEY=AIza...            # Google
+export XAI_API_KEY=xai-...               # xAI (Grok)
+export GROQ_API_KEY=gsk_...              # Groq
+export DEEPSEEK_API_KEY=sk-...           # DeepSeek
+export OLLAMA_BASE_URL=http://localhost:11434/v1   # local, keyless
+export NEAR_AI_API_KEY=...               # verifiable TEE inference (cloud.near.ai)
+export TINFOIL_API_KEY=...               # verifiable TEE inference (tinfoil.sh)
+export VENICE_API_KEY=vapi_...           # no-retention inference
 ```
 
-**Config file** — `~/.privateer/config.json` (global) and/or `./.privateer/config.json` (per project):
-
-```json
-{
-  "defaultModel": "openrouter:anthropic/claude-opus-4.8",
-  "permissionMode": "default",
-  "providers": {
-    "openrouter": { "apiKey": "sk-or-..." },
-    "anthropic":  { "apiKey": "sk-ant-..." }
-  }
-}
-```
+Pick a model with **`/model`** (browse each configured provider's live catalog) or pass one
+directly as `provider/model` — e.g. `openrouter/anthropic/claude-opus-4.8`,
+`ollama/qwen3-coder`, `nearai/zai-org/GLM-5.1-FP8`. Any OpenAI-compatible server (LM Studio,
+vLLM, llama.cpp) works as a custom provider — just give it a base URL.
 
 Override the config location with `PRIVATEER_HOME`.
 
-## Model routing
-
-`defaultModel` handles most turns, but it's often the wrong tool for a particular one —
-it may not accept the file you dropped in, or you'd rather spend a cheaper model on a
-trivial question. The optional **`router`** block lets Privateer switch models per turn
-based on the turn's **data type** and shape:
-
-```json
-{
-  "defaultModel": "openrouter:minimax/minimax-m3",
-  "router": {
-    "vision":   "openrouter:google/gemini-2.5-flash",
-    "document": "openrouter:anthropic/claude-opus-4.8",
-    "audio":    "openrouter:google/gemini-2.5-flash",
-    "video":    "openrouter:google/gemini-2.5-flash",
-    "long":     "openrouter:anthropic/claude-opus-4.8",
-    "fast":     "openrouter:openai/gpt-4o-mini",
-    "longThreshold": 60000,
-    "fastMaxChars": 280,
-    "inlineTextMaxBytes": 65536,
-    "auto": true
-  }
-}
-```
-
-Reference a file in the prompt — drag-drop, paste a path, or `@`-mention — and Privateer
-classifies it by **modality** and routes accordingly:
-
-| Route | Chosen when the turn (or conversation) includes… |
-|---|---|
-| **vision** | an image (`.png .jpg .jpeg .gif .webp`) |
-| **document** | a PDF (`.pdf`) |
-| **audio** | audio (`.mp3 .wav .m4a .ogg .flac`) |
-| **video** | video (`.mp4 .mov .webm .mkv`) |
-| **long** | the estimated context exceeds `longThreshold` tokens (default: half of `contextBudget`) |
-| **fast** | the prompt is ≤ `fastMaxChars` characters (and needs no attachment) |
-| **default** | everything else (`defaultModel`) |
-
-Each attached file collapses to a chip — `[Image #1]`, `[PDF #2]`, `[Audio #3]`,
-`[Video #4]` — while the file itself rides along to the model. **Code/CSV/markdown and
-other text files aren't routed**: they're read and inlined into the prompt (up to
-`inlineTextMaxBytes`; larger ones are left as a path for the agent's read tool).
-
-**Capability requirements outrank `long`/`fast`.** A turn that needs a modality is
-routed to a model that can actually accept it — and routing is **sticky**: once a file
-is in the conversation, later turns stay on a capable model so the attachment is never
-replayed to one that can't read it. A turn that needs **several** modalities at once
-(say an image *and* a PDF) is routed to a model whose support covers all of them. When
-a turn is routed, the transcript shows a line like `↪ routed to gemini-2.5-flash ·
-image input`.
-
-**Hybrid auto-detect** (`"auto": true`, the default): if you reference, say, a PDF but
-haven't set `router.document`, and your `defaultModel` can't read PDFs, Privateer
-auto-selects a capable model from a configured provider. Set the route explicitly to
-control exactly which model is used, or `"auto": false` to disable it (you'll get a
-one-line warning when nothing can handle the modality).
-
-## Data retention (ZDR)
-
-When you route through **OpenRouter**, where your prompts end up depends on which upstream
-endpoint serves the request — some retain data, some don't. Privateer surfaces that for the
-**selected model** so you can see the posture before you send, and optionally enforce it.
-
-**The status-bar shield.** A `⛉ ZDR` badge sits in the status line, colored against the
-model you have selected:
-
-| Badge | Meaning |
-|---|---|
-| 🟢 `⛉ ZDR` | The model has a zero-retention endpoint **and** enforcement is on — the request is pinned to it, so prompts can't be retained. |
-| 🟡 `⛉ ZDR` | A zero-retention endpoint exists, but enforcement is off — a request *may* still land on an endpoint that retains prompts. |
-| 🔴 `⛉ ZDR` | No zero-retention endpoint for this model, or it's blocked by your account's privacy settings — under enforcement the request is rejected outright. |
-| `⛉ ZDR?` (dim) | Posture unknown — no OpenRouter key yet, still loading, or the lookup failed. |
-
-The badge only appears for OpenRouter models; other providers show nothing. The posture is
-derived from two authenticated OpenRouter endpoints — `/endpoints/zdr` (models with at least
-one zero-retention endpoint) and `/models/user` (models your account's privacy settings
-actually permit) — fetched once per account and re-evaluated synchronously as you switch
-models. The same colors annotate every row in the `/model` picker, with a legend explaining
-them.
-
-**Enforcement.** Run **`/zdr`** to toggle enforcement (persisted as
-`providers.openrouter.enforceZdr`). With it on, Privateer pins routing to zero-retention
-endpoints (`provider.zdr` on every request), so yellow models go green — and any model
-*without* a zero-retention endpoint is rejected rather than silently retained. Toggle it off
-to let OpenRouter route freely. Enforcement applies to OpenRouter only; add an OpenRouter key
-with `/login` first.
-
-## Private inference (NEAR AI)
-
-**NEAR AI Cloud** runs every model inside a **Trusted Execution Environment** — an Intel TDX
-confidential VM paired with an NVIDIA confidential-computing GPU. Your prompts are encrypted
-all the way into the enclave (TLS terminates *inside* the TEE, not at a load balancer), so
-the model's inputs, weights, and outputs are invisible to the infrastructure provider, the
-model provider, and NEAR itself. And it's not "trust us": each request can produce a
-**cryptographic attestation** attesting that the inference happened on genuine TEE hardware,
-signed by a key that never leaves the enclave and bound to a nonce you supply. (Privateer's
-`/verify` does a pragmatic check of that report; full validation of the quote chains is done
-with the NEAR Cloud Verifier — see `/verify` below.)
-
-It's a drop-in OpenAI-compatible provider — pick a `nearai:*` model with `/model` (e.g.
-`nearai:zai-org/GLM-5.1-FP8`) and everything else works as usual.
-
-**The status-bar shield.** A `⛉ TEE` badge appears whenever a NEAR AI model is selected,
-colored by the live attestation for that model:
-
-| Badge | Meaning |
-|---|---|
-| 🟢 `⛉ TEE` | A fresh attestation came back bound to our nonce, with a TEE signing key and NVIDIA + Intel hardware evidence — confidential **and** verifiable. |
-| 🟡 `⛉ TEE` | A report returned but couldn't be fully confirmed here (missing signing key, hardware marker, or nonce echo). |
-| 🔴 `⛉ TEE` | No attestation material returned. |
-| `⛉ TEE?` (dim) | Unknown — no NEAR AI key yet, still loading, or the lookup failed. |
-
-**`/verify`.** Run it on a NEAR AI model to fetch the attestation on demand and print the
-verdict, detected hardware, the enclave's signing address, and the nonce. Privateer does a
-pragmatic freshness + presence check suited to a terminal; for full validation of the raw
-NVIDIA/Intel quote chains, take the printed report to the
-[NEAR AI Cloud Verifier](https://github.com/nearai/cloud-verifier).
-
-## Privateer account (billed inference) — what it sees
-
-Instead of bringing your own provider key, run **`/login`** to sign into a Privateer account
-(an app-brokered device flow — you approve a short code in the Privateer app/web, so wallet
-and email accounts work identically and no password or wallet key ever touches the terminal).
-Inference then runs on Privateer's server and is billed to your subscription.
-
-What that means for your data, precisely:
-
-- **The server proxies your prompts; it does not store them.** Like the Privateer apps, the
-  account path sends your prompt to the server, which forwards it to the model provider and
-  streams the reply back. The only thing written server-side is **billing metadata** — model
-  id, token counts, cost — never prompt or response text, and nothing is logged in plaintext.
-- **Same privacy guarantee as the apps: ZDR / TEE, not in-transit E2EE.** The server has to
-  read your prompt to run inference (true of every product, every path). The guarantee isn't
-  "nobody sees it" — it's "nobody *retains* it": OpenRouter routes are pinned to zero-retention
-  endpoints, and the account default is a **NEAR TEE** model where even the provider can't read
-  the prompt. Run **`/verify`** to check the live attestation — a pragmatic freshness + presence
-  check (signing key + hardware markers + your nonce echoed); for full cryptographic validation
-  of the raw quote chains, take the printed report to the NEAR AI Cloud Verifier (see above).
-- **Your local transcript is plaintext on your machine.** Privateer's end-to-end encryption
-  protects data **at rest in Privateer's storage** — it does not (and cannot) encrypt the
-  conversation files this CLI keeps on your own disk under `~/.privateer/`. Treat them like any
-  local shell history.
-- **Your session token is stored unencrypted on disk** at `~/.privateer/credentials.json`,
-  protected only by file permissions (`0600` — readable just by your user). It's a scoped
-  session — not your password or wallet/encryption keys — and it's long-lived: anyone who can
-  read that file (root, a backup, a compromised account) has a usable billed session until it's
-  revoked. It rotates on refresh with server-side reuse detection; revoke it any time from the
-  app (**Settings → Linked terminals**, which now lists individual terminals) or with **`/logout`**.
-- **`/remote-access` streams this terminal's activity to your phone.** When you turn it on, the
-  app can drive this terminal: prompts come down, and the agent's replies **and tool input/output**
-  go up through the Privateer server relay so you can watch and approve actions remotely. Output is
-  size-truncated and run through a best-effort secret redactor before it leaves, but that's a safety
-  net, not a guarantee — terminal output can contain whatever a command prints. The relay is
-  live-only (nothing is archived) and carries no keys; the terminal label sent is a non-PII random
-  tag (no username/host/path). It's **off** until you run `/remote-access on`.
-
-> **Only approve a code you generated yourself.** The login code authorizes *this* terminal to
-> spend on your account. If someone sends you a code and asks you to approve it ("paste this to
-> activate…"), **don't** — approving it hands *them* a billed session on *your* account. A code
-> you didn't just create in your own terminal is an attack, not a convenience.
-
-## Usage
-
-```bash
-privateer                                   # interactive TUI with the default model
-privateer -m openrouter:anthropic/claude-opus-4.8
-privateer -c                                # resume the last session in this dir
-privateer -p "summarize src/"               # headless one-shot, prints to stdout
-```
-
-Run `/model` (no argument) to browse the models each configured provider actually
-offers — the list is fetched live using the API key you entered, then filtered as you
-type. Onboarding ends on the same picker so you choose your default model up front.
-
-You can also pass a model string directly as `provider:model`:
-
-| Example | |
-|---|---|
-| `openrouter:anthropic/claude-opus-4.8` | any model on OpenRouter |
-| `anthropic:claude-opus-4-8` | direct Anthropic |
-| `openai:gpt-5.5` | direct OpenAI |
-| `ollama:qwen3-coder` | local model |
-
-## The prompt
-
-The input is modal — the first character chooses what happens:
-
-| Prefix | Mode |
-|---|---|
-| _(text)_ | a normal prompt to the model |
-| `/` | a slash command — opens an autocomplete menu |
-| `@` | a file mention — fuzzy-completes paths from the cwd |
-| `!` | run a shell command locally and show its output (no model turn) |
-| `#` | append the rest of the line to `PRIVATEER.md` |
-
-Also: **↑/↓** history, **ctrl-r** reverse history search, emacs line editing
-(`ctrl-a/e/u/w`), `ctrl-l` to clear the screen, and **`\`+Enter** for a newline. Messages
-typed while the agent is busy are queued and run in order. `/vim` toggles modal (vim)
-editing. Reference an image file to attach it for vision-capable models — by `@`-mention
-(`@screenshot.png`), or by pasting a path anywhere in the prompt. Absolute paths and paths
-with spaces work too, quoted (`"/Users/me/My Shot.png"`) or backslash-escaped
-(`/Users/me/My\ Shot.png`); a leading `/path/...` is treated as a file, not a command. Each
-referenced image collapses to a short `[Image #1]` chip in the transcript (numbered across
-the session) while the picture itself rides along to the model.
-
-While the agent is working, press **Esc** to interrupt the turn (partial output is kept);
-**Ctrl-C** quits.
-
-## Slash commands
-
-Built-ins (plus any custom commands you add):
-
-| Command | |
-|---|---|
-| `/help` `/doctor` `/config` | help, diagnostics, resolved settings layers |
-| `/model [spec]` `/provider` `/keys` | choose a model, list providers, manage API keys |
-| `/login` `/logout` | sign a Privateer account in/out (see [Privateer account](#privateer-account-billed-inference--what-it-sees)) |
-| `/remote-access [on\|off\|status]` | link this terminal to the Privateer app for phone approval (off by default) |
-| `/permissions [mode]` `/cost` `/context` | permission mode, token usage, context window |
-| `/init` `/memory` `/todo` | write/show `PRIVATEER.md`; show the task list |
-| `/agents` `/mcp [logout]` `/hooks` | inspect sub-agents; MCP status / clear OAuth; hooks |
-| `/skills [list\|info\|install\|remove]` | manage skills (see [Customize & extend](#customize--extend)) |
-| `/routine [list\|pause\|resume\|rm\|run]` | manage scheduled routines |
-| `/output-style [name]` `/vim` `/verbose` | persona, modal editing, full tool output |
-| `/zdr` | toggle OpenRouter zero-data-retention enforcement (see [Data retention](#data-retention-zdr)) |
-| `/verify` | fetch the NEAR AI TEE attestation for the current model (see [Private inference](#private-inference-near-ai)) |
-| `/rewind` `/compact` `/clear` `/export` | restore a checkpoint, compact, clear, save transcript |
-| `/resume` `/sessions` | pick up an earlier session in this directory |
-| `/exit` | quit |
-
-- `/model` — open a picker of each provider's live models (or `/model provider:id` to set one directly).
-- `/init` — the agent explores the repo and writes a `PRIVATEER.md` for you
-  (`/init --stub` just drops an empty template, no model call).
-- `/rewind` — pick an earlier checkpoint and restore the conversation, the files, or both.
-- `/compact` — summarize older history to reclaim context (also happens automatically).
-
-## Tools
-
-`read` · `write` · `edit` · `glob` · `grep` · `bash` · `bash_output` · `kill_shell` ·
-`todo` · `task` · `web_fetch` · `web_search` — plus any tools exposed by connected MCP servers.
-
-The file/search/shell tools are pure-Node (no external binaries required). Mutating tools
-(write/edit/bash) and network tools (web_fetch/web_search, MCP) go through the permission gate.
-`todo` maintains the live task list; `task` delegates an investigation to a sub-agent that
-returns a summary. `bash` can run detached with `run_in_background`; `bash_output` polls a
-background shell's new output and `kill_shell` stops it.
-
-## Customize & extend
-
-Everything below is optional and lives under `.privateer/` (project) or `~/.privateer/`
-(user); project files win. Settings merge across `config.json` → `settings.json` →
-`settings.local.json` (run `/config` to see the resolved chain).
-
-- **Custom commands** — `.privateer/commands/<name>.md`. The body is a prompt template
-  (`$ARGUMENTS`, `$1`…`$9`); optional frontmatter sets `description`/`argument-hint`. They
-  appear in `/help` and `/` autocomplete; subfolders namespace as `dir:name`.
-- **Output styles** — `.privateer/output-styles/<name>.md` swap the agent's persona.
-  Switch with `/output-style <name>` (or `default`).
-- **Sub-agents** — `.privateer/agents/<name>.md` with frontmatter (`description`, `tools`,
-  `model`). Invoke via the `task` tool's `subagent_type`; `/agents` lists them.
-- **Skills** — `.privateer/skills/<name>/SKILL.md` (frontmatter `name`/`description` +
-  instruction body, plus any bundled `scripts/`/`references/` files). The format is
-  Claude Code-compatible, so published Agent Skills drop in unchanged. The agent sees a
-  catalog of names and descriptions and loads a skill's full instructions on demand via
-  the `skill` tool; `/skill-name` invokes one explicitly. Manage with `/skills`
-  (`list`, `info <name>`, `install <owner/repo[/path]> [--project] [--all] [--force]`,
-  `remove <name>`) — install fetches from GitHub with a shallow clone and never executes
-  anything it downloads.
-- **Hooks** — a `hooks` section in `settings.json` runs shell commands on `PreToolUse`,
-  `PostToolUse`, `UserPromptSubmit`, and `Stop`. A hook blocks by exiting `2` or printing
-  `{"decision":"block"}`; `UserPromptSubmit` can inject `additionalContext`. `/hooks` lists them.
-- **MCP servers** — declare them in `.privateer/mcp.json` (`{ "mcpServers": { … } }`).
-  Both **local stdio** servers (`{ "command", "args", "env" }`) and **remote HTTP** servers
-  (`{ "url", "headers?", "transport?" }`) are supported; remote defaults to Streamable HTTP
-  with a fallback to legacy SSE. Their tools are namespaced `server__tool` and gated like the
-  rest. Remote servers authenticate by a static `headers` bearer token, or — when none is set —
-  via **interactive OAuth** (PKCE + dynamic client registration): on a `401` Privateer opens your
-  browser, catches the redirect on a loopback port, and caches the tokens (owner-only) under
-  `~/.privateer/mcp-auth/`. `/mcp` shows each server's connection and auth state;
-  `/mcp logout [server]` clears saved OAuth.
-
-  ```json
-  {
-    "mcpServers": {
-      "fs":     { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "."] },
-      "github": { "url": "https://api.githubcopilot.com/mcp/" },
-      "internal": { "url": "https://mcp.example.com/mcp", "headers": { "Authorization": "Bearer $TOKEN" } }
-    }
-  }
-  ```
-- **Routines** — saved tasks the scheduler daemon (`privateer daemon --detach`) runs
-  unattended, recurring (cron) or one-off. Ask the agent ("summarize world news every
-  morning") and approve; manage with `/routine` (list/pause/resume/rm/run). Runs use a safe
-  read/web toolset by default; a routine's `tools` list can also grant specific MCP tools
-  (`server__tool` or `server__*`) — flagged at approval, since they then run with no one
-  watching. See the [Sheet → WhatsApp recipe](docs/recipes/sheet-to-whatsapp.md) for a
-  full business automation built this way.
-- **Status line** — set `statusLine` to a shell command; it receives session JSON on stdin
-  and its stdout becomes the status line.
+## Private & verifiable inference
+
+**NEAR AI Cloud** and **Tinfoil** run every model inside a **Trusted Execution Environment** —
+a confidential VM where TLS terminates *inside* the enclave, so your prompt's inputs, weights,
+and outputs are invisible to the infrastructure provider, the model provider, and the host
+itself. It isn't "trust us": each request can produce a **cryptographic attestation** proving
+the inference ran on genuine TEE hardware.
+
+- A **status shield** colors the selected model's live posture — 🟢 verified, 🟡 returned but
+  unconfirmed, 🔴 no attestation material.
+- **`/verify`** fetches the attestation on demand and prints the evidence. Privateer does a
+  pragmatic terminal-suited check; take the printed report to the
+  [NEAR AI Cloud Verifier](https://github.com/nearai/cloud-verifier) or the
+  [Tinfoil verifier](https://github.com/tinfoilsh/tinfoil-cli) for full quote-chain validation.
+- **The posture is graded honestly.** A verified enclave (`cryptographic`), a pinned
+  zero-retention route (`observable`), and a provider's retention *promise* (`policy`) are
+  labeled distinctly — a claim never gets to read like a proof.
+
+## The PII gate
+
+Before any prompt is sent to a channel that *isn't* verified-private, Privateer scans it
+**locally** for structured personal data — emails, phone numbers, SSNs, credit-card numbers
+(Luhn-checked), IBANs (mod-97), IP and MAC addresses. If it finds any, it warns and offers to
+**redact** or **send as-is** (or remember your choice for the session). Detection is
+deterministic and on-device — no model ever sees the data in order to find it — and it's
+skipped entirely on an attested TEE or on-device channel, which provably can't read your
+prompt anyway. It's best-effort structured-PII detection, labeled as such — a safety net, not
+a guarantee.
+
+## Privateer account (billed inference)
+
+Instead of bringing your own key, run **`/signin`** to sign into a Privateer account — an
+app-brokered device flow where you approve a short code in the Privateer app, so wallet and
+email accounts work identically and no password or key ever touches the terminal. Inference
+is then billed to your subscription and defaults to a **NEAR TEE** model. Sign out any time
+with `/signout`; manage linked terminals from the app.
+
+> **Only approve a sign-in code you generated yourself.** The code authorizes *this* terminal
+> to spend on your account. If someone sends you a code and asks you to approve it, don't —
+> that hands *them* a billed session on *your* account.
+
+## Approve from your phone
+
+Turn on **`/remote-access`** (off by default) to link this terminal to the Privateer app. The
+app can then drive the terminal — prompts come down, and each proposed action goes up for
+**Allow/Deny** — while execution stays on your machine. The relay is live-only (nothing is
+archived), carries no keys, and output is size-truncated and run through a best-effort secret
+redactor before it leaves.
 
 ## Permission modes
 
@@ -451,40 +190,22 @@ Everything below is optional and lives under `.privateer/` (project) or `~/.priv
 |---|---|
 | `default` | prompt before edits and shell commands |
 | `acceptEdits` | auto-approve file edits; still prompt for shell commands |
-| `bypass` | no prompts (also `--dangerously-skip-permissions` or `--no-quarter`) |
-| `plan` | read-only; the agent presents a plan, then you approve to leave plan mode |
-
-Out of the box the mode is **`acceptEdits`** (naming is a nod to convention: the mode
-called `default` prompts on everything, but isn't the shipped default). Edits are still
-checkpointed, so `/rewind` undoes them; prefer prompting on every edit? Run
-`/permissions default` once — it persists. At an approval prompt: **y** allow once ·
-**a** always · **n** deny. In plan mode, after the agent presents its plan: **a** approve
-and exit plan mode · **k** keep planning.
+| `bypass` | no prompts (destructive commands are *still* blocked) |
+| `plan` | read-only; the agent presents a plan, then you approve to proceed |
 
-## Project context
+Switch with **`/mode`**. Even in `bypass`, a danger filter blocks destructive shell commands,
+and protected files (`.env`, shell rc files…) are guarded — the gate is never fully off.
 
-Create a `PRIVATEER.md` in your repo (via `/init`) to give the agent standing
-context — conventions, architecture notes, anything it should always know.
+## Extend it
 
-## How it compares
-
-There are excellent terminal coding agents already. What this one does differently:
-
-- **Provider-agnostic by construction, not adaptation.** One agent loop over the Vercel AI
-  SDK; OpenRouter, Anthropic, OpenAI, local Ollama, and NEAR AI are interchangeable at
-  `/model` time, including mid-session. No vendor's models are privileged.
-- **Retention posture is a UI element.** ZDR status is visible before you send and
-  enforceable per request (`/zdr`); TEE inference is attestable (`/verify`). Most tools
-  leave this to the provider's terms-of-service page.
-- **Phone approval.** `/remote-access` relays each proposed action to the Privateer app
-  for Allow/Deny while execution stays on your machine — useful for long agent runs you
-  want to supervise from anywhere.
-- **Interop over ecosystem.** Skills use the Agent Skills format, so Claude Code skills
-  drop in unchanged; MCP covers tools. The goal is to reuse what exists, not grow a
-  parallel plugin world.
-
-And, honestly, what it doesn't have: the maturity of the incumbents. It's a young
-codebase — see [Caveats](#caveats) for the sharp edges we know about.
+- **MCP servers** — declare them and their tools become first-class, gated like the rest
+  (local stdio, or remote HTTP with interactive OAuth).
+- **Sub-agents** — delegate investigations to bounded parallel agents that run under the same
+  permission gate.
+- **Routines** — saved tasks the daemon runs unattended; ask the agent to schedule work and
+  approve it once.
+- **Web tools** — private-by-default web search/fetch with pluggable backends (self-hosted
+  SearXNG for fully private search).
 
 ## Develop
 
@@ -493,35 +214,6 @@ npm run typecheck
 npm test
 ```
 
-## Caveats
-
-Privateer's agent core is built provider-agnostic from the ground up; a few areas are
-deliberately simplified for now:
-
-- **Prompt caching and extended thinking are Anthropic-only.** Ephemeral cache breakpoints
-  and the `thinkingBudget` setting apply to direct Anthropic models and OpenRouter routes to
-  `anthropic/*`. Other providers ignore them (a harmless no-op).
-- **Checkpoints are in-memory.** `/rewind` is a within-session undo; snapshots live in memory
-  and aren't persisted across restarts.
-- **Remote MCP OAuth uses a fixed loopback port** (`7777` by default; override with
-  `PRIVATEER_OAUTH_PORT`) so the redirect URI stays stable across runs. Set `PRIVATEER_NO_BROWSER=1`
-  in headless environments to skip the auto-launch and use the printed URL.
-- **Image attachment assumes vision support.** Referenced images are sent as content parts;
-  non-vision models will return an error.
-- **Compaction estimates context size** (~4 chars/token). The summary itself is
-  schema-guided (goals / decisions / files / open threads), with a plain-text fallback, and
-  the most recent messages are always kept verbatim.
-- **`web_search` scrapes DuckDuckGo's keyless HTML endpoint.** It needs no API key but is
-  best-effort and can break if their markup changes. `web_fetch` is robust for known URLs.
-- **Protected files** (`.env`, `.npmrc`, shell rc files, …) always prompt before edit — except
-  in `bypass` mode, which by definition skips all prompts.
-
-## Docs
-
-- [Architecture](docs/ARCHITECTURE.md) — how the provider layer, agent loop, tools, and permissions fit together
-- [Recipe: Sheet → WhatsApp](docs/recipes/sheet-to-whatsapp.md) — an unattended routine + MCP servers messaging new spreadsheet rows
-- [Brand assets](brand/README.md) — the logo and icon set
-
 ## License
 
 [MIT](LICENSE) © Patrick