2b-agent 0.2.1__tar.gz

2b_agent-0.2.1/.github/workflows/release.yml

@@ -0,0 +1,39 @@
+name: release
+
+# Publishes to PyPI when a version tag is pushed. Uses PyPI Trusted Publishing
+# (OIDC) — no API token or secret is stored anywhere. Configure the trusted
+# publisher once on PyPI (owner: dea6cat, repo: 2b-agent, workflow: release.yml,
+# environment: pypi) before the first tag.
+on:
+  push:
+    tags: ["v*"]
+
+jobs:
+  pypi:
+    name: build & publish to PyPI
+    runs-on: ubuntu-latest
+    environment: pypi
+    permissions:
+      contents: read
+      id-token: write          # OIDC identity for Trusted Publishing
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v5
+
+      - name: Guard — the tag must match the package version
+        run: |
+          tag="${GITHUB_REF_NAME#v}"
+          ver="$(sed -n 's/^__version__ *= *"\(.*\)"/\1/p' src/two_b/__init__.py)"
+          echo "tag=$tag  package=$ver"
+          if [ "$tag" != "$ver" ]; then
+            echo "::error::tag v$tag does not match __version__ $ver — bump src/two_b/__init__.py to match before tagging"
+            exit 1
+          fi
+
+      - name: Build sdist + wheel
+        run: uv build
+
+      - name: Publish to PyPI (trusted publishing)
+        run: uv publish

2b_agent-0.2.1/.gitignore

@@ -0,0 +1,7 @@
+__pycache__/
+*.py[cod]
+*.egg-info/
+.venv/
+dist/
+build/
+.DS_Store

2b_agent-0.2.1/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 Alexander Montolio
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

2b_agent-0.2.1/NOTICE

@@ -0,0 +1,4 @@
+2b-agent
+Copyright 2026 Alexander Montolio
+
+This product is licensed under the Apache License, Version 2.0 (see LICENSE).

2b_agent-0.2.1/PKG-INFO

@@ -0,0 +1,336 @@
+Metadata-Version: 2.4
+Name: 2b-agent
+Version: 0.2.1
+Summary: A local-first coding agent that keeps small local models focused instead of hallucinating.
+License-Expression: Apache-2.0
+License-File: LICENSE
+License-File: NOTICE
+Requires-Python: >=3.11
+Requires-Dist: mcp>=1.0
+Requires-Dist: prompt-toolkit>=3.0
+Requires-Dist: rich>=13.0
+Requires-Dist: textual>=0.60
+Description-Content-Type: text/markdown
+
+# 2B
+
+A local-first coding agent for the terminal. It runs your own models over Ollama's
+native protocol, keeps them focused instead of hallucinating, and gives you a full-screen
+TUI — streaming replies, a live plan checklist, narrated tool actions — without ever routing
+your local model through a translation layer that would confuse it.
+
+I named it **2B**, after NieR: Automata. It's built to keep working when the power and the
+internet don't — I live somewhere the grid isn't a guarantee, and I wanted a coding agent that
+doesn't fall apart the moment I'm offline.
+
+> **macOS only.** 2B is built and tested for macOS — the installer is a shell script that leans on
+> Homebrew, and the clipboard integration uses `pbcopy`. It hasn't been tested elsewhere.
+
+---
+
+## Why I built it
+
+I kept being told that small models — Nemotron 3 Nano 4B, `gpt-oss:20b`, the Qwen family — were
+"good enough" for agentic coding. On paper they were. In practice, every off-the-shelf harness I
+tried broke them:
+
+- **opencode** made `gpt-oss` invent tool names that didn't exist and emit fake `<command>` tags as
+  plain text.
+- **Cline**, **Goose**, **Continue**, **OpenHands** each failed in their own way — malformed tool
+  schemas, reasoning collapse, the model narrating tool calls instead of making them.
+
+The models weren't the problem. The harnesses were. Nearly all of them talk to a local model
+through a generic **OpenAI-compatible `/v1` shim** and pile abstraction on top of it. That shim
+measurably degrades a small model's tool selection, and the extra complexity buries whatever
+capability the model actually has.
+
+The one thing that worked cleanly was a ~350-line script I wrote that talked to Ollama's **native
+`/api/chat`** endpoint with a tiny, fixed set of five tools. So I grew that prototype into a real,
+shareable tool. That's 2B.
+
+**The core rule, and the whole point:** all complexity lives on the *host* side. The model's world
+never changes — the same five tools, the same native wire format for whatever provider is active, no
+generic shim. Everything you see below — the TUI, the plan checklist, task management, model
+switching, auto-compaction — is something 2B renders *around* the tool loop, never a new thing the
+model has to understand.
+
+---
+
+## What it does
+
+- **Five tools, and only five.** `list_files`, `read_file`, `search_files`, `edit_file`,
+  `write_file`. That small, concrete surface is exactly what keeps a small model reliable. It
+  explores before it edits — searching for where something lives instead of guessing paths — and
+  prefers exact-snippet edits over rewriting whole files.
+- **Edits that survive small-model drift.** `edit_file` resolves the target host-side in tiers —
+  exact, then whitespace-tolerant, then indentation-agnostic (re-indenting your snippet to the
+  file) — so a model that gets the whitespace slightly wrong still lands the edit instead of
+  bouncing off an exact-match error. It never applies on an ambiguous match, and the tool the model
+  sees is unchanged — all the tolerance lives on the host.
+- **Catches its own mistakes.** After a successful edit, 2B runs the file's checker host-side
+  (`dart analyze`, `ruff` or `py_compile`, …) and folds any new errors straight into the tool
+  result — so a model that just broke the build sees it on the same turn, with no new tool to learn.
+  Bounded so it can't flood a small window, skipped silently when there's no checker, and off with
+  `TWOB_NO_DIAGNOSTICS`.
+- **Finds definitions, not just matches.** `search_files` marks which hit is the *definition* of a
+  symbol and floats it to the top, and `read_file` appends a compact symbol outline with line
+  anchors — so "where is X defined?" is answered by the tools the model already calls, with no
+  navigation tool to learn. When a language server is installed (`dart language-server`, `pyright`,
+  `gopls`, …) it resolves symbols semantically over LSP, spoken as raw stdlib JSON-RPC; a curated MCP
+  resolver is used if one's enabled; with neither, it falls back to a dependency-free regex map.
+  Host-side, schema unchanged; off with `TWOB_NO_LSP`.
+- **Native protocols, never a shim.** Local Ollama models get Ollama's own `/api/chat` with NDJSON
+  streaming. Each cloud provider gets its own native format. Nothing is translated through a
+  lowest-common-denominator layer.
+- **Streaming, full-screen TUI.** A scrolling conversation, a framed input box, a live status line
+  with a spinner, elapsed time, and — for local models — a RAM/GPU readout pulled from Ollama.
+- **Narrated tool actions.** Instead of a wall of raw `read_file {...}`, you see what it's doing in
+  plain language, tied together with a tree gutter and a ✓/✗ per step:
+  ```
+  ├ ✓ Searching for "MemoryScopeLevel" in lib
+  ├ ✓ Reading lib/memory/memory_store.dart
+  └ ✓ Editing lib/memory/memory_store.dart
+  ```
+- **A live plan checklist.** The model writes a short numbered plan before its first tool call; 2B
+  parses it and renders it as a checklist that advances (`□` pending, `■` active, `✓` done) as the
+  work progresses. Purely cosmetic — a wrong guess never breaks anything.
+- **Many providers, one conversation.** Ollama (local and cloud), OpenAI, OpenRouter, Mistral,
+  NVIDIA, Anthropic, and Google Gemini. 2B keeps history in a provider-agnostic form and
+  re-serializes it fresh for whoever's active — so you can switch models *mid-task* with `/model`
+  and keep every bit of context. Start a task on a local Qwen, hand it to Claude when it gets hard,
+  keep going.
+- **Knows the project.** `/init` scans the repo and writes a compact `2B.md` — stack, layout, and a
+  ranked symbol map — that's auto-loaded into context, so the model starts knowing where things are
+  instead of hunting for files. `/map` shows a budget-bounded outline on demand. All bounded, so it
+  never floods a small local window.
+- **Runs things — split by model.** Local models get `run_git` (git only, never a raw shell — no
+  chaining/injection); cloud models get a full `run_command` shell (tests, build, git). Read-only git
+  runs freely; anything that mutates is confirmation-gated and refused in plan mode.
+- **MCP tools, curated.** Pull in tools from MCP servers (dart, mempalace, …) — but **per tool**, not
+  wholesale, because flooding a small model with tools is exactly what breaks it. You enable a server
+  and pick which of its tools the model sees (`/mcp`); local models are capped to a few so their
+  context stays lean. See [MCP servers](#mcp-servers-extra-tools).
+- **Operating modes**, cycled with **Shift+Tab** or set with `/mode`:
+  - **normal** — every write/edit asks first.
+  - **accept edits** — writes apply automatically.
+  - **plan mode** — read-only; `edit_file`/`write_file` *and* MCP tools are refused (they may change
+    state), so the model investigates and returns a plan instead of touching anything.
+- **Auto-compaction.** When a conversation nears the model's context window — which happens fast on
+  small local windows — 2B folds the older turns into a summary and keeps going uninterrupted,
+  instead of hitting the wall. It cuts on a safe boundary so nothing breaks, and shows you
+  "Compacting conversation…" while it does it.
+- **Themes.** `/theme system` (default — transparent, uses your terminal's own background),
+  `/theme light` (a warm parchment palette), `/theme dark` (a dimmed version). Switches live.
+- **Copy that actually works.** Drag to select any text and press **Ctrl+C**, or **Ctrl+Y** / `/copy`
+  to grab the whole last reply. On macOS this goes through `pbcopy`, so it lands on your clipboard
+  even in Terminal.app (which ignores the escape sequence most TUIs rely on).
+- **Multiple tasks.** Queue tasks, background the running one with **Ctrl+B**, foreground it later
+  with `/fg`. A backgrounded task pauses when it needs to write and waits for you.
+- **Undo.** `/undo` reverts the last write or edit — one level, but it's there.
+
+---
+
+## Install
+
+One line — paste it in your terminal:
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/dea6cat/2b-agent/main/install.sh | sh
+```
+
+It installs [`uv`](https://docs.astral.sh/uv/) if you don't have it, installs the `2b` command,
+then — on an interactive terminal — walks you through local model setup:
+
+1. **Optional clean install** — offers to remove other agentic tools that proved unreliable with
+   local models (opencode, Continue, Goose, Cline, OpenHands) and their configs. Off by default;
+   it asks first.
+2. **Grades your machine** — reads your RAM and chip and rates each candidate model
+   (`✓ fits well` / `~ tight` / `✗ needs NGB+`), defaulting to the best one your hardware can run.
+3. **You pick** one or several from the menu.
+4. **Installs Ollama and pulls** what you chose, with a live progress bar.
+5. **Self-tests** each model — tok/s + GPU residency, then a **correctness check that runs a real
+   one-line edit through 2B itself** and verifies the result (`✓ correct` / `✗ wrong`, ~20–90s per
+   model). It only reports — it never removes a model — and `--no-benchmark` skips it. Then it prints
+   how to launch 2B.
+
+Already have Ollama and some models? It skips what you already have — it lists your installed
+models, offers to just use them (pulling nothing), and marks anything in the menu you've already
+got. Your existing setup is left untouched.
+
+Prefer to do it by hand? If you already have `uv`:
+
+```bash
+uv tool install git+https://github.com/dea6cat/2b-agent
+ollama pull qwen3.5:9b        # my default — a good balance on an 18 GB machine
+```
+
+The installer is scriptable too: `--yes` (accept defaults, no prompts), `--clean` / `--no-clean`,
+`--models "qwen3.5:9b qwen3:8b"`, `--no-models`, `--no-benchmark` (skip the correctness check),
+`--fix-path` / `--no-fix-path` (add uv's tool dir to your PATH for you via `uv tool update-shell`,
+or leave it — otherwise it asks, and never edits a profile without consent). Pass them through the
+pipe with `... | sh -s -- --yes --models "qwen3.5:9b"`.
+
+---
+
+## Use it
+
+```bash
+2b                       # start in the current directory, autodetects a local model
+2b "add a docstring to lib/main.dart"   # run one task, then drop into the session
+2b --model qwen3.5:9b    # pin a model
+2b --list-models         # what's available across configured providers
+2b --doctor              # diagnose PATH, Ollama, and the default model, then exit
+2b --update              # upgrade to the latest release (uv tool upgrade)
+2b --rm                  # uninstall 2B and delete its config (asks first); --rm --yes to skip
+```
+
+Then just type what you want done. Type `/` to see the commands.
+
+### Updating
+
+2B installs as a `uv` tool, so updating is one command:
+
+```bash
+2b --update                        # or: uv tool upgrade 2b-agent
+```
+
+Re-running the installer (`curl … | sh`) does the same. 2B also checks for a newer
+release in the background (at most once a day, never blocking startup) and prints a
+one-line notice next launch when one is available — set `TWOB_NO_UPDATE_CHECK=1` to
+turn that off. Releases are tagged `vMAJOR.MINOR.PATCH`; pin one for reproducibility
+with `uv tool install git+https://github.com/dea6cat/2b-agent@v0.2.0`.
+
+### Providers
+
+Local Ollama needs nothing. For anything else, set the matching environment variable and it shows
+up automatically in `/models`:
+
+| Provider   | Environment variable                        |
+| ---------- | ------------------------------------------- |
+| Ollama     | `OLLAMA_API_BASE` (or `OLLAMA_HOST`)        |
+| Ollama Cloud | `OLLAMA_API_KEY`                          |
+| OpenAI     | `OPENAI_API_KEY`                            |
+| OpenRouter | `OPENROUTER_API_KEY`                        |
+| Mistral    | `MISTRAL_API_KEY`                           |
+| NVIDIA     | `NVIDIA_API_KEY`                            |
+| Anthropic  | `ANTHROPIC_API_KEY`                         |
+| Google     | `GEMINI_API_KEY` (or `GOOGLE_API_KEY`)      |
+
+Or connect one from **inside** 2B — `/connect <provider>` prompts for the key with a hidden field
+and saves it to `~/.config/2b/keys.json` (`chmod 600`) so it's there next time; `/connect` on its own
+shows what's connected, and `/disconnect <provider>` removes a saved key. A key exported in your
+shell always takes precedence over a saved one.
+
+Switch models anytime with `/model <name>`. A bare name works when it's unambiguous; otherwise use
+`provider:model` (e.g. `/model anthropic:claude-sonnet-5`).
+
+### Commands
+
+| Command | What it does |
+| --- | --- |
+| `/help` | List commands |
+| `/model [name]` | Show or switch model — context is preserved |
+| `/models [filter]` | List available models, grouped by provider |
+| `/connect [provider] [key]` | Connect a provider (hidden key prompt); bare shows status |
+| `/disconnect <provider>` | Remove a saved provider key |
+| `/init` | Scan the project → write `2B.md` (a compact map auto-loaded into context on new tasks) |
+| `/map [subdir]` | Show a budget-bounded symbol outline of the project |
+| `/mcp` | MCP servers/tools: status, `tools <server>`, `enable`/`disable <server> <tool…\|all>` |
+| `/mode [normal\|accept\|plan]` | Set operating mode (or **Shift+Tab** to cycle) |
+| `/theme [system\|light\|dark]` | Switch color theme |
+| `/context` | Show estimated context usage (auto-compacts near the limit) |
+| `/copy` | Copy the last reply to the clipboard (**Ctrl+Y**) |
+| `/task <desc>` | Queue a task |
+| `/tasks` | List tasks and their status |
+| `/fg <id>` | Foreground a backgrounded task |
+| `/yes` | Toggle accept-edits mode |
+| `/undo` | Revert the last write/edit |
+| `/diff` | Re-show the last diff |
+| `/add <file>` | Pre-load a file into the current task's context |
+| `/clear` | Reset the current task's history |
+| `/quit` | Exit |
+
+### Keyboard
+
+| Key | Action |
+| --- | --- |
+| **Shift+Tab** | Cycle operating mode |
+| **Ctrl+B** | Background the running task |
+| **Ctrl+Y** | Copy the last reply |
+| **Ctrl+C** | Copy the current mouse selection |
+| **Esc** | Stop the current stream/task immediately — back to idle |
+| **Ctrl+D** | Quit |
+| **Tab** | Accept the top `/`-command suggestion |
+
+### MCP servers (extra tools)
+
+2B can pull in tools from [MCP](https://modelcontextprotocol.io) servers (stdio) like `dart` or
+`mempalace`. But its whole reason for existing is that **small local models break when you flood them
+with tools** — so MCP tools are opt-in and **curated per tool**: you enable a server and pick exactly
+which of its tools reach the model. Nothing is exposed until you say so.
+
+Declare servers the usual way — a Claude-Code-style `mcpServers` block in `~/.config/2b/mcp.json` (or
+`./.mcp.json` in a project, which wins):
+
+```json
+{
+  "mcpServers": {
+    "dart": { "command": "dart", "args": ["mcp-server"] }
+  }
+}
+```
+
+Then curate from inside 2B:
+
+```
+/mcp                          # servers and how many tools each has enabled
+/mcp tools dart               # list a server's tools ([x] = enabled)
+/mcp enable dart hot_reload analyze_files
+/mcp enable dart all          # expose everything (careful on small models)
+/mcp disable dart hot_reload  # or /mcp disable dart  to turn the whole server off
+```
+
+Enabled tools appear to the model as `server__tool` (e.g. `dart__hot_reload`) and their results come
+straight back into the loop. Only the tools you enable are ever sent — the model's tool list stays as
+small as you keep it.
+
+### Configuration
+
+- **Context window (local) — sized to your machine.** For a local model 2B works out the largest
+  window your box can run *comfortably* and pins `num_ctx` to it on every request (Ollama otherwise
+  defaults to ~4k regardless of the model). It reads the model's architecture and trained max from
+  `/api/show`, computes the KV-cache cost per token, and fits it into the RAM left after the model
+  weights plus a headroom reserve — so a 16 GB laptop, an 18 GB one, and a 64 GB workstation each get
+  a different, appropriate window (e.g. qwen3.5:9b ≈ 13k on 18 GB), never more than the model was
+  trained for. That number drives auto-compaction (~75%) and the read-a-section threshold. Set
+  `TWOB_CONTEXT_TOKENS` to override (higher if you want to spend more RAM, lower to save it).
+
+---
+
+## Honest caveats
+
+- **It reads and writes wherever you point it.** 2B resolves absolute paths and paths outside the
+  working directory on purpose — it's a personal tool for your machine. Writes are still confirmed
+  in normal mode, and plan mode refuses them entirely.
+- **Switching to a stronger model mid-task hands it a tool-call history it didn't make.** For these
+  five simple tools that's low-risk (the wire format is unambiguously the new provider's own; only
+  the *choices* inside came from a weaker model), but you may see mild "why did I read that file"
+  moments. It is *not* the shim-degradation failure that sank the other harnesses.
+- **It's a full-screen TUI.** That means it lives in the terminal's alternate screen, so a single
+  mouse drag selects what's on screen, not scrolled-off history. For a classic inline REPL, run
+  `2b --classic`.
+
+---
+
+## How it's built
+
+- **Python, standard library first.** The only real dependencies are `rich`, `prompt_toolkit`, and
+  `textual`. Every provider talks over `urllib` — no per-provider SDKs.
+- **A canonical conversation model** that each provider serializes fresh on every request. That
+  re-derivation is exactly what makes switching models mid-task safe.
+- **One worker thread per task**, emitting events into a queue the UI thread drains and renders — so
+  the tool code stays untouched off the main thread and one thread owns the terminal.
+- The five-tool schema is **frozen**. It's what makes small models reliable, and it isn't up for
+  redesign.
+
+Built for local models, kept on task.