chad-code 2.0.3__tar.gz → 2.2.0__tar.gz

{chad_code-2.0.3/src/chad_code.egg-info → chad_code-2.2.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: chad-code
-Version: 2.0.3
+Version: 2.2.0
 Summary: Claude-Code-style coding agent that runs entirely on a 24 GB Apple Silicon Mac (MLX, Qwen3.8-27B)
 License-Expression: MIT
 Project-URL: Repository, https://github.com/nathansutton/chad
@@ -13,11 +13,12 @@ Classifier: Operating System :: MacOS
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
 Classifier: Programming Language :: Python :: 3.13
+Classifier: Programming Language :: Python :: 3.14
 Classifier: Topic :: Software Development :: Code Generators
 Requires-Python: >=3.11
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: mlx==0.32.0; sys_platform == "darwin"
+Requires-Dist: mlx!=0.32.1,<=0.32.2,>=0.32.0; sys_platform == "darwin"
 Requires-Dist: mlx-lm<0.32,>=0.31.3; sys_platform == "darwin"
 Requires-Dist: transformers<5.13,>=5.0
 Requires-Dist: jinja2>=3.1
@@ -43,9 +44,13 @@ Dynamic: license-file
 > *Coding under supervision.*
 
 chad is a single-user coding agent that runs entirely on an Apple Silicon Mac via
-[MLX](https://github.com/ml-explore/mlx). One 27B model and no API key. Every other 
-harness assumes a datacenter on the other end of a socket; chad assumes a laptop, 
-and the whole design falls out of that. (Not affiliated with Anthropic.)
+[MLX](https://github.com/ml-explore/mlx). One 27B model and no API key. (Not affiliated
+with Anthropic.)
+
+Plenty of harnesses run local models now, and [pi](https://pi.dev) is a fantastic default
+for the same reason llama.cpp is: it works with everything. I was steering chad in the
+opposite direction. One model and one set of silicon, taken to the max. Swap out your
+`CHAD_MODEL` and it still runs, you just leave the drafter and the kernels behind.
 
 ## Try it
 
@@ -54,13 +59,13 @@ uvx chad-code          # runs anywhere; the command is still `chad`
 uvx chad-code prove    # offline smoke test: 4 tiny fix-it tasks, verified, timed 🗿
 ```
 
-The first run asks, then downloads the model once (~13 GB) into the shared Hugging
+The first run asks, then downloads the model once (~8 GB) into the shared Hugging
 Face cache. While it downloads, `cd` into a project and think of a scoped first ask:
 *"fix the failing test in `tests/test_x.py`"* lands, *"improve my codebase"* flails.
 
 chad targets 24 GB and nothing smaller. It runs below that and tells you it is doing so,
-but 13 GB of weights sit resident before a single token of context, so a 16 GB Mac gets a
-window too small to work in.
+but 8 GB of weights and drafter plus a ~4 GB prefill transient sit resident before a single
+token of context, so a 16 GB Mac gets a window too small to work in.
 
 The PyPI package is `chad-code`. Bare `chad` is an unrelated squatted package.
 
@@ -73,32 +78,42 @@ The PyPI package is `chad-code`. Bare `chad` is an unrelated squatted package.
 
 ## Same model, same Mac, stock engine
 
-The question worth answering: what do you gain over pointing a generic local-model tool at
-the same weights? Qwen3.8-27B at the same `UD-Q3_K_XL` recipe (Unsloth's GGUF for
-llama.cpp, chad's MLX conversion of the same bit map), the same M4 Pro (24 GB), one engine
-resident at a time, each measured with its own benchmark on a 512-token prompt and a
+What do you gain over pointing a generic local-model tool at the same weights?
+Qwen3.8-27B at the `UD-Q3_K_XL` recipe (Unsloth's GGUF for llama.cpp, chad's MLX
+conversion of the same bit map; chad runs it with `--model`), the same M4 Pro (24 GB), one
+engine resident at a time, each measured with its own benchmark on a 512-token prompt and a
 128-token generation.
 
 | Engine | Prefill (512-tok prompt) | Decode (128 tok) | Speculative decoding |
 |---|---|---|---|
-| llama.cpp `llama-bench` (stock, build 10470) | 102 tok/s | 10.9 tok/s | none for this model |
-| **chad**, serial (`CHAD_NO_DFLASH=1`) | 99 tok/s | 18.1 tok/s | off |
-| **chad**, default | 98 tok/s | **62 tok/s**¹ | DFlash2 block drafter |
-
-A 200-token function body takes roughly 18 seconds at 10.9 tok/s and 3 at 62. That is the
-difference between a batch job and a pair programmer, and closing it is what the project is
-for.
-
-Ollama does not get its own row: it is llama.cpp underneath with no speculative decoding for
-this model, and on the same GGUF (0.32.15, Modelfile `FROM` only) it measures 96 tok/s
+| llama.cpp `llama-bench` (stock, build 10470) | 102 tok/s | 10.9 tok/s | off in this benchmark |
+| llama.cpp `llama-server` (build 10917), serial | 97 tok/s | 11.3 tok/s | off |
+| llama.cpp `llama-server` (build 10917) | 95 tok/s | 11.1 tok/s² | DFlash2 drafter (Q4_K_M GGUF) |
+| **chad**, serial (`CHAD_NO_DFLASH=1`) | 100 tok/s | 17.9 tok/s | off |
+| **chad** | 101 tok/s | **62.9 tok/s**¹ | DFlash2 block drafter |
+
+A 200-token function body takes roughly 18 seconds at 10.9 tok/s and 3 at 63. You wait for
+the first one and you talk to the second.
+
+The weights chad ships by default are the same model in Prism ML's **ternary** build
+(every projection Hadamard-rotated and stored at 2 bits): 7.2 GB resident instead of 12.3,
+the same 64 tok/s drafted and 21 serial on this Mac, and the 5 GB it gives back is
+context, roughly a **150k-token window** where the 3-bit landed near 56k. The like-for-like
+table above stays on the 3-bit recipe because that is the quant llama.cpp has a GGUF of.
+
+Ollama does not get its own row: it is llama.cpp underneath, measured without speculative
+decoding, and on the same GGUF (0.32.15, Modelfile `FROM` only) it measures 96 tok/s
 prefill and the same **10.9** decode.
 
-¹ 62 is a ceiling: `chad-bench`'s prompt is tiled code the drafter reads easily. Replayed
-against ten real mid-session contexts from `~/.chad/sessions` (12–19k tokens, tool results in
+¹ 63 is a ceiling: `chad-bench`'s prompt is tiled code the drafter reads easily. Replayed
+against ten real mid-session contexts from `~/.chad/sessions` (12-19k tokens, tool results in
 place, 384-token decodes) the same engine measures **31.7 tok/s median / 21.4 floor** greedy
 against 14.8 serial, and **27.6 / 17.7** thinking against 13.9. That ~2× is what a session
-lives at. llama.cpp's Metal path for this hybrid architecture was not profiled, so read its
-row as what a fitted engine buys, not as a verdict on llama.cpp.
+lives at.
+
+² llama.cpp has run DFlash2 since build 10658. Here it accepts 96.5% of drafted tokens and still
+gains nothing: verifying 8 tokens costs ~6.4 serial steps on this GGUF and Mac
+([details](docs/benchmarks.md#same-model-same-mac-stock-engine)).
 
 Method, the longer runs and the caveats are in
 [Throughput & performance](docs/benchmarks.md#same-model-same-mac-stock-engine); the rows are
@@ -110,17 +125,17 @@ committed under [`benchmarks/stock/_runs/`](benchmarks/stock/); reproduce them w
 chad owns its inference loop instead of talking to a server, and the engine is fitted to the
 one checkpoint it ships:
 
-- **DFlash2 block speculation.** A bundled 1.9B drafter ([z-lab's DFlash2](https://huggingface.co/z-lab/Qwen3.8-27B-DFlash2),
+- DFlash2 block speculation. A bundled 1.9B drafter ([z-lab's DFlash2](https://huggingface.co/z-lab/Qwen3.8-27B-DFlash2),
   ported to MLX and quantized) proposes a whole block of tokens from the main model's own
   hidden states. The main model verifies the block in one batched forward, and exact
   rejection sampling keeps every emitted token the model's own.
-- **A persistent prefix KV cache.** The transcript is kept a strict token-prefix of the live
+- A persistent prefix KV cache. The transcript is kept a strict token-prefix of the live
   cache, so a follow-up step prefills the ~16 tokens it appended instead of the 5,000 it
-  already read: **~0.75 s per step instead of ~50 s**. Any server with prompt caching gets
+  already read: **~0.55 s per step instead of ~48 s**. Any server with prompt caching gets
   the easy case; the work is holding it true across compaction, truncated turns and restarts.
   The system prefix is checkpointed to disk, so the second session anywhere starts warm
   (75.6 s → 5.5 s to the first tool call).
-- **Fused Metal kernels.** Quantized-KV attention, a small-M matmul for speculative verify,
+- Fused Metal kernels. Quantized-KV attention, a small-M matmul for speculative verify,
   and a compiled single-token layer step, chosen per machine at load time, no knobs.
 
 ## What chad gives up
@@ -146,16 +161,21 @@ budget; everything else is deliberately plain.
 
 `uv run chad` launches a full-screen terminal UI (built on prompt_toolkit):
 
-- **shift-tab cycles permission modes**: `normal` (confirm each bash/write/edit) →
+- shift-tab cycles permission modes: `normal` (confirm each bash/write/edit) →
   `auto-accept edits` (edits land silently, **terminal commands still ask**) → `yolo`
   (nothing asks) → `plan mode` (read-only: investigate and propose a numbered plan) → back.
-- **type-ahead message queue.** Keep typing while the agent works; messages run in order.
-- **ctrl-c interrupts the running turn** without killing the session. **↑prefilled /
+  A finished plan lands in `./plans/`; ctrl-g (or `/accept`) clears the context and starts
+  implementing it ([details](docs/configuration.md#plan-mode)).
+- Type-ahead message queue. Keep typing while the agent works; messages run in order.
+- ctrl-c interrupts the running turn without killing the session. **↑prefilled /
   ↓generated** token counts show an advancing **%** on an unavoidable full re-prefill, so it
   is never silent.
-- **`@file` / `@dir` mentions** and **`!command` shell passthrough.** Pull a file into
+- `@file` / `@dir` mentions and `!command` shell passthrough. Pull a file into
   context inline, or run a shell command without invoking the model.
-- **voice mode, all local.** `/speech`, then ctrl-t to talk: Parakeet-on-MLX transcribes into
+- Standing project instructions. A `CLAUDE.md` (or `AGENTS.md`) in the working directory is
+  appended to the system prompt, and `/init` reads the project and writes one for you
+  ([details](docs/configuration.md#project-instructions-claudemd--agentsmd)).
+- Voice mode, all local. `/speech`, then ctrl-t to talk: Parakeet-on-MLX transcribes into
   the input box for you to review before Enter sends it, and replies are read aloud via macOS
   `say`. A word table teaches it your identifiers. Needs the `speech` extra
   ([details](docs/configuration.md#voice-mode-speech)).
@@ -169,12 +189,13 @@ budget; everything else is deliberately plain.
 | `--plan` | start in read-only plan mode (investigate and propose, edits blocked) |
 | `--yolo` | auto-approve bash/write/edit (skip confirm prompts) |
 | `--no-think` | skip the model's `<think>` blocks, faster on well-scoped work |
+| `--think-budget N` | soft-cap each step's `<think>` at N tokens, force-close it and carry on (off by default) |
+| `--backend llama` | run the same harness against a remote llama.cpp server, with `--base-url`, `--tokenizer` and `--api-key-env` ([details](docs/configuration.md#alternate-backend-remote)) |
 | `--model` | `auto` (the shipped default), or any HF repo id / local model dir |
 | `--repl` | plain line REPL instead of the TUI |
 
-Three subcommands, each with its own `--help`: `chad prove` (the offline smoke test),
-`chad serve` ([serve this Mac's model to a container or the LAN](docs/configuration.md#serving-the-local-model-to-a-container-chad-serve)),
-and `chad levers` (print the result-channel lever registry as JSON, for A/B ablation).
+Two subcommands, each with its own `--help`: `chad prove` (the offline smoke test) and
+`chad levers` (print the result-channel lever registry as JSON, for A/B ablation).
 
 A headless task (positional, or piped with no TTY) auto-approves mutating tools and runs
 greedy (temp 0). Every conversation is persisted under `~/.chad/sessions/`, and every resume
@@ -187,20 +208,24 @@ chad ships exactly one, downloaded once into the shared Hugging Face cache
 
 | Model | Quant | Footprint |
 |---|---|---|
-| [Qwen3.8-27B `UD-Q3_K_XL-DFlash2`](https://huggingface.co/nathansutton/Qwen3.8-27B-UD-Q3_K_XL-DFlash2-MLX) | 3-bit group-64 body, 5-bit `lm_head`, bundled 4-bit DFlash2 drafter | ~13 GB resident, 262k native context |
+| [Qwen3.8-27B `Ternary-Bonsai-2-DFlash2`](https://huggingface.co/nathansutton/Qwen3.8-27B-Ternary-Bonsai-2-DFlash2-MLX) (default) | Prism ML's ternary build: Hadamard-rotated, 2-bit group-128, levels {−s, 0, +s}; bundled 4-bit DFlash2 drafter | ~8 GB resident, 262k native context, ~150k usable on 24 GB |
+| [Qwen3.8-27B `UD-Q3_K_XL-DFlash2`](https://huggingface.co/nathansutton/Qwen3.8-27B-UD-Q3_K_XL-DFlash2-MLX) via `--model` | 3-bit group-64 body, 5-bit `lm_head`, the same bundled drafter | ~13 GB resident, ~56k usable on 24 GB |
 
 Qwen3.8-27B is **dense** (64 layers: 48 GatedDeltaNet + 16 full attention), so every
-parameter is on the critical path for every token and the quant is where decode speed comes
-from. The bits go where held-out perplexity says they pay: `lm_head` is a second full
-1.27B-param tensor and is held at 5-bit, while `embed_tokens` is a lookup table whose error
-never compounds through a matmul, so it is cheapest. The name follows
-[Unsloth's convention](https://docs.unsloth.ai/) (`UD-…`), though the quant is MLX group-64
-affine, not a llama.cpp k-quant. The drafter ships in the same repo, pre-quantized.
+parameter is on the critical path for every token and the quant is where both decode speed
+and context come from: the governor prices a gigabyte of weights at about 29k tokens of
+window. The ternary build is Prism ML's [Bonsai](https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-mlx-2bit)
+pack of this checkpoint, repacked text-only with the base tokenizer and the drafter
+bundled (Apache-2.0; created using Bonsai by Prism ML). Its projections are stored in a
+rotated basis that an ordinary MLX loader would silently get wrong, so chad carries its own
+loader for it and fits the decode fast-path, the verify kernel and the drafter to it. What it
+costs: on code, teacher-forced perplexity is 4.49 against the 3-bit's 3.99 (+12%), while
+the private eval tiers tie at 56/56. The 3-bit stays one flag away.
 
 `--model <repo or local dir>` runs different weights through the same engine and stays a
 first-class escape hatch. The drafter, the fused-attention coverage, the decode fastpath and
-the context governor are all fitted to the shipped checkpoint, so other weights run slower;
-they do not break.
+the context governor are all fitted to the two checkpoints above, so other weights run
+slower; they do not break.
 
 ## Installing & upgrading
 
@@ -220,7 +245,7 @@ uv run chad "add a --json flag to main.py and update the tests"   # one-shot, he
 uv run chad -c               # resume this directory's last conversation
 ```
 
-**Optional extras.** Two features are opt-in because they pull deps not every install wants:
+Two features are opt-in because they pull deps not every install wants:
 `speech` (voice mode: a mic library, no torch) and `highlight` (syntax colour in diffs and
 previews). An extra rides on the install spec, not on a separate command, so how you add it
 depends on how you installed chad:
@@ -234,12 +259,12 @@ uv sync --extra speech                        # from a clone
 `/speech` in the TUI prints whichever of those matches your install, so you never have to
 work it out from here.
 
-**Upgrading** depends on how you installed: `uv tool upgrade chad-code`, `uvx --refresh
+Upgrading depends on how you installed: `uv tool upgrade chad-code`, `uvx --refresh
 chad-code`, or `git pull && uv sync` for a clone. What changed lands in
 [`CHANGELOG.md`](CHANGELOG.md). Model weights are versioned separately, so a code upgrade
 never re-downloads the model.
 
-**Development.** `uv sync` once, then `uv run pytest -q`. The fast unit gate loads **no model
+For development, `uv sync` once, then `uv run pytest -q`. The fast unit gate loads **no model
 weights**, runs in seconds, and is what CI runs. For throughput on your own machine, use
 `uv run chad-bench` (see [Throughput & performance](docs/benchmarks.md)).
 
@@ -247,11 +272,12 @@ weights**, runs in seconds, and is what CI runs. For throughput on your own mach
 
 chad speaks the same two extension formats as Claude Code:
 
-- **[Agent Skills](https://agentskills.io).** Drop a `SKILL.md` folder in
-  `./.claude/skills/` and it becomes a slash command: `/ship`, `/investigate the flaky test`.
-  Skills cost nothing until you run one, because chad puts no skill catalog in the system
-  prompt.
-- **[MCP servers](https://modelcontextprotocol.io).** Configure stdio or HTTP servers in
+- [Agent Skills](https://agentskills.io). Drop a `SKILL.md` folder in `./.agents/skills/`
+  or `./.claude/skills/` (or under `~/` for every project) and it becomes a slash command:
+  `/ship`, `/investigate the flaky test`. Skills cost nothing until you run one, because
+  chad puts no skill catalog in the system prompt. Precedence and the full list of roots
+  are in the [Configuration reference](docs/configuration.md#agent-skills-agentskillsio).
+- [MCP servers](https://modelcontextprotocol.io). Configure stdio or HTTP servers in
   `./.mcp.json` to expose external tools (GitHub, Postgres, Linear, Slack, …) alongside
   chad's builtins, with static-token and OAuth auth.
 
@@ -259,16 +285,17 @@ Both are covered in full in the [Configuration reference](docs/configuration.md)
 
 ## Documentation
 
-- **[Design & internals](docs/design.md)** covers why prefill is the bill, the persistent
+- [Design & internals](docs/design.md) covers why prefill is the bill, the persistent
   prefix cache, the trimmable/append-only trade, why the tool surface is five tools, and the
   ideas borrowed from other agents.
-- **[Throughput & performance](docs/benchmarks.md)** has the prefill, decode and warm-step
+- [Throughput & performance](docs/benchmarks.md) has the prefill, decode and warm-step
   numbers you can reproduce with `chad-bench`, the stock-engine comparison, and what the
   cross-session warm start is worth.
-- **[Configuration reference](docs/configuration.md)** documents Agent Skills, MCP servers,
-  the context window, every environment variable, and the safety opt-outs.
-- **[Troubleshooting](docs/troubleshooting.md)** is the symptom→knob map for when a session
+- [Configuration reference](docs/configuration.md) documents project instructions, Agent
+  Skills, MCP servers, plan mode, the slash commands, the context window, every environment
+  variable, and the safety opt-outs.
+- [Troubleshooting](docs/troubleshooting.md) maps symptoms to knobs for when a session
   rambles, loops, or slows.
-- **[Contributing](CONTRIBUTING.md)** says what lands easily and what needs a conversation
+- [Contributing](CONTRIBUTING.md) says what lands easily and what needs a conversation
   first.
 

{chad_code-2.0.3 → chad_code-2.2.0}/README.md

@@ -8,9 +8,13 @@
 > *Coding under supervision.*
 
 chad is a single-user coding agent that runs entirely on an Apple Silicon Mac via
-[MLX](https://github.com/ml-explore/mlx). One 27B model and no API key. Every other 
-harness assumes a datacenter on the other end of a socket; chad assumes a laptop, 
-and the whole design falls out of that. (Not affiliated with Anthropic.)
+[MLX](https://github.com/ml-explore/mlx). One 27B model and no API key. (Not affiliated
+with Anthropic.)
+
+Plenty of harnesses run local models now, and [pi](https://pi.dev) is a fantastic default
+for the same reason llama.cpp is: it works with everything. I was steering chad in the
+opposite direction. One model and one set of silicon, taken to the max. Swap out your
+`CHAD_MODEL` and it still runs, you just leave the drafter and the kernels behind.
 
 ## Try it
 
@@ -19,13 +23,13 @@ uvx chad-code          # runs anywhere; the command is still `chad`
 uvx chad-code prove    # offline smoke test: 4 tiny fix-it tasks, verified, timed 🗿
 ```
 
-The first run asks, then downloads the model once (~13 GB) into the shared Hugging
+The first run asks, then downloads the model once (~8 GB) into the shared Hugging
 Face cache. While it downloads, `cd` into a project and think of a scoped first ask:
 *"fix the failing test in `tests/test_x.py`"* lands, *"improve my codebase"* flails.
 
 chad targets 24 GB and nothing smaller. It runs below that and tells you it is doing so,
-but 13 GB of weights sit resident before a single token of context, so a 16 GB Mac gets a
-window too small to work in.
+but 8 GB of weights and drafter plus a ~4 GB prefill transient sit resident before a single
+token of context, so a 16 GB Mac gets a window too small to work in.
 
 The PyPI package is `chad-code`. Bare `chad` is an unrelated squatted package.
 
@@ -38,32 +42,42 @@ The PyPI package is `chad-code`. Bare `chad` is an unrelated squatted package.
 
 ## Same model, same Mac, stock engine
 
-The question worth answering: what do you gain over pointing a generic local-model tool at
-the same weights? Qwen3.8-27B at the same `UD-Q3_K_XL` recipe (Unsloth's GGUF for
-llama.cpp, chad's MLX conversion of the same bit map), the same M4 Pro (24 GB), one engine
-resident at a time, each measured with its own benchmark on a 512-token prompt and a
+What do you gain over pointing a generic local-model tool at the same weights?
+Qwen3.8-27B at the `UD-Q3_K_XL` recipe (Unsloth's GGUF for llama.cpp, chad's MLX
+conversion of the same bit map; chad runs it with `--model`), the same M4 Pro (24 GB), one
+engine resident at a time, each measured with its own benchmark on a 512-token prompt and a
 128-token generation.
 
 | Engine | Prefill (512-tok prompt) | Decode (128 tok) | Speculative decoding |
 |---|---|---|---|
-| llama.cpp `llama-bench` (stock, build 10470) | 102 tok/s | 10.9 tok/s | none for this model |
-| **chad**, serial (`CHAD_NO_DFLASH=1`) | 99 tok/s | 18.1 tok/s | off |
-| **chad**, default | 98 tok/s | **62 tok/s**¹ | DFlash2 block drafter |
-
-A 200-token function body takes roughly 18 seconds at 10.9 tok/s and 3 at 62. That is the
-difference between a batch job and a pair programmer, and closing it is what the project is
-for.
-
-Ollama does not get its own row: it is llama.cpp underneath with no speculative decoding for
-this model, and on the same GGUF (0.32.15, Modelfile `FROM` only) it measures 96 tok/s
+| llama.cpp `llama-bench` (stock, build 10470) | 102 tok/s | 10.9 tok/s | off in this benchmark |
+| llama.cpp `llama-server` (build 10917), serial | 97 tok/s | 11.3 tok/s | off |
+| llama.cpp `llama-server` (build 10917) | 95 tok/s | 11.1 tok/s² | DFlash2 drafter (Q4_K_M GGUF) |
+| **chad**, serial (`CHAD_NO_DFLASH=1`) | 100 tok/s | 17.9 tok/s | off |
+| **chad** | 101 tok/s | **62.9 tok/s**¹ | DFlash2 block drafter |
+
+A 200-token function body takes roughly 18 seconds at 10.9 tok/s and 3 at 63. You wait for
+the first one and you talk to the second.
+
+The weights chad ships by default are the same model in Prism ML's **ternary** build
+(every projection Hadamard-rotated and stored at 2 bits): 7.2 GB resident instead of 12.3,
+the same 64 tok/s drafted and 21 serial on this Mac, and the 5 GB it gives back is
+context, roughly a **150k-token window** where the 3-bit landed near 56k. The like-for-like
+table above stays on the 3-bit recipe because that is the quant llama.cpp has a GGUF of.
+
+Ollama does not get its own row: it is llama.cpp underneath, measured without speculative
+decoding, and on the same GGUF (0.32.15, Modelfile `FROM` only) it measures 96 tok/s
 prefill and the same **10.9** decode.
 
-¹ 62 is a ceiling: `chad-bench`'s prompt is tiled code the drafter reads easily. Replayed
-against ten real mid-session contexts from `~/.chad/sessions` (12–19k tokens, tool results in
+¹ 63 is a ceiling: `chad-bench`'s prompt is tiled code the drafter reads easily. Replayed
+against ten real mid-session contexts from `~/.chad/sessions` (12-19k tokens, tool results in
 place, 384-token decodes) the same engine measures **31.7 tok/s median / 21.4 floor** greedy
 against 14.8 serial, and **27.6 / 17.7** thinking against 13.9. That ~2× is what a session
-lives at. llama.cpp's Metal path for this hybrid architecture was not profiled, so read its
-row as what a fitted engine buys, not as a verdict on llama.cpp.
+lives at.
+
+² llama.cpp has run DFlash2 since build 10658. Here it accepts 96.5% of drafted tokens and still
+gains nothing: verifying 8 tokens costs ~6.4 serial steps on this GGUF and Mac
+([details](docs/benchmarks.md#same-model-same-mac-stock-engine)).
 
 Method, the longer runs and the caveats are in
 [Throughput & performance](docs/benchmarks.md#same-model-same-mac-stock-engine); the rows are
@@ -75,17 +89,17 @@ committed under [`benchmarks/stock/_runs/`](benchmarks/stock/); reproduce them w
 chad owns its inference loop instead of talking to a server, and the engine is fitted to the
 one checkpoint it ships:
 
-- **DFlash2 block speculation.** A bundled 1.9B drafter ([z-lab's DFlash2](https://huggingface.co/z-lab/Qwen3.8-27B-DFlash2),
+- DFlash2 block speculation. A bundled 1.9B drafter ([z-lab's DFlash2](https://huggingface.co/z-lab/Qwen3.8-27B-DFlash2),
   ported to MLX and quantized) proposes a whole block of tokens from the main model's own
   hidden states. The main model verifies the block in one batched forward, and exact
   rejection sampling keeps every emitted token the model's own.
-- **A persistent prefix KV cache.** The transcript is kept a strict token-prefix of the live
+- A persistent prefix KV cache. The transcript is kept a strict token-prefix of the live
   cache, so a follow-up step prefills the ~16 tokens it appended instead of the 5,000 it
-  already read: **~0.75 s per step instead of ~50 s**. Any server with prompt caching gets
+  already read: **~0.55 s per step instead of ~48 s**. Any server with prompt caching gets
   the easy case; the work is holding it true across compaction, truncated turns and restarts.
   The system prefix is checkpointed to disk, so the second session anywhere starts warm
   (75.6 s → 5.5 s to the first tool call).
-- **Fused Metal kernels.** Quantized-KV attention, a small-M matmul for speculative verify,
+- Fused Metal kernels. Quantized-KV attention, a small-M matmul for speculative verify,
   and a compiled single-token layer step, chosen per machine at load time, no knobs.
 
 ## What chad gives up
@@ -111,16 +125,21 @@ budget; everything else is deliberately plain.
 
 `uv run chad` launches a full-screen terminal UI (built on prompt_toolkit):
 
-- **shift-tab cycles permission modes**: `normal` (confirm each bash/write/edit) →
+- shift-tab cycles permission modes: `normal` (confirm each bash/write/edit) →
   `auto-accept edits` (edits land silently, **terminal commands still ask**) → `yolo`
   (nothing asks) → `plan mode` (read-only: investigate and propose a numbered plan) → back.
-- **type-ahead message queue.** Keep typing while the agent works; messages run in order.
-- **ctrl-c interrupts the running turn** without killing the session. **↑prefilled /
+  A finished plan lands in `./plans/`; ctrl-g (or `/accept`) clears the context and starts
+  implementing it ([details](docs/configuration.md#plan-mode)).
+- Type-ahead message queue. Keep typing while the agent works; messages run in order.
+- ctrl-c interrupts the running turn without killing the session. **↑prefilled /
   ↓generated** token counts show an advancing **%** on an unavoidable full re-prefill, so it
   is never silent.
-- **`@file` / `@dir` mentions** and **`!command` shell passthrough.** Pull a file into
+- `@file` / `@dir` mentions and `!command` shell passthrough. Pull a file into
   context inline, or run a shell command without invoking the model.
-- **voice mode, all local.** `/speech`, then ctrl-t to talk: Parakeet-on-MLX transcribes into
+- Standing project instructions. A `CLAUDE.md` (or `AGENTS.md`) in the working directory is
+  appended to the system prompt, and `/init` reads the project and writes one for you
+  ([details](docs/configuration.md#project-instructions-claudemd--agentsmd)).
+- Voice mode, all local. `/speech`, then ctrl-t to talk: Parakeet-on-MLX transcribes into
   the input box for you to review before Enter sends it, and replies are read aloud via macOS
   `say`. A word table teaches it your identifiers. Needs the `speech` extra
   ([details](docs/configuration.md#voice-mode-speech)).
@@ -134,12 +153,13 @@ budget; everything else is deliberately plain.
 | `--plan` | start in read-only plan mode (investigate and propose, edits blocked) |
 | `--yolo` | auto-approve bash/write/edit (skip confirm prompts) |
 | `--no-think` | skip the model's `<think>` blocks, faster on well-scoped work |
+| `--think-budget N` | soft-cap each step's `<think>` at N tokens, force-close it and carry on (off by default) |
+| `--backend llama` | run the same harness against a remote llama.cpp server, with `--base-url`, `--tokenizer` and `--api-key-env` ([details](docs/configuration.md#alternate-backend-remote)) |
 | `--model` | `auto` (the shipped default), or any HF repo id / local model dir |
 | `--repl` | plain line REPL instead of the TUI |
 
-Three subcommands, each with its own `--help`: `chad prove` (the offline smoke test),
-`chad serve` ([serve this Mac's model to a container or the LAN](docs/configuration.md#serving-the-local-model-to-a-container-chad-serve)),
-and `chad levers` (print the result-channel lever registry as JSON, for A/B ablation).
+Two subcommands, each with its own `--help`: `chad prove` (the offline smoke test) and
+`chad levers` (print the result-channel lever registry as JSON, for A/B ablation).
 
 A headless task (positional, or piped with no TTY) auto-approves mutating tools and runs
 greedy (temp 0). Every conversation is persisted under `~/.chad/sessions/`, and every resume
@@ -152,20 +172,24 @@ chad ships exactly one, downloaded once into the shared Hugging Face cache
 
 | Model | Quant | Footprint |
 |---|---|---|
-| [Qwen3.8-27B `UD-Q3_K_XL-DFlash2`](https://huggingface.co/nathansutton/Qwen3.8-27B-UD-Q3_K_XL-DFlash2-MLX) | 3-bit group-64 body, 5-bit `lm_head`, bundled 4-bit DFlash2 drafter | ~13 GB resident, 262k native context |
+| [Qwen3.8-27B `Ternary-Bonsai-2-DFlash2`](https://huggingface.co/nathansutton/Qwen3.8-27B-Ternary-Bonsai-2-DFlash2-MLX) (default) | Prism ML's ternary build: Hadamard-rotated, 2-bit group-128, levels {−s, 0, +s}; bundled 4-bit DFlash2 drafter | ~8 GB resident, 262k native context, ~150k usable on 24 GB |
+| [Qwen3.8-27B `UD-Q3_K_XL-DFlash2`](https://huggingface.co/nathansutton/Qwen3.8-27B-UD-Q3_K_XL-DFlash2-MLX) via `--model` | 3-bit group-64 body, 5-bit `lm_head`, the same bundled drafter | ~13 GB resident, ~56k usable on 24 GB |
 
 Qwen3.8-27B is **dense** (64 layers: 48 GatedDeltaNet + 16 full attention), so every
-parameter is on the critical path for every token and the quant is where decode speed comes
-from. The bits go where held-out perplexity says they pay: `lm_head` is a second full
-1.27B-param tensor and is held at 5-bit, while `embed_tokens` is a lookup table whose error
-never compounds through a matmul, so it is cheapest. The name follows
-[Unsloth's convention](https://docs.unsloth.ai/) (`UD-…`), though the quant is MLX group-64
-affine, not a llama.cpp k-quant. The drafter ships in the same repo, pre-quantized.
+parameter is on the critical path for every token and the quant is where both decode speed
+and context come from: the governor prices a gigabyte of weights at about 29k tokens of
+window. The ternary build is Prism ML's [Bonsai](https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-mlx-2bit)
+pack of this checkpoint, repacked text-only with the base tokenizer and the drafter
+bundled (Apache-2.0; created using Bonsai by Prism ML). Its projections are stored in a
+rotated basis that an ordinary MLX loader would silently get wrong, so chad carries its own
+loader for it and fits the decode fast-path, the verify kernel and the drafter to it. What it
+costs: on code, teacher-forced perplexity is 4.49 against the 3-bit's 3.99 (+12%), while
+the private eval tiers tie at 56/56. The 3-bit stays one flag away.
 
 `--model <repo or local dir>` runs different weights through the same engine and stays a
 first-class escape hatch. The drafter, the fused-attention coverage, the decode fastpath and
-the context governor are all fitted to the shipped checkpoint, so other weights run slower;
-they do not break.
+the context governor are all fitted to the two checkpoints above, so other weights run
+slower; they do not break.
 
 ## Installing & upgrading
 
@@ -185,7 +209,7 @@ uv run chad "add a --json flag to main.py and update the tests"   # one-shot, he
 uv run chad -c               # resume this directory's last conversation
 ```
 
-**Optional extras.** Two features are opt-in because they pull deps not every install wants:
+Two features are opt-in because they pull deps not every install wants:
 `speech` (voice mode: a mic library, no torch) and `highlight` (syntax colour in diffs and
 previews). An extra rides on the install spec, not on a separate command, so how you add it
 depends on how you installed chad:
@@ -199,12 +223,12 @@ uv sync --extra speech                        # from a clone
 `/speech` in the TUI prints whichever of those matches your install, so you never have to
 work it out from here.
 
-**Upgrading** depends on how you installed: `uv tool upgrade chad-code`, `uvx --refresh
+Upgrading depends on how you installed: `uv tool upgrade chad-code`, `uvx --refresh
 chad-code`, or `git pull && uv sync` for a clone. What changed lands in
 [`CHANGELOG.md`](CHANGELOG.md). Model weights are versioned separately, so a code upgrade
 never re-downloads the model.
 
-**Development.** `uv sync` once, then `uv run pytest -q`. The fast unit gate loads **no model
+For development, `uv sync` once, then `uv run pytest -q`. The fast unit gate loads **no model
 weights**, runs in seconds, and is what CI runs. For throughput on your own machine, use
 `uv run chad-bench` (see [Throughput & performance](docs/benchmarks.md)).
 
@@ -212,11 +236,12 @@ weights**, runs in seconds, and is what CI runs. For throughput on your own mach
 
 chad speaks the same two extension formats as Claude Code:
 
-- **[Agent Skills](https://agentskills.io).** Drop a `SKILL.md` folder in
-  `./.claude/skills/` and it becomes a slash command: `/ship`, `/investigate the flaky test`.
-  Skills cost nothing until you run one, because chad puts no skill catalog in the system
-  prompt.
-- **[MCP servers](https://modelcontextprotocol.io).** Configure stdio or HTTP servers in
+- [Agent Skills](https://agentskills.io). Drop a `SKILL.md` folder in `./.agents/skills/`
+  or `./.claude/skills/` (or under `~/` for every project) and it becomes a slash command:
+  `/ship`, `/investigate the flaky test`. Skills cost nothing until you run one, because
+  chad puts no skill catalog in the system prompt. Precedence and the full list of roots
+  are in the [Configuration reference](docs/configuration.md#agent-skills-agentskillsio).
+- [MCP servers](https://modelcontextprotocol.io). Configure stdio or HTTP servers in
   `./.mcp.json` to expose external tools (GitHub, Postgres, Linear, Slack, …) alongside
   chad's builtins, with static-token and OAuth auth.
 
@@ -224,16 +249,17 @@ Both are covered in full in the [Configuration reference](docs/configuration.md)
 
 ## Documentation
 
-- **[Design & internals](docs/design.md)** covers why prefill is the bill, the persistent
+- [Design & internals](docs/design.md) covers why prefill is the bill, the persistent
   prefix cache, the trimmable/append-only trade, why the tool surface is five tools, and the
   ideas borrowed from other agents.
-- **[Throughput & performance](docs/benchmarks.md)** has the prefill, decode and warm-step
+- [Throughput & performance](docs/benchmarks.md) has the prefill, decode and warm-step
   numbers you can reproduce with `chad-bench`, the stock-engine comparison, and what the
   cross-session warm start is worth.
-- **[Configuration reference](docs/configuration.md)** documents Agent Skills, MCP servers,
-  the context window, every environment variable, and the safety opt-outs.
-- **[Troubleshooting](docs/troubleshooting.md)** is the symptom→knob map for when a session
+- [Configuration reference](docs/configuration.md) documents project instructions, Agent
+  Skills, MCP servers, plan mode, the slash commands, the context window, every environment
+  variable, and the safety opt-outs.
+- [Troubleshooting](docs/troubleshooting.md) maps symptoms to knobs for when a session
   rambles, loops, or slows.
-- **[Contributing](CONTRIBUTING.md)** says what lands easily and what needs a conversation
+- [Contributing](CONTRIBUTING.md) says what lands easily and what needs a conversation
   first.