v0.4.0, the parser fixes

Two Read calls in one reply used to become one Read call. The parser deduplicated tool calls by tool
name only, so when a model asked to read two files at once, the second request vanished and nobody got
an error. It now deduplicates on name plus arguments (#64). That was the first of eight fixes to
proxy/server.py that went in on Oct 9, most of them the same kind of thing: the server quietly
throwing away something the model or Claude Code sent.

The rest of that batch:

Gemma 4 kept losing its number and boolean arguments. It writes strings as key:<|"|>value<|"|> and
everything else bare, and once a string matched, the bare values were never read, so Edit lost
replace_all and Read lost offset and limit (#65). Gemma 4 arguments are now read as Gemma's own
JSON dialect first, which also fixes nested values like TodoWrite's todos and dashed keys like Grep's
-i, with the old regexes kept as a fallback (#68). Arguments sent as a JSON string, OpenAI style, are
decoded instead of arriving as empty input (#68).

A JSON array inside <tool_call> tags used to crash the request with a 500. It is now read as a list of
calls (#66).

When Claude Code's Read tool opened an image, the whole base64 payload went into the prompt as text.
One screenshot was enough to push a 32 GB Mac into swap. Images and documents in a tool result now
become a short marker (#67).

Errors now come back in the Anthropic shape ({"type": "error", "error": {"type": "api_error"}}), and a
request body that isn't valid JSON gets a 400 instead of a dropped connection (#69).

stop_sequences is honored on both the buffered and the streaming path, a failed tool's is_error
reaches the model as Error: ..., and a request that declares no tools no longer gets tool calls
parsed out of its JSON answer (#70).

The live stream now filters <think> blocks the same way the buffered path does, and the tool-call
retry runs on a throwaway cache so it no longer sees the conversation twice or changes the cache the
next request reuses (#71).

Earlier in the cycle @zwolf25 fixed two code-mode bugs from #54 (#56). Code mode's tool filter dropped
Skill, so slash commands could never call the skill they named. And fenced JSON tool calls shaped as
an array, which is what Hermes 4 14B emits, were skipped. Both fixed, and that PR added
scripts/test_parse_tool_calls.py.

Tests you can run without a model

There are now five test scripts that stub out MLX and need no model and no running server:

python3 scripts/test_parse_tool_calls.py
python3 scripts/test_convert_messages.py
python3 scripts/test_generate_response.py
python3 scripts/test_http_errors.py
python3 scripts/test_stream_and_retry.py

On this commit they report 38 passing checks between them (20, 1, 6, 4 and 7).
scripts/test_mlx_server.py still needs a server on localhost:4000.

Keep going when Claude runs out

bin/keepgoing finds the last Claude Code conversation in the current folder and carries it on with
another model: a free OpenRouter model if you add a key, or one of the local models through this repo's
server. setup.sh links it in and adds a keep going shell command (scripts/keepgoing-shell.sh).
scripts/statusline-keepgoing.sh is an optional status line that reminds you it's there. Picking a
cloud model in that menu sends the conversation to OpenRouter; the local ones don't leave the Mac.

New model lineup in setup

setup.sh now picks by memory:

Memory	Model
96 GB and up	Qwen 3.8 27B, 8-bit (lmstudio-community/Qwen3.8-27B-MLX-8bit)
64 to 95 GB	Gemma 4 31B, 4-bit
32 to 63 GB	Gemma 4 12B, 4-bit
16 to 31 GB	Hermes 4 14B, 4-bit
under 16 GB	Gemma 4 E4B, 4-bit (labeled: it failed our Claude Code smoke test)

Qwen 3.8 27B 8-bit passed the Claude Code smoke test (write hello.py, run it, report) in 38 s on an
M5 Max 128 GB. On Agent-12 with the DFlash 2 drafter it went 12/12 easy in 122 s and 7/8 hard in
401 s (Sep 19 rerun). Those Agent-12 numbers come from Agent-12's own harness, not from inside Claude
Code. scripts/doctor.sh gives the same recommendations as setup now.

If you're upgrading

bash scripts/download-and-import.sh qwen now downloads Qwen 3.8 27B 8-bit (~29 GB). It used to
download Qwen 3.5 122B. gemma12 and hermes are new shortcuts. llama still works for existing
setups.

The Llama 70B.command launcher is gone and Qwen 3.8 Code.command replaces it. Qwen 3.5 122B,
Qwen 3.5 4B and DeepSeek through ds4 are no longer offered by setup.

Docs

The README was rewritten for people arriving for the first time, and the old long README moved,
unchanged apart from relative links, to docs/FULL-GUIDE.md. New CONTRIBUTING.md. A section on
other ways to run Claude Code against other models (Ollama's Anthropic API, claude-code-router,
claude-code-proxy). The lsof privacy check in the README now lists network connections only (#62).
@pratikgx fixed a broken link in docs/PHONE-CONTROL.md (#60). bin/fia is a single entry point that
runs the existing launchers.

This release stands on Anthropic's Claude Code, Apple's MLX and mlx-lm, and the model teams at Google
(Gemma), Qwen, Nous Research (Hermes), lmstudio-community and mlx-community for the builds. Thanks to
@zwolf25 and @pratikgx for the PRs, and to everyone who filed issues.