augint-shell 0.78.0__tar.gz → 0.80.0__tar.gz

{augint_shell-0.78.0 → augint_shell-0.80.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: augint-shell
-Version: 0.78.0
+Version: 0.80.0
 Summary: Launch AI coding tools and local LLMs in Docker containers
 Author: svange
 Requires-Dist: docker>=7.0.0
@@ -96,29 +96,67 @@ ai-shell opencode
 
 ## Configuration
 
-Optional `ai-shell.toml` in your project root:
-
-```toml
-[container]
-image = "svange/augint-shell"
-image_tag = "latest"
-extra_env = { MY_VAR = "value" }
-
-[llm]
-primary_model = "qwen3-coder:30b-a3b-q4_K_M"
-fallback_model = "huihui_ai/llama3.3-abliterated"
-context_size = 32768
-ollama_port = 11434
-webui_port = 3000
+Optional `.ai-shell.yaml` in your project root (YAML is the default; TOML is
+also accepted — see `ai-shell init` for the full generated template with
+per-section rationale):
+
+```yaml
+container:
+  image: svange/augint-shell
+  image_tag: latest
+  extra_env:
+    MY_VAR: value
+
+llm:
+  primary_chat_model: qwen3.5:27b
+  secondary_chat_model: huihui_ai/qwen3.5-abliterated:27b
+  primary_coding_model: qwen3-coder:30b-a3b-q4_K_M
+  secondary_coding_model: huihui_ai/qwen3-coder-abliterated:30b-a3b-instruct-q4_K_M
+  context_size: 32768
+  ollama_port: 11434
+  webui_port: 3000
+  extra_models: []   # additional Ollama tags to pull alongside the 4 slots
 ```
 
-Global config at `~/.config/ai-shell/config.toml` is also supported.
+Global config at `~/.ai-shell.yaml` or `~/.config/ai-shell/config.yaml` is
+also supported.
+
+> The previous `primary_model` / `fallback_model` keys were removed. They were
+> role-ambiguous (chat vs. coding). If you had them set, move them to the
+> matching slot above. ai-shell will refuse to start with those legacy keys
+> present and print a migration hint.
 
 `ai-shell` does not manage tool-specific config files for Codex, OpenCode, or
 Aider. Use `augint-opencodex` or the tools' native config files for those, and
 use `ai-shell` for container/runtime settings such as AWS profiles, local LLM
 ports, and Claude options.
 
+### Local LLM stack
+
+Four role-specific model slots, each sized for an RTX 4090 (24 GiB VRAM). All
+four defaults together total ~74 GB on disk.
+
+| Slot | Default | Size | Role | Routed to |
+|---|---|---|---|---|
+| `primary_chat_model` | `qwen3.5:27b` | 17 GB | Best chat model that fits a 4090 | Open WebUI default |
+| `secondary_chat_model` | `huihui_ai/qwen3.5-abliterated:27b` | 17 GB | Best uncensored chat (abliterated Qwen3.5) | Open WebUI (selectable) |
+| `primary_coding_model` | `qwen3-coder:30b-a3b-q4_K_M` | 19 GB | Best agentic coder with explicit Ollama tools badge | OpenCode / Aider default |
+| `secondary_coding_model` | `huihui_ai/qwen3-coder-abliterated:30b-a3b-instruct-q4_K_M` | 19 GB | Best uncensored coder (abliterated Qwen3-Coder) | OpenCode (selectable) |
+
+Each pair shares a base model — primary is the standard aligned release;
+secondary is the huihui.ai abliterated variant (refusal directions neutralized
+via weight surgery, benchmark quality preserved). Switching primary <->
+secondary within a slot keeps tool formats and context semantics identical.
+
+`ai-shell llm pull` / `ai-shell llm setup` downloads all 4 slots plus any
+`extra_models` entries, deduped.
+
+**Three caveats worth knowing:**
+
+1. **Qwen3.5 Ollama tool calling is broken** ([ollama #14493](https://github.com/ollama/ollama/issues/14493), open). This does not affect Open WebUI's default chat with web search and RAG — those run server-side in WebUI without touching Ollama's tools API. It does affect agent CLIs routed through Ollama's `/v1/chat/completions` tools array, which is why the chat slots are Qwen3.5 and the coding slots are Qwen3-Coder (explicit tools badge, working parser).
+2. **Ollama `num_ctx` defaults to 4096** for every model, well below what modern agent prompts need (Claude Code sends ~35K tokens). `context_size` in your config is applied via Modelfile override during `llm setup` — leave it at 32768 unless you have a reason.
+3. **Qwen3-Coder tool-count cliff**: reliable native `tool_calls` emission below ~5 registered tools; above that the model may emit XML inside content and some parsers miss it. Keep agent tool sets tight.
+
 ## How It Works
 
 - Pulls a pre-built Docker image from Docker Hub (`svange/augint-shell`)

{augint_shell-0.78.0 → augint_shell-0.80.0}/README.md

@@ -83,29 +83,67 @@ ai-shell opencode
 
 ## Configuration
 
-Optional `ai-shell.toml` in your project root:
-
-```toml
-[container]
-image = "svange/augint-shell"
-image_tag = "latest"
-extra_env = { MY_VAR = "value" }
-
-[llm]
-primary_model = "qwen3-coder:30b-a3b-q4_K_M"
-fallback_model = "huihui_ai/llama3.3-abliterated"
-context_size = 32768
-ollama_port = 11434
-webui_port = 3000
+Optional `.ai-shell.yaml` in your project root (YAML is the default; TOML is
+also accepted — see `ai-shell init` for the full generated template with
+per-section rationale):
+
+```yaml
+container:
+  image: svange/augint-shell
+  image_tag: latest
+  extra_env:
+    MY_VAR: value
+
+llm:
+  primary_chat_model: qwen3.5:27b
+  secondary_chat_model: huihui_ai/qwen3.5-abliterated:27b
+  primary_coding_model: qwen3-coder:30b-a3b-q4_K_M
+  secondary_coding_model: huihui_ai/qwen3-coder-abliterated:30b-a3b-instruct-q4_K_M
+  context_size: 32768
+  ollama_port: 11434
+  webui_port: 3000
+  extra_models: []   # additional Ollama tags to pull alongside the 4 slots
 ```
 
-Global config at `~/.config/ai-shell/config.toml` is also supported.
+Global config at `~/.ai-shell.yaml` or `~/.config/ai-shell/config.yaml` is
+also supported.
+
+> The previous `primary_model` / `fallback_model` keys were removed. They were
+> role-ambiguous (chat vs. coding). If you had them set, move them to the
+> matching slot above. ai-shell will refuse to start with those legacy keys
+> present and print a migration hint.
 
 `ai-shell` does not manage tool-specific config files for Codex, OpenCode, or
 Aider. Use `augint-opencodex` or the tools' native config files for those, and
 use `ai-shell` for container/runtime settings such as AWS profiles, local LLM
 ports, and Claude options.
 
+### Local LLM stack
+
+Four role-specific model slots, each sized for an RTX 4090 (24 GiB VRAM). All
+four defaults together total ~74 GB on disk.
+
+| Slot | Default | Size | Role | Routed to |
+|---|---|---|---|---|
+| `primary_chat_model` | `qwen3.5:27b` | 17 GB | Best chat model that fits a 4090 | Open WebUI default |
+| `secondary_chat_model` | `huihui_ai/qwen3.5-abliterated:27b` | 17 GB | Best uncensored chat (abliterated Qwen3.5) | Open WebUI (selectable) |
+| `primary_coding_model` | `qwen3-coder:30b-a3b-q4_K_M` | 19 GB | Best agentic coder with explicit Ollama tools badge | OpenCode / Aider default |
+| `secondary_coding_model` | `huihui_ai/qwen3-coder-abliterated:30b-a3b-instruct-q4_K_M` | 19 GB | Best uncensored coder (abliterated Qwen3-Coder) | OpenCode (selectable) |
+
+Each pair shares a base model — primary is the standard aligned release;
+secondary is the huihui.ai abliterated variant (refusal directions neutralized
+via weight surgery, benchmark quality preserved). Switching primary <->
+secondary within a slot keeps tool formats and context semantics identical.
+
+`ai-shell llm pull` / `ai-shell llm setup` downloads all 4 slots plus any
+`extra_models` entries, deduped.
+
+**Three caveats worth knowing:**
+
+1. **Qwen3.5 Ollama tool calling is broken** ([ollama #14493](https://github.com/ollama/ollama/issues/14493), open). This does not affect Open WebUI's default chat with web search and RAG — those run server-side in WebUI without touching Ollama's tools API. It does affect agent CLIs routed through Ollama's `/v1/chat/completions` tools array, which is why the chat slots are Qwen3.5 and the coding slots are Qwen3-Coder (explicit tools badge, working parser).
+2. **Ollama `num_ctx` defaults to 4096** for every model, well below what modern agent prompts need (Claude Code sends ~35K tokens). `context_size` in your config is applied via Modelfile override during `llm setup` — leave it at 32768 unless you have a reason.
+3. **Qwen3-Coder tool-count cliff**: reliable native `tool_calls` emission below ~5 registered tools; above that the model may emit XML inside content and some parsers miss it. Keep agent tool sets tight.
+
 ## How It Works
 
 - Pulls a pre-built Docker image from Docker Hub (`svange/augint-shell`)

{augint_shell-0.78.0 → augint_shell-0.80.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "augint-shell"
-version = "0.78.0"
+version = "0.80.0"
 description = "Launch AI coding tools and local LLMs in Docker containers"
 authors = [{name = "svange"}]
 readme = "README.md"

{augint_shell-0.78.0 → augint_shell-0.80.0}/src/ai_shell/__init__.py

@@ -1,6 +1,6 @@
 """augint-shell (ai-shell) - Launch AI coding tools and local LLMs in Docker containers."""
 
-__version__ = "0.78.0"
+__version__ = "0.80.0"
 
 __all__ = [
     "__version__",

{augint_shell-0.78.0 → augint_shell-0.80.0}/src/ai_shell/cli/commands/llm.py

@@ -119,8 +119,8 @@ def _validate_models_or_abort(*model_refs: str) -> None:
     for ref in missing:
         console.print(f"  - [cyan]{ref}[/cyan]  (tags: {_tag_list_url(ref)})")
     console.print(
-        "\nUpdate [bold]primary_model[/bold] / [bold]fallback_model[/bold] in "
-        "your ai-shell config to a valid tag and retry."
+        "\nUpdate the relevant [bold]*_chat_model[/bold] / [bold]*_coding_model[/bold] "
+        "entry (or [bold]extra_models[/bold]) in your ai-shell config to a valid tag and retry."
     )
     raise click.Abort()
 
@@ -398,15 +398,13 @@ def llm_pull(ctx):
     manager = _get_manager(ctx)
     config = manager.config
 
-    _validate_models_or_abort(config.primary_model, config.fallback_model)
+    models = config.models_to_pull
+    _validate_models_or_abort(*models)
 
-    console.print(f"[bold]Pulling primary model: {config.primary_model}...[/bold]")
-    output = manager.exec_in_ollama(["ollama", "pull", config.primary_model])
-    console.print(output)
-
-    console.print(f"\n[bold]Pulling fallback model: {config.fallback_model}...[/bold]")
-    output = manager.exec_in_ollama(["ollama", "pull", config.fallback_model])
-    console.print(output)
+    for model in models:
+        console.print(f"[bold]Pulling {model}...[/bold]")
+        output = manager.exec_in_ollama(["ollama", "pull", model])
+        console.print(output)
 
     console.print("\n[bold]Available models:[/bold]")
     output = manager.exec_in_ollama(["ollama", "list"])
@@ -426,7 +424,8 @@ def llm_setup(ctx, webui: bool, voice: bool, no_voice: bool, n8n: bool, all_: bo
     manager = _get_manager(ctx)
     config = manager.config
 
-    _validate_models_or_abort(config.primary_model, config.fallback_model)
+    models = config.models_to_pull
+    _validate_models_or_abort(*models)
 
     console.print("[bold]Starting LLM stack...[/bold]")
     _warn_if_low_memory()
@@ -452,13 +451,10 @@ def llm_setup(ctx, webui: bool, voice: bool, no_voice: bool, n8n: bool, all_: bo
         console.print("[bold red]Ollama failed to start after 20s[/bold red]")
         raise click.Abort()
 
-    console.print(f"\n[bold]Pulling primary model: {config.primary_model}...[/bold]")
-    output = manager.exec_in_ollama(["ollama", "pull", config.primary_model])
-    console.print(output)
-
-    console.print(f"\n[bold]Pulling fallback model: {config.fallback_model}...[/bold]")
-    output = manager.exec_in_ollama(["ollama", "pull", config.fallback_model])
-    console.print(output)
+    for model in models:
+        console.print(f"\n[bold]Pulling {model}...[/bold]")
+        output = manager.exec_in_ollama(["ollama", "pull", model])
+        console.print(output)
 
     console.print("\n[bold green]============================================[/bold green]")
     console.print("[bold green] Setup complete![/bold green]")
@@ -469,9 +465,13 @@ def llm_setup(ctx, webui: bool, voice: bool, no_voice: bool, n8n: bool, all_: bo
         console.print(f"  Open WebUI:  http://localhost:{config.webui_port}")
     if n8n:
         console.print(f"  n8n:         http://localhost:{config.n8n_port}")
-    console.print(f"\n  Primary model:  {config.primary_model}")
-    console.print(f"  Fallback model: {config.fallback_model}")
-    console.print(f"  Context window: {config.context_size} tokens")
+    console.print(f"\n  Primary chat:      {config.primary_chat_model}")
+    console.print(f"  Secondary chat:    {config.secondary_chat_model}")
+    console.print(f"  Primary coding:    {config.primary_coding_model}")
+    console.print(f"  Secondary coding:  {config.secondary_coding_model}")
+    if config.extra_models:
+        console.print(f"  Extra models:      {', '.join(config.extra_models)}")
+    console.print(f"  Context window:    {config.context_size} tokens")
     console.print("[bold green]============================================[/bold green]")
 
 
@@ -530,9 +530,13 @@ def llm_status(ctx):
         console.print(f"  n8n:                http://{lan}:{config.n8n_port}")
 
     console.print("\n[bold]Configuration:[/bold]")
-    console.print(f"  Primary model:   {config.primary_model}")
-    console.print(f"  Fallback model:  {config.fallback_model}")
-    console.print(f"  Context window:  {config.context_size} tokens")
+    console.print(f"  Primary chat:      {config.primary_chat_model}")
+    console.print(f"  Secondary chat:    {config.secondary_chat_model}")
+    console.print(f"  Primary coding:    {config.primary_coding_model}")
+    console.print(f"  Secondary coding:  {config.secondary_coding_model}")
+    if config.extra_models:
+        console.print(f"  Extra models:      {', '.join(config.extra_models)}")
+    console.print(f"  Context window:    {config.context_size} tokens")
 
     vram = get_vram_info()
     if vram is not None:

{augint_shell-0.78.0 → augint_shell-0.80.0}/src/ai_shell/cli/commands/tools.py

@@ -1154,6 +1154,11 @@ def opencode(
     manager.ensure_tool_fresh(name, "opencode")
 
     cmd = ["/root/.opencode/bin/opencode"]
+    if not use_bedrock:
+        # Default OpenCode to the primary coding slot (benchmark-optimized,
+        # explicit Ollama tools badge). Users can switch to the secondary
+        # (uncensored) slot in the OpenCode model picker at runtime.
+        cmd.extend(["--model", f"ollama/{config.primary_coding_model}"])
     console.print(f"[bold]Launching opencode{bedrock_label} in {name}...[/bold]")
     manager.exec_interactive(name, cmd, extra_env=exec_env)
 
@@ -1165,7 +1170,7 @@ def opencode(
 def aider(ctx, safe, extra_args):
     """Launch aider with local LLM in the dev container."""
     manager, name, exec_env, config = _get_manager(ctx)
-    aider_model = f"ollama_chat/{config.primary_model}"
+    aider_model = f"ollama_chat/{config.primary_coding_model}"
     cmd = ["aider", "--model", aider_model]
     if not safe:
         cmd.append("--yes-always")

{augint_shell-0.78.0 → augint_shell-0.80.0}/src/ai_shell/config.py

@@ -24,13 +24,15 @@ from ai_shell import __version__
 from ai_shell.defaults import (
     DEFAULT_CONTEXT_SIZE,
     DEFAULT_DEV_PORTS,
-    DEFAULT_FALLBACK_MODEL,
     DEFAULT_IMAGE,
     DEFAULT_KOKORO_PORT,
     DEFAULT_KOKORO_VOICE,
     DEFAULT_N8N_PORT,
     DEFAULT_OLLAMA_PORT,
-    DEFAULT_PRIMARY_MODEL,
+    DEFAULT_PRIMARY_CHAT_MODEL,
+    DEFAULT_PRIMARY_CODING_MODEL,
+    DEFAULT_SECONDARY_CHAT_MODEL,
+    DEFAULT_SECONDARY_CODING_MODEL,
     DEFAULT_WEBUI_PORT,
 )
 
@@ -47,9 +49,15 @@ class AiShellConfig:
     project_name: str = ""
     project_dir: Path = field(default_factory=Path.cwd)
 
-    # LLM
-    primary_model: str = DEFAULT_PRIMARY_MODEL
-    fallback_model: str = DEFAULT_FALLBACK_MODEL
+    # LLM model slots. Primary = best-available; secondary = best uncensored
+    # alternative. Chat slots are routed to Open WebUI, coding slots to
+    # OpenCode / Aider. `extra_models` is a free-form list of additional
+    # Ollama tags to pull alongside the 4 slots (deduped).
+    primary_chat_model: str = DEFAULT_PRIMARY_CHAT_MODEL
+    secondary_chat_model: str = DEFAULT_SECONDARY_CHAT_MODEL
+    primary_coding_model: str = DEFAULT_PRIMARY_CODING_MODEL
+    secondary_coding_model: str = DEFAULT_SECONDARY_CODING_MODEL
+    extra_models: list[str] = field(default_factory=list)
     context_size: int = DEFAULT_CONTEXT_SIZE
     ollama_port: int = DEFAULT_OLLAMA_PORT
     webui_port: int = DEFAULT_WEBUI_PORT
@@ -85,6 +93,28 @@ class AiShellConfig:
         """Return deduplicated, sorted list of dev container ports to expose."""
         return sorted(set(DEFAULT_DEV_PORTS + self.extra_ports))
 
+    @property
+    def models_to_pull(self) -> list[str]:
+        """Return the full deduped list of Ollama model tags to pull.
+
+        The 4 slots in order, followed by any ``extra_models``. Duplicates
+        are removed while preserving first-occurrence order.
+        """
+        ordered = [
+            self.primary_chat_model,
+            self.secondary_chat_model,
+            self.primary_coding_model,
+            self.secondary_coding_model,
+            *self.extra_models,
+        ]
+        seen: set[str] = set()
+        deduped: list[str] = []
+        for model in ordered:
+            if model and model not in seen:
+                seen.add(model)
+                deduped.append(model)
+        return deduped
+
 
 def load_config(
     project_override: str | None = None,
@@ -151,6 +181,39 @@ def _load_config_file(path: Path) -> dict:
         return tomllib.load(f)
 
 
+_LEGACY_LLM_KEY_HINT = {
+    "primary_model": (
+        "renamed to `primary_coding_model` (coding) or `primary_chat_model` "
+        "(chat). The new config uses 4 role-specific slots; pick the one "
+        "that matches your intent. See the generated .ai-shell.yaml for the "
+        "full layout."
+    ),
+    "fallback_model": (
+        "removed. The previous `fallback_model` was role-ambiguous. Use "
+        "`secondary_chat_model` and `secondary_coding_model` instead "
+        "(both default to the best uncensored variants). See the generated "
+        ".ai-shell.yaml for the full layout."
+    ),
+}
+
+
+def _reject_legacy_llm_keys(llm_section: dict, path: Path) -> None:
+    """Raise on deprecated `primary_model` / `fallback_model` keys.
+
+    These were removed when the llm config split into 4 role-specific slots
+    (primary/secondary x chat/coding). Silently aliasing them would corrupt
+    intent — e.g. the old `fallback_model` meant different things to chat and
+    coding users. Fail loudly with migration guidance.
+    """
+    bad = [k for k in _LEGACY_LLM_KEY_HINT if k in llm_section]
+    if not bad:
+        return
+    lines = [f"\nDeprecated llm key(s) found in {path}:"]
+    for key in bad:
+        lines.append(f"  - `{key}`: {_LEGACY_LLM_KEY_HINT[key]}")
+    raise ValueError("\n".join(lines))
+
+
 def _apply_config(config: AiShellConfig, path: Path) -> None:
     """Apply settings from a YAML or TOML config file."""
     try:
@@ -176,10 +239,17 @@ def _apply_config(config: AiShellConfig, path: Path) -> None:
 
     # [llm] section
     llm = data.get("llm", {})
-    if "primary_model" in llm:
-        config.primary_model = llm["primary_model"]
-    if "fallback_model" in llm:
-        config.fallback_model = llm["fallback_model"]
+    _reject_legacy_llm_keys(llm, path)
+    if "primary_chat_model" in llm:
+        config.primary_chat_model = llm["primary_chat_model"]
+    if "secondary_chat_model" in llm:
+        config.secondary_chat_model = llm["secondary_chat_model"]
+    if "primary_coding_model" in llm:
+        config.primary_coding_model = llm["primary_coding_model"]
+    if "secondary_coding_model" in llm:
+        config.secondary_coding_model = llm["secondary_coding_model"]
+    if "extra_models" in llm:
+        config.extra_models.extend(str(m) for m in llm["extra_models"])
     if "context_size" in llm:
         config.context_size = int(llm["context_size"])
     if "ollama_port" in llm:
@@ -214,14 +284,29 @@ def _apply_config(config: AiShellConfig, path: Path) -> None:
         config.skip_updates = bool(container["skip_updates"])
 
 
+_LEGACY_ENV_VARS = {
+    "AI_SHELL_PRIMARY_MODEL": ("AI_SHELL_PRIMARY_CODING_MODEL or AI_SHELL_PRIMARY_CHAT_MODEL"),
+    "AI_SHELL_FALLBACK_MODEL": ("AI_SHELL_SECONDARY_CHAT_MODEL or AI_SHELL_SECONDARY_CODING_MODEL"),
+}
+
+
 def _apply_env_vars(config: AiShellConfig) -> None:
     """Apply AI_SHELL_* environment variable overrides."""
+    bad_env = [k for k in _LEGACY_ENV_VARS if os.environ.get(k) is not None]
+    if bad_env:
+        lines = ["\nDeprecated AI_SHELL_* env var(s) set:"]
+        for key in bad_env:
+            lines.append(f"  - {key}: use {_LEGACY_ENV_VARS[key]} instead")
+        raise ValueError("\n".join(lines))
+
     env_map: dict[str, tuple[str, type]] = {
         "AI_SHELL_IMAGE": ("image", str),
         "AI_SHELL_IMAGE_TAG": ("image_tag", str),
         "AI_SHELL_PROJECT": ("project_name", str),
-        "AI_SHELL_PRIMARY_MODEL": ("primary_model", str),
-        "AI_SHELL_FALLBACK_MODEL": ("fallback_model", str),
+        "AI_SHELL_PRIMARY_CHAT_MODEL": ("primary_chat_model", str),
+        "AI_SHELL_SECONDARY_CHAT_MODEL": ("secondary_chat_model", str),
+        "AI_SHELL_PRIMARY_CODING_MODEL": ("primary_coding_model", str),
+        "AI_SHELL_SECONDARY_CODING_MODEL": ("secondary_coding_model", str),
         "AI_SHELL_CONTEXT_SIZE": ("context_size", int),
         "AI_SHELL_OLLAMA_PORT": ("ollama_port", int),
         "AI_SHELL_WEBUI_PORT": ("webui_port", int),

{augint_shell-0.78.0 → augint_shell-0.80.0}/src/ai_shell/container.py

@@ -357,6 +357,11 @@ class ContainerManager:
         environment = {
             "OLLAMA_BASE_URL": f"http://{OLLAMA_CONTAINER}:11434",
             "WEBUI_AUTH": "false",
+            # DEFAULT_MODELS is a PersistentConfig: env seeds the DB on first
+            # boot and UI edits win after that. Point new chats at the
+            # primary chat slot; users can pick the secondary (uncensored)
+            # from the model dropdown.
+            "DEFAULT_MODELS": self.config.primary_chat_model,
         }
         if voice_enabled:
             environment.update(

{augint_shell-0.78.0 → augint_shell-0.80.0}/src/ai_shell/defaults.py

@@ -57,8 +57,13 @@ WEBUI_IMAGE = "ghcr.io/open-webui/open-webui:main"
 KOKORO_IMAGE_CPU = "ghcr.io/remsky/kokoro-fastapi-cpu:latest"
 KOKORO_IMAGE_GPU = "ghcr.io/remsky/kokoro-fastapi-gpu:latest"
 N8N_IMAGE = "docker.n8n.io/n8nio/n8n"
-DEFAULT_PRIMARY_MODEL = "qwen3-coder:30b-a3b-q4_K_M"
-DEFAULT_FALLBACK_MODEL = "huihui_ai/llama3.3-abliterated"
+# Model slots (RTX 4090-sized, validated April 2026). Primary = best available for
+# the role; secondary = best uncensored alternative. See README "Local LLM stack"
+# and the generated .ai-shell.yaml for per-slot rationale and caveats.
+DEFAULT_PRIMARY_CHAT_MODEL = "qwen3.5:27b"
+DEFAULT_SECONDARY_CHAT_MODEL = "huihui_ai/qwen3.5-abliterated:27b"
+DEFAULT_PRIMARY_CODING_MODEL = "qwen3-coder:30b-a3b-q4_K_M"
+DEFAULT_SECONDARY_CODING_MODEL = "huihui_ai/qwen3-coder-abliterated:30b-a3b-instruct-q4_K_M"
 DEFAULT_CONTEXT_SIZE = 32768
 DEFAULT_OLLAMA_PORT = 11434
 DEFAULT_WEBUI_PORT = 3000

augint_shell-0.80.0/src/ai_shell/templates/ai-shell.yaml

@@ -0,0 +1,72 @@
+# =============================================================================
+# .ai-shell.yaml - Project configuration for ai-shell
+# =============================================================================
+# Priority (highest wins): CLI flags > env vars > this file > global config > defaults
+# Global config: ~/.ai-shell.yaml (applies to all projects)
+# Full docs:     https://github.com/svange/augint-shell#local-llm-stack
+
+# -----------------------------------------------------------------------------
+# llm - Local LLM stack (Ollama + Open WebUI + optional TTS / n8n)
+# -----------------------------------------------------------------------------
+# Four role-specific model slots. Primary = best-available; secondary =
+# best uncensored (abliterated) alternative of the same base. Chat slots
+# route to Open WebUI (DEFAULT_MODELS); coding slots route to OpenCode /
+# Aider (--model). `ai-shell llm pull` pulls all 4 slots plus any
+# `extra_models` entries, deduped. See README for per-slot rationale and
+# caveats (Qwen3.5 Ollama tool-call bug, num_ctx trap, tool-count cliff).
+llm:
+  primary_chat_model: qwen3.5:27b
+  secondary_chat_model: huihui_ai/qwen3.5-abliterated:27b
+  primary_coding_model: qwen3-coder:30b-a3b-q4_K_M
+  secondary_coding_model: huihui_ai/qwen3-coder-abliterated:30b-a3b-instruct-q4_K_M
+  context_size: 32768
+  # Additional Ollama tags to pull alongside the 4 slots (deduped).
+  # Uncomment any line to enable -- indentation is already correct.
+  extra_models:
+    # - llama3.1:8b                         # ~5 GB   fast general chat
+    # - llama3.2:latest                     # ~2 GB   Llama 3.2 3B, very fast
+    # - dolphin3:8b                         # ~5 GB   uncensored Llama 3.1 8B
+    # - qwen3:30b-a3b-instruct-2507-q4_K_M  # ~19 GB  Qwen3 MoE chat alt (~196 tok/s)
+    # - qwen2.5-coder:32b-q4_k_m            # ~19 GB  Qwen2.5-Coder (previous gen)
+    # - qwen2.5-coder:14b-instruct          # ~9 GB   smaller Qwen2.5-Coder
+
+# -----------------------------------------------------------------------------
+# aws - AWS profile + region. Uncomment to override defaults.
+# -----------------------------------------------------------------------------
+# ai_profile:      AWS profile for infra tools (terraform/cdk). Sets AWS_PROFILE
+#                  in the container.
+# bedrock_profile: AWS profile for Bedrock LLM calls (--aws mode). Often a
+#                  different account than ai_profile.
+# region:          Region for Bedrock. Default: us-east-1
+# Auth: ~/.aws is bind-mounted read-write. `aws sso login` on the host as needed.
+#
+# aws:
+#   ai_profile: my-infra-account
+#   bedrock_profile: my-ai-account
+#   region: us-east-1
+
+# -----------------------------------------------------------------------------
+# claude - Claude Code backend selection. Uncomment to use Bedrock instead of
+# the Anthropic API. Equivalent per-session flag: `ai-shell claude --aws`.
+# -----------------------------------------------------------------------------
+# claude:
+#   provider: aws
+
+# -----------------------------------------------------------------------------
+# container - Docker image, env, mounts, ports. Uncomment to override.
+# -----------------------------------------------------------------------------
+# image / image_tag:  Override the default svange/augint-shell:latest image.
+# extra_env:          Additional env vars injected into the dev container.
+# extra_volumes:      Additional bind mounts ("/host:/container" or ":/path:ro").
+# ports:              Extra host ports to expose on the dev container.
+#
+# container:
+#   image: svange/augint-shell
+#   image_tag: latest
+#   extra_env:
+#     MY_VAR: value
+#   extra_volumes:
+#     - /host/path:/container/path
+#   ports:
+#     - 9000
+#     - 9229

augint_shell-0.78.0/src/ai_shell/templates/ai-shell.yaml

@@ -1,207 +0,0 @@
-# =============================================================================
-# .ai-shell.yaml - Project configuration for ai-shell
-# =============================================================================
-# Uncomment and modify settings you want to override.
-# Priority (highest wins): CLI flags > env vars > this file > global config > defaults
-# Global config: ~/.config/ai-shell/config.yaml (same format, applies to all projects)
-# Docs: https://github.com/svange/augint-shell
-
-# =============================================================================
-# aws - Amazon Web Services configuration
-# =============================================================================
-# ai_profile: AWS profile for the AI's working environment (aws cli, terraform,
-#   cdk). Sets AWS_PROFILE in the container. This is the account the AI uses
-#   when running infrastructure commands.
-#   Override with env var: AI_SHELL_AI_PROFILE
-#
-# bedrock_profile: AWS profile for Bedrock LLM API calls. Often a different
-#   account than ai_profile. Overrides AWS_PROFILE specifically for AI tool
-#   processes launched with `--aws`.
-#   Override with env var: AI_SHELL_BEDROCK_PROFILE
-#   Override per-session with: --profile <name> on the CLI
-#
-# region: AWS region for Bedrock API calls. Default: us-east-1
-#   Override with env var: AI_SHELL_AWS_REGION
-#
-# Authentication: ~/.aws is bind-mounted into the container (read-write).
-# SSO, credential files, and config are available automatically.
-# If SSO tokens expire, run 'aws sso login --profile <name>' on the host.
-#
-# aws:
-#   ai_profile: my-infra-account
-#   bedrock_profile: my-ai-account
-#   region: us-east-1
-
-# =============================================================================
-# claude - Claude Code settings
-# =============================================================================
-# provider: API backend for Claude Code.
-#   "anthropic" - Direct Anthropic API (default, uses ~/.claude credentials)
-#   "aws"       - Amazon Bedrock (uses bedrock_profile from aws section)
-#
-# When provider is "aws":
-#   - CLAUDE_CODE_USE_BEDROCK=1 is set in the environment
-#   - AWS_PROFILE is set to bedrock_profile for Claude's process
-#   - Quick switch with CLI: ai-shell claude --aws
-#   - Override per-session:   ai-shell claude --aws --profile <name>
-#   - Tip: pin Bedrock model versions with ANTHROPIC_DEFAULT_SONNET_MODEL env var
-#
-# Override with env var: AI_SHELL_CLAUDE_PROVIDER
-#
-# claude:
-#   provider: aws
-
-# Codex runtime note:
-#   ai-shell does not manage Codex's own config file, but `ai-shell codex --aws`
-#   launches Codex with Bedrock by injecting `CLAUDE_CODE_USE_BEDROCK=1` and
-#   setting `AWS_PROFILE` to `aws.bedrock_profile` (or `--profile` if passed).
-#   Local-LLM Codex configuration is not managed in this file.
-
-# =============================================================================
-# container - Docker container settings
-# =============================================================================
-# image: Docker image (default: svange/augint-shell)
-# image_tag: Image tag (default: current ai-shell version)
-# extra_env: Additional environment variables for the container
-# extra_volumes: Additional bind mounts ("/host/path:/container/path" or ":/path:ro")
-# ports: Additional ports to expose (extends the default dev port set)
-#
-# container:
-#   image: svange/augint-shell
-#   image_tag: latest
-#   extra_env:
-#     MY_VAR: value
-#   extra_volumes:
-#     - /host/path:/container/path
-#   ports:
-#     - 9000
-#     - 9229
-
-# =============================================================================
-# llm - Local LLM settings (Ollama + Open WebUI)
-# =============================================================================
-# primary_model: Default Ollama model for inference
-# fallback_model: Backup model if primary unavailable
-# context_size: Context window in tokens (default: 32768)
-# ollama_port: Host port for Ollama API (default: 11434)
-# webui_port: Host port for Open WebUI (default: 3000)
-# n8n_port: Host port for n8n workflow automation (default: 5678)
-#
-# Models are downloaded automatically by `ai-shell llm setup`, which:
-#   1. Starts the Ollama container (GPU-enabled if an NVIDIA card is detected)
-#   2. Runs `ollama pull <primary_model>` and `ollama pull <fallback_model>`
-#   3. Applies the context-window Modelfile so both models run with num_ctx set
-#
-# To pull a model manually at any time:
-#   ai-shell llm shell        # opens a bash shell inside the Ollama container
-#   ollama pull <model>       # then run any ollama command directly
-#
-# ai-shell only manages the local LLM endpoint and generic runtime settings.
-# Tool-specific config files for Codex, OpenCode, and Aider should be managed
-# separately, for example via augint-opencodex.
-# Keep tool-specific provider, model, auth, and permission settings out of this
-# file. Put those in the generated tool config files instead.
-#
-# ─── RTX 4090 model guide (24 GiB VRAM) ─────────────────────────────────────
-#
-# General-chat / assistant models
-# ─────────────────────────────────
-# qwen3.5:27b                  ~15 GiB  Q4_K_M  fits on 4090 with headroom
-# qwen3.5:27b-q5_k_m           ~19 GiB  Q5_K_M  fits, higher quality
-# qwen3.5:32b-q4_k_m           ~19 GiB  Q4_K_M  tight but fits on 4090
-#
-# Uncensored / instruction-following variants
-# ────────────────────────────────────────────
-# dolphin3:8b                      ~5 GiB   Dolphin 3.0 (uncensored Llama 3.1 8B)
-# huihui_ai/llama3.3-abliterated   ~16 GiB  Llama 3.3 70B abliterated (uncensored chat)
-# llama3.1:8b                      ~5 GiB   Meta Llama 3.1 8B instruct
-# llama3.1:8b-instruct-q4_k_m      ~5 GiB   Q4_K_M quantized
-# llama3.2:latest                  ~2 GiB   Meta Llama 3.2 3B (fast/small)
-#
-# Coding-heavy models
-# ────────────────────
-# qwen2.5-coder:32b-q4_k_m        ~19 GiB  Q4_K_M  top coding quality on 4090
-# qwen3:14b-q4_k_m               ~9 GiB   Q4_K_M  fast coder with good accuracy
-# qwen3:32b-q4_k_m               ~19 GiB  Q4_K_M  best local coding on 4090
-# qwen3-coder:30b-a3b-q4_K_M     ~19 GiB  Q4_K_M  Qwen3-Coder 30B A3B (Mixture-of-Experts, ~3B active)
-#
-# Quick-start pull commands (run inside `ai-shell llm shell`):
-#   ollama pull qwen3.5:27b
-#   ollama pull qwen3-coder:30b-a3b-q4_K_M
-#   ollama pull huihui_ai/llama3.3-abliterated
-#   ollama pull dolphin3:8b
-#   ollama pull llama3.1:8b
-#   ollama pull qwen2.5-coder:32b-q4_k_m
-#   ollama pull qwen3:14b-q4_k_m
-#   ollama pull qwen3:32b-q4_k_m
-#
-# ─── RTX 5070 Ti model guide (12 GiB VRAM) ──────────────────────────────────
-#
-# Chat models
-# ────────────
-# qwen3.5:9b                              ~7 GiB   Q4_K_M  best chat, 256K ctx, multimodal
-# huihui_ai/qwen3.5-abliterated:9b        ~7 GiB   Q4_K    uncensored Qwen3.5 (abliterated)
-#
-# Coding models
-# ──────────────
-# qwen2.5-coder:14b-instruct              ~9 GiB   Q4_K_M  largest dedicated coder that fits
-# huihui_ai/qwen3.5-abliterated:9b-OmniCoder  ~6 GiB  Q4_K  uncensored coding variant
-#
-# Quick-start pull commands (run inside `ai-shell llm shell`):
-#   ollama pull qwen3.5:9b
-#   ollama pull huihui_ai/qwen3.5-abliterated:9b
-#   ollama pull qwen2.5-coder:14b-instruct
-#   ollama pull huihui_ai/qwen3.5-abliterated:9b-OmniCoder
-#
-# After pulling, set primary_model (and fallback_model) below, then run:
-#   ai-shell llm setup        # pulls models + applies context-window config
-#
-# llm:
-#   primary_model: qwen3-coder:30b-a3b-q4_K_M
-#   fallback_model: huihui_ai/llama3.3-abliterated
-#   context_size: 32768
-#   ollama_port: 11434
-#   webui_port: 3000
-#   n8n_port: 5678
-#
-# --- Example: 4090 coding-focused setup ---
-# llm:
-#   primary_model: qwen2.5-coder:32b-q4_k_m
-#   fallback_model: qwen3:14b-q4_k_m
-#   context_size: 32768
-#
-# --- Example: 4090 general-chat setup ---
-# llm:
-#   primary_model: qwen3.5:27b
-#   fallback_model: llama3.1:8b
-#   context_size: 32768
-#
-# --- Example: 4090 uncensored setup ---
-# llm:
-#   primary_model: dolphin3:8b
-#   fallback_model: llama3.1:8b
-#   context_size: 32768
-#
-# --- Example: 5070 Ti coding-focused setup ---
-# llm:
-#   primary_model: qwen2.5-coder:14b-instruct
-#   fallback_model: qwen3.5:9b
-#   context_size: 32768
-#
-# --- Example: 5070 Ti general-chat setup ---
-# llm:
-#   primary_model: qwen3.5:9b
-#   fallback_model: huihui_ai/qwen3.5-abliterated:9b
-#   context_size: 32768
-#
-# --- Example: 5070 Ti uncensored-chat setup ---
-# llm:
-#   primary_model: huihui_ai/qwen3.5-abliterated:9b
-#   fallback_model: dolphin3:8b
-#   context_size: 32768
-#
-# --- Example: 5070 Ti uncensored-coding setup ---
-# llm:
-#   primary_model: huihui_ai/qwen3.5-abliterated:9b-OmniCoder
-#   fallback_model: qwen2.5-coder:14b-instruct
-#   context_size: 32768