This is an automated email from the ASF dual-hosted git repository.

davsclaus pushed a commit to branch main
in repository https://gitbox.apache.org/repos/asf/camel.git


The following commit(s) were added to refs/heads/main by this push:
     new 270ea3b9ee7c CAMEL-26962: camel-jbang - the AI clients work with the 
current models (#29907)
270ea3b9ee7c is described below

commit 270ea3b9ee7cc82fabe6f665cc6c19b8d6e99588
Author: Claus Ibsen <[email protected]>
AuthorDate: Sun Nov 15 10:21:35 2026 +0100

    CAMEL-26962: camel-jbang - the AI clients work with the current models 
(#29907)
    
    
    https://issues.apache.org/jira/browse/CAMEL-26962
    
    The AI clients of the Camel CLI (`camel ask`, `camel explain`, `camel 
overview --ai` and the AI panel) call Anthropic through `LlmClient`. Every 
request had a `temperature`. Claude Opus 5.5 refuses it, and Claude Sonnet 5.5 
refuses a value other than the default. From API version `2026-10-21`, every 
model refuses the field.
    
    ## Changes
    
    - No `temperature`, `top_p` or `top_k` in the requests to Anthropic. The 
other providers keep the temperature.
    - API version `2026-10-21`: in the `anthropic-version` header, and on 
Vertex AI in the `anthropic_version` of the body.
    - The default model is `claude-sonnet-5-5`.
    - `max_tokens` is at least 16,000 for Anthropic. The current models think 
before they answer, and the thinking counts toward `max_tokens`, so the 4096 
that the commands ask for could cut long answers off.
    - Thinking blocks: in a tool loop, the answers go back as they came, 
thinking blocks included, as the API asks. The thinking blocks of earlier 
questions are left out. The AI panel compacts that part of the history, and the 
API refuses a thinking block when the history before it changed. Leaving out 
the thinking blocks at the start of the history is allowed.
    - Prompt caching in a tool loop: a breakpoint after the system prompt, so 
the tools and the system prompt stay cached when the history changes, and the 
automatic breakpoint at the end of the request. One-shot calls (`camel 
explain`, the overviews) stay uncached: there a cache write would only cost 
more.
    - Token usage: Anthropic's `input_tokens` counts only the part of the 
prompt that is not cached. The prompt size now adds the tokens read from the 
cache and written to it, as the other providers count it.
    - A refusal with no content is shown as a refusal, not as an empty answer.
    - No Vertex AI name for Claude Sonnet 4.5, which retires on Nov 30 2026. 
The full name `claude-sonnet-4-5@20250929` still works.
    - The default Gemini model was `gemini-2.0-flash`, which Google has shut 
down. It is now the `gemini-flash-latest` alias, which Google moves to each new 
Flash model, with two weeks' notice.
    - No command uses `llama3.2` when no model is given. With Ollama, `camel 
ask`, `camel explain`, `camel overview`, `camel harden` and the AI panel take 
the installed model best at tool calling: qwen3.6 first, then qwen3.8 and the 
other families, never an embedding model. When none can be used, they ask for 
`qwen3.6:35b-a3b`, so the error of Ollama names the model to pull.
    - A small model is one below 24B, not 14B: the Doctor and `camel ask` warn 
below it. The 27B models measured on Camel tasks are above it (Ollama reports 
them as 26.9B).
    - Only a model on this machine gets the temperature, as the AI tools were 
tested with it. A hosted model (OpenAI, Gemini, watsonx, Anthropic) gets its 
own default sampling: the reasoning models of OpenAI refuse any other value.
    - The default OpenAI model is `gpt-6-luna`, which OpenAI names as the 
replacement of `gpt-4o-mini` (shut down on Jan 14 2027). OpenAI itself gets 
`reasoning_effort: none`, as its Chat Completions call tools only with it; a 
model that refuses the field gets the request again without it.
    - Docs and the AI setup guides:
      - The in-app AI setup guide, and the guide the AI panel shows when it 
finds no provider, are short: three ways (a local model, a coding agent over 
ACP, an API key) and links to the user manual. The manual is the one place with 
the full setup and the list of models.
      - The list of local models has only the models measured on Camel tasks: 
qwen3.6 35B-A3B (GGUF and MLX, the MLX build recommended on Apple Silicon with 
Ollama 0.40 or later) and qwen3.8 27B. The old and small models (llama3.3, 
hermes3, qwen2.5) and the "at least 14B" advice are gone.
      - The `--temperature` help of `camel explain` and `camel overview` says 
that it is not sent to Anthropic. The AI panel page says that the panel turns 
prompt caching on with Anthropic.
      - IBM Bob is named with the coding agents, and Vertex AI is called by its 
new name, Gemini Enterprise Agent Platform.
    
    ## Tests
    
    Unit tests with the fake LLM server check:
    - the request: no temperature, the API version header, `max_tokens`, and 
both cache breakpoints;
    - the token usage sum;
    - the thinking blocks: those of a tool loop go back as they came, and none 
from before the question;
    - no cache for an explanation;
    - the refusal, and the default models of Anthropic, Gemini and OpenAI;
    - the choice of the Ollama model (never `llama3.2` while a better one is 
installed) and the 24B line;
    - the temperature only for a model on this machine, and `reasoning_effort: 
none` for OpenAI itself;
    - that the AI panel's tool loop keeps the answer as it came.
    
    Merged with main twice: `camel-jbang-commands-metadata.json` is 
regenerated, `cli-help.txt` follows the new `--model` and `--temperature` texts 
of `camel explain`, `camel overview` and `camel harden`, and `AiPanelTest` 
keeps the tests of #30003. The reason the content blocks of a tool loop go back 
unchanged is in the Javadoc of `LlmClient.Message.assistant`, not in a comment 
in `AiPanel.runAgentLoop`, which keeps that method within its size in 
`SourceSizeTest`.
    
    Not tested against the hosted APIs: there is no API key on the test 
machine. The rules are from the documentation of each provider: the Claude API 
migration guide (https://platform.claude.com/docs/en/api/migration-2026-10-21), 
the Google AI Studio models page (https://aistudio.google.com/docs/models), and 
the OpenAI models and deprecations pages 
(https://developers.openai.com/api/docs/models/gpt-6-luna, 
https://developers.openai.com/api/docs/deprecations).
    
    Co-Authored-By: Claude Opus 5.5 <[email protected]>
---
 .../ROOT/pages/camel-jbang-ai-providers.adoc       |  31 +-
 .../ROOT/pages/camel-jbang-tui-ai-ways.adoc        |   6 +-
 .../modules/ROOT/pages/camel-jbang-tui-ai.adoc     |   6 +-
 .../ROOT/pages/camel-jbang-tui-local-models.adoc   |  12 +-
 .../ROOT/pages/camel-jbang-tui-settings.adoc       |   2 +-
 .../pages/jbang-commands/camel-jbang-explain.adoc  |   4 +-
 .../pages/jbang-commands/camel-jbang-harden.adoc   |   4 +-
 .../pages/jbang-commands/camel-jbang-overview.adoc |   4 +-
 .../META-INF/camel-jbang-commands-metadata.json    |   6 +-
 .../apache/camel/dsl/jbang/core/commands/Ask.java  |  15 +-
 .../camel/dsl/jbang/core/commands/Explain.java     |  14 +-
 .../camel/dsl/jbang/core/commands/Harden.java      |  33 ++-
 .../camel/dsl/jbang/core/commands/LlmClient.java   | 324 +++++++++++++++------
 .../camel/dsl/jbang/core/commands/Overview.java    |  14 +-
 .../dsl/jbang/core/common/OllamaDoctorSupport.java |  23 +-
 .../camel/dsl/jbang/core/commands/AskTest.java     |   3 +
 .../dsl/jbang/core/commands/LlmClientChatTest.java | 144 ++++++++-
 .../jbang/core/commands/LlmClientGeminiTest.java   |   6 +
 .../core/commands/LlmClientListModelsTest.java     |  23 +-
 .../jbang/core/commands/LlmClientVertexTest.java   |   5 +-
 .../jbang/core/common/OllamaDoctorSupportTest.java |  11 +
 .../src/test/resources/golden/cli-help.txt         |  29 +-
 .../src/test/resources/golden/cli-parse.txt        |   6 +-
 .../camel/dsl/jbang/core/commands/tui/AiPanel.java |  73 +----
 .../jbang/core/commands/tui/DocViewerPopup.java    | 123 +++-----
 .../dsl/jbang/core/commands/tui/DoctorPopup.java   |   3 +-
 .../dsl/jbang/core/commands/tui/AiPanelTest.java   |  52 ++++
 27 files changed, 631 insertions(+), 345 deletions(-)

diff --git a/docs/user-manual/modules/ROOT/pages/camel-jbang-ai-providers.adoc 
b/docs/user-manual/modules/ROOT/pages/camel-jbang-ai-providers.adoc
index f201d6eec784..5804992570e1 100644
--- a/docs/user-manual/modules/ROOT/pages/camel-jbang-ai-providers.adoc
+++ b/docs/user-manual/modules/ROOT/pages/camel-jbang-ai-providers.adoc
@@ -9,7 +9,8 @@ This page explains how the provider is chosen, and how to run a 
model on your ow
 The provider is detected from the environment. The first match wins:
 
 1. `ANTHROPIC_API_KEY` -> Anthropic API
-2. `CLOUD_ML_REGION` + `ANTHROPIC_VERTEX_PROJECT_ID` -> Anthropic models on 
Google Vertex AI
+2. `CLOUD_ML_REGION` + `ANTHROPIC_VERTEX_PROJECT_ID` -> Claude on Google 
Cloud, in the Gemini Enterprise Agent
+   Platform (formerly Vertex AI)
 3. `AZURE_OPENAI_API_KEY` + `AZURE_OPENAI_ENDPOINT` -> Azure OpenAI (uses the 
`api-key` header; optional
    `AZURE_OPENAI_DEPLOYMENT_NAME` and `AZURE_OPENAI_API_VERSION`)
 4. `GEMINI_API_KEY` -> Google Gemini native API 
(`generativelanguage.googleapis.com`).
@@ -25,9 +26,9 @@ Override the detection with explicit options:
 [source,bash]
 ----
 camel ask "check health" --api-type=anthropic --api-key=sk-...
-camel ask "check health" --api-type=openai --model=gpt-4
-camel ask "check health" --api-type=gemini --model=gemini-2.0-flash
-camel ask "check health" --api-type=ollama --model=llama3.1
+camel ask "check health" --api-type=openai --model=gpt-6-luna
+camel ask "check health" --api-type=gemini --model=gemini-3.8-flash
+camel ask "check health" --api-type=ollama --model=qwen3.6:35b-a3b
 ----
 
 With `--url`, the CLI uses that endpoint and no other: when it does not 
answer, the command fails instead of going
@@ -71,24 +72,22 @@ The CLI checks what models are available and auto-selects a 
suitable one.
 ==== Model requirements
 
 `camel ask` and the TUI F8 panel rely on tool calling to inspect your running 
Camel process.
-Models smaller than ~14B do not reliably invoke tools and answer from training 
knowledge instead.
-Use at least a 14B model. Prefer a mixture-of-experts model such as 
`qwen3.6:35b-a3b`: with only
-3B parameters active per token it processes the tool-heavy prompt many times 
faster than a dense
-27B/32B model, so answers start in seconds instead of a minute.
+Small models do not call the tools reliably and answer from training knowledge 
instead. The table lists the
+models that were measured on Camel tasks: building, running and fixing routes 
with the tools of the CLI.
+Prefer the mixture-of-experts model `qwen3.6:35b-a3b`: with only 3B parameters 
active per token it processes the
+tool-heavy prompt many times faster than a dense model, so answers start in 
seconds instead of a minute.
 
 [options="header"]
 |===
-| Model | RAM (Q4) | Notes
-| `qwen3.6:35b-a3b` | ~23 GB | Recommended: fastest prompt processing, needs 
32 GB+
+| Model | RAM | Notes
+| `qwen3.6:35b-a3b` | ~23 GB | Recommended: the fastest prompt processing, 
needs 32 GB or more
 | `qwen3.6:35b-mlx` | ~24 GB | The same model as an MLX build: recommended on 
Apple Silicon with Ollama 0.40 or later
-| `qwen2.5:14b` | ~9 GB | Minimum for 16 GB machines
-| `qwen3.6:27b` | ~18 GB | Strong dense model, several times slower prompt 
processing
-| `qwen2.5:32b` | ~20 GB | Good quality, slow prompt processing
-| `hermes3:70b` | ~43 GB | Excellent tool calling, needs 64 GB+
-| `llama3.3:70b` | ~43 GB | Best open model, needs 64 GB+
+| `qwen3.8:27b-mlx` | ~23 GB | A dense model: it reads the prompt about 7 
times slower than `qwen3.6:35b-mlx`, and
+passed fewer of the Camel tasks
 |===
 
-On Apple Silicon, all RAM is unified — a 64 GB M-series Mac can run 
`llama3.3:70b` comfortably alongside the OS and other dev tools.
+With less than 32 GB of memory, use a coding agent over ACP or a hosted 
provider instead of a local model.
+
 On a Mac with Apple Silicon, use the MLX build `qwen3.6:35b-mlx` with Ollama 
0.40 or later. It is the same
 35B-A3B model, in the format of Apple's MLX engine, which Ollama 0.40 uses on 
Apple Silicon and which reuses the cached
 prompt between the questions of a conversation. Measured on an M4 Pro with 
Ollama 0.40.3 and an 11K-token prompt,
diff --git a/docs/user-manual/modules/ROOT/pages/camel-jbang-tui-ai-ways.adoc 
b/docs/user-manual/modules/ROOT/pages/camel-jbang-tui-ai-ways.adoc
index c60fd4ea12a3..13e237c8bdaa 100644
--- a/docs/user-manual/modules/ROOT/pages/camel-jbang-tui-ai-ways.adoc
+++ b/docs/user-manual/modules/ROOT/pages/camel-jbang-tui-ai-ways.adoc
@@ -22,7 +22,7 @@ image::jbang/camel-tui-ai-ways.svg[Four ways to use AI with 
the Camel TUI: the b
 
 | *2. Coding agent in the panel*
 | The AI panel (*F8*) in the TUI
-| An agent the TUI starts over ACP (Claude Code, Codex, opencode, Qwen, ...), 
with its own model and login
+| An agent the TUI starts over ACP (Claude Code, Codex, opencode, IBM Bob, 
Qwen, ...), with its own model and login
 | The TUI's MCP server, started for the agent on a random localhost port
 | `camel tui`, then *F8*, *Ctrl+P* and an `acp:` provider
 
@@ -64,7 +64,7 @@ image::jbang/camel-tui-fix-with-ai.gif[A local model fixes a 
failing route from
 == 2. A coding agent in the panel (ACP)
 
 The same panel can hand the conversation to a coding agent instead of a model. 
Press *Ctrl+P* in the panel and pick
-an `acp:` provider (`acp:claude`, `acp:codex`, `acp:opencode`, ...). The TUI 
starts the agent as a subprocess and talks
+an `acp:` provider (`acp:claude`, `acp:codex`, `acp:opencode`, `acp:bob`, 
...). The TUI starts the agent as a subprocess and talks
 to it over the https://agentclientprotocol.com[Agent Client Protocol] (ACP). 
The agent brings its own model, login,
 settings and skills; the TUI shows its answers, tool calls and permission 
questions in the panel.
 
@@ -139,6 +139,6 @@ panel lists those calls as *integration* next to the 
panel's own (*TUI ask*).
 == Which one should I pick?
 
 * New to the TUI, or want to try a local model: *1*.
-* Already pay for a coding agent such as Claude Code or Codex and want it in 
the TUI: *2*.
+* Already pay for a coding agent such as Claude Code, Codex or IBM Bob and 
want it in the TUI: *2*.
 * Already work in an agent and want it to see and drive your running 
integrations: *3*.
 * No TUI, just Camel knowledge for your agent: *4*.
diff --git a/docs/user-manual/modules/ROOT/pages/camel-jbang-tui-ai.adoc 
b/docs/user-manual/modules/ROOT/pages/camel-jbang-tui-ai.adoc
index 3589fc185d11..2393f3e820b2 100644
--- a/docs/user-manual/modules/ROOT/pages/camel-jbang-tui-ai.adoc
+++ b/docs/user-manual/modules/ROOT/pages/camel-jbang-tui-ai.adoc
@@ -248,5 +248,7 @@ also shows whether the provider reused its cached prompt. 
For Ollama it prints t
 (`prefill`) and generation (`gen`) time reported by the server -- a prefill of 
a fraction of a second
 on a 10k-token prompt means the cache was hit, while a prefill of several 
seconds means the whole
 prompt was processed again. For OpenAI, Anthropic and Gemini it prints the 
number of `cached` input
-tokens the provider reported. Use it to check that follow-up questions are 
cheap before blaming the
-model for being slow.
+tokens the provider reported. With Anthropic the panel turns prompt caching 
on: within five minutes
+of the previous request, the tools and the system prompt come from the cache, 
and so does the history
+on each further step of a question. Use it to check that follow-up questions 
are cheap before blaming
+the model for being slow.
diff --git 
a/docs/user-manual/modules/ROOT/pages/camel-jbang-tui-local-models.adoc 
b/docs/user-manual/modules/ROOT/pages/camel-jbang-tui-local-models.adoc
index 6a8564782215..8a003b455f4b 100644
--- a/docs/user-manual/modules/ROOT/pages/camel-jbang-tui-local-models.adoc
+++ b/docs/user-manual/modules/ROOT/pages/camel-jbang-tui-local-models.adoc
@@ -27,8 +27,8 @@ camel tui
 On a Mac with Apple Silicon and Ollama 0.40 or later, pull the MLX build 
`qwen3.6:35b-mlx` instead: the same model,
 with faster prompt processing and answers on the M chips.
 
-Then press *F8*. How to install Ollama, which models work well (at least 14B, 
with tool calling; a
-mixture-of-experts model such as `qwen3.6:35b-a3b` is the fastest) and how the 
context window is chosen
+Then press *F8*. How to install Ollama, which models work well (the ones 
measured on Camel tasks; the
+mixture-of-experts model `qwen3.6:35b-a3b` is the fastest) and how the context 
window is chosen
 are described in xref:camel-jbang-ai-providers.adoc#_ollama[AI Providers and 
Local Models], shared with
 `camel ask` and the other AI commands.
 
@@ -146,9 +146,11 @@ measured prompt; the panel title shows the fill as `ctx 
34%`.
 
 *Choosing the model.* With no `camel.tui.ai.model` set the panel takes 
`llama3.2` when it is installed,
 otherwise the first installed model; run `/model <name>` in the panel or set 
*AI Model* in
-*F2 -> Settings* to pin one. The model must support tool calling and should 
have at least 14B
-parameters; a mixture-of-experts model such as `qwen3.6:35b-a3b` prefills 
several times faster than a
-dense model of similar quality, which is what matters for a tool-heavy prompt. 
*F2 -> Run Doctor* shows
+*F2 -> Settings* to pin one. The model must support tool calling; small models 
do not call the tools
+reliably, so pick one of the models measured on Camel tasks, listed in
+xref:camel-jbang-ai-providers.adoc#_model_requirements[Model requirements]. 
The mixture-of-experts model
+`qwen3.6:35b-a3b` prefills several times faster than a dense model of similar 
quality, which is what
+matters for a tool-heavy prompt. *F2 -> Run Doctor* shows
 whether Ollama was found, which models are installed and whether they are 
large enough.
 
 *Where to look.* The <<_ollama_tab,Ollama tab>> is the instrument for all of 
the above: tokens per second live and
diff --git a/docs/user-manual/modules/ROOT/pages/camel-jbang-tui-settings.adoc 
b/docs/user-manual/modules/ROOT/pages/camel-jbang-tui-settings.adoc
index 835f43a16fd4..54078016fff7 100644
--- a/docs/user-manual/modules/ROOT/pages/camel-jbang-tui-settings.adoc
+++ b/docs/user-manual/modules/ROOT/pages/camel-jbang-tui-settings.adoc
@@ -66,7 +66,7 @@ The Doctor checks your development environment and reports 
issues:
 * JBang installation
 * Maven repository connectivity
 * Docker/Podman availability
-* Ollama: whether it runs, and the models it has (the AI panel needs one of 
14B or more)
+* Ollama: whether it runs, and the models it has (the AI panel needs one of 
24B or more)
 * AI and MCP: the AI provider and model set up, and the MCP server of the 
session
 * Common port conflicts (8080, 8443, 9090)
 * Disk space in temp directory
diff --git 
a/docs/user-manual/modules/ROOT/pages/jbang-commands/camel-jbang-explain.adoc 
b/docs/user-manual/modules/ROOT/pages/jbang-commands/camel-jbang-explain.adoc
index 20af5bdeea54..dc29b31fbe47 100644
--- 
a/docs/user-manual/modules/ROOT/pages/jbang-commands/camel-jbang-explain.adoc
+++ 
b/docs/user-manual/modules/ROOT/pages/jbang-commands/camel-jbang-explain.adoc
@@ -23,11 +23,11 @@ camel explain [options]
 | `--api-type` | API type: 'ollama', 'openai' (OpenAI-compatible), 'anthropic' 
(Anthropic/Vertex AI), or 'watsonx' (IBM watsonx.ai) |  | ApiType
 | `--catalog-context` | Include Camel Catalog descriptions in the prompt |  | 
boolean
 | `--format` | Output format (text, markdown) | text | String
-| `--model` | Model to use | llama3.2 | String
+| `--model` | Model to use. Default: the provider's default; with Ollama the 
installed model best at tool calling, such as qwen3.6 |  | String
 | `--show-prompt` | Show the prompt sent to the LLM |  | boolean
 | `--stream` | Stream the response as it's generated (shows progress) | true | 
boolean
 | `--system-prompt` | Custom system prompt |  | String
-| `--temperature` | Temperature for response generation (0.0-2.0) | 0.7 | 
double
+| `--temperature` | Temperature for response generation (0.0-2.0), for a model 
on this machine: a hosted model uses its own default | 0.7 | double
 | `--timeout` | Timeout in seconds for LLM response | 120 | int
 | `--url` | LLM API endpoint URL. Auto-detected from 'camel infra' for Ollama 
if not specified. Also reads AZURE_OPENAI_ENDPOINT or WATSONX_URL env vars |  | 
String
 | `--verbose,-v` | Include detailed technical information |  | boolean
diff --git 
a/docs/user-manual/modules/ROOT/pages/jbang-commands/camel-jbang-harden.adoc 
b/docs/user-manual/modules/ROOT/pages/jbang-commands/camel-jbang-harden.adoc
index 928d2495c738..18854c26e244 100644
--- a/docs/user-manual/modules/ROOT/pages/jbang-commands/camel-jbang-harden.adoc
+++ b/docs/user-manual/modules/ROOT/pages/jbang-commands/camel-jbang-harden.adoc
@@ -23,11 +23,11 @@ camel harden [options]
 | `--api-type` | API type: 'ollama' or 'openai' (OpenAI-compatible) | ollama | 
ApiType
 | `--catalog-context` | Include Camel Catalog descriptions in the prompt |  | 
boolean
 | `--format` | Output format (text, markdown) | text | String
-| `--model` | Model to use | llama3.2 | String
+| `--model` | Model to use. Default: with Ollama the installed model best at 
tool calling, such as qwen3.6; with an OpenAI-compatible server the first model 
it lists |  | String
 | `--show-prompt` | Show the prompt sent to the LLM |  | boolean
 | `--stream` | Stream the response as it's generated (shows progress) | true | 
boolean
 | `--system-prompt` | Custom system prompt |  | String
-| `--temperature` | Temperature for response generation (0.0-2.0) | 0.7 | 
double
+| `--temperature` | Temperature for response generation (0.0-2.0), for a model 
on this machine: a hosted model uses its own default | 0.7 | double
 | `--timeout` | Timeout in seconds for LLM response | 120 | int
 | `--url` | LLM API endpoint URL. Auto-detected from 'camel infra' for Ollama 
if not specified. |  | String
 | `--verbose,-v` | Include detailed security recommendations with code 
examples |  | boolean
diff --git 
a/docs/user-manual/modules/ROOT/pages/jbang-commands/camel-jbang-overview.adoc 
b/docs/user-manual/modules/ROOT/pages/jbang-commands/camel-jbang-overview.adoc
index 6bcc7a6f1af1..44084481bebc 100644
--- 
a/docs/user-manual/modules/ROOT/pages/jbang-commands/camel-jbang-overview.adoc
+++ 
b/docs/user-manual/modules/ROOT/pages/jbang-commands/camel-jbang-overview.adoc
@@ -24,10 +24,10 @@ camel overview [options]
 | `--api-type` | API type: 'ollama', 'openai' (OpenAI-compatible), 'anthropic' 
(Anthropic/Vertex AI), or 'watsonx' (IBM watsonx.ai) |  | ApiType
 | `--apply-descriptions` | Put the AI-assisted route descriptions and notes of 
camel-summary.md into the route sources, so they become part of the routes |  | 
boolean
 | `--format` | Output format (markdown, json) | markdown | String
-| `--model` | Model to use | llama3.2 | String
+| `--model` | Model to use. Default: the provider's default; with Ollama the 
installed model best at tool calling, such as qwen3.6 |  | String
 | `--save` | Save the overview as camel-summary.md in the project directory, 
keeping its AI-assisted sections |  | boolean
 | `--show-prompt` | Show the prompt sent to the LLM |  | boolean
-| `--temperature` | Temperature for response generation (0.0-2.0) | 0.3 | 
double
+| `--temperature` | Temperature for response generation (0.0-2.0), for a model 
on this machine: a hosted model uses its own default | 0.3 | double
 | `--timeout` | Timeout in seconds for LLM response | 300 | int
 | `--url` | LLM API endpoint URL. Auto-detected from 'camel infra' for Ollama 
if not specified. Also reads AZURE_OPENAI_ENDPOINT or WATSONX_URL env vars |  | 
String
 | `-h,--help` | Display this help and exit |  | boolean
diff --git 
a/dsl/camel-jbang/camel-jbang-core/src/generated/resources/META-INF/camel-jbang-commands-metadata.json
 
b/dsl/camel-jbang/camel-jbang-core/src/generated/resources/META-INF/camel-jbang-commands-metadata.json
index c4b8e2ec8a85..40724a559901 100644
--- 
a/dsl/camel-jbang/camel-jbang-core/src/generated/resources/META-INF/camel-jbang-commands-metadata.json
+++ 
b/dsl/camel-jbang/camel-jbang-core/src/generated/resources/META-INF/camel-jbang-commands-metadata.json
@@ -13,17 +13,17 @@
     { "name": "doc", "fullName": "doc", "description": "Shows documentation 
for kamelet, component, EIP, and other Camel resources", "sourceClass": 
"org.apache.camel.dsl.jbang.core.commands.catalog.CatalogDoc", "options": [ { 
"names": "--camel-version", "description": "To use a different Camel version 
than the default version", "javaType": "java.lang.String", "type": "string" }, 
{ "names": "--download", "description": "Whether to allow automatic downloading 
JAR dependencies (over the int [...]
     { "name": "doctor", "fullName": "doctor", "description": "Checks the 
environment and reports potential issues", "sourceClass": 
"org.apache.camel.dsl.jbang.core.commands.Doctor", "options": [ { "names": 
"--json", "description": "Output in JSON Format: the status (OK, WARN or FAIL) 
of each check", "javaType": "boolean", "type": "boolean" }, { "names": 
"-h,--help", "description": "Display this help and exit", "javaType": 
"boolean", "type": "boolean" } ] },
     { "name": "eval", "fullName": "eval", "description": "Evaluate Camel 
expressions and scripts", "sourceClass": 
"org.apache.camel.dsl.jbang.core.commands.EvalCommand", "options": [ { "names": 
"-h,--help", "description": "Display this help and exit", "javaType": 
"boolean", "type": "boolean" } ], "subcommands": [ { "name": "expression", 
"fullName": "eval expression", "description": "Evaluates Camel expression", 
"sourceClass": "org.apache.camel.dsl.jbang.core.commands.action.EvalExpressio 
[...]
-    { "name": "explain", "fullName": "explain", "description": "Explain what a 
Camel route does using AI\/LLM", "sourceClass": 
"org.apache.camel.dsl.jbang.core.commands.Explain", "options": [ { "names": 
"--api-key", "description": "API key for authentication. Also reads 
ANTHROPIC_API_KEY, OPENAI_API_KEY, WATSONX_APIKEY, or LLM_API_KEY env vars", 
"javaType": "java.lang.String", "type": "string" }, { "names": "--api-type", 
"description": "API type: 'ollama', 'openai' (OpenAI-compatible), ' [...]
+    { "name": "explain", "fullName": "explain", "description": "Explain what a 
Camel route does using AI\/LLM", "sourceClass": 
"org.apache.camel.dsl.jbang.core.commands.Explain", "options": [ { "names": 
"--api-key", "description": "API key for authentication. Also reads 
ANTHROPIC_API_KEY, OPENAI_API_KEY, WATSONX_APIKEY, or LLM_API_KEY env vars", 
"javaType": "java.lang.String", "type": "string" }, { "names": "--api-type", 
"description": "API type: 'ollama', 'openai' (OpenAI-compatible), ' [...]
     { "name": "export", "fullName": "export", "description": "Export to other 
runtimes (Camel Main, Spring Boot, or Quarkus)", "sourceClass": 
"org.apache.camel.dsl.jbang.core.commands.Export", "options": [ { "names": 
"--build-property", "description": "Maven build properties, ex. 
--build-property=prop1=foo", "javaType": "java.util.List", "type": "array" }, { 
"names": "--camel-spring-boot-version", "description": "Camel version to use 
with Spring Boot", "javaType": "java.lang.String", "ty [...]
     { "name": "get", "fullName": "get", "description": "Get status of Camel 
integrations", "sourceClass": 
"org.apache.camel.dsl.jbang.core.commands.process.CamelStatus", "options": [ { 
"names": "--watch", "description": "Execute periodically and showing output 
fullscreen", "javaType": "boolean", "type": "boolean" }, { "names": 
"-h,--help", "description": "Display this help and exit", "javaType": 
"boolean", "type": "boolean" } ], "subcommands": [ { "name": "about", 
"fullName": "get about" [...]
-    { "name": "harden", "fullName": "harden", "description": "Suggest security 
hardening for Camel routes using AI\/LLM", "sourceClass": 
"org.apache.camel.dsl.jbang.core.commands.Harden", "options": [ { "names": 
"--api-key", "description": "API key for authentication. Also reads 
OPENAI_API_KEY or LLM_API_KEY env vars", "javaType": "java.lang.String", 
"type": "string" }, { "names": "--api-type", "description": "API type: 'ollama' 
or 'openai' (OpenAI-compatible)", "defaultValue": "ollama", [...]
+    { "name": "harden", "fullName": "harden", "description": "Suggest security 
hardening for Camel routes using AI\/LLM", "sourceClass": 
"org.apache.camel.dsl.jbang.core.commands.Harden", "options": [ { "names": 
"--api-key", "description": "API key for authentication. Also reads 
OPENAI_API_KEY or LLM_API_KEY env vars", "javaType": "java.lang.String", 
"type": "string" }, { "names": "--api-type", "description": "API type: 'ollama' 
or 'openai' (OpenAI-compatible)", "defaultValue": "ollama", [...]
     { "name": "hawtio", "fullName": "hawtio", "description": "Launch Hawtio 
web console", "sourceClass": 
"org.apache.camel.dsl.jbang.core.commands.process.Hawtio", "options": [ { 
"names": "--host", "description": "Hostname to bind the Hawtio web console to", 
"defaultValue": "127.0.0.1", "javaType": "java.lang.String", "type": "string" 
}, { "names": "--open-url,--openUrl", "description": "To automatic open Hawtio 
web console in the web browser (without an interactive terminal the default  
[...]
     { "name": "infra", "fullName": "infra", "description": "List and Run 
external services for testing and prototyping", "sourceClass": 
"org.apache.camel.dsl.jbang.core.commands.infra.InfraCommand", "options": [ { 
"names": "--json", "description": "Output in JSON Format", "javaType": 
"boolean", "type": "boolean" }, { "names": "-h,--help", "description": "Display 
this help and exit", "javaType": "boolean", "type": "boolean" } ], 
"subcommands": [ { "name": "get", "fullName": "infra get", " [...]
     { "name": "init", "fullName": "init", "description": "Creates a new Camel 
integration", "sourceClass": "org.apache.camel.dsl.jbang.core.commands.Init", 
"options": [ { "names": "--clean-dir,--clean-directory", "description": 
"Whether to clean directory first (deletes all files in directory)", 
"javaType": "boolean", "type": "boolean" }, { "names": "--dir,--directory", 
"description": "Directory relative path where the new Camel integration will be 
saved", "defaultValue": ".", "javaType" [...]
     { "name": "jolokia", "fullName": "jolokia", "description": "Attach Jolokia 
JVM Agent to a running Camel integration", "sourceClass": 
"org.apache.camel.dsl.jbang.core.commands.process.Jolokia", "options": [ { 
"names": "--port", "description": "To use a specific port number when attaching 
Jolokia JVM Agent (default a free port is found in range 8778-9999)", 
"javaType": "int", "type": "integer" }, { "names": "--stop", "description": 
"Stops the Jolokia JVM Agent in the running Camel inte [...]
     { "name": "log", "fullName": "log", "description": "Tail logs from running 
Camel integrations", "sourceClass": 
"org.apache.camel.dsl.jbang.core.commands.action.CamelLogAction", "options": [ 
{ "names": "--duration", "description": "How long to keep watching or 
following, in millis or as a duration such as 30s (default: until Enter is 
pressed)", "javaType": "long", "type": "integer" }, { "names": "--find", 
"description": "Find and highlight matching text (ignore case).", "javaType": 
"j [...]
     { "name": "nano", "fullName": "nano", "description": "Nano editor to edit 
file", "sourceClass": "org.apache.camel.dsl.jbang.core.commands.Nano", 
"options": [ { "names": "-h,--help", "description": "Display this help and 
exit", "javaType": "boolean", "type": "boolean" } ] },
-    { "name": "overview", "fullName": "overview", "description": "High-level 
overview of the integrations in a project, optionally explained by AI\/LLM", 
"sourceClass": "org.apache.camel.dsl.jbang.core.commands.Overview", "options": 
[ { "names": "--ai", "description": "Ask an LLM to explain the project (an 
overview, the capabilities, labels and notes of routes that have none) and save 
it in camel-summary.md, marked as AI-assisted", "javaType": "boolean", "type": 
"boolean" }, { "names": " [...]
+    { "name": "overview", "fullName": "overview", "description": "High-level 
overview of the integrations in a project, optionally explained by AI\/LLM", 
"sourceClass": "org.apache.camel.dsl.jbang.core.commands.Overview", "options": 
[ { "names": "--ai", "description": "Ask an LLM to explain the project (an 
overview, the capabilities, labels and notes of routes that have none) and save 
it in camel-summary.md, marked as AI-assisted", "javaType": "boolean", "type": 
"boolean" }, { "names": " [...]
     { "name": "plugin", "fullName": "plugin", "description": "Manage plugins 
that add sub-commands to this CLI", "sourceClass": 
"org.apache.camel.dsl.jbang.core.commands.plugin.PluginCommand", "options": [ { 
"names": "-h,--help", "description": "Display this help and exit", "javaType": 
"boolean", "type": "boolean" } ], "subcommands": [ { "name": "add", "fullName": 
"plugin add", "description": "Add new plugin", "sourceClass": 
"org.apache.camel.dsl.jbang.core.commands.plugin.PluginAdd", "o [...]
     { "name": "ps", "fullName": "ps", "description": "List running Camel 
integrations", "sourceClass": 
"org.apache.camel.dsl.jbang.core.commands.process.ListProcess", "options": [ { 
"names": "--duration", "description": "How long to keep watching or following, 
in millis or as a duration such as 30s (default: until Enter is pressed)", 
"javaType": "long", "type": "integer" }, { "names": "--json", "description": 
"Output in JSON Format", "javaType": "boolean", "type": "boolean" }, { "names": 
[...]
     { "name": "restart", "fullName": "restart", "description": "Restarts 
running Camel integrations (stop + re-launch)", "sourceClass": 
"org.apache.camel.dsl.jbang.core.commands.process.RestartProcess", "options": [ 
{ "names": "--timeout", "description": "How long to wait for the integration to 
stop, and again to start, in millis or as a duration such as 30s", 
"defaultValue": "60000", "javaType": "long", "type": "integer" }, { "names": 
"-h,--help", "description": "Display this help and e [...]
diff --git 
a/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/Ask.java
 
b/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/Ask.java
index 8bca3919aa8c..0373e1749d71 100644
--- 
a/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/Ask.java
+++ 
b/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/Ask.java
@@ -30,6 +30,7 @@ import org.aesh.command.option.Option;
 import org.apache.camel.dsl.jbang.core.commands.ai.AnswerChecks;
 import org.apache.camel.dsl.jbang.core.commands.ai.ToolConfirms;
 import org.apache.camel.dsl.jbang.core.common.EnvironmentHelper;
+import org.apache.camel.dsl.jbang.core.common.OllamaDoctorSupport;
 import org.apache.camel.dsl.jbang.core.common.RuntimeHelper;
 import org.apache.camel.dsl.jbang.run.common.Printer;
 import org.apache.camel.util.json.JsonObject;
@@ -54,7 +55,6 @@ import org.jline.terminal.TerminalBuilder;
                            "  camel ask                                   
(interactive chat)" })
 public class Ask extends CamelCommand {
 
-    private static final String DEFAULT_MODEL = "llama3.2";
     private static final String NO_PROCESS
             = "No running Camel process connected. Start one with: camel run 
<file>";
 
@@ -117,7 +117,7 @@ public class Ask extends CamelCommand {
     }
 
     /** Below this size (billions of parameters) a model's answers about a 
running integration are often wrong. */
-    static final double SMALL_MODEL_BILLIONS = 8;
+    static final double SMALL_MODEL_BILLIONS = 
OllamaDoctorSupport.MIN_MODEL_BILLIONS;
 
     // the tools the model does not get, such as starting a chaos exploration 
that camel cmd chaos-explore started
     Set<String> hiddenTools = Set.of();
@@ -153,11 +153,9 @@ public class Ask extends CamelCommand {
                 .withMaxTokens(4096)
                 .withVerbose(verbose)
                 .withPrinter(printer());
+        // without a model, each provider uses its default; Ollama the 
installed model best at tool calling
         if (model != null && !model.isBlank()) {
             client.withExplicitModel(model);
-        } else {
-            // the placeholder that Ollama replaces with the installed model 
best at tool calling
-            client.withModel(DEFAULT_MODEL);
         }
 
         if (!client.detectEndpoint()) {
@@ -264,7 +262,8 @@ public class Ask extends CamelCommand {
             totalUsage = totalUsage.add(response.usage());
 
             if (response.toolCalls() != null && 
!response.toolCalls().isEmpty()) {
-                
messages.add(LlmClient.Message.assistantWithToolCalls(response.text(), 
response.toolCalls()));
+                // as it came: Anthropic wants the thinking blocks of a tool 
loop back unchanged
+                messages.add(LlmClient.Message.assistant(response));
 
                 List<LlmClient.ToolResult> results = new ArrayList<>();
                 for (LlmClient.ToolCall toolCall : response.toolCalls()) {
@@ -388,8 +387,8 @@ public class Ask extends CamelCommand {
     }
 
     /**
-     * A note about a small model, or null: in a test, the llama3.2 (3B) that 
the CLI names as its default got three of
-     * four questions about a running integration wrong (CAMEL-25676).
+     * A note about a small model, or null: in a test, llama3.2 (3B), then the 
default of the CLI, got three of four
+     * questions about a running integration wrong (CAMEL-25676).
      */
     static String smallModelNote(String name, double billions) {
         if (billions <= 0 || billions >= SMALL_MODEL_BILLIONS) {
diff --git 
a/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/Explain.java
 
b/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/Explain.java
index 31280ebce01a..d0b95fb32184 100644
--- 
a/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/Explain.java
+++ 
b/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/Explain.java
@@ -62,8 +62,6 @@ public class Explain extends CamelCommand {
         }
     }
 
-    private static final String DEFAULT_MODEL = "llama3.2";
-
     private static final List<String> COMMON_COMPONENTS = Arrays.asList(
             "kafka", "http", "https", "file", "timer", "direct", "seda",
             "log", "mock", "rest", "rest-api", "sql", "jms", "activemq",
@@ -102,8 +100,10 @@ public class Explain extends CamelCommand {
             description = "API key for authentication. Also reads 
ANTHROPIC_API_KEY, OPENAI_API_KEY, WATSONX_APIKEY, or LLM_API_KEY env vars")
     String apiKey;
 
-    @Option(name = "model", description = "Model to use", defaultValue = 
DEFAULT_MODEL)
-    String model = DEFAULT_MODEL;
+    @Option(name = "model",
+            description = "Model to use. Default: the provider's default; with 
Ollama the installed model best at"
+                          + " tool calling, such as qwen3.6")
+    String model;
 
     @Option(name = "verbose", shortName = 'v', hasValue = false, description = 
"Include detailed technical information")
     boolean verbose;
@@ -121,7 +121,9 @@ public class Explain extends CamelCommand {
     @Option(name = "show-prompt", hasValue = false, description = "Show the 
prompt sent to the LLM")
     boolean showPrompt;
 
-    @Option(name = "temperature", description = "Temperature for response 
generation (0.0-2.0)", defaultValue = "0.7")
+    @Option(name = "temperature",
+            description = "Temperature for response generation (0.0-2.0), for 
a model on this machine: a hosted model uses its own default",
+            defaultValue = "0.7")
     double temperature = 0.7;
 
     @Option(name = "system-prompt", description = "Custom system prompt")
@@ -152,7 +154,7 @@ public class Explain extends CamelCommand {
                 .withUrl(url)
                 .withApiType(apiType)
                 .withApiKey(apiKey)
-                .withModel(model)
+                .withExplicitModel(model)
                 .withTimeout(timeout)
                 .withTemperature(temperature)
                 .withStream(stream)
diff --git 
a/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/Harden.java
 
b/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/Harden.java
index 7c2c7cfa640a..4af04e713baf 100644
--- 
a/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/Harden.java
+++ 
b/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/Harden.java
@@ -63,7 +63,7 @@ import org.apache.camel.util.json.Jsoner;
                            "Examples:",
                            "  camel harden hello.java",
                            "  camel harden hello.yaml --format=markdown",
-                           "  camel harden hello.java --model=gpt-4" })
+                           "  camel harden hello.java --model=qwen3.6:35b-a3b" 
})
 public class Harden extends CamelCommand {
 
     public static class FormatCompletionCandidates extends 
CompletionCandidates {
@@ -78,7 +78,6 @@ public class Harden extends CamelCommand {
     }
 
     private static final String DEFAULT_OLLAMA_URL = "http://localhost:11434";;
-    private static final String DEFAULT_MODEL = "llama3.2";
     private static final int CONNECT_TIMEOUT_SECONDS = 10;
     private static final int HEALTH_CHECK_TIMEOUT_SECONDS = 5;
 
@@ -148,8 +147,10 @@ public class Harden extends CamelCommand {
     @Option(name = "api-key", description = "API key for authentication. Also 
reads OPENAI_API_KEY or LLM_API_KEY env vars")
     String apiKey;
 
-    @Option(name = "model", description = "Model to use", defaultValue = 
DEFAULT_MODEL)
-    String model = DEFAULT_MODEL;
+    @Option(name = "model",
+            description = "Model to use. Default: with Ollama the installed 
model best at tool calling, such as qwen3.6;"
+                          + " with an OpenAI-compatible server the first model 
it lists")
+    String model;
 
     @Option(name = "verbose", shortName = 'v', hasValue = false,
             description = "Include detailed security recommendations with code 
examples")
@@ -168,7 +169,9 @@ public class Harden extends CamelCommand {
     @Option(name = "show-prompt", hasValue = false, description = "Show the 
prompt sent to the LLM")
     boolean showPrompt;
 
-    @Option(name = "temperature", description = "Temperature for response 
generation (0.0-2.0)", defaultValue = "0.7")
+    @Option(name = "temperature",
+            description = "Temperature for response generation (0.0-2.0), for 
a model on this machine: a hosted model uses its own default",
+            defaultValue = "0.7")
     double temperature = 0.7;
 
     @Option(name = "system-prompt", description = "Custom system prompt")
@@ -218,6 +221,9 @@ public class Harden extends CamelCommand {
         }
 
         String resolvedApiKey = resolveApiKey();
+        if (model == null || model.isBlank()) {
+            model = defaultModel(endpoint, resolvedApiKey);
+        }
         printConfiguration(endpoint, resolvedApiKey);
 
         for (String file : files) {
@@ -229,6 +235,18 @@ public class Harden extends CamelCommand {
         return 0;
     }
 
+    /**
+     * The model when none is given, chosen as the other AI commands choose 
it: with Ollama the installed model best at
+     * tool calling, with an OpenAI-compatible server the first model it 
lists, and with OpenAI its default.
+     */
+    private String defaultModel(String endpoint, String resolvedApiKey) {
+        LlmClient client = LlmClient.create()
+                .withApiType(apiType == ApiType.ollama ? 
LlmClient.ApiType.ollama : LlmClient.ApiType.openai)
+                .withUrl(endpoint)
+                .withApiKey(resolvedApiKey);
+        return client.detectEndpoint() ? client.model() : null;
+    }
+
     private void printConfiguration(String endpoint, String resolvedApiKey) {
         printer().printDiagnostic("LLM Configuration:");
         printer().printDiagnostic("  URL: " + endpoint);
@@ -573,7 +591,10 @@ public class Harden extends CamelCommand {
         JsonObject request = new JsonObject();
         request.put("model", model);
         request.put("messages", messages);
-        request.put("temperature", temperature);
+        // a hosted model gets its own default sampling, as with the other AI 
commands
+        if (LlmClient.isLocalUrl(endpoint)) {
+            request.put("temperature", temperature);
+        }
 
         String apiUrl = normalizeOpenAiUrl(endpoint);
 
diff --git 
a/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/LlmClient.java
 
b/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/LlmClient.java
index 11d280f58737..346fc200445b 100644
--- 
a/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/LlmClient.java
+++ 
b/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/LlmClient.java
@@ -31,6 +31,7 @@ import java.util.ArrayList;
 import java.util.Collection;
 import java.util.HashMap;
 import java.util.List;
+import java.util.Locale;
 import java.util.Map;
 import java.util.Objects;
 import java.util.stream.Collectors;
@@ -52,11 +53,36 @@ public class LlmClient {
     private static final String DEFAULT_OLLAMA_URL = "http://localhost:11434";;
     private static final String OLLAMA_ROOT_RESPONSE = "Ollama is running";
     private static final String DEFAULT_ANTHROPIC_URL = 
"https://api.anthropic.com";;
-    private static final String ANTHROPIC_VERSION = "2023-06-01";
-    private static final String VERTEX_ANTHROPIC_VERSION = "vertex-2023-10-16";
-    private static final String DEFAULT_ANTHROPIC_MODEL = "claude-sonnet-4-6";
-    private static final String DEFAULT_OPENAI_MODEL = "gpt-4o-mini";
-    private static final String DEFAULT_OLLAMA_MODEL = "llama3.2";
+    // the anthropic-version header, and on Vertex AI the anthropic_version of 
the body
+    private static final String ANTHROPIC_VERSION = "2026-10-21";
+    private static final String DEFAULT_ANTHROPIC_MODEL = "claude-sonnet-5-5";
+    /**
+     * The smallest {@code max_tokens} of a request to Anthropic. The current 
Claude models think before they answer,
+     * and the thinking counts toward {@code max_tokens}, so the 4096 the 
commands ask for would cut long answers off.
+     * 16,000 is about the most for a request that does not stream.
+     */
+    private static final int ANTHROPIC_MIN_MAX_TOKENS = 16_000;
+    // OpenAI's fast, low-cost model, the replacement it names for gpt-4o-mini 
(shut down on Jan 14 2027). Its Chat
+    // Completions call tools only with reasoning_effort none, see 
turnThinkingOff.
+    private static final String DEFAULT_OPENAI_MODEL = "gpt-6-luna";
+    /**
+     * The model name the CLI commands once passed for "no model chosen", and 
still read that way: Ollama replaces it
+     * with the installed model best at tool calling, and the other providers 
use their own default. The model is too
+     * small for Camel, so it is never chosen on purpose.
+     */
+    private static final String PLACEHOLDER_MODEL = "llama3.2";
+    /**
+     * The model asked of Ollama when no installed model can be used: the one 
recommended for Camel, so that the error
+     * of Ollama names the model to pull.
+     */
+    private static final String FALLBACK_OLLAMA_MODEL = "qwen3.6:35b-a3b";
+    /**
+     * The installed Ollama model used when none is chosen, by family in this 
order: the ones measured on Camel tasks
+     * first, then others that call tools well.
+     */
+    private static final List<String> PREFERRED_OLLAMA_MODELS = List.of(
+            "qwen3.6", "qwen3.8", "qwen3.5", "qwen3", "nemotron-3-nano", 
"mistral-nemo", "qwen2.5", "granite4.1",
+            "llama3.1", "llama3.3", "mistral");
     /**
      * Keep the model (and its prompt cache) loaded between turns of a 
conversation. Ollama's default of five minutes is
      * shorter than a slow local answer plus the time the user spends reading 
it, after which the next request pays for
@@ -82,13 +108,14 @@ public class LlmClient {
     /** Last-resort Azure deployment segment when URL normalization runs 
before model detection completes. */
     private static final String DEFAULT_AZURE_DEPLOYMENT_FALLBACK = "gpt-4o";
     private static final String DEFAULT_GEMINI_URL = 
"https://generativelanguage.googleapis.com/v1beta";;
-    private static final String DEFAULT_GEMINI_MODEL = "gemini-2.0-flash";
+    // the alias Google moves to each new Flash model, with two weeks' notice, 
so the default does not go out of
+    // service as gemini-2.0-flash did
+    private static final String DEFAULT_GEMINI_MODEL = "gemini-flash-latest";
     private static final int CONNECT_TIMEOUT_SECONDS = 10;
     private static final int HEALTH_CHECK_TIMEOUT_SECONDS = 5;
 
     // Pre-4.6 Claude models require a @date suffix on Vertex AI
     private static final Map<String, String> VERTEX_MODEL_MAP = Map.of(
-            "claude-sonnet-4-5", "claude-sonnet-4-5@20250929",
             "claude-opus-4-5", "claude-opus-4-5@20251101",
             "claude-haiku-4-5", "claude-haiku-4-5@20251001");
 
@@ -120,7 +147,17 @@ public class LlmClient {
     public record ToolResult(String toolCallId, String content) {
     }
 
-    public record Message(String role, String content, List<ToolCall> 
toolCalls, List<ToolResult> toolResults) {
+    /**
+     * A message of the conversation. {@code contentBlocks} is an answer of 
Anthropic as it came, thinking blocks
+     * included, which the API wants back unchanged in a tool loop (see {@link 
#assistant(ChatResponse)}); null for the
+     * other messages.
+     */
+    public record Message(String role, String content, List<ToolCall> 
toolCalls, List<ToolResult> toolResults,
+            JsonArray contentBlocks) {
+
+        public Message(String role, String content, List<ToolCall> toolCalls, 
List<ToolResult> toolResults) {
+            this(role, content, toolCalls, toolResults, null);
+        }
 
         public static Message user(String text) {
             return new Message("user", text, null, null);
@@ -130,6 +167,14 @@ public class LlmClient {
             return new Message("assistant", text, calls, null);
         }
 
+        /**
+         * The answer of the model for the history: its text and tool calls, 
and with Anthropic its content blocks as
+         * they came, as Anthropic wants the thinking blocks of a tool loop 
back unchanged.
+         */
+        public static Message assistant(ChatResponse response) {
+            return new Message("assistant", response.text(), 
response.toolCalls(), null, response.contentBlocks());
+        }
+
         public static Message toolResults(List<ToolResult> results) {
             return new Message("tool", null, null, results);
         }
@@ -175,8 +220,17 @@ public class LlmClient {
         }
     }
 
+    /**
+     * The answer of a chat request. {@code contentBlocks} is the content of 
an Anthropic answer as it came, thinking
+     * blocks included; null for the other providers.
+     */
     public record ChatResponse(String text, List<ToolCall> toolCalls, String 
stopReason, boolean streamed,
-            TokenUsage usage) {
+            TokenUsage usage, JsonArray contentBlocks) {
+
+        public ChatResponse(String text, List<ToolCall> toolCalls, String 
stopReason, boolean streamed,
+                            TokenUsage usage) {
+            this(text, toolCalls, stopReason, streamed, usage, null);
+        }
     }
 
     public static String formatTokens(int tokens) {
@@ -257,9 +311,11 @@ public class LlmClient {
      * far slower than hosted ones, so callers use this to trim what they send 
per request.
      */
     public boolean isLocalEndpoint() {
-        if (apiType == ApiType.ollama) {
-            return true;
-        }
+        return apiType == ApiType.ollama || isLocalUrl(url);
+    }
+
+    /** Whether the host of the URL is a loopback address, as the server of a 
model on this machine has. */
+    public static boolean isLocalUrl(String url) {
         if (url == null) {
             return false;
         }
@@ -272,6 +328,15 @@ public class LlmClient {
         }
     }
 
+    /**
+     * Whether a request carries the temperature: only for a model on this 
machine, where the tests of the AI tools ran
+     * with it. A hosted model gets its own default sampling: the reasoning 
models of OpenAI refuse any other value, and
+     * Anthropic refuses the field.
+     */
+    boolean sendsTemperature() {
+        return isLocalEndpoint();
+    }
+
     // -- Builder --
 
     public static LlmClient create() {
@@ -392,19 +457,19 @@ public class LlmClient {
                 if (openAiAuthMode == OpenAiAuthMode.api_key
                         || (url != null && isAzureOpenAiEndpoint(url))) {
                     resolveAzureOpenAiModel();
-                } else if (model == null || model.isBlank() || 
DEFAULT_OLLAMA_MODEL.equals(model)) {
+                } else if (model == null || model.isBlank() || 
PLACEHOLDER_MODEL.equals(model)) {
                     model = isOpenAiCompatibleServer() ? 
resolveOpenAiCompatibleModel() : DEFAULT_OPENAI_MODEL;
                 }
             }
             case gemini -> {
-                if (model == null || model.isBlank() || 
DEFAULT_OLLAMA_MODEL.equals(model)) {
+                if (model == null || model.isBlank() || 
PLACEHOLDER_MODEL.equals(model)) {
                     model = DEFAULT_GEMINI_MODEL;
                 }
             }
             case ollama -> {
                 if (model == null || model.isBlank()) {
                     model = resolveDefaultOllamaModel();
-                } else if (DEFAULT_OLLAMA_MODEL.equals(model) && 
!explicitModel) {
+                } else if (PLACEHOLDER_MODEL.equals(model) && !explicitModel) {
                     resolveOllamaModel();
                 }
             }
@@ -670,7 +735,9 @@ public class LlmClient {
         JsonObject request = new JsonObject();
         request.put("model", model);
         request.put("messages", messages);
-        request.put("temperature", temperature);
+        if (sendsTemperature()) {
+            request.put("temperature", temperature);
+        }
         turnThinkingOff(request);
 
         String resolvedKey = resolveApiKey();
@@ -721,11 +788,15 @@ public class LlmClient {
             request.put("contents", contents);
         }
         JsonObject generationConfig = new JsonObject();
-        generationConfig.put("temperature", temperature);
+        if (sendsTemperature()) {
+            generationConfig.put("temperature", temperature);
+        }
         if (maxTokens > 0) {
             generationConfig.put("maxOutputTokens", maxTokens);
         }
-        request.put("generationConfig", generationConfig);
+        if (!generationConfig.isEmpty()) {
+            request.put("generationConfig", generationConfig);
+        }
         JsonArray declarations = buildGeminiFunctionDeclarations(tools);
         if (declarations != null) {
             JsonObject toolsObj = new JsonObject();
@@ -1014,7 +1085,9 @@ public class LlmClient {
         JsonObject request = new JsonObject();
         request.put("model", model);
         request.put("messages", jsonMessages);
-        request.put("temperature", temperature);
+        if (sendsTemperature()) {
+            request.put("temperature", temperature);
+        }
         if (jsonTools != null) {
             request.put("tools", jsonTools);
         }
@@ -1027,27 +1100,30 @@ public class LlmClient {
         return parseOpenAiChatResponse(response);
     }
 
-    // ---- Thinking off with a local OpenAI-compatible server ----
+    // ---- Thinking off with an OpenAI-compatible server ----
 
     // the fields for thinking a server refused, left out from then on
     private volatile boolean reasoningEffortRefused;
     private volatile boolean templateKwargsRefused;
 
     /**
-     * Turns thinking off in a request to an OpenAI-compatible server on this 
machine, as {@code "think": false} does on
-     * Ollama's own API. Local servers (Ollama's /v1, llama.cpp server, vLLM, 
LM Studio) render the model's own chat
-     * template, and the Qwen 3.x templates think by default, which makes 
every answer many times slower. Ollama reads
-     * {@code reasoning_effort}, llama.cpp reads {@code chat_template_kwargs}, 
and vLLM reads both. Hosted endpoints get
-     * neither field, as they may refuse fields they do not know.
+     * Turns thinking off in a request to an OpenAI-compatible server, as 
{@code "think": false} does on Ollama's own
+     * API. Local servers (Ollama's /v1, llama.cpp server, vLLM, LM Studio) 
render the model's own chat template, and
+     * the Qwen 3.x templates think by default, which makes every answer many 
times slower. Ollama reads
+     * {@code reasoning_effort}, llama.cpp reads {@code chat_template_kwargs}, 
and vLLM reads both. OpenAI itself and
+     * Azure get {@code reasoning_effort} only: the reasoning models of OpenAI 
call tools on Chat Completions only with
+     * {@code none}. A model that does not know the field refuses it, and gets 
the request again without it (see
+     * {@link #sendWithFewerThinkingFields}).
      */
     void turnThinkingOff(JsonObject request) {
-        if (apiType == ApiType.openai && isLocalEndpoint()) {
-            if (!reasoningEffortRefused) {
-                request.put("reasoning_effort", "none");
-            }
-            if (!templateKwargsRefused) {
-                request.put("chat_template_kwargs", new 
JsonObject(Map.of("enable_thinking", false)));
-            }
+        if (apiType != ApiType.openai) {
+            return;
+        }
+        if (!reasoningEffortRefused) {
+            request.put("reasoning_effort", "none");
+        }
+        if (isLocalEndpoint() && !templateKwargsRefused) {
+            request.put("chat_template_kwargs", new 
JsonObject(Map.of("enable_thinking", false)));
         }
     }
 
@@ -1591,8 +1667,17 @@ public class LlmClient {
     private ChatResponse chatAnthropicFormat(String systemPrompt, 
List<Message> messages, List<ToolDef> tools) {
         JsonArray jsonMessages = new JsonArray();
 
-        for (Message msg : messages) {
-            if (msg.toolCalls() != null && !msg.toolCalls().isEmpty()) {
+        int question = lastQuestion(messages);
+        for (int i = 0; i < messages.size(); i++) {
+            Message msg = messages.get(i);
+            JsonArray blocks = msg.contentBlocks() != null ? 
anthropicBlocks(msg.contentBlocks(), i > question) : null;
+            if (blocks != null) {
+                // an answer of the model as it came
+                JsonObject assistantMsg = new JsonObject();
+                assistantMsg.put("role", "assistant");
+                assistantMsg.put("content", blocks);
+                jsonMessages.add(assistantMsg);
+            } else if (msg.toolCalls() != null && !msg.toolCalls().isEmpty()) {
                 // assistant message with tool_use blocks
                 JsonObject assistantMsg = new JsonObject();
                 assistantMsg.put("role", "assistant");
@@ -1662,27 +1747,87 @@ public class LlmClient {
 
     // ---- Anthropic helpers ----
 
+    /**
+     * The index of the last message the user wrote: the question, or a 
request after an answer such as a correction.
+     * Tool results are not one.
+     */
+    private static int lastQuestion(List<Message> messages) {
+        for (int i = messages.size() - 1; i >= 0; i--) {
+            Message m = messages.get(i);
+            if ("user".equals(m.role()) && (m.toolResults() == null || 
m.toolResults().isEmpty())) {
+                return i;
+            }
+        }
+        return -1;
+    }
+
+    /**
+     * The content blocks of an earlier answer for the next request, or null 
when none are left. The answers since the
+     * last question go back as they came: in a tool loop the API wants the 
thinking blocks back unchanged. The answers
+     * before it go back without their thinking blocks. The callers change 
that part of the history (the AI panel
+     * compacts old tool results and drops old turns, a retry removes a turn), 
and the current models refuse a thinking
+     * block when the history before it changed. Thinking blocks may be left 
out from the start of the history, so the
+     * later ones stay valid.
+     */
+    private static JsonArray anthropicBlocks(JsonArray blocks, boolean 
sinceLastQuestion) {
+        JsonArray kept = new JsonArray();
+        for (Object block : blocks) {
+            if (sinceLastQuestion || !isThinkingBlock(block)) {
+                kept.add(block);
+            }
+        }
+        return kept.isEmpty() ? null : kept;
+    }
+
+    private static boolean isThinkingBlock(Object block) {
+        return block instanceof JsonObject b
+                && ("thinking".equals(b.getString("type")) || 
"redacted_thinking".equals(b.getString("type")));
+    }
+
     private JsonObject buildAnthropicRequest(String systemPrompt, JsonArray 
messages, JsonArray tools) {
         JsonObject request = new JsonObject();
         if (isVertexAi()) {
             // Vertex AI: model is in the URL, version goes in body
-            request.put("anthropic_version", VERTEX_ANTHROPIC_VERSION);
+            request.put("anthropic_version", ANTHROPIC_VERSION);
         } else {
             // Direct Anthropic API: model goes in body
             request.put("model", model);
         }
-        request.put("max_tokens", maxTokens > 0 ? maxTokens : 4096);
+        request.put("max_tokens", Math.max(maxTokens, 
ANTHROPIC_MIN_MAX_TOKENS));
+        // no temperature: the current models refuse any value but the 
default, and API version 2026-10-21 refuses the
+        // field on every model
+        boolean toolLoop = tools != null && !tools.isEmpty();
         if (systemPrompt != null) {
-            request.put("system", systemPrompt);
+            request.put("system", toolLoop ? cachedText(systemPrompt) : 
systemPrompt);
         }
         request.put("messages", messages);
-        request.put("temperature", temperature);
-        if (tools != null && !tools.isEmpty()) {
+        if (toolLoop) {
             request.put("tools", tools);
+            // each step of a tool loop sends the tools, the system prompt and 
the history again. The breakpoint after
+            // the system prompt keeps those two in the cache when the history 
before the question changes (the AI
+            // panel compacts it); the automatic one moves to the end of each 
request for the history. A single
+            // question, as an explanation is, would only pay for writing the 
cache.
+            request.put("cache_control", ephemeral());
         }
         return request;
     }
 
+    private static JsonObject ephemeral() {
+        JsonObject cacheControl = new JsonObject();
+        cacheControl.put("type", "ephemeral");
+        return cacheControl;
+    }
+
+    private static JsonArray cachedText(String text) {
+        JsonObject block = new JsonObject();
+        block.put("type", "text");
+        block.put("text", text);
+        block.put("cache_control", ephemeral());
+        JsonArray blocks = new JsonArray();
+        blocks.add(block);
+        return blocks;
+    }
+
     private String resolveAnthropicUrl() {
         if (isVertexAi()) {
             String vertexModel = resolveVertexModel(model);
@@ -1887,9 +2032,10 @@ public class LlmClient {
         TokenUsage usage = extractAnthropicUsage(response);
         JsonArray contentBlocks = (JsonArray) response.get("content");
         if (contentBlocks == null) {
-            return new ChatResponse(null, List.of(), stopReason, false, usage);
+            contentBlocks = new JsonArray();
         }
 
+        // by the type of each block: an answer can begin with thinking 
blocks, which have no text for the user
         StringBuilder text = new StringBuilder();
         List<ToolCall> toolCalls = new ArrayList<>();
         for (Object obj : contentBlocks) {
@@ -1905,7 +2051,11 @@ public class LlmClient {
             }
         }
         String textContent = text.length() > 0 ? text.toString() : null;
-        return new ChatResponse(textContent, toolCalls, stopReason, false, 
usage);
+        if (textContent == null && toolCalls.isEmpty() && 
"refusal".equals(stopReason)) {
+            // the safety classifiers of the current models can decline, with 
no content
+            textContent = "The model declined to answer (stop reason: 
refusal).";
+        }
+        return new ChatResponse(textContent, toolCalls, stopReason, false, 
usage, contentBlocks);
     }
 
     // ---- Token usage extraction ----
@@ -1931,9 +2081,12 @@ public class LlmClient {
         if (usage == null) {
             return TokenUsage.EMPTY;
         }
-        int input = getIntValue(usage, "input_tokens");
+        // input_tokens is only the part of the prompt that was neither read 
from the cache nor written to it; the
+        // prompt is the sum, as the other providers count it
+        int cached = getIntValue(usage, "cache_read_input_tokens");
+        int input = getIntValue(usage, "input_tokens") + getIntValue(usage, 
"cache_creation_input_tokens") + cached;
         int output = getIntValue(usage, "output_tokens");
-        return new TokenUsage(input, output, input + output, 
getIntValue(usage, "cache_read_input_tokens"), 0, 0);
+        return new TokenUsage(input, output, input + output, cached, 0, 0);
     }
 
     private static int getIntValue(JsonObject obj, String key) {
@@ -2243,7 +2396,7 @@ public class LlmClient {
             apiType = ApiType.anthropic;
             apiKey = key;
             url = DEFAULT_ANTHROPIC_URL;
-            if (model == null || "llama3.2".equals(model)) {
+            if (model == null || PLACEHOLDER_MODEL.equals(model)) {
                 model = DEFAULT_ANTHROPIC_MODEL;
             }
             return true;
@@ -2258,7 +2411,7 @@ public class LlmClient {
             apiType = ApiType.anthropic;
             vertexRegion = region;
             vertexProjectId = project;
-            if (model == null || "llama3.2".equals(model)) {
+            if (model == null || PLACEHOLDER_MODEL.equals(model)) {
                 model = DEFAULT_ANTHROPIC_MODEL;
             }
             return true;
@@ -2339,7 +2492,7 @@ public class LlmClient {
     }
 
     private boolean isGenericPlaceholderModel(String configuredModel) {
-        return DEFAULT_OLLAMA_MODEL.equals(configuredModel) || 
DEFAULT_OPENAI_MODEL.equals(configuredModel);
+        return PLACEHOLDER_MODEL.equals(configuredModel) || 
DEFAULT_OPENAI_MODEL.equals(configuredModel);
     }
 
     /**
@@ -2403,20 +2556,36 @@ public class LlmClient {
     }
 
     /**
-     * Picks a default Ollama model when none was configured (for example 
after a provider switch). Rather than assume a
-     * hardcoded model that may not be installed, this queries the installed 
models and prefers {@code llama3.2} when
-     * present, otherwise the first installed model. Falls back to the {@code 
llama3.2} placeholder only when the list
-     * cannot be retrieved, so the model is never left null.
+     * Picks the Ollama model when none was configured (for example after a 
provider switch): the installed model best
+     * at tool calling (see {@link #bestOllamaModel}). When the installed 
models cannot be listed or none can be used,
+     * the model recommended for Camel is asked for, so the model is never 
left null and the error of Ollama says what
+     * to pull.
      */
     private String resolveDefaultOllamaModel() {
-        List<String> available = listOllamaModels();
-        if (available.isEmpty()) {
-            return DEFAULT_OLLAMA_MODEL;
+        String best = bestOllamaModel(listOllamaModels());
+        return best != null ? best : FALLBACK_OLLAMA_MODEL;
+    }
+
+    /**
+     * The installed model to use when none was chosen: the first of {@link 
#PREFERRED_OLLAMA_MODELS} by family, else
+     * the first installed model that is not an embedding model; null when 
there is none. Never the placeholder llama3.2
+     * while a model of a preferred family is installed: it is too small to 
call the tools of Camel reliably.
+     */
+    static String bestOllamaModel(List<String> installed) {
+        if (installed == null || installed.isEmpty()) {
+            return null;
+        }
+        for (String family : PREFERRED_OLLAMA_MODELS) {
+            for (String name : installed) {
+                if (name.equals(family) || name.startsWith(family + ":")) {
+                    return name;
+                }
+            }
         }
-        return available.stream()
-                .filter(name -> name.equals(DEFAULT_OLLAMA_MODEL) || 
name.startsWith(DEFAULT_OLLAMA_MODEL + ":"))
+        return installed.stream()
+                .filter(name -> 
!name.toLowerCase(Locale.ROOT).contains("embed"))
                 .findFirst()
-                .orElse(available.get(0));
+                .orElse(null);
     }
 
     /**
@@ -2447,45 +2616,10 @@ public class LlmClient {
     }
 
     private void resolveOllamaModel() {
-        try {
-            HttpRequest request = HttpRequest.newBuilder()
-                    .uri(URI.create(url + "/api/tags"))
-                    .timeout(Duration.ofSeconds(HEALTH_CHECK_TIMEOUT_SECONDS))
-                    .GET()
-                    .build();
-            HttpResponse<String> response = httpClient.send(request, 
HttpResponse.BodyHandlers.ofString());
-            if (response.statusCode() != 200) {
-                return;
-            }
-            JsonObject json = (JsonObject) Jsoner.deserialize(response.body());
-            JsonArray models = (JsonArray) json.get("models");
-            if (models == null || models.isEmpty()) {
-                return;
-            }
-
-            List<String> available = new ArrayList<>();
-            for (Object obj : models) {
-                JsonObject m = (JsonObject) obj;
-                String name = m.getString("name");
-                if (name != null) {
-                    available.add(name);
-                }
-            }
-
-            List<String> preferred
-                    = List.of("qwen3.6", "qwen3.5", "qwen3", 
"nemotron-3-nano", "mistral-nemo",
-                            "qwen2.5", "granite4.1", "llama3.1", "llama3.3", 
"mistral");
-            for (String pref : preferred) {
-                for (String avail : available) {
-                    if (avail.equals(pref) || avail.startsWith(pref + ":")) {
-                        model = avail;
-                        printer.println("Auto-selected model: " + model + " 
(better tool-calling support)");
-                        return;
-                    }
-                }
-            }
-        } catch (Exception e) {
-            // best-effort, keep default
+        String best = bestOllamaModel(listOllamaModels());
+        if (best != null) {
+            model = best;
+            printer.println("Auto-selected model: " + model + " (better 
tool-calling support)");
         }
     }
 
diff --git 
a/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/Overview.java
 
b/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/Overview.java
index 75b0e324bd42..4cfd9e4b274f 100644
--- 
a/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/Overview.java
+++ 
b/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/commands/Overview.java
@@ -64,8 +64,6 @@ public class Overview extends CamelCommand {
         }
     }
 
-    private static final String DEFAULT_MODEL = "llama3.2";
-
     @Argument(description = "Project directory (default: the current 
directory)")
     String directory;
 
@@ -98,13 +96,17 @@ public class Overview extends CamelCommand {
             description = "API key for authentication. Also reads 
ANTHROPIC_API_KEY, OPENAI_API_KEY, WATSONX_APIKEY, or LLM_API_KEY env vars")
     String apiKey;
 
-    @Option(name = "model", description = "Model to use", defaultValue = 
DEFAULT_MODEL)
-    String model = DEFAULT_MODEL;
+    @Option(name = "model",
+            description = "Model to use. Default: the provider's default; with 
Ollama the installed model best at"
+                          + " tool calling, such as qwen3.6")
+    String model;
 
     @Option(name = "timeout", description = "Timeout in seconds for LLM 
response", defaultValue = "300")
     int timeout = 300;
 
-    @Option(name = "temperature", description = "Temperature for response 
generation (0.0-2.0)", defaultValue = "0.3")
+    @Option(name = "temperature",
+            description = "Temperature for response generation (0.0-2.0), for 
a model on this machine: a hosted model uses its own default",
+            defaultValue = "0.3")
     double temperature = 0.3;
 
     @Option(name = "show-prompt", hasValue = false, description = "Show the 
prompt sent to the LLM")
@@ -166,7 +168,7 @@ public class Overview extends CamelCommand {
                 .withUrl(url)
                 .withApiType(apiType)
                 .withApiKey(apiKey)
-                .withModel(model)
+                .withExplicitModel(model)
                 .withTimeout(timeout)
                 .withTemperature(temperature)
                 .withStream(false)
diff --git 
a/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/common/OllamaDoctorSupport.java
 
b/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/common/OllamaDoctorSupport.java
index b3ca9888eea3..9342b3d26f23 100644
--- 
a/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/common/OllamaDoctorSupport.java
+++ 
b/dsl/camel-jbang/camel-jbang-core/src/main/java/org/apache/camel/dsl/jbang/core/common/OllamaDoctorSupport.java
@@ -28,6 +28,13 @@ public final class OllamaDoctorSupport {
 
     private static final String DEFAULT_OLLAMA_DISPLAY = "localhost:11434";
 
+    /**
+     * The smallest model, in billions of parameters, that the AI panel and 
camel ask count as large enough. Smaller
+     * models do not call the tools of Camel reliably. The smallest model 
measured on Camel tasks has 27B; Ollama
+     * reports such models as 26.9B, so the line is lower.
+     */
+    public static final int MIN_MODEL_BILLIONS = 24;
+
     private OllamaDoctorSupport() {
     }
 
@@ -101,8 +108,8 @@ public final class OllamaDoctorSupport {
     }
 
     /**
-     * Returns true when the model tag suggests fewer than 14B parameters. 
Tool-calling in the TUI F8 panel requires at
-     * least 14B.
+     * Returns true when the model tag suggests fewer than {@link 
#MIN_MODEL_BILLIONS} parameters, too few for the tool
+     * calling of the AI panel.
      */
     public static boolean isSmallModel(String name) {
         int colon = name.lastIndexOf(':');
@@ -111,7 +118,7 @@ public final class OllamaDoctorSupport {
             int b = tag.indexOf('b');
             try {
                 int params = Integer.parseInt(tag.substring(0, b));
-                return params < 14;
+                return params < MIN_MODEL_BILLIONS;
             } catch (NumberFormatException e) {
                 return false;
             }
@@ -120,7 +127,7 @@ public final class OllamaDoctorSupport {
     }
 
     /**
-     * Returns true when the model tag confirms >= 14B parameters.
+     * Returns true when the model tag confirms at least {@link 
#MIN_MODEL_BILLIONS} parameters.
      */
     static boolean isLargeModel(String name) {
         int colon = name.lastIndexOf(':');
@@ -128,7 +135,7 @@ public final class OllamaDoctorSupport {
         if (tag.matches("\\d+b.*")) {
             int b = tag.indexOf('b');
             try {
-                return Integer.parseInt(tag.substring(0, b)) >= 14;
+                return Integer.parseInt(tag.substring(0, b)) >= 
MIN_MODEL_BILLIONS;
             } catch (NumberFormatException e) {
                 return false;
             }
@@ -137,9 +144,9 @@ public final class OllamaDoctorSupport {
     }
 
     /**
-     * Returns true when no model in the list is known to be >= 14B. Models 
without an explicit numeric size tag (e.g.
-     * {@code llama3.2:latest}) are treated as unknown and counted as "not 
large", so callers should warn when this
-     * returns true.
+     * Returns true when no model in the list is known to have at least {@link 
#MIN_MODEL_BILLIONS} parameters. Models
+     * without an explicit numeric size tag (e.g. {@code llama3.2:latest}) are 
treated as unknown and counted as "not
+     * large", so callers should warn when this returns true.
      */
     public static boolean allModelsSmall(List<String> models) {
         return models != null && !models.isEmpty() && 
models.stream().noneMatch(OllamaDoctorSupport::isLargeModel);
diff --git 
a/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/AskTest.java
 
b/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/AskTest.java
index a5d91d56263c..59b3f61bf531 100644
--- 
a/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/AskTest.java
+++ 
b/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/AskTest.java
@@ -117,6 +117,9 @@ class AskTest {
         assertThat(Ask.smallModelNote("granite4:3b", 3.0)).contains("(3B 
parameters)");
         assertThat(Ask.smallModelNote("qwen3.6:35b-mlx", 35.1)).isNull();
         assertThat(Ask.smallModelNote("unknown", -1)).isNull();
+        // 14B is small for the tools of Camel; Ollama reports the 27B models 
as 26.9B
+        assertThat(Ask.smallModelNote("qwen2.5:14b", 14.8)).startsWith("Note: 
qwen2.5:14b is a small model");
+        assertThat(Ask.smallModelNote("qwen3.8:27b-mlx", 26.9)).isNull();
     }
 
     @Test
diff --git 
a/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/LlmClientChatTest.java
 
b/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/LlmClientChatTest.java
index 24b6efcdd1c2..ec1fdc544574 100644
--- 
a/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/LlmClientChatTest.java
+++ 
b/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/LlmClientChatTest.java
@@ -18,18 +18,21 @@ package org.apache.camel.dsl.jbang.core.commands;
 
 import java.io.IOException;
 import java.net.ServerSocket;
+import java.util.ArrayList;
 import java.util.List;
 import java.util.Map;
 
 import org.apache.camel.dsl.jbang.core.commands.LlmClient.ApiType;
 import org.apache.camel.dsl.jbang.core.commands.LlmClient.ChatResponse;
 import org.apache.camel.dsl.jbang.core.commands.LlmClient.Message;
+import org.apache.camel.dsl.jbang.core.commands.LlmClient.TokenUsage;
 import org.apache.camel.dsl.jbang.core.commands.LlmClient.ToolCall;
 import org.apache.camel.dsl.jbang.core.commands.LlmClient.ToolDef;
 import org.apache.camel.dsl.jbang.core.commands.LlmClient.ToolResult;
 import org.apache.camel.dsl.jbang.run.common.StringPrinter;
 import org.apache.camel.util.json.JsonArray;
 import org.apache.camel.util.json.JsonObject;
+import org.apache.camel.util.json.Jsoner;
 import org.junit.jupiter.api.AfterEach;
 import org.junit.jupiter.api.BeforeEach;
 import org.junit.jupiter.api.Test;
@@ -201,16 +204,35 @@ class LlmClientChatTest {
     }
 
     @Test
-    void aHostedServerGetsNoFieldsForThinking() {
-        JsonObject request = new JsonObject();
+    void openAiItselfGetsReasoningEffortOnly() {
+        JsonObject openAi = new JsonObject();
+        JsonObject azure = new JsonObject();
+        JsonObject watsonx = new JsonObject();
 
-        
LlmClient.create().withApiType(ApiType.openai).withUrl("https://api.openai.com/v1";).turnThinkingOff(request);
+        
LlmClient.create().withApiType(ApiType.openai).withUrl("https://api.openai.com/v1";).turnThinkingOff(openAi);
         
LlmClient.create().withApiType(ApiType.openai).withUrl("https://myresource.openai.azure.com";)
-                .turnThinkingOff(request);
+                .turnThinkingOff(azure);
         
LlmClient.create().withApiType(ApiType.watsonx).withUrl("https://us-south.ml.cloud.ibm.com";)
-                .turnThinkingOff(request);
+                .turnThinkingOff(watsonx);
 
-        assertThat(request).isEmpty();
+        // the reasoning models of OpenAI call tools on Chat Completions only 
with reasoning_effort none; the field of
+        // the local chat templates is for local servers
+        assertThat(openAi).containsEntry("reasoning_effort", 
"none").doesNotContainKey("chat_template_kwargs");
+        assertThat(azure).containsEntry("reasoning_effort", 
"none").doesNotContainKey("chat_template_kwargs");
+        assertThat(watsonx).isEmpty();
+    }
+
+    @Test
+    void onlyAModelOnThisMachineGetsTheTemperature() {
+        // the tests of the AI tools ran with it locally; a hosted model gets 
its own default sampling
+        
assertThat(LlmClient.create().withApiType(ApiType.openai).withUrl("https://api.openai.com/v1";).sendsTemperature())
+                .isFalse();
+        
assertThat(LlmClient.create().withApiType(ApiType.gemini).withUrl("https://generativelanguage.googleapis.com";)
+                .sendsTemperature()).isFalse();
+        
assertThat(LlmClient.create().withApiType(ApiType.openai).withUrl("http://localhost:1234";).sendsTemperature())
+                .isTrue();
+        
assertThat(LlmClient.create().withApiType(ApiType.ollama).withUrl("http://ollama.lan:11434";).sendsTemperature())
+                .isTrue();
     }
 
     @Test
@@ -273,9 +295,11 @@ class LlmClientChatTest {
         server.reply("/v1/messages", 200, """
                 
{"stop_reason":"tool_use","content":[{"type":"text","text":"Let me look"},
                  
{"type":"tool_use","id":"toolu_1","name":"get_routes","input":{"name":"orders"}}],
-                 
"usage":{"input_tokens":30,"output_tokens":7,"cache_read_input_tokens":20}}""");
+                 
"usage":{"input_tokens":30,"output_tokens":7,"cache_read_input_tokens":20,
+                  "cache_creation_input_tokens":5}}""");
 
-        ChatResponse response = client(ApiType.anthropic).withApiKey("sk-ant")
+        // as the commands set them
+        ChatResponse response = 
client(ApiType.anthropic).withApiKey("sk-ant").withTemperature(0.3).withMaxTokens(4096)
                 .chatWithTools("You help", List.of(Message.user("Hi")), TOOLS);
 
         assertThat(response.text()).isEqualTo("Let me look");
@@ -283,21 +307,115 @@ class LlmClientChatTest {
         ToolCall call = response.toolCalls().get(0);
         assertThat(call.id()).isEqualTo("toolu_1");
         assertThat(call.arguments().getString("name")).isEqualTo("orders");
-        assertThat(response.usage().totalTokens()).isEqualTo(37);
+        // the prompt is what came from the cache, what went into it, and the 
rest
+        assertThat(response.usage().inputTokens()).isEqualTo(55);
+        assertThat(response.usage().totalTokens()).isEqualTo(62);
         assertThat(response.usage().cachedTokens()).isEqualTo(20);
 
         
assertThat(server.lastHeaders().getFirst("x-api-key")).isEqualTo("sk-ant");
-        
assertThat(server.lastHeaders().getFirst("anthropic-version")).isEqualTo("2023-06-01");
+        
assertThat(server.lastHeaders().getFirst("anthropic-version")).isEqualTo("2026-10-21");
         JsonObject request = server.lastRequest();
-        // the system prompt apart from the messages, and the tools with their 
input schema
-        assertThat(request.getString("system")).isEqualTo("You help");
-        assertThat(request.getInteger("max_tokens")).isEqualTo(4096);
+        // the current models refuse the sampling parameters
+        assertThat(request).doesNotContainKeys("temperature", "top_p", 
"top_k");
+        // room for the thinking before the answer
+        assertThat(request.getInteger("max_tokens")).isEqualTo(16000);
+        // the system prompt apart from the messages, with a cache breakpoint 
after it, and the automatic breakpoint
+        JsonObject system = (JsonObject) ((JsonArray) 
request.get("system")).get(0);
+        assertThat(system.getString("text")).isEqualTo("You help");
+        JsonObject breakpoint = system.getMap("cache_control");
+        assertThat(breakpoint).containsEntry("type", "ephemeral");
+        JsonObject automatic = request.getMap("cache_control");
+        assertThat(automatic).containsEntry("type", "ephemeral");
+        // the tools with their input schema
         JsonObject tool = (JsonObject) ((JsonArray) 
request.get("tools")).get(0);
         assertThat(tool.getString("name")).isEqualTo("get_routes");
         JsonObject schema = tool.getMap("input_schema");
         assertThat(schema).containsEntry("type", "object");
     }
 
+    @Test
+    void anthropicGetsTheThinkingOfAToolLoopBackAsItCame() {
+        String answer = """
+                [{"type":"thinking","thinking":"","signature":"sig-1"},
+                 {"type":"redacted_thinking","data":"secret"},
+                 {"type":"text","text":"Let me look"},
+                 
{"type":"tool_use","id":"toolu_1","name":"get_routes","input":{"name":"orders"}}]""";
+        server.reply("/v1/messages", 200,
+                "{\"stop_reason\":\"tool_use\",\"content\":" + answer + "}",
+                
"{\"stop_reason\":\"end_turn\",\"content\":[{\"type\":\"text\",\"text\":\"Done\"}]}");
+        LlmClient client = client(ApiType.anthropic);
+        List<Message> messages = new ArrayList<>();
+        messages.add(Message.user("How are the routes?"));
+
+        ChatResponse response = client.chatWithTools("You help", messages, 
TOOLS);
+        messages.add(Message.assistant(response));
+        messages.add(Message.toolResults(List.of(new ToolResult("toolu_1", 
"orders: Started"))));
+        client.chatWithTools("You help", messages, TOOLS);
+
+        // the text by the type of the blocks, after the thinking
+        assertThat(response.text()).isEqualTo("Let me look");
+        JsonObject assistant = (JsonObject) ((JsonArray) 
server.lastRequest().get("messages")).get(1);
+        assertThat(assistant.getString("role")).isEqualTo("assistant");
+        
assertThat(assistant.get("content")).isEqualTo(Jsoner.deserialize(answer, new 
JsonArray()));
+    }
+
+    @Test
+    void anthropicGetsNoThinkingFromBeforeTheQuestion() {
+        JsonArray blocks = Jsoner.deserialize("""
+                [{"type":"thinking","thinking":"","signature":"sig-1"},
+                 {"type":"text","text":"Let me look"},
+                 
{"type":"tool_use","id":"toolu_1","name":"get_routes","input":{"name":"orders"}}]""",
+                new JsonArray());
+        ToolCall call = new ToolCall("toolu_1", "get_routes", new 
JsonObject(Map.of("name", "orders")));
+        ChatResponse toolUse = new ChatResponse("Let me look", List.of(call), 
"tool_use", false, TokenUsage.EMPTY, blocks);
+        server.reply("/v1/messages", 200, 
"{\"stop_reason\":\"end_turn\",\"content\":[{\"type\":\"text\",\"text\":\"Done\"}]}");
+
+        // the AI panel compacts the history before the question, and a 
thinking block refuses an earlier change
+        client(ApiType.anthropic).chatWithTools("You help", List.of(
+                Message.user("How are the routes?"),
+                Message.assistant(toolUse),
+                Message.toolResults(List.of(new ToolResult("toolu_1", "orders: 
Started"))),
+                Message.assistantWithToolCalls("They are started", List.of()),
+                Message.user("And the orders route?")), TOOLS);
+
+        JsonArray earlier = (JsonArray) ((JsonObject) ((JsonArray) 
server.lastRequest().get("messages")).get(1))
+                .get("content");
+        assertThat(earlier).hasSize(2);
+        assertThat(((JsonObject) 
earlier.get(0)).getString("type")).isEqualTo("text");
+        assertThat(((JsonObject) 
earlier.get(1)).getString("id")).isEqualTo("toolu_1");
+    }
+
+    @Test
+    void anExplanationWithAnthropicIsNotCached() {
+        server.reply("/v1/messages", 200, 
"{\"stop_reason\":\"end_turn\",\"content\":[{\"type\":\"text\",\"text\":\"Done\"}]}");
+
+        String answer = 
client(ApiType.anthropic).withTemperature(0.7).generate("You explain", "The 
route");
+
+        // a single question would only pay for writing the cache
+        assertThat(answer).isEqualTo("Done");
+        JsonObject request = server.lastRequest();
+        assertThat(request.getString("system")).isEqualTo("You explain");
+        assertThat(request).doesNotContainKeys("cache_control", "temperature");
+    }
+
+    @Test
+    void aRefusalOfAnthropicIsSaid() {
+        server.reply("/v1/messages", 200, 
"{\"stop_reason\":\"refusal\",\"content\":[]}");
+
+        ChatResponse response = client(ApiType.anthropic).chatWithTools("You 
help", List.of(Message.user("Hi")), TOOLS);
+
+        assertThat(response.text()).isEqualTo("The model declined to answer 
(stop reason: refusal).");
+        assertThat(response.stopReason()).isEqualTo("refusal");
+    }
+
+    @Test
+    void theDefaultModelOfAnthropic() {
+        LlmClient client = 
LlmClient.create().withApiType(ApiType.anthropic).withUrl(server.url()).withPrinter(printer);
+
+        assertThat(client.detectEndpoint()).isTrue();
+        assertThat(client.model()).isEqualTo("claude-sonnet-5-5");
+    }
+
     @Test
     void theConversationInTheFormatOfOpenAi() {
         server.reply("/v1/chat/completions", 200, 
"{\"choices\":[{\"message\":{\"content\":\"Done\"}}]}");
diff --git 
a/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/LlmClientGeminiTest.java
 
b/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/LlmClientGeminiTest.java
index 277e365ec315..845182efebd9 100644
--- 
a/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/LlmClientGeminiTest.java
+++ 
b/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/LlmClientGeminiTest.java
@@ -42,6 +42,12 @@ class LlmClientGeminiTest {
         }
     }
 
+    @Test
+    void theDefaultModelIsTheAliasOfTheLatestFlash() {
+        // gemini-2.0-flash, the default before, is shut down
+        
assertThat(LlmClient.normalizeGeminiModelId(null)).isEqualTo("gemini-flash-latest");
+    }
+
     @Test
     void normalizeGeminiModelIdStripsModelsPrefix() {
         
assertThat(LlmClient.normalizeGeminiModelId("models/gemini-2.0-flash")).isEqualTo("gemini-2.0-flash");
diff --git 
a/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/LlmClientListModelsTest.java
 
b/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/LlmClientListModelsTest.java
index a90baba72d04..d48df0e67b2a 100644
--- 
a/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/LlmClientListModelsTest.java
+++ 
b/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/LlmClientListModelsTest.java
@@ -28,6 +28,7 @@ import org.junit.jupiter.api.AfterEach;
 import org.junit.jupiter.api.Test;
 
 import static org.junit.jupiter.api.Assertions.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertNull;
 import static org.junit.jupiter.api.Assertions.assertTrue;
 
 /**
@@ -168,7 +169,7 @@ class LlmClientListModelsTest {
 
         assertTrue(client.detectEndpoint());
         assertEquals(LlmClient.ApiType.openai, client.apiType());
-        assertEquals("gpt-4o-mini", client.model(), "OpenAI detection must 
leave a usable default model, not null");
+        assertEquals("gpt-6-luna", client.model(), "OpenAI detection must 
leave a usable default model, not null");
     }
 
     @Test
@@ -211,7 +212,8 @@ class LlmClientListModelsTest {
 
         assertTrue(client.detectEndpoint());
         assertEquals(LlmClient.ApiType.ollama, client.apiType());
-        assertEquals("llama3.2", client.model(), "Ollama detection must leave 
a usable default model, not null");
+        assertEquals("qwen3.6:35b-a3b", client.model(),
+                "Ollama detection must leave a usable model, not null: the one 
recommended, which Ollama says to pull");
     }
 
     @Test
@@ -224,12 +226,23 @@ class LlmClientListModelsTest {
     }
 
     @Test
-    void prefersInstalledLlama32OllamaModelWhenNoneConfigured() throws 
IOException {
-        String baseUrl = 
startOllamaServer("{\"models\":[{\"name\":\"qwen3\"},{\"name\":\"llama3.2:latest\"}]}");
+    void neverPrefersLlama32WhenNoneConfigured() throws IOException {
+        String baseUrl = 
startOllamaServer("{\"models\":[{\"name\":\"llama3.2:latest\"},{\"name\":\"qwen3\"}]}");
         LlmClient client = 
LlmClient.create().withApiType(LlmClient.ApiType.ollama).withUrl(baseUrl);
 
         assertTrue(client.detectEndpoint());
-        assertEquals("llama3.2:latest", client.model(), "must prefer an 
installed llama3.2 over the first entry");
+        assertEquals("qwen3", client.model(), "llama3.2 is too small for 
Camel; a model that calls tools well wins");
+    }
+
+    @Test
+    void prefersTheModelsMeasuredOnCamelTasks() {
+        assertEquals("qwen3.6:35b-mlx", LlmClient.bestOllamaModel(
+                List.of("embeddinggemma:300m", "llama3.2:latest", 
"qwen3.8:27b-mlx", "qwen3.6:35b-mlx")));
+        assertEquals("qwen3.8:27b-mlx", 
LlmClient.bestOllamaModel(List.of("granite4:3b", "qwen3.8:27b-mlx")));
+        // none of a preferred family: the first one that is not an embedding 
model
+        assertEquals("saluki:27b", 
LlmClient.bestOllamaModel(List.of("embeddinggemma:300m", "saluki:27b")));
+        
assertNull(LlmClient.bestOllamaModel(List.of("granite-embedding:30m")));
+        assertNull(LlmClient.bestOllamaModel(List.of()));
     }
 
     /**
diff --git 
a/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/LlmClientVertexTest.java
 
b/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/LlmClientVertexTest.java
index dfa76a5cf343..002315037ab7 100644
--- 
a/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/LlmClientVertexTest.java
+++ 
b/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/commands/LlmClientVertexTest.java
@@ -27,7 +27,6 @@ class LlmClientVertexTest {
 
     @Test
     void mapsPreFourSixModels() {
-        assertEquals("claude-sonnet-4-5@20250929", 
LlmClient.resolveVertexModel("claude-sonnet-4-5"));
         assertEquals("claude-haiku-4-5@20251001", 
LlmClient.resolveVertexModel("claude-haiku-4-5"));
         assertEquals("claude-opus-4-5@20251101", 
LlmClient.resolveVertexModel("claude-opus-4-5"));
     }
@@ -37,12 +36,14 @@ class LlmClientVertexTest {
         assertEquals("claude-sonnet-4-6", 
LlmClient.resolveVertexModel("claude-sonnet-4-6"));
         assertEquals("claude-opus-4-6", 
LlmClient.resolveVertexModel("claude-opus-4-6"));
         assertEquals("claude-opus-4-7", 
LlmClient.resolveVertexModel("claude-opus-4-7"));
+        // the default model
+        assertEquals("claude-sonnet-5-5", 
LlmClient.resolveVertexModel("claude-sonnet-5-5"));
     }
 
     @Test
     void passesThroughAlreadyVersionedModels() {
         assertEquals("claude-haiku-4-5@20251001", 
LlmClient.resolveVertexModel("claude-haiku-4-5@20251001"));
-        assertEquals("claude-sonnet-4-5@20250929", 
LlmClient.resolveVertexModel("claude-sonnet-4-5@20250929"));
+        assertEquals("claude-opus-4-5@20251101", 
LlmClient.resolveVertexModel("claude-opus-4-5@20251101"));
     }
 
     @Test
diff --git 
a/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/common/OllamaDoctorSupportTest.java
 
b/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/common/OllamaDoctorSupportTest.java
index f9832e580ae3..ff2c8aa22325 100644
--- 
a/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/common/OllamaDoctorSupportTest.java
+++ 
b/dsl/camel-jbang/camel-jbang-core/src/test/java/org/apache/camel/dsl/jbang/core/common/OllamaDoctorSupportTest.java
@@ -128,6 +128,17 @@ class OllamaDoctorSupportTest {
         assertThat(OllamaDoctorSupport.isSmallModel("llama3.1:70b")).isFalse();
     }
 
+    @Test
+    void theLineIsAt24B() {
+        // a 14B model is too small for the tools of Camel; the 27B models 
measured on them are large enough
+        assertThat(OllamaDoctorSupport.isSmallModel("qwen2.5:14b")).isTrue();
+        assertThat(OllamaDoctorSupport.isSmallModel("gpt-oss:20b")).isTrue();
+        assertThat(OllamaDoctorSupport.isSmallModel("devstral:24b")).isFalse();
+        
assertThat(OllamaDoctorSupport.isSmallModel("qwen3.8:27b-mlx")).isFalse();
+        assertThat(OllamaDoctorSupport.allModelsSmall(List.of("qwen2.5:14b", 
"qwen3.6:35b-a3b"))).isFalse();
+        assertThat(OllamaDoctorSupport.allModelsSmall(List.of("qwen2.5:14b", 
"mistral-nemo:12b"))).isTrue();
+    }
+
     @Test
     void isSmallModelReturnsFalseForNonNumericTag() {
         
assertThat(OllamaDoctorSupport.isSmallModel("llama3.2:latest")).isFalse();
diff --git 
a/dsl/camel-jbang/camel-jbang-core/src/test/resources/golden/cli-help.txt 
b/dsl/camel-jbang/camel-jbang-core/src/test/resources/golden/cli-help.txt
index 1bd030f2733e..1788e9b2f9ae 100644
--- a/dsl/camel-jbang/camel-jbang-core/src/test/resources/golden/cli-help.txt
+++ b/dsl/camel-jbang/camel-jbang-core/src/test/resources/golden/cli-help.txt
@@ -3580,8 +3580,9 @@ Options:
       --api-key=<api-key>   API key for authentication. Also reads
                             ANTHROPIC_API_KEY, OPENAI_API_KEY, WATSONX_APIKEY,
                             or LLM_API_KEY env vars
-      --model=<model>       Model to use
-                            Default: llama3.2
+      --model=<model>       Model to use. Default: the provider's default; with
+                            Ollama the installed model best at tool calling,
+                            such as qwen3.6
   -v, --verbose             Include detailed technical information
       --format=<format>     Output format (text, markdown)
                             Default: text
@@ -3590,7 +3591,9 @@ Options:
       --catalog-context     Include Camel Catalog descriptions in the prompt
       --show-prompt         Show the prompt sent to the LLM
       --temperature=<temperature>
-                            Temperature for response generation (0.0-2.0)
+                            Temperature for response generation (0.0-2.0), for 
a
+                            model on this machine: a hosted model uses its own
+                            default
                             Default: 0.7
       --system-prompt=<system-prompt>
                             Custom system prompt
@@ -4902,8 +4905,9 @@ Options:
                             Default: ollama
       --api-key=<api-key>   API key for authentication. Also reads
                             OPENAI_API_KEY or LLM_API_KEY env vars
-      --model=<model>       Model to use
-                            Default: llama3.2
+      --model=<model>       Model to use. Default: with Ollama the installed
+                            model best at tool calling, such as qwen3.6; with 
an
+                            OpenAI-compatible server the first model it lists
   -v, --verbose             Include detailed security recommendations with code
                             examples
       --format=<format>     Output format (text, markdown)
@@ -4913,7 +4917,9 @@ Options:
       --catalog-context     Include Camel Catalog descriptions in the prompt
       --show-prompt         Show the prompt sent to the LLM
       --temperature=<temperature>
-                            Temperature for response generation (0.0-2.0)
+                            Temperature for response generation (0.0-2.0), for 
a
+                            model on this machine: a hosted model uses its own
+                            default
                             Default: 0.7
       --system-prompt=<system-prompt>
                             Custom system prompt
@@ -4925,7 +4931,7 @@ Options:
 Examples:
   camel harden hello.java
   camel harden hello.yaml --format=markdown
-  camel harden hello.java --model=gpt-4
+  camel harden hello.java --model=qwen3.6:35b-a3b
 
 == camel hawtio --help
 Launch Hawtio web console
@@ -5242,12 +5248,15 @@ Options:
       --api-key=<api-key>   API key for authentication. Also reads
                             ANTHROPIC_API_KEY, OPENAI_API_KEY, WATSONX_APIKEY,
                             or LLM_API_KEY env vars
-      --model=<model>       Model to use
-                            Default: llama3.2
+      --model=<model>       Model to use. Default: the provider's default; with
+                            Ollama the installed model best at tool calling,
+                            such as qwen3.6
       --timeout=<timeout>   Timeout in seconds for LLM response
                             Default: 300
       --temperature=<temperature>
-                            Temperature for response generation (0.0-2.0)
+                            Temperature for response generation (0.0-2.0), for 
a
+                            model on this machine: a hosted model uses its own
+                            default
                             Default: 0.3
       --show-prompt         Show the prompt sent to the LLM
   <directory>               Project directory (default: the current directory)
diff --git 
a/dsl/camel-jbang/camel-jbang-core/src/test/resources/golden/cli-parse.txt 
b/dsl/camel-jbang/camel-jbang-core/src/test/resources/golden/cli-parse.txt
index a57d0ff59142..519c1c88e139 100644
--- a/dsl/camel-jbang/camel-jbang-core/src/test/resources/golden/cli-parse.txt
+++ b/dsl/camel-jbang/camel-jbang-core/src/test/resources/golden/cli-parse.txt
@@ -2870,7 +2870,7 @@
   catalogContext=false
   files=["v"]
   format="text"
-  model="llama3.2"
+  model=null
   showPrompt=false
   stream=true
   systemPrompt=null
@@ -4297,7 +4297,7 @@
   catalogContext=false
   files=["v"]
   format="text"
-  model="llama3.2"
+  model=null
   showPrompt=false
   stream=true
   systemPrompt=null
@@ -4574,7 +4574,7 @@
   applyDescriptions=false
   directory=null
   format="markdown"
-  model="llama3.2"
+  model=null
   save=false
   showPrompt=false
   temperature=0.3
diff --git 
a/dsl/camel-jbang/camel-jbang-plugin-tui/src/main/java/org/apache/camel/dsl/jbang/core/commands/tui/AiPanel.java
 
b/dsl/camel-jbang/camel-jbang-plugin-tui/src/main/java/org/apache/camel/dsl/jbang/core/commands/tui/AiPanel.java
index 2f44a1852b3e..95d9339fc509 100644
--- 
a/dsl/camel-jbang/camel-jbang-plugin-tui/src/main/java/org/apache/camel/dsl/jbang/core/commands/tui/AiPanel.java
+++ 
b/dsl/camel-jbang/camel-jbang-plugin-tui/src/main/java/org/apache/camel/dsl/jbang/core/commands/tui/AiPanel.java
@@ -1318,7 +1318,7 @@ class AiPanel {
             }
 
             if (response.toolCalls() != null && 
!response.toolCalls().isEmpty()) {
-                
messages.add(LlmClient.Message.assistantWithToolCalls(response.text(), 
response.toolCalls()));
+                messages.add(LlmClient.Message.assistant(response));
 
                 List<LlmClient.ToolResult> results = new ArrayList<>();
                 for (LlmClient.ToolCall toolCall : response.toolCalls()) {
@@ -2376,77 +2376,36 @@ class AiPanel {
     }
 
     private String buildAiSetupGuide() {
+        // short: the documentation online is the one place with the full 
setup and the models that work well
         return """
                 ## AI Assistant — Getting Started
 
                 No LLM provider was detected. Choose one of the options below, 
then press **F8** to reopen this panel.
 
-                > **Tool calling is required.** This panel inspects your Camel 
process by
-                > invoking built-in tools. Models smaller than ~14B do not 
reliably call
-                > tools and will answer from training knowledge instead — use 
at least 14B.
-                > Prefer a mixture-of-experts model such as qwen3.6:35b-a3b: 
it processes
-                > the tool-heavy prompt many times faster than a dense 27B/32B 
model.
+                ### Option A: a local model with Ollama (no API key)
 
-                ---
-
-                ### Option A: Local — Ollama (no API key needed)
-
-                Run models entirely on your machine — no data leaves your host.
+                Install Ollama natively (`brew install ollama` on macOS, `curl 
-fsSL https://ollama.com/install.sh | sh`
+                on Linux), then pull a model:
 
                 ```
-                # macOS
-                brew install ollama
-
-                # Linux
-                curl -fsSL https://ollama.com/install.sh | sh
-
-                # then on both:
-                ollama serve             # start the daemon (skip if 
auto-started)
-                ollama pull qwen3.6:35b-a3b  # recommended
+                ollama pull qwen3.6:35b-mlx  # on a Mac with Apple Silicon 
(Ollama 0.40 or later)
+                ollama pull qwen3.6:35b-a3b  # on other machines
                 ```
 
-                Ollama is auto-detected at `localhost:11434` — no 
configuration needed.
-
-                **Models that work well** (tool-calling capable, ≥14B):
-
-                | Model | RAM | Notes |
-                |---|---|---|
-                | qwen3.6:35b-a3b | ~23 GB | Recommended: only 3B active per 
token, fastest prompt processing |
-                | qwen2.5:14b | ~9 GB | Minimum for 16 GB machines |
-                | qwen3.6:27b | ~18 GB | Strong dense model, several times 
slower prompt processing |
-                | qwen2.5:32b | ~20 GB | Good quality, slow prompt processing |
-                | hermes3:70b  | ~43 GB | Excellent tool calling, needs 64 GB+ 
|
-                | llama3.3:70b | ~43 GB | Best open model, needs 64 GB+ |
+                It needs 32 GB of memory or more: smaller models do not call 
the tools reliably.
 
-                **Tip:** Install Ollama natively — `camel infra run ollama` 
uses Docker and
-                loses GPU acceleration (Metal on macOS, CUDA on Linux), making 
inference
-                much slower. Native install is always preferred for 
development use.
+                ### Option B: a coding agent you already use
 
-                ---
+                Press **Ctrl+P** and pick an `acp:` provider: Claude Code, 
Codex, opencode, IBM Bob, Qwen Code or
+                DeepSeek Harness. The agent keeps its own login and model.
 
-                ### Option B: Cloud provider (API key required)
-
-                Set one environment variable before starting the TUI:
-
-                | Variable | Provider |
-                |---|---|
-                | `ANTHROPIC_API_KEY` | Claude |
-                | `OPENAI_API_KEY` | OpenAI (GPT-4o etc.) |
-                | `GEMINI_API_KEY` | Gemini |
-                | `AZURE_OPENAI_API_KEY` + `AZURE_OPENAI_ENDPOINT` | Azure 
OpenAI |
-                | `WATSONX_APIKEY` | IBM watsonx.ai |
-
-                For any **OpenAI-compatible** server (LM Studio, vLLM, 
llama.cpp, GPT4All, …):
-
-                ```
-                export LLM_API_KEY=any-value
-                export LLM_BASE_URL=http://localhost:1234
-                ```
+                ### Option C: an API key
 
-                `OPENAI_BASE_URL` is also supported as an alternative to 
`LLM_BASE_URL`.
+                Set `ANTHROPIC_API_KEY`, `OPENAI_API_KEY`, `GEMINI_API_KEY` or 
the key of another provider before you
+                start the TUI, or press **Ctrl+P** to pick a provider now.
 
-                Or press **Ctrl+P** to select and configure a provider now.
-                """;
+                The full guide, with every provider and the models that work 
well:
+                """ + "<" + DocViewerPopup.AI_PROVIDERS_DOCS_URL + ">\n";
     }
 
     /**
diff --git 
a/dsl/camel-jbang/camel-jbang-plugin-tui/src/main/java/org/apache/camel/dsl/jbang/core/commands/tui/DocViewerPopup.java
 
b/dsl/camel-jbang/camel-jbang-plugin-tui/src/main/java/org/apache/camel/dsl/jbang/core/commands/tui/DocViewerPopup.java
index 848094c2bd52..42e2b795f5d3 100644
--- 
a/dsl/camel-jbang/camel-jbang-plugin-tui/src/main/java/org/apache/camel/dsl/jbang/core/commands/tui/DocViewerPopup.java
+++ 
b/dsl/camel-jbang/camel-jbang-plugin-tui/src/main/java/org/apache/camel/dsl/jbang/core/commands/tui/DocViewerPopup.java
@@ -56,6 +56,13 @@ class DocViewerPopup {
 
     /** The documentation of screenshots and demo recordings, with all the 
tape commands. */
     static final String RECORDING_DOCS_URL = 
"https://camel.apache.org/manual/camel-jbang-tui-recording.html";;
+    /**
+     * The AI setup in full: the ways to use AI, the providers with the local 
models that work well, and the coding
+     * agents. The guides of the TUI stay short and point here, so the 
documentation is the one place to keep current.
+     */
+    static final String AI_WAYS_DOCS_URL = 
"https://camel.apache.org/manual/camel-jbang-tui-ai-ways.html";;
+    static final String AI_PROVIDERS_DOCS_URL = 
"https://camel.apache.org/manual/camel-jbang-ai-providers.html";;
+    static final String AI_AGENTS_DOCS_URL = 
"https://camel.apache.org/manual/camel-jbang-tui-ai-agents.html";;
 
     private boolean showViewer;
     private boolean showPicker;
@@ -400,102 +407,40 @@ class DocViewerPopup {
             status = "**Detected:** Ollama at " + 
OllamaDoctorSupport.formatDisplayHost(ollama.baseUrl())
                      + " with " + 
OllamaDoctorSupport.modelCountLabel(ollama.models());
         } else if (ollama.running()) {
-            status = "**Detected:** Ollama is running but has no models. Run 
`ollama pull qwen3.6:35b-a3b`.";
+            status = "**Detected:** Ollama is running but has no models. Run 
`ollama pull qwen3.6:35b-a3b` "
+                     + "(on a Mac with Apple Silicon: `qwen3.6:35b-mlx`).";
         } else {
-            status = "**Status:** No AI provider detected. Set an API key, 
start Ollama, or pick a coding agent "
-                     + "with Ctrl+P in the panel (section 5), then press F8.";
+            status = "**Status:** No AI provider detected. Start Ollama, pick 
a coding agent with Ctrl+P in the "
+                     + "panel, or set an API key, then press F8.";
         }
         String url = "http://localhost:"; + mcpPort + "/mcp";
         openMarkdown("Setup AI",
                 "# Setup AI\n\n"
                                  + status + "\n\n"
-                                 + "Press **F8** to open the AI prompt panel. 
The panel answers questions about your running\n"
-                                 + "integrations by calling built-in tools 
(routes, logs, errors, traces, health).\n\n"
-                                 + "## 1. Choose a provider\n\n"
-                                 + "The panel auto-detects a provider from 
environment variables, in this order:\n\n"
-                                 + "| Environment | Provider |\n"
-                                 + "|-------------|----------|\n"
-                                 + "| `ANTHROPIC_API_KEY` | Anthropic Claude 
|\n"
-                                 + "| `CLOUD_ML_REGION` + 
`ANTHROPIC_VERTEX_PROJECT_ID` | Vertex AI |\n"
-                                 + "| `AZURE_OPENAI_API_KEY` + 
`AZURE_OPENAI_ENDPOINT` | Azure OpenAI |\n"
-                                 + "| `GEMINI_API_KEY` | Google Gemini |\n"
-                                 + "| `OPENAI_API_KEY` | OpenAI |\n"
-                                 + "| `WATSONX_APIKEY` | IBM watsonx.ai |\n"
-                                 + "| `LLM_API_KEY` + `LLM_BASE_URL` | Any 
OpenAI-compatible server |\n"
-                                 + "| Ollama at `localhost:11434` | Local 
Ollama (no key needed) |\n\n"
-                                 + "Export the variable before starting the 
TUI, for example:\n\n"
-                                 + "    export ANTHROPIC_API_KEY=sk-ant-...\n"
-                                 + "    camel tui\n\n"
-                                 + "To pin a provider, model or URL regardless 
of the environment, open\n"
-                                 + "**F2 → Settings** and set *AI Provider*, 
*AI Model* and *AI Base URL*\n"
-                                 + "(stored as `camel.tui.ai.provider`, 
`camel.tui.ai.model`, `camel.tui.ai.url`).\n\n"
-                                 + "## 2. Local AI with Ollama (no API 
key)\n\n"
-                                 + "Install Ollama natively for GPU 
acceleration, pull a model, then press F8:\n\n"
-                                 + "    brew install ollama                    
      # macOS\n"
-                                 + "    curl -fsSL 
https://ollama.com/install.sh | sh  # Linux\n"
-                                 + "    ollama pull qwen3.6:35b-a3b\n\n"
-                                 + "The AI panel relies on tool calling. 
Models smaller than ~14B do not call tools\n"
-                                 + "reliably and answer from training data 
instead. Use at least 14B. Prefer a\n"
-                                 + "mixture-of-experts model such as 
`qwen3.6:35b-a3b`: with only 3B parameters active\n"
-                                 + "per token it processes the tool-heavy 
prompt many times faster than a dense 27B/32B\n"
-                                 + "model, so answers start in seconds instead 
of a minute.\n\n"
-                                 + "| Model | RAM | Notes |\n"
-                                 + "|-------|-----|-------|\n"
-                                 + "| `qwen3.6:35b-a3b` | ~23 GB | 
Recommended: fastest prompt processing, needs 32 GB+ |\n"
-                                 + "| `qwen2.5:14b` | ~9 GB | Minimum for 16 
GB machines |\n"
-                                 + "| `qwen3.6:27b` | ~18 GB | Strong dense 
model, several times slower prompt processing |\n"
-                                 + "| `qwen2.5:32b` | ~20 GB | Good quality, 
slow prompt processing |\n"
-                                 + "| `hermes3:70b` | ~43 GB | Excellent tool 
calling, needs 64 GB+ |\n"
-                                 + "| `llama3.3:70b` | ~43 GB | Best open 
model, needs 64 GB+ |\n\n"
-                                 + "`camel infra run ollama` runs Ollama in 
Docker without GPU acceleration, which is\n"
-                                 + "much slower. Prefer the native install for 
development.\n\n"
-                                 + "On Apple Silicon, use the default (GGUF) 
tags rather than the `-mlx` tags: the Ollama\n"
-                                 + "MLX engine cannot yet reuse the cached 
prompt for Qwen 3.x models, so every question\n"
-                                 + "re-processes the whole prompt.\n\n"
-                                 + "## 3. OpenAI-compatible local servers\n\n"
-                                 + "LM Studio, vLLM, llama.cpp, GPT4All and 
similar servers work through `LLM_BASE_URL`:\n\n"
-                                 + "    export LLM_API_KEY=any-value\n"
-                                 + "    export 
LLM_BASE_URL=http://localhost:1234\n";
-                                 + "    camel tui\n\n"
-                                 + "`OPENAI_BASE_URL` is accepted as an 
alternative to `LLM_BASE_URL`. The model is the first one\n"
-                                 + "the server lists on `/v1/models`; pick 
another with `/model <name>` in the panel or\n"
-                                 + "*AI Model* in **F2 -> Settings**. The 
model must support tool calling.\n\n"
-                                 + "## 4. Using the AI panel\n\n"
-                                 + "- **F8** opens and closes the panel; 
**Enter** sends the prompt\n"
-                                 + "- **Ctrl+P** (or `/provider`) switches 
provider or model for the session\n"
-                                 + "- `/model <name>` switches the model, 
`/clear` resets the conversation\n"
-                                 + "- `/run`, `/infra` and `/send` run Camel 
CLI commands from the prompt\n"
-                                 + "- **Ctrl+U** toggles the usage view with 
token consumption\n"
-                                 + "- `/help` lists all slash commands; 
**↑/↓** recalls earlier prompts\n\n"
-                                 + "Check **F2 → Run Doctor** to verify the 
detected provider and Ollama models.\n\n"
-                                 + "## 5. Use your coding agent inside the 
panel (ACP)\n\n"
-                                 + "If you already use Claude Code, Codex, 
OpenCode, IBM Bob, Qwen Code or DeepSeek Harness,\n"
-                                 + "the panel can hand your questions to it 
over the Agent Client Protocol. No API key is\n"
-                                 + "needed in the TUI: the agent keeps its own 
login, model, memory and skills.\n\n"
-                                 + "Press **Ctrl+P** in the panel and pick a 
preset:\n\n"
-                                 + "| Provider | Runs | Before the first 
question |\n"
-                                 + 
"|----------|------|---------------------------|\n"
-                                 + "| `acp:claude` | `npx -y 
@agentclientprotocol/claude-agent-acp` | log in with the `claude` CLI |\n"
-                                 + "| `acp:codex` | `npx -y 
@agentclientprotocol/codex-acp` | `codex login` |\n"
-                                 + "| `acp:opencode` | `opencode acp` | 
`opencode auth login` |\n"
-                                 + "| `acp:bob` | `bob acp` | set 
`BOBSHELL_API_KEY` or run `bob` once |\n"
-                                 + "| `acp:qwen` | `qwen --acp` | set 
`OPENAI_API_KEY` and `OPENAI_BASE_URL` |\n"
-                                 + "| `acp:dsh` | `npx -y @deepseek-ai/dsh 
--profile acp` | configure the model key in DeepSeek Harness |\n"
-                                 + "| `acp:custom` | *ACP Command* in **F2 → 
Settings** | depends on the agent |\n\n"
-                                 + "The `npx` presets need Node.js 22 or 
newer. The agent starts with your first question and\n"
-                                 + "gets the TUI's tools through an MCP server 
that is started on demand, so `--mcp` is not\n"
-                                 + "required. Read-only tools are approved 
automatically; a tool that changes something, or\n"
-                                 + "a file edit or command of the agent's own, 
opens a permission popup first (**Enter**\n"
-                                 + "allows, **Esc** rejects, **Ctrl+C** 
cancels the turn). File edits go through the same\n"
-                                 + "`/write confirm|auto|live` modes as the 
built-in providers. The agent's skills and\n"
-                                 + "commands appear as `/agent:<name>`. Set 
*AI Provider* to the preset in **F2 → Settings**\n"
-                                 + "to make it the default.\n\n"
-                                 + "## 6. Connect an agent from outside over 
MCP (optional)\n\n"
-                                 + "Start the TUI with `--mcp` to expose an 
MCP server, then connect an agent such as\n"
-                                 + "Claude Code from its own terminal:\n\n"
+                                 + "Press **F8** to open the AI panel. It 
answers questions about your running integrations "
+                                 + "by calling built-in tools (routes, logs, 
errors, traces, health). Pick one of three ways:\n\n"
+                                 + "1. **A local model with Ollama**, with no 
API key and nothing leaving your machine. Install "
+                                 + "Ollama natively, then pull 
`qwen3.6:35b-mlx` on a Mac with Apple Silicon (Ollama 0.40 or "
+                                 + "later), or `qwen3.6:35b-a3b` on other 
machines. It needs 32 GB of memory or more: smaller "
+                                 + "models do not call the tools reliably.\n"
+                                 + "2. **A coding agent you already use**: 
Claude Code, Codex, opencode, IBM Bob, Qwen Code or "
+                                 + "DeepSeek Harness. Press **Ctrl+P** in the 
panel and pick an `acp:` provider. The agent keeps "
+                                 + "its own login and model, so the TUI needs 
no API key.\n"
+                                 + "3. **An API key** in the environment 
before you start the TUI, such as "
+                                 + "`export ANTHROPIC_API_KEY=sk-ant-...`. 
`OPENAI_API_KEY`, `GEMINI_API_KEY` and others work "
+                                 + "too, and `LLM_BASE_URL` points at an 
OpenAI-compatible server such as LM Studio.\n\n"
+                                 + "Pin a provider, model or URL in **F2 → 
Settings**, and see what was found with "
+                                 + "**F2 → Run Doctor**. In the panel, `/help` 
lists the commands.\n\n"
+                                 + "The full guide is online: all providers, 
the models that work well, and the coding agents.\n\n"
+                                 + "- Ways to use AI: <" + AI_WAYS_DOCS_URL + 
">\n"
+                                 + "- Providers and local models: <" + 
AI_PROVIDERS_DOCS_URL + ">\n"
+                                 + "- Coding agents over ACP and MCP: <" + 
AI_AGENTS_DOCS_URL + ">\n\n"
+                                 + "## Connect an agent from outside over 
MCP\n\n"
+                                 + "Start the TUI with `--mcp`, then connect 
an agent such as Claude Code from its own "
+                                 + "terminal:\n\n"
                                  + "    claude mcp add --transport http 
camel-tui " + url + "\n\n"
-                                 + "The footer turns green when the agent 
connects. See **F2 → AI & MCP → MCP Info**\n"
-                                 + "for the available tools and a `.mcp.json` 
example.\n");
+                                 + "The footer turns green when the agent 
connects. **F2 → AI & MCP → MCP Info** shows the "
+                                 + "tools and a `.mcp.json` example.\n");
     }
 
     void openMcpInfo() {
diff --git 
a/dsl/camel-jbang/camel-jbang-plugin-tui/src/main/java/org/apache/camel/dsl/jbang/core/commands/tui/DoctorPopup.java
 
b/dsl/camel-jbang/camel-jbang-plugin-tui/src/main/java/org/apache/camel/dsl/jbang/core/commands/tui/DoctorPopup.java
index 4b17721d2580..60118c9371dd 100644
--- 
a/dsl/camel-jbang/camel-jbang-plugin-tui/src/main/java/org/apache/camel/dsl/jbang/core/commands/tui/DoctorPopup.java
+++ 
b/dsl/camel-jbang/camel-jbang-plugin-tui/src/main/java/org/apache/camel/dsl/jbang/core/commands/tui/DoctorPopup.java
@@ -390,7 +390,8 @@ class DoctorPopup {
                     Style.EMPTY.dim())));
             if (allSmall) {
                 result.add(Line.from(Span.styled(
-                        "                    F8 needs ≥14B — run: ollama pull 
qwen3.6:35b-a3b",
+                        "                    F8 needs ≥" + 
OllamaDoctorSupport.MIN_MODEL_BILLIONS
+                                                 + "B — run: ollama pull 
qwen3.6:35b-a3b",
                         Style.EMPTY.dim())));
             }
         } else {
diff --git 
a/dsl/camel-jbang/camel-jbang-plugin-tui/src/test/java/org/apache/camel/dsl/jbang/core/commands/tui/AiPanelTest.java
 
b/dsl/camel-jbang/camel-jbang-plugin-tui/src/test/java/org/apache/camel/dsl/jbang/core/commands/tui/AiPanelTest.java
index 0d94b4999cdc..a24dd65997ee 100644
--- 
a/dsl/camel-jbang/camel-jbang-plugin-tui/src/test/java/org/apache/camel/dsl/jbang/core/commands/tui/AiPanelTest.java
+++ 
b/dsl/camel-jbang/camel-jbang-plugin-tui/src/test/java/org/apache/camel/dsl/jbang/core/commands/tui/AiPanelTest.java
@@ -31,7 +31,9 @@ import dev.tamboui.tui.event.KeyCode;
 import dev.tamboui.tui.event.KeyEvent;
 import dev.tamboui.tui.event.KeyModifiers;
 import org.apache.camel.dsl.jbang.core.commands.LlmClient;
+import org.apache.camel.util.json.JsonArray;
 import org.apache.camel.util.json.JsonObject;
+import org.apache.camel.util.json.Jsoner;
 import org.junit.jupiter.api.Test;
 
 import static org.apache.camel.dsl.jbang.core.commands.tui.TuiAwait.await;
@@ -1297,6 +1299,56 @@ class AiPanelTest {
         }
     }
 
+    // ---- the answers of a tool loop go back as they came (CAMEL-26962) ----
+
+    @Test
+    void theToolLoopSendsTheAnswerBackAsItCame() throws Exception {
+        AiPanel panel = new AiPanel();
+        panel.setToolRegistryForTesting(new TuiToolRegistry(null));
+        ThinkingLlmClient client = new ThinkingLlmClient();
+        panel.setClientForTesting(client);
+        panel.open();
+        type(panel, "how are the routes");
+        panel.handleKeyEvent(KeyEvent.ofKey(KeyCode.ENTER, KeyModifiers.NONE));
+        await().atMost(10, TimeUnit.SECONDS).until(() -> 
!panel.isAgentThreadRunningForTesting());
+
+        // Anthropic wants the thinking blocks of a tool loop back unchanged
+        assertEquals(ThinkingLlmClient.BLOCKS, client.sentBack);
+    }
+
+    /** Thinks before its tool call, as the current Claude models do, and 
answers after the tool result. */
+    private static final class ThinkingLlmClient extends LlmClient {
+
+        static final JsonArray BLOCKS = Jsoner.deserialize(
+                
"[{\"type\":\"thinking\",\"thinking\":\"\",\"signature\":\"sig-1\"},"
+                                                           + 
"{\"type\":\"tool_use\",\"id\":\"toolu_1\","
+                                                           + 
"\"name\":\"tui_list_integrations\",\"input\":{}}]",
+                new JsonArray());
+
+        volatile JsonArray sentBack;
+
+        ThinkingLlmClient() {
+            withModel("test-model");
+            withApiType(ApiType.anthropic);
+        }
+
+        @Override
+        public boolean detectEndpoint() {
+            return true;
+        }
+
+        @Override
+        public ChatResponse chatWithTools(String systemPrompt, List<Message> 
messages, List<ToolDef> tools) {
+            Message last = messages.get(messages.size() - 1);
+            if (last.toolResults() == null) {
+                ToolCall call = new ToolCall("toolu_1", 
"tui_list_integrations", new JsonObject());
+                return new ChatResponse(null, List.of(call), "tool_use", 
false, TokenUsage.EMPTY, BLOCKS);
+            }
+            sentBack = messages.get(messages.size() - 2).contentBlocks();
+            return new ChatResponse("The routes are started.", List.of(), 
"end_turn", false, TokenUsage.EMPTY);
+        }
+    }
+
     // ---- what ends a turn is shown and logged, also an Error and a failure 
with no message ----
 
     @Test

Reply via email to