Say on the page that a worker needs llama.cpp, and that it can run without one

distinct_server/presentation.py

@@ -1346,6 +1346,13 @@ details[open] .c-libhelp__toggle::after{content:" -";}
              color:var(--c-text-2) !important; user-select:all;}
 .c-join__note{margin:0; font-size:var(--t-micro); color:var(--c-muted) !important;
               line-height:1.55; max-width:56ch;}
+/* The one outbound link in the setup panel. Underlined, because Gradio's own
+   prose rules strip underlines from anchors inside its container and a link
+   that reads as body text is a link nobody clicks -- and this one is the
+   difference between a worker that runs models and one that does not. */
+.c-join__link{color:var(--c-accent) !important; text-decoration:underline !important;
+              text-underline-offset:2px; font-weight:600;}
+.c-join__link:hover{color:var(--c-sage-deep) !important;}
 .c-join__note code{font-family:var(--mono); font-size:11px;
                    background:var(--c-fill); border-radius:4px; padding:1px 5px;}
 .c-join__note .c-join__code{display:inline-block; padding:2px 8px;

distinct_server/render.py

@@ -336,7 +336,35 @@ def join_this_server(
         "a <code>setup.py</code> this project does not have.</p>"
     )
 
-    # STEP THREE: the run, with this server's address and this browser's own
+    # STEP THREE EXISTS BECAUSE THE PAGE SENT PEOPLE INTO A WALL IT NEVER
+    # MENTIONED.
+    #
+    # The panel used to be clone, install, run. A volunteer who followed it
+    # exactly got through all three and then met "This worker has no llama.cpp
+    # to run a model with" from the agent -- a requirement the page had never
+    # named, after a wait long enough to feel like the install had worked.
+    # Their words: the setup "is not the same as what is on the website".
+    #
+    # llama.cpp genuinely is a separate program, and it genuinely is optional:
+    # a worker with no runner still pairs and still offers every tool and skill
+    # in the library, which is most of what this network does. So this is a
+    # step with two honest endings rather than a prerequisite dressed as a
+    # warning, and the run command in step four follows whichever was chosen.
+    runner = (
+        '<p class="c-join__note">A model needs a program to run it, and that program '
+        "is <strong>llama.cpp</strong> — a separate project, not part of this one. "
+        'Download a build from <a class="c-join__link" href="https://github.com/'
+        'ggml-org/llama.cpp/releases" target="_blank" rel="noopener noreferrer">'
+        "github.com/ggml-org/llama.cpp/releases</a>, unzip it, and put "
+        "<code>llama-server</code> on your PATH — or drop it in a <code>runtime</code> "
+        "folder beside this one, where the worker looks first.</p>"
+        '<p class="c-join__note">Not interested in running models? Add '
+        "<code>--demo-runner</code> to the command in the next step. The worker still "
+        "pairs, and still runs every tool and skill in the library on your machine; "
+        "only the model's own replies are absent. You can add llama.cpp later.</p>"
+    )
+
+    # STEP FOUR: the run, with this server's address and this browser's own
     # library defaults already in it.
     code_note = (
         f'<p class="c-join__note">Pairing code, one use and ten minutes: '
@@ -369,11 +397,16 @@ def join_this_server(
         '<div class="c-step-num__body"><p class="c-step-num__head">Install it</p>'
         f"{install}</div></li>"
         '<li class="c-step-num"><span class="c-step-num__n">3</span>'
+        '<div class="c-step-num__body">'
+        '<p class="c-step-num__head">Get something to run the models with</p>'
+        f"{runner}</div></li>"
+        '<li class="c-step-num"><span class="c-step-num__n">4</span>'
         '<div class="c-step-num__body"><p class="c-step-num__head">Point it at this server</p>'
         f'<code class="c-join__address">{esc(first)}</code>{alternates}{code_note}'
         f"{run_it}"
         f"{picture}</div></li>"
-        # STEP FOUR EXISTS BECAUSE PEOPLE ASK FOR IT AFTER READING STEP THREE.
+        # THE LAST STEP EXISTS BECAUSE PEOPLE ASK FOR IT AFTER READING THE ONE
+        # BEFORE IT.
         #
         # Sharing was one clause at the end of the previous step, and the
         # question "how do several people use my machine?" kept coming back,
@@ -381,7 +414,7 @@ def join_this_server(
         # point of running a worker at all, so it is a step of its own, and
         # it says the non-obvious part out loud: the code is not spent by
         # the first person who enters it.
-        '<li class="c-step-num"><span class="c-step-num__n">4</span>'
+        '<li class="c-step-num"><span class="c-step-num__n">5</span>'
         '<div class="c-step-num__body">'
         '<p class="c-step-num__head">Share it with whoever should use it</p>'
         '<p class="c-join__note">When it pairs, the worker prints an <strong>access '

gitignore-new.txt

@@ -1,18 +0,0 @@
-__pycache__/
-*.py[cod]
-.env
-.DS_Store
-
-# Build output from `pip install -e .`. Regenerated on every install, and a
-# tracked copy means every volunteer's install shows up as a diff.
-*.egg-info/
-
-# Scratch left behind by the apply-and-push helpers used during development.
-APPLY*.bat
-TIDY*.bat
-AGENT-SETUP.bat
-push-log*.txt
-distinct-update*.zip
-agent-setup.log
-runtime-outputs/
-screenshots/