Context window setting is now live, per agent tab

CHANGELOG.md

@@ -24,6 +24,15 @@ recorded by the `MandoCode` submodule.
   period-correct white dialog on the silver desktop.
 
 ### Fixed
+- **The context window setting now actually reaches the model.** It was exported as
+  `OLLAMA_CONTEXT_LENGTH` only when MandoCode launched the Ollama daemon itself — anyone whose
+  daemon was already running (the tray app, most commonly) silently got the daemon's own default
+  instead, making the Settings field, the "context window sized to Nk tokens" line, and the
+  per-model auto-sizing all cosmetic. The window now rides on every chat request as `num_ctx`,
+  which outranks the tray app's slider and the daemon default, applies from the next message with
+  no restart, and — as a bonus the old design could never offer — is genuinely per-agent: two tabs
+  can run different windows against the same daemon. The Settings caption and README stop calling
+  it app-wide, and `0` still means "let Ollama decide."
 - **W98 chat prompts are readable again.** Your own prompts rendered in the theme's gold, which
   resolves to a dark mustard `#806000` — 3.21:1 on a silver window, under the accessibility floor
   and hard going for anyone with less-than-perfect sight. W98 prompts now use black window text
@@ -33,6 +42,13 @@ recorded by the `MandoCode` submodule.
   and is now white underlined at 4.77:1. Other themes are untouched.
 
 ### Added
+- **Conversations compact themselves before the context window overflows** (pinned harness
+  update). Local Ollama never rejects an oversized prompt — it silently drops the oldest tokens,
+  system prompt first, which surfaced as "Model returned an empty response" at the end of a
+  tool-heavy turn on a small model. The harness now estimates each outgoing prompt (history plus
+  every tool schema riding along, MCP servers included) before sending, and when it nears the
+  window it folds older history into a recap first and says so in the reply — leaving thinking
+  models the generation headroom they spend reasoning before any visible answer appears.
 - **Undo for the notes assistant.** A gold undo arrow appears in the note header after the assistant
   inserts or replaces text, putting the note back exactly as it was. Ctrl+Z can't do this job —
   assigning the editor's text resets the TextBox's own undo history, so the one edit you *didn't*

README.md

@@ -148,9 +148,11 @@ approval in one agent would stop another's from ever rendering; and `ChatControl
 (not `+=`) five approval delegates, so on shared services the last agent constructed silently
 steals every approval.
 
-Two settings can't be per-agent and are labelled app-wide in the UI: **Appearance** is a property
-of the window (and lives outside the shared config), and **Context window** is applied as
-`OLLAMA_CONTEXT_LENGTH` when MandoCode starts the Ollama daemon — one daemon, one context window.
+One setting can't be per-agent and is labelled app-wide in the UI: **Appearance** is a property
+of the window (and lives outside the shared config). **Context window** used to share that label —
+it was applied as `OLLAMA_CONTEXT_LENGTH` at daemon start, one daemon, one window — but it now
+rides on every chat request as `num_ctx`, so each agent's own value governs its own conversations,
+like any other per-agent setting.
 
 ### Settings and the config file
 

src/MandoCode.Desktop/MainWindow.xaml

@@ -1027,11 +1027,11 @@
                                            Minimum="1" SpinButtonPlacementMode="Compact" SmallChange="1024"
                                            ValueChanged="Setting_NumberChanged" HorizontalAlignment="Stretch"
                                            ToolTipService.ToolTip="Upper limit on how long a single response can be. Raise it if answers get cut off mid-code; lower it for shorter, faster replies."/>
-                                <NumberBox x:Name="S_ContextLength" Header="Context window (tokens, 0 = Ollama default) — app-wide"
+                                <NumberBox x:Name="S_ContextLength" Header="Context window (tokens, 0 = Ollama default)"
                                            Tag="contextLength" Minimum="0" SpinButtonPlacementMode="Compact" SmallChange="2048"
                                            ValueChanged="Setting_NumberChanged" HorizontalAlignment="Stretch"
                                            ToolTipService.ToolTip="How much conversation and file content the model can see at once. Larger windows track big files and long chats but use much more RAM/VRAM and respond slower. 0 keeps Ollama's default."/>
-                                <TextBlock Text="Applied when MandoCode starts Ollama (as OLLAMA_CONTEXT_LENGTH). One Ollama server backs every agent, so unlike the settings above this one is app-wide, not per-agent."
+                                <TextBlock Text="Sent with every request (num_ctx), so it applies from this agent's next message — no restart needed. Auto-sized to the model's hardware tier when you switch models; set it here to override."
                                            Opacity="0.5" FontSize="11" TextWrapping="Wrap" Margin="0,-6,0,0"/>
                                 <ComboBox x:Name="S_Streaming" Header="Response streaming" SelectionChanged="Streaming_Changed"
                                           HorizontalAlignment="Stretch"

src/MandoCode.Desktop/Services/ConfigCoordinator.cs

@@ -12,14 +12,13 @@ namespace MandoCode.Desktop.Services;
 /// current settings"; it is <see cref="Defaults"/>, the starting point every NEW agent is seeded
 /// from. "Make Default for New Agents" is the one action that writes it.
 ///
-/// Two things are deliberately app-wide rather than per-agent, because the machinery underneath
-/// them is:
-///   • MCP servers — OS processes owned by a single shared McpClientManager. An agent can turn
-///     MCP on or off for itself (AIService.AttachMcpPluginsAsync honours its own EnableMcp), but
-///     the server SET is one list. Edits land on Defaults and mirror into every live agent, so
-///     per-agent autoApprove lookups agree with what the MCP page shows.
-///   • ContextLength — applied as OLLAMA_CONTEXT_LENGTH when MandoCode starts the Ollama daemon.
-///     One daemon, one context window.
+/// One thing is deliberately app-wide rather than per-agent, because the machinery underneath
+/// it is: MCP servers — OS processes owned by a single shared McpClientManager. An agent can turn
+/// MCP on or off for itself (AIService.AttachMcpPluginsAsync honours its own EnableMcp), but
+/// the server SET is one list. Edits land on Defaults and mirror into every live agent, so
+/// per-agent autoApprove lookups agree with what the MCP page shows.
+/// (ContextLength used to be on this list — env-var-scoped to the one daemon — but it now rides
+/// on every request as num_ctx, so each agent's own value governs its own conversations.)
 ///
 /// A clone's Save() must never be called: it would write that agent's session settings over
 /// everybody's defaults. The method is public and non-virtual on a type in the read-only harness

src/MandoCode.Desktop/ViewModels/ChatController.cs

@@ -1173,7 +1173,7 @@ private async Task ApplyModelSwitchAsync(string modelTag)
         if (recommendedCtx > 0 && recommendedCtx != _config.ContextLength)
         {
             _config.ContextLength = recommendedCtx;
-            _transcript.Append(_html.Dim($"Context window sized to {recommendedCtx / 1024}k tokens for this model tier (takes effect when MandoCode starts Ollama)."));
+            _transcript.Append(_html.Dim($"Context window sized to {recommendedCtx / 1024}k tokens for this model tier (applies from your next message)."));
         }
 
         _busy.Start("Switching model...");