Expose GLM reasoning-effort tiers in the model picker

README.md

@@ -52,6 +52,27 @@ This starts a local server and opens OpenCode in your default browser. Sessions
 
 See the [OpenCode Web docs](https://opencode.ai/docs/web/) for more.
 
+### GLM reasoning-effort tiers
+
+`glm-latest` (GLM-5.2) reasons by default. To trade thinking depth for speed, the
+model picker (`ctrl+x m`) exposes the same model at several reasoning-effort tiers —
+GLM-5.2 collapses low/medium into "high", so these are the only distinct levels:
+
+| Picker entry | `reasoning_effort` | Use it for |
+|---|---|---|
+| `glm-latest` | *(gateway default)* | the default — thinking on |
+| `glm-max`    | `max`  | deepest reasoning |
+| `glm-high`   | `high` | strong reasoning, faster |
+| `glm-fast`   | `none` | no thinking, fastest replies |
+
+<figure>
+<img alt="Switching GLM reasoning-effort tiers from the OpenCode model picker" src="demo/glm-effort-picker.gif" />
+<figcaption>Picking a GLM reasoning-effort tier in OpenCode (<code>ctrl+x m</code>)</figcaption>
+</figure>
+
+The tiers are defined in [`coding-agents/opencode/settings/juspay.nix`](coding-agents/opencode/settings/juspay.nix);
+all target the same gateway model (`glm-latest`) and differ only in `reasoningEffort`.
+
 ## Coding Agent Setup
 
 This repo uses [APM](https://microsoft.github.io/apm/) for coding agent configuration. `.claude/` and `.opencode/` are **vendored** — committed to git and kept in sync by a CI check (`apm-sync` workflow).

coding-agents/opencode/settings/juspay.nix

@@ -1,10 +1,31 @@
 let
-  # All models accept text+image input, text output; name = attr key
-  mkModel = name: { context, output }: {
-    inherit name;
-    modalities = { input = [ "text" "image" ]; output = [ "text" ]; };
-    limit = { inherit context output; };
-  };
+  # All models accept text+image input, text output; name = attr key.
+  # reasoningEffort (optional) marks the model as a reasoning model and is
+  # forwarded as the OpenAI-compatible `reasoning_effort` request field —
+  # @ai-sdk/openai-compatible maps the camelCase key to snake_case for us.
+  # id (optional) sets the wire model id when the attr key differs from the
+  # gateway's model name — used to expose one gateway model at several
+  # reasoning-effort tiers.
+  mkModel = name: { context, output, reasoningEffort ? null, id ? null }:
+    let
+      base = {
+        inherit name;
+        modalities = { input = [ "text" "image" ]; output = [ "text" ]; };
+        limit = { inherit context output; };
+      } // (if id == null then { } else { inherit id; });
+    in
+    if reasoningEffort == null then base
+    else base // {
+      reasoning = true;
+      options = { inherit reasoningEffort; };
+    };
+
+  # GLM-5.2 (glm-latest and its effort-tier siblings) share a 1M-token context
+  # window (the old 202752 was a copy-paste default). output stays at 32000:
+  # OpenCode caps the wire max_tokens at 32000 for this custom-provider model
+  # regardless of this field (verified), so a higher value would only shrink the
+  # usable input budget without raising the real output limit.
+  glmLimits = { context = 1000000; output = 32000; };
 
   models = builtins.mapAttrs mkModel {
     open-large              = { context = 202752;  output = 32000; };
@@ -18,7 +39,15 @@ let
     gemini-3-pro-preview    = { context = 1048576; output = 65535; };
     gemini-3-flash-preview  = { context = 1048576; output = 65535; };
     minimax-m2              = { context = 202752;  output = 32000; };
-    glm-latest              = { context = 202752;  output = 32000; };
+    # glm-latest is GLM-5.2. Its default reasoning is left untouched (no
+    # reasoning_effort sent — the gateway default, which is thinking-on). The
+    # effort tiers below are sibling picker entries that target the same gateway
+    # model via `id`; GLM-5.2 collapses low/medium into "high", so max / high /
+    # off are the only distinct levels.
+    glm-latest              = glmLimits;
+    glm-max                 = glmLimits // { reasoningEffort = "max";  id = "glm-latest"; };
+    glm-high                = glmLimits // { reasoningEffort = "high"; id = "glm-latest"; };
+    glm-fast                = glmLimits // { reasoningEffort = "none"; id = "glm-latest"; };
     kimi-latest             = { context = 262000;  output = 32000; };
   };
 in