Fix: use reasoningEffort instead of variant for GLM reasoning control

Z.AI's GLM-5.2 API uses reasoning_effort as a top-level parameter
(values: max, high, low), NOT OpenCode's variant field.

OpenCode passes unrecognized agent frontmatter fields through to the
provider as model options. Setting reasoningEffort: 'max' on an agent
gets sent to Z.AI as reasoning_effort='max' — the correct parameter.

Changes:
- models.sh: inject reasoningEffort instead of variant
- models.sh: strip function cleans model, variant, AND reasoningEffort
- install.sh: updated warning text to explain the mapping

Source repo agents are unaffected (template has no model/effort fields).
Deployed targets need re-install to get the corrected field.

Author: Cipher-85

.opencode/install.sh

@@ -133,11 +133,10 @@ fi
 # ── Validate models ──────────────────────────────────────────────
 printf '\n── Validating models ──────────────────────────────────────\n'
 
-# Warn about variants if provided
+# Warn about reasoning effort if provided
 if [ -n "$opus_variant$sonnet_variant$haiku_variant" ]; then
-  printf '  ⚠  Variants are provider-specific. If your provider does not support\n' >&2
-  printf '     the variant parameter, subagent tasks will fail with "Invalid API\n' >&2
-  printf '     parameter". If this happens, re-run without --variant-* flags.\n\n' >&2
+  printf '  Note: reasoning effort maps to the provider'\''s reasoning_effort API parameter.\n'
+  printf '  Z.AI/GLM accepts: max, high, low. Other providers may differ.\n\n'
 fi
 
 validation_failed=0

.opencode/lib/models.sh

@@ -60,12 +60,12 @@ else:
 " 2>/dev/null
 }
 
-# Inject or replace model: in an agent file's frontmatter
-# Usage: ccgs_inject_model <agent_file> <model_id> [variant]
-# Note: variant is provider-specific and may cause "Invalid API parameter"
-# errors if the provider doesn't support it. Leave empty to omit.
+# Inject or replace model: and reasoningEffort: in an agent file's frontmatter
+# Usage: ccgs_inject_model <agent_file> <model_id> [effort]
+# Note: reasoningEffort maps to the provider's reasoning_effort API parameter.
+# Z.AI/GLM accepts: "max", "high", "low". Other providers may differ.
 ccgs_inject_model() {
-  local file="$1" model_id="$2" variant="${3:-}"
+  local file="$1" model_id="$2" effort="${3:-}"
   python3 -c "
 import sys, re, yaml
 
@@ -74,14 +74,15 @@ with open('$file') as f:
 
 m = re.match(r'^(---\n)(.*?)(\n---)', txt, re.S)
 if not m:
-    print(f'ERROR: no frontmatter in $file', file=sys.stderr)
+    print(f'ERROR: no frontmatter in {file}', file=sys.stderr)
     sys.exit(1)
 
 fm = yaml.safe_load(m.group(2))
 fm['model'] = '$model_id'
-if '$variant':
-    fm['variant'] = '$variant'
+if '$effort':
+    fm['reasoningEffort'] = '$effort'
 else:
+    fm.pop('reasoningEffort', None)
     fm.pop('variant', None)
 
 # Re-dump frontmatter preserving key order
@@ -109,12 +110,10 @@ if not m:
 
 fm = yaml.safe_load(m.group(2))
 changed = False
-if 'model' in fm:
-    del fm['model']
-    changed = True
-if 'variant' in fm:
-    del fm['variant']
-    changed = True
+for key in ('model', 'variant', 'reasoningEffort'):
+    if key in fm:
+        del fm[key]
+        changed = True
 if not changed:
     sys.exit(0)