feat(agent): default to GLM-5.3 with Kimi override

.github/ISSUE_TEMPLATE/agent-change.md

@@ -24,4 +24,4 @@ title: "[Agent] "
 
 <!-- Compatibility, performance, security, and how to revert the change if needed. -->
 
-<!-- A maintainer applies agent-candidate when the issue is ready. The agent opens a PR; merging is manual. -->
+<!-- A maintainer applies agent-candidate when the issue is ready. The agent uses GLM-5.3 by default; apply agent-kimi first for harder or image-based work. The agent opens a PR; merging is manual. -->

.github/workflows/agent-issue-trigger.yml

@@ -41,8 +41,16 @@ jobs:
           issue_json="$(gh api "repos/opensensor/django-forms-workflows/issues/${ISSUE_NUMBER}")"
           issue_title="$(jq -r .title <<< "$issue_json")"
           issue_body="$(jq -r '.body // "" | .[0:16000]' <<< "$issue_json")"
+          model="glm-5.3"
+          spec="automation/agent-harness/glm.yaml"
+          if jq -e '.labels | any(.name == "agent-kimi")' <<< "$issue_json" >/dev/null; then
+            model="kimi-k3"
+            spec="automation/agent-harness/kimi.yaml"
+          fi
+          echo "Launching issue #${ISSUE_NUMBER} with ${model}"
+          echo "Issue #${ISSUE_NUMBER}: ${model}" >> "$GITHUB_STEP_SUMMARY"
           session_json="$(doctl harness-runtime create \
-            --spec automation/agent-harness/kimi.yaml \
+            --spec "$spec" \
             --name "$session_name" \
             --interactive=false -o json)"
           session_id="$(jq -r '.session_id // empty' <<< "$session_json")"

automation/agent-harness/README.md

@@ -1,6 +1,6 @@
 # Issue agent pilot
 
-A maintainer can apply `agent-candidate` to a scoped public issue to start one DigitalOcean Harness Runtime session. The session uses Kimi K3 with automatic tool approval, edits a clone of this repository, and pushes `agent/issue-<issue>-<workflow run>` to the dedicated `matteius/django-forms-workflows` fork. GitHub Actions is disabled on that fork. A signed push webhook asks the shared PR broker to open a regular pull request against `opensensor/django-forms-workflows` main. The repository's existing PR CI then runs; a maintainer reviews and merges manually. The agent cannot push to the base repository.
+A maintainer can apply `agent-candidate` to a scoped public issue to start one DigitalOcean Harness Runtime session. The session uses GLM-5.3 by default with automatic tool approval. Apply `agent-kimi` before `agent-candidate` for harder or image-based work to select Kimi K3. The model is selected once when the session starts. The agent edits a clone of this repository and pushes `agent/issue-<issue>-<workflow run>` to the dedicated `matteius/django-forms-workflows` fork. GitHub Actions is disabled on that fork. A signed push webhook asks the shared PR broker to open a regular pull request against `opensensor/django-forms-workflows` main. The repository's existing PR CI then runs; a maintainer reviews and merges manually. The agent cannot push to the base repository.
 
 The launcher exits after submitting the prompt; it does not keep a runner occupied while the agent works. Applying the label again launches another session. Remove the label to prevent a delayed fork push from opening a PR. Use the `Agent-ready change` issue template to give the agent clear acceptance criteria and validation steps.
 

automation/agent-harness/glm.yaml

@@ -0,0 +1,13 @@
+name: django-forms-public-issue-glm
+agent: opencode
+size: mars-2vcpu-4gb
+idle_timeout: 30m
+repos:
+  - opensensor/django-forms-workflows
+env:
+  HARNESS_INFERENCE_MODEL: glm-5.3
+secrets:
+  HARNESS_INFERENCE_API_KEY: ${HARNESS_INFERENCE_API_KEY}
+  GIT_DEPLOY_KEY_B64: ${GIT_DEPLOY_KEY_B64}
+permissions:
+  default: allow