Switch polecat + crew default model to Gemini 3.8 Flash

[anuveyatsu]

Switches the default worker model from GLM 5.3 Flash to Gemini 3.8 Flash.

Why

GLM looped and no-op'd on real beads (the msf Africa-map bead cost $3.26 twice and landed nothing). Gemini 3.8 Flash is the capability call.

Billing — stays one bill (verified)

Reached through the same Cloudflare AI Gateway, so credentials and attribution are unchanged. Google AI Studio is enabled on the gateway and on Unified Billing, so Gemini draws the same prepaid Cloudflare balance GLM did — one bill, no BYOK Google key on the nodes. Confirmed on staging 2026-09-14:

• workgraph-staging-oss: wholesale=true, byok_only=false, workers_ai_billing_mode=unified
• google-ai-studio/gemini-3.8-flash routes HTTP 200 through the /compat endpoint with only the gateway token (no per-request provider key)

Changes

• infra/ansible/roles/execution_cell/defaults/main.yml (the live default): wg_model_roles polecat + crew → google-ai-studio/gemini-3.8-flash; added its wg_model_limits entry (context 1000000, output 8192 — output cap required by OpenCode; context far above its ~19,900-token system prompt). Rationale comment rewritten.
• internal/runner/plan.go: registered google-ai-studio/gemini-3.8-flash in GatewayModels so it's known to validation and the bake-off. The built-in DefaultModels fallback stays anthropic/claude-sonnet-5 on purpose — it pairs with the default claude runtime, and the live default comes from models.json (the ansible file), not the fallback.

Runtime stays opencode. Any run can override with -runtime claude -model anthropic/claude-sonnet-5.

Verification

go vet ./internal/runner/...   # clean
go test ./internal/runner/... ./cmd/runner/...   # ok

On merge, deploy.yml rewrites /etc/workgraph/models.json and restarts the dispatchers. Post-deploy, a dispatch should show model: google-ai-studio/gemini-3.8-flash in workgraph_bead, tagged in usage_records.

🤖 Generated with Claude Code

Summary by CodeRabbit

• New Features

  • Added support for Gemini 3.8 Flash in OpenCode runs.
  • Configured Gemini 3.8 Flash with a 1,000,000-token context limit and 8,192-token output limit.

• Configuration

  • Updated default work-role model settings to use Gemini 3.8 Flash.
  • Retained support for the existing GLM model limits.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: dfe43c88-36c8-47eb-b7ab-afcf12d1736f

📥 Commits

Reviewing files that changed from the base of the PR and between 7a01a60 and dc9ff70.

📒 Files selected for processing (2)

• infra/ansible/roles/execution_cell/defaults/main.yml
• internal/runner/plan.go

Included review availability: Your plan provides up to 1 included review per hour; 0 remain after this review.

---

📝 Walkthrough

Walkthrough

The default polecat and crew models now use Gemini 3.8 Flash. Ansible and runner model-limit tables define a 1,000,000-token context limit and an 8,192-token output limit.

Changes

Gemini model support

Layer / File(s)	Summary
Model selection and runtime limits infra/ansible/roles/execution_cell/defaults/main.yml, internal/runner/plan.go	The default polecat and crew models change to google-ai-studio/gemini-3.8-flash. The Ansible and runner limit tables add context 1000000 and output 8192 for this model. The existing GLM limit entry remains.

Priority: ⬇️ Low

Estimated code review effort: 2 (Simple) | ~10 minutes

Change: Feature

Merge Risk: ⚪ Minimal · up to dc9ff

The default model change and runtime limits are aligned across the configured layers; no current merge-blocking risk is identified.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely describes the main change: switching the default model for polecat and crew to Gemini 3.8 Flash.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 1…
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch switch-agents-to-gemini-3.8-flash

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}