Configuration
Configure Gem Team in .gem-team.yaml at the project root. Start with this minimal example:
orchestrator:
max_concurrent_agents: 2
git_commit_on_gate_pass: false
model_routing:
enabled: false
tiers:
premium: "gpt-5.6-sol (openai)"
explore: "gpt-5.6-luna (openai)"
quality:
visual_regression_enabled: false
visual_diff_threshold: 0.95
a11y_audit_level: none
devops:
approval_required_for: [production]
deployment_strategy: rolling
auto_rollback_on_failure: false
testing:
screenshot_on_failure: true
See the fully commented default for every available setting.
Settings Reference
Model Routing
Model routing chooses a configured model for each delegated agent. Values use the model (provider) format.
| Setting | Type | Default | Description |
|---|---|---|---|
model_routing.enabled | boolean | false | Enable tier-based model selection for delegated agents |
model_routing.tiers.premium | string | gpt-5.6-sol (openai) | Model for planning, diagnosis, and review |
model_routing.tiers.explore | string | gpt-5.6-luna (openai) | Model for exploration and bounded execution |
Agent tiers
Premium: gem-planner, gem-debugger, and gem-reviewer.
Explore: gem-researcher, gem-implementer, gem-browser-tester,
gem-mobile-tester, gem-devops, gem-documentation-writer,
gem-skill-creator, and gem-code-simplifier.
The Orchestrator is not assigned a tier. With routing disabled, delegation uses the normal selection process.
Quality settings are defaults, not hard requirements. Task criteria and explicit user requests take precedence; otherwise, .gem-team.yaml takes precedence over agent defaults.
Orchestrator
| Setting | Type | Default | Description |
|---|---|---|---|
orchestrator.max_concurrent_agents | number | 2 | Maximum parallel agent executions |
orchestrator.git_commit_on_gate_pass | boolean | false | Commit wave output when an invoked gate passes |
Complexity and reviewer selection are evidence-driven workflow rules, not configuration switches. Phase 0 assigns complexity from the request and supplied evidence. The Planner mechanically self-checks every MEDIUM/HIGH plan before writing it; explicit requests, validation failures, HIGH/high-risk work, and critic signals route to the Reviewer.
The reviewer contract uses independent review_mode, review_target, and
review_scope axes. For critic mode, provide critic_subject and
critic_context under handoff as documented in Reviewer routing.
Quality
| Setting | Type | Default | Description |
|---|---|---|---|
quality.visual_regression_enabled | boolean | false | Enable screenshot comparison checks |
quality.visual_diff_threshold | number | 0.95 | Visual comparison threshold from 0.0 to 1.0 |
quality.a11y_audit_level | enum | none | A11y audit depth: none, basic, or full |
DevOps
| Setting | Type | Default | Description |
|---|---|---|---|
devops.approval_required_for | enum | [production] | Environments requiring explicit approval |
devops.deployment_strategy | enum | rolling | Strategy: rolling, blue_green, or canary |
devops.auto_rollback_on_failure | boolean | false | Attempt rollback after deployment failure |
Testing
| Setting | Type | Default | Description |
|---|---|---|---|
testing.screenshot_on_failure | boolean | true | Capture a screenshot when browser/UI tests fail |