feat(sdd): configurable models for judgment-day agents by danielgap · Pull Request #383 · Gentleman-Programming/gentle-ai

danielgap · 2026-04-26T18:12:47Z

🔗 Linked Issue

Closes #208

🏷️ PR Type

type:feature — New feature (non-breaking change that adds functionality)

📝 Summary

Adds 3 configurable judgment-day agents (jd-judge-a, jd-judge-b, jd-fix-agent) to the OpenCode overlay system, enabling users to assign different models per JD agent through the TUI model picker.

Key design decisions:

Named agents in SDD overlay — JD agents live in sdd-overlay-multi.json alongside SDD phases, since the JD skill is distributed with SDD
Root model exclusion — JD agents are intentionally excluded from inheriting the root model, enabling diversity of perspective between judges (different models = different viewpoints)
4 presets: Sonnet, Opus, Haiku, and Diversity (Judge A=Opus, Judge B=Haiku, Fix=Sonnet)
Profile-safe — JD agents are excluded from profile flows (they're global workflow agents, not SDD phases)

📂 Changes

File / Area	What Changed
`internal/assets/opencode/sdd-overlay-multi.json`	3 JD agent definitions + orchestrator delegation permissions
`internal/opencode/models.go`	`JDPhases()` and `ConfigurableAgentPhases()` for dynamic agent discovery
`internal/components/sdd/inject.go`	Model assignment rows, reasons, `isJDAgent()` helper, root model exclusion
`internal/model/claude_model.go`	JD entries in 3 presets + `ClaudeModelPresetDiversity()`
`internal/components/sdd/read_assignments.go`	`buildConfigurableAgentSet()` using `ConfigurableAgentPhases()`
`internal/components/sdd/profiles.go`	JD global permissions in `GenerateProfileOverlay()`, reserved profile names
`internal/tui/screens/model_picker.go`	JD rows, separator with focus indicator, `ModelPickerRowsForProfile()`
`internal/tui/screens/claude_model_picker.go`	JD in `claudePhases`/labels, Diversity preset wired in
`internal/tui/screens/profile_create.go`	Uses `ModelPickerRowsForProfile()` to exclude JD from profiles
`internal/tui/model.go`	`ForProfile` flag for profile create/edit flows
`internal/assets/skills/judgment-day/SKILL.md`	Conditional delegation (named agents for OpenCode, ad-hoc for others)
`internal/assets/claude/agents/jd-*.md`	3 Claude agent files with `{{CLAUDE_MODEL}}` sentinel
`internal/assets/kiro/agents/jd-*.md`	3 Kiro agent files with `{{KIRO_MODEL}}` sentinel
`internal/components/uninstall/service.go`	JD agents in `sddPhaseAgents` cleanup list

🧪 Test Plan

Unit Tests

go test ./...

E2E Tests (Docker required)

cd e2e && ./docker-test.sh

Unit tests pass (go test ./...)
E2E tests pass (cd e2e && ./docker-test.sh)
Manually tested locally — full integration test with temp HOME:
- Built binary from branch, ran sync --sdd-mode multi against temp config
- Verified 15 agents in output (build + 11 SDD + 3 JD)
- Verified JD agents: mode=subagent, hidden=true, no model inherited from root
- Verified orchestrator permissions: jd-judge-a: allow, jd-judge-b: allow, jd-fix-agent: allow
- Verified TUI model picker shows JD rows with separator
- Verified presets (Sonnet, Opus, Haiku, Diversity) assign models correctly

🤖 Automated Checks

The following checks run automatically on this PR:

Check	Status	Description
Check Issue Reference	⏳	PR body must contain `Closes/Fixes/Resolves #N`
Check Issue Has `status:approved`	⏳	Linked issue must have been approved before work began
Check PR Has `type:*` Label	⏳	Exactly one `type:*` label must be applied
Unit Tests	⏳	`go test ./...` must pass
E2E Tests	⏳	`cd e2e && ./docker-test.sh` must pass

✅ Contributor Checklist

PR is linked to an issue with status:approved
I have added the appropriate type:* label to this PR
Unit tests pass (go test ./...)
E2E tests pass (cd e2e && ./docker-test.sh)
I have updated documentation if necessary
My commits follow Conventional Commits format
My commits do not include Co-Authored-By trailers

💬 Notes for Reviewers

The 9 commits are organized in dependency order — each builds on the previous one
A judgment-day adversarial review was run on both the proposal AND the implementation (0 criticals)
The Diversity preset (Opus for deep reasoning on Judge A, Haiku for fast pattern matching on Judge B, Sonnet for Fix) is the recommended default for perspective diversity
No changes to single-mode overlay — JD delegation contradicts single-mode's "Do NOT delegate" instruction

Add jd-judge-a, jd-judge-b, and jd-fix-agent as hidden sub-agents in the multi-mode overlay. Update orchestrator permissions to allow delegation to these agents. Refs: Gentleman-Programming#208

JDPhases returns the judgment-day sub-agent names (jd-judge-a, jd-judge-b, jd-fix-agent). ConfigurableAgentPhases combines SDD phases and JD agents for the TUI model picker. Refs: Gentleman-Programming#208

Add jd-judge-a, jd-judge-b, jd-fix-agent to claudeModelAssignmentRowOrder and claudeModelAssignmentReasons in inject.go. Add JD entries to all three existing presets (balanced, performance, economy) in claude_model.go. Add new ClaudeModelPresetDiversity() preset for perspective diversity between judges (opus/haiku/sonnet split). Refs: Gentleman-Programming#208

Add isJDAgent helper and modify injectModelAssignments case 3 to skip root model fallback for judgment-day agents. This preserves independent model configuration for diversity of perspective between judges. Refs: Gentleman-Programming#208

Pattern 1 now documents both OpenCode named-agent delegation (jd-judge-a, jd-judge-b, jd-fix-agent) and generic delegate() for agents without named sub-agent support. Pattern 4 and Rules section updated to reference both delegation modes.

@Builtin

Create jd-judge-a, jd-judge-b, and jd-fix-agent agent files for both Claude and Kiro backends. Claude agents use {{CLAUDE_MODEL}} placeholder with explicit tool allowlists. Kiro agents use {{KIRO_MODEL}} with @Builtin and @engram tool references.

Add jd-judge-a, jd-judge-b, and jd-fix-agent to sddPhaseAgents so they are properly cleaned up during uninstall.

…filtering, diversity preset

Copilot

Pull request overview

Adds configurable Judgment Day (JD) workflow agents to the OpenCode SDD multi-mode overlay and exposes per-JD-agent model selection in the TUI (including a new “Diversity” preset), while keeping JD agents global (excluded from profile-scoped model assignment).

Changes:

Introduces 3 named JD agents (jd-judge-a, jd-judge-b, jd-fix-agent) in the OpenCode multi overlay with orchestrator delegation permissions.
Extends model discovery/configuration to include JD agents (new JDPhases() / ConfigurableAgentPhases()), updates injection/assignment reading, and adds root-model propagation exclusion for JD agents.
Updates TUI model pickers and presets (Claude) to support configuring JD agent models; updates JD skill docs and adds agent definition assets (Claude/Kiro).

Reviewed changes

Copilot reviewed 21 out of 21 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
internal/tui/screens/profile_create.go	Uses profile-filtered model-picker rows to exclude JD agents from profile flows.
internal/tui/screens/model_picker_test.go	Updates expected row count to include separator + JD rows.
internal/tui/screens/model_picker.go	Adds JD rows + separator, ForProfile filtering, and JD assignment mapping/rendering.
internal/tui/screens/claude_model_picker.go	Adds “Diversity” preset and includes JD phases/labels in Claude custom mode.
internal/tui/model.go	Flags model picker as profile-scoped and uses filtered rows during profile creation/edit.
internal/opencode/models.go	Adds `JDPhases()` and `ConfigurableAgentPhases()` for shared agent discovery.
internal/model/claude_model.go	Extends presets with JD entries and adds `ClaudeModelPresetDiversity()`.
internal/components/uninstall/service.go	Includes JD agents in SDD-related cleanup list.
internal/components/sdd/read_assignments.go	Expands valid assignment set to include JD agents (via `ConfigurableAgentPhases()`).
internal/components/sdd/profiles_test.go	Updates permission expectations to account for JD permissions only in multi overlay.
internal/components/sdd/profiles.go	Reserves JD names from profile naming and grants JD permissions in generated profile overlays.
internal/components/sdd/inject_test.go	Updates multi-overlay agent count expectations to include JD agents.
internal/components/sdd/inject.go	Adds JD rows to model assignment tables and excludes JD agents from root model propagation.
internal/assets/skills/judgment-day/SKILL.md	Documents named-agent delegation when supported; fallback guidance for other agents.
internal/assets/opencode/sdd-overlay-multi.json	Adds JD agent definitions and allows orchestrator delegation to them.
internal/assets/kiro/agents/jd-judge-b.md	Adds Kiro JD Judge B agent definition.
internal/assets/kiro/agents/jd-judge-a.md	Adds Kiro JD Judge A agent definition.
internal/assets/kiro/agents/jd-fix-agent.md	Adds Kiro JD Fix Agent definition.
internal/assets/claude/agents/jd-judge-b.md	Adds Claude JD Judge B agent definition.
internal/assets/claude/agents/jd-judge-a.md	Adds Claude JD Judge A agent definition.
internal/assets/claude/agents/jd-fix-agent.md	Adds Claude JD Fix Agent definition.

Comments suppressed due to low confidence (1)

internal/tui/screens/model_picker.go:106

ModelPickerRows/ModelPickerRowsForProfile hard-code slice capacities (15/11). Since the row list length depends on opencode.SDDPhases()/JDPhases(), the fixed numbers will drift as phases change. Prefer computing capacity from the phase slice lengths (or just omit the capacity) to avoid future off-by-one assumptions.

func ModelPickerRows() []string {
	rows := make([]string, 0, 15)
	rows = append(rows, SDDOrchestratorPhase)
	rows = append(rows, "Set all phases")
	rows = append(rows, opencode.SDDPhases()...)
	rows = append(rows, "--- Judgment Day ---")
	rows = append(rows, opencode.JDPhases()...)
	return rows
}

// ModelPickerRowsForProfile returns model picker rows for profile creation.
// JD agents are excluded because they are global (not profile-scoped).
func ModelPickerRowsForProfile() []string {
	rows := make([]string, 0, 11)
	rows = append(rows, SDDOrchestratorPhase)
	rows = append(rows, "Set all phases")
	rows = append(rows, opencode.SDDPhases()...)
	return rows

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-04-26T18:18:25Z

+// sddPhaseSet is the set of valid agent names that may appear in
+// opencode.json. It includes SDD phases, JD agents, and the orchestrator.
+var sddPhaseSet = buildConfigurableAgentSet()

-func buildSDDPhaseSet() map[string]bool {
-	phases := opencode.SDDPhases()
+func buildConfigurableAgentSet() map[string]bool {
+	phases := opencode.ConfigurableAgentPhases() // SDD + JD phases
 	set := make(map[string]bool, len(phases)+1)


The variable name sddPhaseSet is now misleading: it includes JD agents and the orchestrator (i.e., more than SDD phases). Renaming it (e.g., configurableAgentSet / agentNameSet) would reduce confusion when reading the filtering logic in ReadCurrentModelAssignments.

Copilot · 2026-04-26T18:18:25Z

 	ClaudePresetPerformance: "Maximum quality: opus for architecture, planning & verification phases",
 	ClaudePresetEconomy:     "Cost-optimised: sonnet for all phases, haiku for archiving",
+	ClaudePresetDiversity:   "Diversity: Opus for Judge A, Haiku for Judge B, Sonnet for fixes",
 	ClaudePresetCustom:      "Pick the model alias for each SDD phase individually",


The "custom" preset description says it lets users pick a model for each SDD phase, but the custom phase list now also includes JD agents (and "default"). Update the description so it matches what the UI actually configures.

Suggested change

ClaudePresetCustom: "Pick the model alias for each SDD phase individually",

ClaudePresetCustom: "Pick the model alias for each SDD phase, JD agent, and general delegation entry individually",

Copilot · 2026-04-26T18:18:25Z

 var reservedProfileNames = map[string]bool{
 	"default":          true,
 	"sdd-orchestrator": true,
+	"jd-judge-a":      true,
+	"jd-judge-b":      true,
+	"jd-fix-agent":    true,
 }


JD agent names are duplicated as string literals in reservedProfileNames. Since there is already a central list in opencode.JDPhases(), consider building this map using that list (plus the existing reserved names) to avoid future drift if JD agents are renamed/added.

Suggested change

var reservedProfileNames = map[string]bool{

"default": true,

"sdd-orchestrator": true,

"jd-judge-a": true,

"jd-judge-b": true,

"jd-fix-agent": true,

}

// Keep explicitly reserved non-JD names here, and derive JD agent names from

// the canonical opencode.JDPhases() list to avoid drift.

var reservedProfileNames = func() map[string]bool {

names := map[string]bool{

"default": true,

"sdd-orchestrator": true,

}

for _, name := range opencode.JDPhases() {

names[name] = true

}

return names

}()

Copilot · 2026-04-26T18:18:26Z

@@ -212,8 +232,16 @@ func handleModelNav(
 				assignments[phase] = assignment
 			}
 			state.AllPhasesModel = assignment
+		case state.SelectedPhaseIdx == separatorIdx:
+			// Separator row ("--- Judgment Day ---") — no action, skip.
+		case state.SelectedPhaseIdx > separatorIdx:
+			// JD agent rows: map to JDPhases() after separator.
+			jdIdx := state.SelectedPhaseIdx - separatorIdx - 1
+			if jdIdx < len(jdPhases) {
+				assignments[jdPhases[jdIdx]] = assignment
+			}


The "--- Judgment Day ---" separator row is treated as a selectable phase (it appears in ModelPickerRows), but selecting a model for it becomes a no-op in handleModelNav. In the current flow, pressing Enter on that row will still open provider/model selection and then silently apply nothing, which is confusing UX and looks like a broken selection.

Consider making the separator row non-actionable (e.g., skip it in the ScreenModelPicker enter logic / cursor navigation), or represent rows with metadata so the UI can render separators without allowing selection.

Copilot · 2026-04-26T18:18:26Z

@@ -279,8 +307,15 @@ func renderPhaseList(
 	b.WriteString(styles.SubtextStyle.Render("Current assignments:"))
 	b.WriteString("\n\n")

-	rows := ModelPickerRows()
+	var rows []string
+	if state.ForProfile {
+		rows = ModelPickerRowsForProfile()
+	} else {
+		rows = ModelPickerRows()
+	}
 	phases := opencode.SDDPhases()
+	jdPhases := opencode.JDPhases()
+	separatorIdx := 2 + len(phases)

 	for idx, row := range rows {
 		focused := idx == cursor
@@ -306,8 +341,29 @@ func renderPhaseList(
 			} else {
 				label = fmt.Sprintf("%-20s (not set)", row)
 			}
+		case idx == separatorIdx:
+			// Separator row — render as a visual divider with subtle indicator when focused.
+			if focused {
+				b.WriteString(styles.SubtextStyle.Render("▸ " + row) + "\n")
+			} else {
+				b.WriteString(styles.SubtextStyle.Render("  " + row) + "\n")
+			}
+			continue
+		case idx > separatorIdx:
+			// JD agent rows
+			jdIdx := idx - separatorIdx - 1
+			if jdIdx < len(jdPhases) {
+				phase := jdPhases[jdIdx]
+				assignment, ok := assignments[phase]
+				if ok && assignment.ProviderID != "" {
+					provName, modelName := resolveNames(assignment, state)
+					label = fmt.Sprintf("%-20s %s / %s", row, provName, modelName)
+				} else {
+					label = fmt.Sprintf("%-20s (default)", row)
+				}
+			}
 		default:


New JD-row mapping and separator behavior in handleModelNav/renderPhaseList isn’t covered by tests (e.g., selecting the separator should not change assignments, and selecting each JD row should set the expected jd-* key). Since this logic is index-based, a small row-order change could silently mis-assign models.

Add unit tests that exercise the separator index and the JD indices (first/last JD row) to lock in the mapping.

… separator, tests

danielgap added 9 commits April 26, 2026 17:30

feat: add 3 JD agent definitions to sdd-overlay-multi.json

0b33a38

Add jd-judge-a, jd-judge-b, and jd-fix-agent as hidden sub-agents in the multi-mode overlay. Update orchestrator permissions to allow delegation to these agents. Refs: Gentleman-Programming#208

feat: add JDPhases() and ConfigurableAgentPhases() to opencode

87e52f6

JDPhases returns the judgment-day sub-agent names (jd-judge-a, jd-judge-b, jd-fix-agent). ConfigurableAgentPhases combines SDD phases and JD agents for the TUI model picker. Refs: Gentleman-Programming#208

feat: add JD agents to uninstall cleanup list

911aa2d

Add jd-judge-a, jd-judge-b, and jd-fix-agent to sddPhaseAgents so they are properly cleaned up during uninstall.

feat: add JD agents to sddPhaseSet, profile perms, and TUI model picker

0b1753e

fix: address judgment-day review findings — separator focus, profile …

03351f0

…filtering, diversity preset

Copilot AI review requested due to automatic review settings April 26, 2026 18:12

Copilot started reviewing on behalf of danielgap April 26, 2026 18:13 View session

Copilot AI reviewed Apr 26, 2026

View reviewed changes

fix: address Copilot PR review — rename, dynamic init, non-selectable…

a6f3511

… separator, tests

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(sdd): configurable models for judgment-day agents#383

feat(sdd): configurable models for judgment-day agents#383
danielgap wants to merge 10 commits intoGentleman-Programming:mainfrom
danielgap:feat/issue-208-judgment-day-models

danielgap commented Apr 26, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Apr 26, 2026

Uh oh!

Copilot AI Apr 26, 2026

Uh oh!

Copilot AI Apr 26, 2026

Uh oh!

Copilot AI Apr 26, 2026

Uh oh!

Copilot AI Apr 26, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

	ClaudePresetCustom: "Pick the model alias for each SDD phase individually",
	ClaudePresetCustom: "Pick the model alias for each SDD phase, JD agent, and general delegation entry individually",

Conversation

danielgap commented Apr 26, 2026

🔗 Linked Issue

🏷️ PR Type

📝 Summary

📂 Changes

🧪 Test Plan

🤖 Automated Checks

✅ Contributor Checklist

💬 Notes for Reviewers

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Apr 26, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 26, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 26, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 26, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 26, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants