feat: add model variant support to execute plan panel

README.md

@@ -432,7 +432,7 @@ After the architect presents a summary, the user chooses an execution mode from
 | `Execute here` | When preserving current context matters |
 | `Loop` | Safer autonomous iteration |
 
-The dialog also lets you pick the execution model and auditor model at launch time. Those selections are remembered per project and pre-filled on later launches.
+The dialog also lets you pick the execution model and auditor model at launch time. Those selections are remembered per project and pre-filled on later launches. Optional **variant selectors** accompany each model selector, letting you choose provider-specific reasoning or thinking-effort levels (e.g., `low`, `high`, `max`) when the model exposes them. Variant selections are also persisted per project.
 
 Execution is immediate — there are no additional LLM calls between approval and execution. The system intercepts the user's approval answer, reads the cached plan, and dispatches it programmatically to the code agent. The architect never processes the approval response.
 

docs/api/README.md

@@ -342,6 +342,16 @@ Two model selectors are available:
 - Same model selection interface
 - Defaults to last-used selection, falling back to `config.auditorModel` → `config.executionModel`
 
+#### Variant Selection
+
+Each model selector has a companion **variant selector** that lets you choose a provider-specific reasoning or thinking-effort level. Variants are optional and provider/model-dependent — they appear only when the selected model exposes them via OpenCode model metadata.
+
+- Variants are provided by OpenCode model metadata. Examples include reasoning/thinking effort levels such as `low`, `high`, or `max` when exposed by the provider.
+- **Use default** leaves OpenCode, the agent, and the model defaults in control.
+- **Execution variants** affect code prompts for all launch modes (`New session`, `Execute here`, and `Loop`).
+- **Auditor variants** affect loop audit and final-audit prompts.
+- Variants are persisted per-project alongside model selections (same 30-day TTL).
+
 #### Persistence
 
 Your selections are automatically saved in `tui_preferences` after launch:
@@ -478,6 +488,10 @@ Model selection follows this priority order:
 3. `config.executionModel`
 4. Platform default
 
+**For variants:**
+1. Dialog selection persisted per project
+2. OpenCode/agent/model default
+
 ### Troubleshooting
 
 - **No plan found** — Ensure the architect output included the forge plan markers, or open the Plan Viewer for the current session.

docs/architecture.md

@@ -48,7 +48,7 @@ The TUI plugin provides a sidebar widget that displays:
 
 - Active and recent loops
 - Plan viewer with inline editing (view/edit/execute/export tabs)
-- Execution dialog with mode and model selection
+- Execution dialog with mode, model, and variant selection
 - Loop details dialog with session statistics
 - Command palette integration (`Forge: Show loops`, `Forge: View plan`, `Forge: Execute plan`)
 - Model selection dialog with recent model tracking

docs/modules.md

@@ -54,7 +54,7 @@ The TUI plugin providing sidebar widget and dialog system. Communicates with the
 
 - Exports `{ id: 'oc-forge', tui }`
 - Registers commands: `forge.plan.view`, `forge.plan.load`
-- Provides plan viewer, execution dialog, loop details, model selection
+- Provides plan viewer, execution dialog, loop details, model and variant selection
 
 Source: [src/tui.tsx](../src/tui.tsx)
 

package.json

@@ -1,6 +1,6 @@
 {
   "name": "opencode-forge",
-  "version": "0.4.5",
+  "version": "0.4.6",
   "type": "module",
   "oc-plugin": [
     "server",

src/agents/architect.ts

@@ -63,7 +63,7 @@ The plugin auto-captures marked plans from your assistant responses into SQL sto
     - Start with a short unmarked summary containing **Intention**, **Goal**, and **Approach**. Keep it brief: 1-3 sentences for intention/goal and 2-4 bullets for approach.
     - After the summary, wrap exactly one final plan with \`<!-- forge-plan:start -->\` and \`<!-- forge-plan:end -->\` markers (each on its own line)
     - Do NOT wrap only summaries, design options, or partial drafts
-    - The marked plan body must follow the existing detailed plan format: Objective, Loop Name, Phases with file targets/edits/acceptance criteria/verification, Decisions, Conventions, Key Context
+    - The marked plan body must follow the existing detailed plan format: Objective, a machine-readable \`Loop Name: short-slug\` line, Phases with file targets/edits/acceptance criteria/verification, Decisions, Conventions, Key Context
     - The marked plan must be extremely detailed and execution-ready: name exact files, exact symbols/functions/types to change, concrete data shapes, command wiring, expected control flow, error handling, and validation steps
     - Every phase must include explicit implementation instructions, precise edits per file, acceptance criteria, and targeted verification commands or assertions the code agent can run
 4. **Approve** — After the marked plan is output and auto-captured, call the question tool to get explicit approval with these options:
@@ -76,7 +76,7 @@ The plugin auto-captures marked plans from your assistant responses into SQL sto
 
 Present plans with:
 - **Objective**: What we're building and why
-- **Loop Name**: A short, machine-friendly name (1-3 words) that captures the plan's main intent. This will be used for worktree/session naming. Example: "Loop Name: auth-refactor" or "Loop Name: api-validation"
+- **Loop Name**: A short, machine-friendly name (1-3 words) that captures the plan's main intent. This will be used for worktree/session naming. Emit it as a plain machine-readable line, not a markdown heading or bullet: \`Loop Name: auth-refactor\` or \`Loop Name: api-validation\`. Place it near the top of the marked plan immediately after the objective.
 - **Phases**: Ordered implementation steps. Use exactly one \`<!-- forge-section -->\` marker per executable phase. Place it immediately before that phase's \`## Phase ...\` heading. Never place it before \`### Files\`, \`### Edits\`, \`### Acceptance Criteria\`, or \`### Verification\` — those are subsections inside the current phase. For every phase, specify the exact files affected, the precise code-level edits to make, sample change examples (such as function signature updates, new branches, or new exports), the existing symbols/modules being integrated with, concrete acceptance criteria, and phase-specific verification. Use \`### Files\`, \`### Edits\`, \`### Acceptance Criteria\`, and \`### Verification\` as subsections inside each phase. Shared blocks (\`## Decisions\`, \`## Conventions\`, \`## Key Context\`) go after all sections without a preceding marker.
 
   **Valid shape:**

src/hooks/forge-session-attach.ts

@@ -116,6 +116,8 @@ async function attachForgeSession(
       title?: string
       executionModel?: string
       auditorModel?: string
+      executionVariant?: string
+      auditorVariant?: string
       planSource?: 'stored' | 'inline'
       planText?: string
       initialPromptOwner?: 'server' | 'tui'
@@ -270,6 +272,8 @@ async function attachForgeSession(
           hostSessionId: resolvedHostSessionId,
           executionModel: cfg.executionModel,
           auditorModel: cfg.auditorModel,
+          executionVariant: cfg.executionVariant,
+          auditorVariant: cfg.auditorVariant,
           maxIterations: cfg.maxIterations ?? 50,
           sandboxEnabled: cfg.sandboxEnabled ?? false,
           planText,
@@ -278,7 +282,7 @@ async function attachForgeSession(
            startWatchdog: true,
            sendInitialPrompt,
          },
-        )
+       )
         if (!result.ok && result.code === 'conflict') {
           const row = deps.execDeps.loopsRepo.get(sessionProjectId, loopName)
           const removalAction = row?.status === 'cancelled' || row?.status === 'errored' || row?.status === 'stalled'

src/hooks/plan-approval.ts

@@ -1,7 +1,7 @@
 import type { ToolContext } from '../tools/types'
 import type { Hooks } from '@opencode-ai/plugin'
 import { parseModelString, retryWithModelFallback } from '../utils/model-fallback'
-import { extractPlanTitle, extractLoopNames, PLAN_EXECUTION_LABELS, type PlanExecutionLabel } from '../utils/plan-execution'
+import { extractPlanExecutionMetadata, PLAN_EXECUTION_LABELS, type PlanExecutionLabel } from '../utils/plan-execution'
 import { buildStartLoopCommand, createForgeExecutionService, type ForgeExecutionRequestContext } from '../services/execution'
 import { captureLatestPlanForSession } from '../services/plan-capture'
 
@@ -96,7 +96,7 @@ const processedApprovalCalls = new WeakMap<ToolContext, Set<string>>()
 const claimedApprovalPlans = new Set<string>()
 
 export { LOOP_BLOCKED_TOOLS }
-export { extractPlanTitle }
+export { extractPlanTitle } from '../utils/plan-execution'
 
 function isActiveLoopToolSession(state: { active?: boolean; sessionId?: string }, sessionID: string): boolean {
   return state.active === true && state.sessionId === sessionID
@@ -262,7 +262,7 @@ export function createToolExecuteAfterHook(ctx: ToolContext): Hooks['tool.execut
             return
           }
           const planText = plan.content
-          const title = extractPlanTitle(planText)
+          const { title, executionName } = extractPlanExecutionMetadata(planText)
 
           if (matchedLabel && !claimApprovalCall(ctx, input, matchedLabel, plan.key)) {
             markApprovalHandled(output, true)
@@ -330,7 +330,6 @@ export function createToolExecuteAfterHook(ctx: ToolContext): Hooks['tool.execut
           }
 
           if (matchedLabel === 'Loop') {
-            const { executionName } = extractLoopNames(planText)
             const uniqueLoopName = loop.generateUniqueLoopName(executionName)
 
             logger.log(`Plan approval: "${matchedLabel}" — scheduling dispatch IIFE for loop "${uniqueLoopName}"`)

src/index.ts

@@ -491,6 +491,7 @@ READ-ONLY mode: no file edits, no destructive commands. Search and analyze only.
 
 When emitting the final plan:
 - Wrap the plan in \`<!-- forge-plan:start -->\` and \`<!-- forge-plan:end -->\` (each on its own line)
+- Include one plain machine-readable \`Loop Name: short-slug\` line near the top of the marked plan, immediately after the objective. Do not emit loop name as a markdown heading or bullet.
 - Use exactly one \`<!-- forge-section -->\` marker per executable phase; place it immediately before that phase's \`## Phase\` heading
 - Do not insert \`<!-- forge-section -->\` before \`### Files\`, \`### Edits\`, \`### Acceptance Criteria\`, or \`### Verification\`
 - Shared \`## Decisions\` / \`## Conventions\` / \`## Key Context\` blocks go after all sections (no preceding marker)

src/loop/runtime.ts

@@ -183,6 +183,7 @@ export function createLoop(deps: LoopRuntimeDeps): Loop {
     promptText: string
     agent: 'code' | 'auditor-loop'
     model?: { providerID: string; modelID: string } | null
+    variant?: string
   }): Promise<{ error?: unknown; usedModel?: { providerID: string; modelID: string } | undefined }> {
     const { loopName, sessionId, promptText, agent } = input
 
@@ -200,7 +201,7 @@ export function createLoop(deps: LoopRuntimeDeps): Loop {
               worktreeDir: freshState.worktreeDir,
               workspaceId: freshState.workspaceId,
               prompt: promptText,
-              ...(model ? { auditorModel: model } : {}),
+              ...(model ? { auditorModel: model, ...(input.variant ? { auditorVariant: input.variant } : {}) } : {}),
             })
             return result.ok ? { data: true } : { error: result.error }
           })
@@ -234,7 +235,7 @@ export function createLoop(deps: LoopRuntimeDeps): Loop {
             ...(freshState.workspaceId ? { workspace: freshState.workspaceId } : {}),
             agent: 'code',
             parts: [{ type: 'text' as const, text: promptText }],
-            ...(model ? { model } : {}),
+            ...(model ? { model, ...(input.variant ? { variant: input.variant } : {}) } : {}),
           })
         })
       } catch (err) {
@@ -282,14 +283,17 @@ export function createLoop(deps: LoopRuntimeDeps): Loop {
       }>
 
       const lastMessage = messages.length > 0 ? messages[messages.length - 1] : null
-      const lastAssistant = [...messages].reverse().find((m) => m.info.role === 'assistant')
 
-      if (!lastAssistant) {
-        const role = lastMessage?.info.role ?? 'none'
-        logger.log(`Loop: no assistant message found in session ${sessionId}, last message role: ${role}`)
-        return { text: null, error: null, lastMessageRole: role }
+      if (!lastMessage) {
+        return { text: null, error: null, lastMessageRole: 'none' }
       }
 
+      if (lastMessage.info.role !== 'assistant') {
+        logger.log(`Loop: no assistant message found in session ${sessionId}, last message role: ${lastMessage.info.role ?? 'unknown'}`)
+        return { text: null, error: null, lastMessageRole: lastMessage.info.role ?? 'unknown' }
+      }
+
+      const lastAssistant = lastMessage
       if (lastAssistant.info.finish && lastAssistant.info.finish !== 'stop') {
         logger.log(`Loop: assistant message in session ${sessionId} is not final yet (finish=${lastAssistant.info.finish})`)
         return { text: null, error: null, lastMessageRole: `assistant:${lastAssistant.info.finish}` }
@@ -700,6 +704,7 @@ export function createLoop(deps: LoopRuntimeDeps): Loop {
       promptText: continuationPrompt,
       agent: 'code',
       model: loopModel,
+      variant: currentState.executionVariant,
     })
 
     if (promptResultError) {
@@ -762,6 +767,7 @@ export function createLoop(deps: LoopRuntimeDeps): Loop {
       promptText: continuationPrompt,
       agent: 'code',
       model: loopModel,
+      variant: state.executionVariant,
     })
     if (error) {
       logger.error(`rotateToCodingAfterAuditFailure: failed to send continuation prompt`, error)
@@ -804,6 +810,7 @@ export function createLoop(deps: LoopRuntimeDeps): Loop {
         promptText: recoveryPrompt,
         agent: 'code',
         model: resolveLoopModel(currentConfig, loopService, loopName),
+        variant: freshState.executionVariant,
       })
       if (promptResultError) {
         clearPromptPending(loopName, logger)
@@ -1057,6 +1064,7 @@ export function createLoop(deps: LoopRuntimeDeps): Loop {
     workspaceId?: string
     prompt: string
     auditorModel?: { providerID: string; modelID: string }
+    auditorVariant?: string
   }): Promise<{ ok: true } | { ok: false; error: unknown }> {
     const result = await promptAuditSession(v2Client, input)
     if (result.ok) return result
@@ -1069,7 +1077,7 @@ export function createLoop(deps: LoopRuntimeDeps): Loop {
         body: {
           agent: 'auditor-loop',
           parts: [{ type: 'text' as const, text: input.prompt }],
-          ...(input.auditorModel ? { model: input.auditorModel } : {}),
+          ...(input.auditorModel ? { model: input.auditorModel, ...(input.auditorVariant ? { variant: input.auditorVariant } : {}) } : {}),
         },
       })
       if (legacyResult.error) return { ok: false, error: legacyResult.error }
@@ -1149,6 +1157,7 @@ export function createLoop(deps: LoopRuntimeDeps): Loop {
       promptText: finalAuditPrompt,
       agent: 'auditor-loop',
       model: auditorModel,
+      variant: currentState.auditorVariant,
     })
 
     if (finalAuditPromptErr) {
@@ -1281,6 +1290,7 @@ export function createLoop(deps: LoopRuntimeDeps): Loop {
           sessionId: rotatedSessionId,
           promptText: continuationPrompt,
           agent: 'code',
+          variant: currentState.executionVariant,
         })
         if (promptErr) {
           logger.error(`Loop: failed to send continuation prompt after audit creation failure`, promptErr)
@@ -1315,6 +1325,7 @@ export function createLoop(deps: LoopRuntimeDeps): Loop {
       promptText: loopService.buildAuditPrompt(currentState),
       agent: 'auditor-loop',
       model: auditorModel,
+      variant: currentState.auditorVariant,
     })
 
     if (auditPromptErr) {
@@ -1327,6 +1338,8 @@ export function createLoop(deps: LoopRuntimeDeps): Loop {
             worktreeDir: currentState.worktreeDir,
             workspaceId: currentState.workspaceId,
             prompt: loopService.buildAuditPrompt(currentState),
+            auditorModel,
+            auditorVariant: currentState.auditorVariant,
           })
           if (retryResult.ok) {
             logger.log(`Loop: recovered audit prompt after workspace re-bind for ${loopName}`)
@@ -1345,6 +1358,8 @@ export function createLoop(deps: LoopRuntimeDeps): Loop {
               worktreeDir: fresh.worktreeDir,
               workspaceId: fresh.workspaceId,
               prompt: loopService.buildAuditPrompt(fresh),
+              auditorModel,
+              auditorVariant: fresh.auditorVariant,
             })
             if (!retry.ok) throw retry.error
           })
@@ -1699,6 +1714,7 @@ export function createLoop(deps: LoopRuntimeDeps): Loop {
         sessionId: newCodeSessionId,
         promptText: continuationPrompt,
         agent: 'code',
+        variant: currentState.executionVariant,
       })
       if (promptErr) {
         logger.error(`Loop: failed to send rewind continuation prompt for ${loopName}`, promptErr)

src/loop/service.ts

@@ -111,6 +111,8 @@ export function rowToLoopState(row: LoopRow, large: LoopLargeFields | null): Loo
     currentSectionIndex: row.currentSectionIndex,
     totalSections: row.totalSections,
     finalAuditDone: row.finalAuditDone === 1,
+    executionVariant: row.executionVariant ?? undefined,
+    auditorVariant: row.auditorVariant ?? undefined,
   }
 }
 
@@ -156,6 +158,8 @@ export function createLoopService(
       currentSectionIndex: state.currentSectionIndex,
       totalSections: state.totalSections,
       finalAuditDone: state.finalAuditDone ? 1 : 0,
+      executionVariant: state.executionVariant ?? null,
+      auditorVariant: state.auditorVariant ?? null,
     }
   }
 

src/loop/state.ts

@@ -29,6 +29,8 @@ interface LoopStateBase {
   currentSectionIndex: number
   totalSections: number
   finalAuditDone: boolean
+  executionVariant?: string
+  auditorVariant?: string
 }
 
 export interface CodingState extends LoopStateBase {
@@ -76,6 +78,8 @@ export function loopRowToState(row: LoopRow, large?: LoopLargeFields | null): Lo
     currentSectionIndex: row.currentSectionIndex,
     totalSections: row.totalSections,
     finalAuditDone: row.finalAuditDone === 1,
+    executionVariant: row.executionVariant ?? undefined,
+    auditorVariant: row.auditorVariant ?? undefined,
   }
 
   switch (row.phase) {
@@ -117,5 +121,7 @@ export function loopStateToRow(state: LoopState, projectId: string): Omit<LoopRo
     currentSectionIndex: state.currentSectionIndex,
     totalSections: state.totalSections,
     finalAuditDone: state.finalAuditDone ? 1 : 0,
+    executionVariant: state.executionVariant ?? null,
+    auditorVariant: state.auditorVariant ?? null,
   }
 }