fix: resolve 3 bugs - subagent model override, empty plan completion, deep task refusal

- #2741: Pass inheritedModel as fallback in subagent-resolver when user hasn't
  configured an override, ensuring custom provider models take priority
- #2648: Fix getPlanProgress to treat plans with 0 checkboxes as incomplete
  instead of complete (total > 0 && completed === total)
- #2779: Relax Hephaestus single-task guard to accept multi-step sub-tasks
  from Atlas delegation, only rejecting genuinely independent tasks

Fixes #2741, fixes #2648, fixes #2779

src/agents/hephaestus/gpt-5-3-codex.ts

@@ -152,7 +152,19 @@ Asking the user is the LAST resort after exhausting creative alternatives.
 - "I'll do X" / "I recommend X" then ending turn → You COMMITTED to X. DO X NOW before ending.
 - Explaining findings without acting on them → ACT on your findings immediately.
 
-**CORRECT:**
+|**CORRECT:**
+|- Keep going until COMPLETELY done
+|- Run verification (lint, tests, build) WITHOUT asking
+|- Make decisions. Course-correct only on CONCRETE failure
+|- Note assumptions in final message, not as questions mid-work
+|- Need context? Fire explore/librarian in background IMMEDIATELY — continue only with non-overlapping work while they search
+|- User asks "did you do X?" and you didn't → Acknowledge briefly, DO X immediately
+|- User asks a question implying work → Answer briefly, DO the implied work in the same turn
+|- You wrote a plan in your response → EXECUTE the plan before ending turn — plans are starting lines, not finish lines
+|
+### Task Scope Clarification
+|
+You handle multi-step sub-tasks of a SINGLE GOAL. What you receive is ONE goal that may require multiple steps to complete — this is your primary use case. Only reject when given MULTIPLE INDEPENDENT goals in one request.
 - Keep going until COMPLETELY done
 - Run verification (lint, tests, build) WITHOUT asking
 - Make decisions. Course-correct only on CONCRETE failure

src/agents/hephaestus/gpt-5-4.ts

@@ -111,7 +111,21 @@ When blocked: try a different approach → decompose the problem → challenge a
 - "I'll do X" / "I recommend X" then ending turn → You COMMITTED to X. DO X NOW before ending.
 - Explaining findings without acting on them → ACT on your findings immediately.
 
-**CORRECT:**
+|**CORRECT:**
+|- Keep going until COMPLETELY done
+|- Run verification (lint, tests, build) WITHOUT asking
+|- Make decisions. Course-correct only on CONCRETE failure
+|- Note assumptions in final message, not as questions mid-work
+|- Need context? Fire explore/librarian in background IMMEDIATELY — continue only with non-overlapping work while they search
+|- User asks "did you do X?" and you didn't → Acknowledge briefly, DO X immediately
+|- User asks a question implying work → Answer briefly, DO the implied work in the same turn
+|- You wrote a plan in your response → EXECUTE the plan before ending turn — plans are starting lines, not finish lines
+|
+### Task Scope Clarification
+|
+You handle multi-step sub-tasks of a SINGLE GOAL. What you receive is ONE goal that may require multiple steps to complete — this is your primary use case. Only reject when given MULTIPLE INDEPENDENT goals in one request.
+|
+## Hard Constraints
 - Keep going until COMPLETELY done
 - Run verification (lint, tests, build) WITHOUT asking
 - Make decisions. Course-correct only on CONCRETE failure

src/agents/hephaestus/gpt.ts

@@ -105,7 +105,18 @@ Asking the user is the LAST resort after exhausting creative alternatives.
 - "I noticed Y, should I fix it?" → FIX IT OR NOTE IN FINAL MESSAGE.
 - Stopping after partial implementation → 100% OR NOTHING.
 
-**CORRECT:**
+|**CORRECT:**
+|- Keep going until COMPLETELY done
+|- Run verification (lint, tests, build) WITHOUT asking
+|- Make decisions. Course-correct only on CONCRETE failure
+|- Note assumptions in final message, not as questions mid-work
+|- Need context? Fire explore/librarian in background IMMEDIATELY — continue only with non-overlapping work while they search
+|
+### Task Scope Clarification
+|
+You handle multi-step sub-tasks of a SINGLE GOAL. What you receive is ONE goal that may require multiple steps to complete — this is your primary use case. Only reject when given MULTIPLE INDEPENDENT goals in one request.
+|
+## Hard Constraints
 - Keep going until COMPLETELY done
 - Run verification (lint, tests, build) WITHOUT asking
 - Make decisions. Course-correct only on CONCRETE failure

src/features/background-agent/spawner.ts

@@ -135,7 +135,9 @@ export async function startTask(
   promptWithModelSuggestionRetry(client, {
     path: { id: sessionID },
     body: {
-      agent: input.agent,
+      // When a model is explicitly provided, omit the agent name so opencode's
+      // built-in agent fallback chain does not override the user-specified model.
+      ...(launchModel ? {} : { agent: input.agent }),
       ...(launchModel ? { model: launchModel } : {}),
       ...(launchVariant ? { variant: launchVariant } : {}),
       system: input.skillContent,
@@ -220,7 +222,9 @@ export async function resumeTask(
   client.session.promptAsync({
     path: { id: task.sessionID },
     body: {
-      agent: task.agent,
+      // When a model is explicitly provided, omit the agent name so opencode's
+      // built-in agent fallback chain does not override the user-specified model.
+      ...(resumeModel ? {} : { agent: task.agent }),
       ...(resumeModel ? { model: resumeModel } : {}),
       ...(resumeVariant ? { variant: resumeVariant } : {}),
       tools: {

src/features/boulder-state/storage.test.ts

@@ -481,7 +481,7 @@ describe("boulder-state", () => {
       expect(progress.isComplete).toBe(true)
     })
 
-    test("should return isComplete true for empty plan", () => {
+    test("should return isComplete false for empty plan", () => {
       // given - plan with no checkboxes
       const planPath = join(TEST_DIR, "empty-plan.md")
       writeFileSync(planPath, "# Plan\nNo tasks here")
@@ -491,7 +491,7 @@ describe("boulder-state", () => {
 
       // then
       expect(progress.total).toBe(0)
-      expect(progress.isComplete).toBe(true)
+      expect(progress.isComplete).toBe(false)
     })
 
     test("should handle non-existent file", () => {

src/features/boulder-state/storage.ts

@@ -186,7 +186,7 @@ export function getPlanProgress(planPath: string): PlanProgress {
     return {
       total,
       completed,
-      isComplete: total === 0 || completed === total,
+      isComplete: total > 0 && completed === total,
     }
   } catch {
     return { total: 0, completed: 0, isComplete: true }

src/hooks/atlas/system-reminder-templates.ts

@@ -218,21 +218,31 @@ ${createSystemDirective(SystemDirectiveTypes.SINGLE_TASK_ONLY)}
 
 **STOP. READ THIS BEFORE PROCEEDING.**
 
-If you were NOT given **exactly ONE atomic task**, you MUST:
+If you were given **multiple genuinely independent goals** (unrelated tasks, parallel workstreams, separate features), you MUST:
 1. **IMMEDIATELY REFUSE** this request
-2. **DEMAND** the orchestrator provide a single, specific task
+2. **DEMAND** the orchestrator provide a single goal
 
-**Your response if multiple tasks detected:**
-> "I refuse to proceed. You provided multiple tasks. An orchestrator's impatience destroys work quality.
+**What counts as multiple independent tasks (REFUSE):**
+- "Implement feature A. Also, add feature B."
+- "Fix bug X. Then refactor module Y. Also update the docs."
+- Multiple unrelated changes bundled into one request
+
+**What is a single task with sequential steps (PROCEED):**
+- A single goal broken into numbered steps (e.g., "Implement X by: 1. finding files, 2. adding logic, 3. writing tests")
+- Multi-step context where all steps serve ONE objective
+- Orchestrator-provided context explaining approach for a single deliverable
+
+**Your response if genuinely independent tasks are detected:**
+> "I refuse to proceed. You provided multiple independent tasks. Each task needs full attention.
 > 
-> PROVIDE EXACTLY ONE TASK. One file. One change. One verification.
+> PROVIDE EXACTLY ONE GOAL. One deliverable. One clear outcome.
 > 
-> Your rushing will cause: incomplete work, missed edge cases, broken tests, wasted context."
+> Batching unrelated tasks causes: incomplete work, missed edge cases, broken tests, wasted context."
 
 **WARNING TO ORCHESTRATOR:**
-- Your hasty batching RUINS deliverables
-- Each task needs FULL attention and PROPER verification  
-- Batch delegation = sloppy work = rework = wasted tokens
+- Bundling unrelated tasks RUINS deliverables
+- Each independent goal needs FULL attention and PROPER verification
+- Batch delegation of separate concerns = sloppy work = rework = wasted tokens
 
-**REFUSE multi-task requests. DEMAND single-task clarity.**
+**REFUSE genuinely multi-task requests. ALLOW single-goal multi-step workflows.**
 `

src/tools/delegate-task/constants.ts

@@ -261,12 +261,16 @@ You are NOT an interactive assistant. You are an autonomous problem-solver.
 4. DO NOT ask clarifying questions - the goal is already defined
 
 **Autonomous executor mindset**:
-- You receive a GOAL, not step-by-step instructions
+- You receive a GOAL. When the goal includes numbered steps or phases, treat them as one atomic task broken into sub-steps - NOT as separate independent tasks.
 - Figure out HOW to achieve the goal yourself
 - Thorough research before any action
 - Fix hairy problems that require deep understanding
 - Work independently without frequent check-ins
 
+**Single vs. multi-step context**:
+- Sub-steps of ONE goal (e.g., "Step 1: analyze X, Step 2: implement Y, Step 3: test Z" for a single feature) = execute all steps, they are phases of one atomic task.
+- Genuinely independent tasks (e.g., "Task A: refactor module X" AND "Task B: fix unrelated bug Y") = flag and refuse, require separate delegations.
+
 **Approach**:
 - Explore extensively, understand deeply, then act decisively
 - Prefer comprehensive solutions over quick patches

src/tools/delegate-task/subagent-resolver.ts

@@ -17,7 +17,8 @@ export async function resolveSubagentExecution(
   args: DelegateTaskArgs,
   executorCtx: ExecutorContext,
   parentAgent: string | undefined,
-  categoryExamples: string
+  categoryExamples: string,
+  inheritedModel?: string
 ): Promise<{ agentToUse: string; categoryModel: { providerID: string; modelID: string; variant?: string } | undefined; fallbackChain?: FallbackEntry[]; error?: string }> {
   const { client, agentOverrides, userCategories } = executorCtx
 
@@ -116,7 +117,7 @@ Create the work plan directly - that's your job as the planning agent.`,
         : undefined
 
       const resolution = resolveModelForDelegateTask({
-        userModel: agentOverride?.model,
+        userModel: agentOverride?.model ?? inheritedModel,
         userFallbackModels: normalizedAgentFallbackModels,
         categoryDefaultModel: matchedAgentModelStr,
         fallbackChain: agentRequirement?.fallbackChain,

src/tools/delegate-task/sync-prompt-sender.ts

@@ -56,7 +56,9 @@ export async function sendSyncPrompt(
   const promptArgs = {
     path: { id: input.sessionID },
     body: {
-      agent: input.agentToUse,
+      // When a custom model is configured, omit the agent name so opencode's
+      // built-in agent fallback chain does not override the user-specified model.
+      ...(input.categoryModel ? {} : { agent: input.agentToUse }),
       system: input.systemContent,
       tools,
       parts: [createInternalAgentTextPart(effectivePrompt)],

src/tools/delegate-task/tools.ts

@@ -226,7 +226,7 @@ export function createDelegateTask(options: DelegateTaskToolOptions): ToolDefini
           return executeUnstableAgentTask(args, ctx, options, parentContext, agentToUse, categoryModel, systemContent, actualModel)
         }
       } else {
-        const resolution = await resolveSubagentExecution(args, options, parentContext.agent, categoryExamples)
+        const resolution = await resolveSubagentExecution(args, options, parentContext.agent, categoryExamples, inheritedModel)
         if (resolution.error) {
           return resolution.error
         }