@kkelly-offical/kkcode 0.1.7 → 0.2.3-preview.1

package/src/session/longagent-plan.mjs

@@ -1,365 +1,376 @@
-import { newId } from "../core/types.mjs"
-import { processTurnLoop } from "./loop.mjs"
-import { stripFence, parseJsonLoose } from "./longagent-utils.mjs"
-
-function normalizeFileList(value) {
-  if (!Array.isArray(value)) return []
-  return [...new Set(value
-    .map((v) => String(v || "").trim())
-    .filter(Boolean)
-    .slice(0, 80))]
-}
-
-function normalizeStringList(value) {
-  if (!Array.isArray(value)) return []
-  return value
-    .map((v) => String(v || "").trim())
-    .filter(Boolean)
-    .slice(0, 50)
-}
-
-function normalizeTask(task, stageId, defaults = {}) {
-  const baseId = String(task?.taskId || task?.id || "").trim()
-  const taskId = baseId || `${stageId}_task_${newId("t").slice(-6)}`
-  const prompt = String(task?.prompt || "").trim()
-  if (!prompt) return null
-  const timeoutMs = Number(task?.timeoutMs || defaults.timeoutMs || 600000)
-  const maxRetries = Number(task?.maxRetries ?? defaults.maxRetries ?? 2)
-  const complexity = ["low", "medium", "high"].includes(task?.complexity) ? task.complexity : "medium"
-  const dependsOn = normalizeStringList(task?.dependsOn || [])
-  return {
-    taskId,
-    prompt,
-    subagentType: task?.subagentType ? String(task.subagentType) : undefined,
-    category: task?.category ? String(task.category) : undefined,
-    plannedFiles: normalizeFileList(task?.plannedFiles),
-    acceptance: normalizeStringList(task?.acceptance),
-    dependsOn,
-    complexity,
-    timeoutMs: Number.isFinite(timeoutMs) && timeoutMs >= 1000 ? timeoutMs : 600000,
-    maxRetries: Number.isFinite(maxRetries) && maxRetries >= 0 ? maxRetries : 2
-  }
-}
-
-function normalizeStage(stage, defaults = {}, idx = 0) {
-  const stageId = String(stage?.stageId || stage?.id || `stage_${idx + 1}`).trim() || `stage_${idx + 1}`
-  const name = String(stage?.name || `Stage ${idx + 1}`).trim() || `Stage ${idx + 1}`
-  const tasks = Array.isArray(stage?.tasks)
-    ? stage.tasks.map((t) => normalizeTask(t, stageId, defaults)).filter(Boolean)
-    : []
-  return {
-    stageId,
-    name,
-    passRule: "all_success",
-    tasks
-  }
-}
-
-export function defaultStagePlan(objective, defaults = {}) {
-  const stageId = "stage_1"
-  return {
-    planId: newId("plan"),
-    objective: String(objective || "").trim(),
-    stages: [
-      {
-        stageId,
-        name: "Execution",
-        passRule: "all_success",
-        tasks: [
-          {
-            taskId: `${stageId}_task_1`,
-            prompt: String(objective || "").trim(),
-            plannedFiles: [],
-            acceptance: ["Task objective is fully usable"],
-            timeoutMs: Number(defaults.timeoutMs || 600000),
-            maxRetries: Number(defaults.maxRetries ?? 2)
-          }
-        ]
-      }
-    ]
-  }
-}
-
-export function validateAndNormalizeStagePlan(input, { objective = "", defaults = {} } = {}) {
-  if (!input || typeof input !== "object") {
-    return { plan: defaultStagePlan(objective, defaults), errors: ["plan is not object"] }
-  }
-
-  const plan = {
-    planId: String(input.planId || newId("plan")),
-    objective: String(input.objective || objective || "").trim(),
-    stages: Array.isArray(input.stages)
-      ? input.stages.map((s, idx) => normalizeStage(s, defaults, idx))
-      : []
-  }
-
-  const errors = []
-  if (!plan.objective) errors.push("objective is empty")
-  if (!plan.stages.length) errors.push("no stages")
-  for (const stage of plan.stages) {
-    if (!stage.tasks.length) errors.push(`stage "${stage.stageId}" has no tasks`)
-    // Early file isolation check — detect overlapping file ownership at plan time
-    const ownership = new Map()
-    for (const task of stage.tasks) {
-      for (const file of task.plannedFiles || []) {
-        if (ownership.has(file)) {
-          errors.push(`stage "${stage.stageId}": file "${file}" claimed by "${ownership.get(file)}" and "${task.taskId}"`)
-        } else {
-          ownership.set(file, task.taskId)
-        }
-      }
-    }
-  }
-
-  // Cross-stage dependency check: later stages should not own files already owned by earlier stages
-  const globalOwnership = new Map()
-  for (const stage of plan.stages) {
-    for (const task of stage.tasks) {
-      for (const file of task.plannedFiles || []) {
-        if (globalOwnership.has(file)) {
-          const prev = globalOwnership.get(file)
-          errors.push(`file "${file}" appears in stage "${prev}" and "${stage.stageId}" — split into dependency chain or deduplicate`)
-        } else {
-          globalOwnership.set(file, stage.stageId)
-        }
-      }
-    }
-  }
-
-  // Quality score: penalize tasks with no files or no acceptance criteria
-  let qualityScore = 100
-  let totalTasks = 0
-  for (const stage of plan.stages) {
-    for (const task of stage.tasks) {
-      totalTasks += 1
-      if (!task.plannedFiles.length) qualityScore -= 15
-      if (!task.acceptance.length) qualityScore -= 10
-    }
-  }
-  qualityScore = Math.max(0, Math.min(100, qualityScore))
-
-  if (errors.length) {
-    return {
-      plan: defaultStagePlan(objective || plan.objective, defaults),
-      errors,
-      qualityScore: 0
-    }
-  }
-  return { plan, errors: [], qualityScore }
-}
-
-export async function runIntakeDialogue({
-  objective,
-  model,
-  providerType,
-  sessionId,
-  configState,
-  baseUrl = null,
-  apiKeyEnv = null,
-  agent = null,
-  signal = null,
-  maxRounds = 6
-}) {
-  const rounds = Math.max(1, Number(maxRounds || 6))
-  const transcript = []
-  let summary = ""
-
-  for (let i = 1; i <= rounds; i++) {
-    const prompt = [
-      "You are the INTAKE ANALYST for a production-grade parallel coding pipeline.",
-      "Multiple independent sub-agents will execute tasks IN PARALLEL — they cannot communicate with each other during execution.",
-      "Your job: resolve ALL ambiguities NOW so that parallel agents produce compatible, integrable code.",
-      "",
-      "## Analysis Categories (address ALL that apply)",
-      "",
-      "### 1. SCOPE & BOUNDARIES",
-      "- What files/modules/directories are IN scope for modification?",
-      "- What existing code must NOT be changed (public APIs, shared interfaces, config schemas)?",
-      "- Are there related features that should be explicitly excluded to prevent scope creep?",
-      "",
-      "### 2. TECHNOLOGY & PATTERNS",
-      "- What language version, runtime, and framework constraints apply?",
-      "- What existing patterns in the codebase MUST be followed (error handling, logging, naming, async style)?",
-      "- What existing utilities/helpers MUST be reused instead of reimplemented?",
-      "- Are there specific libraries to use or avoid?",
-      "",
-      "### 3. INTERFACE CONTRACTS",
-      "- What are the exact function signatures, parameter types, and return types?",
-      "- What data schemas are involved (DB models, API payloads, config shapes)?",
-      "- What error types should be thrown/returned and how should callers handle them?",
-      "- What events/hooks/callbacks are part of the contract?",
-      "",
-      "### 4. QUALITY CONSTRAINTS",
-      "- What test coverage is expected (unit, integration, e2e)?",
-      "- What backward compatibility guarantees must be maintained?",
-      "- Are there performance budgets (latency, memory, bundle size)?",
-      "- What security considerations apply (input validation, auth, secrets)?",
-      "",
-      "### 5. DEPENDENCY ORDER",
-      "- Which components must exist before others can be built?",
-      "- Which components are independent and can be built in parallel?",
-      "- Are there shared types/interfaces that must be defined first?",
-      "",
-      "## Output Rules",
-      "- For each question, provide your BEST ASSUMPTION as the answer based on the objective and codebase context.",
-      "- Be specific: 'use HS256 JWT with 24h expiry' not 'implement auth'.",
-      "- Set enough=true ONLY when you have sufficient clarity to generate a concrete file-level execution plan.",
-      "",
-      "Return STRICT JSON (no markdown wrapping):",
-      `{"enough":boolean,"summary":"technical summary with ALL resolved assumptions and concrete decisions","qa":[{"q":"specific question","a":"concrete answer with implementation detail"}]}`,
-      "",
-      `Round: ${i}/${rounds}`,
-      `Objective: ${objective}`,
-      summary ? `Previous summary: ${summary}` : ""
-    ].filter(Boolean).join("\n")
-
-    const out = await processTurnLoop({
-      prompt,
-      mode: "ask",
-      model,
-      providerType,
-      sessionId,
-      configState,
-      baseUrl,
-      apiKeyEnv,
-      agent,
-      signal,
-      output: { write: () => {} },
-      allowQuestion: true
-    })
-
-    const parsed = parseJsonLoose(out.reply)
-    if (parsed && Array.isArray(parsed.qa)) {
-      for (const item of parsed.qa.slice(0, 10)) {
-        const q = String(item?.q || "").trim()
-        const a = String(item?.a || "").trim()
-        if (q || a) transcript.push({ q, a })
-      }
-      summary = String(parsed.summary || "").trim() || summary
-      const enough = Boolean(parsed.enough)
-      if (enough && i >= 2) break
-      continue
-    }
-
-    const fallbackLine = String(out.reply || "").trim()
-    if (fallbackLine) {
-      transcript.push({ q: `Round ${i} synthesis`, a: fallbackLine })
-      summary = fallbackLine
-    }
-    if (i >= 2) break
-  }
-
-  return {
-    transcript,
-    summary: summary || String(objective || "").trim()
-  }
-}
-
-export async function buildStagePlan({
-  objective,
-  intakeSummary = "",
-  model,
-  providerType,
-  sessionId,
-  configState,
-  baseUrl = null,
-  apiKeyEnv = null,
-  agent = null,
-  signal = null,
-  defaults = {}
-}) {
-  const plannerPrompt = [
-    "You are the EXECUTION PLANNER for a production-grade parallel coding pipeline.",
-    "Generate a stage plan that will be executed by independent sub-agents running in parallel.",
-    "",
-    "Return STRICT JSON ONLY (no markdown wrapping, no explanation) with this schema:",
-    '{"planId":"...","objective":"...","stages":[{"stageId":"...","name":"...","passRule":"all_success","tasks":[{"taskId":"...","prompt":"...","plannedFiles":["..."],"acceptance":["..."],"timeoutMs":600000,"maxRetries":2,"complexity":"low|medium|high"}]}]}',
-    "",
-    "## Execution Model (understand this before planning)",
-    "",
-    "1. An ARCHITECT agent creates ALL plannedFiles as scaffolds with detailed inline comments (signatures + logic descriptions, no implementation)",
-    "2. Each task is assigned to an INDEPENDENT sub-agent that reads the scaffold and implements real code",
-    "3. Sub-agents within the same stage run IN PARALLEL — they CANNOT see each other's work or communicate",
-    "4. Stages run SEQUENTIALLY — stage N+1 starts only after ALL tasks in stage N succeed",
-    "5. If a task fails, it is retried up to maxRetries times. If all retries fail, the stage fails.",
-    "",
-    "## File Assignment Rules (violations cause runtime failures)",
-    "",
-    "- Files that import from each other MUST be in the SAME task (parallel agents cannot resolve cross-task imports)",
-    "- A module and its test file MUST be in the same task (tests need the implementation to exist)",
-    "- A component and its type definitions MUST be in the same task",
-    "- Each task should own 2-8 files. 1 file → merge with a related task. >10 files → split into smaller tasks",
-    "- NO file may appear in multiple tasks within a stage or across stages",
-    "- Shared types/interfaces that multiple tasks depend on → put in stage 1 as a dedicated 'shared types' task",
-    "",
-    "## Stage Ordering Strategy",
-    "",
-    "- Stage 1: Shared types, interfaces, config schemas, utility functions (things others import from)",
-    "- Stage 2: Core business logic modules (independent of each other, depend on stage 1)",
-    "- Stage 3: Integration layer (routes, controllers, orchestrators that wire stage 2 modules together)",
-    "- Stage 4: Tests, validation, documentation (if not already co-located with their modules)",
-    "- Minimize stage count — only create a new stage when there is a REAL import dependency from a previous stage",
-    "- If all tasks are independent, use a SINGLE stage with multiple parallel tasks",
-    "",
-    "## Task Prompt Requirements (this is what the sub-agent sees)",
-    "",
-    "The sub-agent's ONLY context is: (1) your task prompt, (2) the scaffold file with inline comments, (3) the file ownership list.",
-    "The sub-agent has NO access to the original user objective, blueprint, or other tasks' prompts.",
-    "",
-    "Each task prompt MUST include:",
-    "1. WHAT to implement — specific behavior with concrete details, not vague goals",
-    "2. HOW to implement — key algorithms, data structures, patterns to use",
-    "3. INTEGRATION — what this task exports (function names, signatures), what format other tasks expect",
-    "4. ERROR HANDLING — what errors to throw/catch, how to handle edge cases (null, empty, invalid input)",
-    "5. TESTING — if test files are in this task, what test cases to write (happy path, error cases, edge cases)",
-    "",
-    "BAD prompt: 'implement the auth module'",
-    "GOOD prompt: 'Implement src/auth/jwt.mjs: generateToken(userId, role) returns signed JWT string using HS256 with 24h expiry, reading secret from process.env.JWT_SECRET. Throw AuthError if secret is missing. verifyToken(token) returns decoded payload or throws TokenExpiredError/InvalidTokenError. Export both functions as named exports. Test file: test happy path (valid token roundtrip), expired token, invalid signature, missing secret.'",
-    "",
-    "## Acceptance Criteria Rules",
-    "",
-    "- MUST be machine-verifiable commands or assertions:",
-    "  - 'node --check src/auth/jwt.mjs' (syntax check)",
-    "  - 'node --test test/auth.test.mjs' (test execution)",
-    "  - 'grep -q \"export function generateToken\" src/auth/jwt.mjs' (API existence)",
-    "- NEVER subjective: 'code is clean', 'implementation is correct', 'works as expected'",
-    "- FINAL stage MUST include: 'all modified files parse without syntax errors' AND 'test suite passes'",
-    "",
-    "## Complexity Rating",
-    "",
-    "- low: simple CRUD, config changes, straightforward utility functions (< 100 lines total)",
-    "- medium: business logic with multiple code paths, moderate error handling (100-500 lines)",
-    "- high: complex algorithms, concurrent operations, extensive edge cases (> 500 lines)",
-    "",
-    `## Objective`,
-    objective,
-    intakeSummary ? `\n## Intake Summary\n${intakeSummary}` : ""
-  ].filter(Boolean).join("\n")
-
-  const out = await processTurnLoop({
-    prompt: plannerPrompt,
-    mode: "plan",
-    model,
-    providerType,
-    sessionId,
-    configState,
-    baseUrl,
-    apiKeyEnv,
-    agent,
-    signal,
-    output: { write: () => {} },
-    allowQuestion: true
-  })
-
-  const parsed = parseJsonLoose(out.reply)
-  if (!parsed) {
-    return {
-      plan: defaultStagePlan(objective, defaults),
-      errors: ["planner returned unparseable response — falling back to single-stage plan"],
-      qualityScore: 0,
-      rawReply: out.reply
-    }
-  }
-  const { plan, errors, qualityScore } = validateAndNormalizeStagePlan(parsed, { objective, defaults })
-  return { plan, errors, qualityScore, rawReply: out.reply }
-}
+import { newId } from "../core/types.mjs"
+import { processTurnLoop } from "./loop.mjs"
+import { stripFence, parseJsonLoose } from "./longagent-utils.mjs"
+
+function normalizeFileList(value) {
+  if (!Array.isArray(value)) return []
+  return [...new Set(value
+    .map((v) => String(v || "").trim())
+    .filter(Boolean)
+    .slice(0, 80))]
+}
+
+function normalizeStringList(value) {
+  if (!Array.isArray(value)) return []
+  return value
+    .map((v) => String(v || "").trim())
+    .filter(Boolean)
+    .slice(0, 50)
+}
+
+function normalizeTask(task, stageId, defaults = {}) {
+  const baseId = String(task?.taskId || task?.id || "").trim()
+  const taskId = baseId || `${stageId}_task_${newId("t").slice(-6)}`
+  const prompt = String(task?.prompt || "").trim()
+  if (!prompt) return null
+  const timeoutMs = Number(task?.timeoutMs || defaults.timeoutMs || 600000)
+  const maxRetries = Number(task?.maxRetries ?? defaults.maxRetries ?? 2)
+  const complexity = ["low", "medium", "high"].includes(task?.complexity) ? task.complexity : "medium"
+  const dependsOn = normalizeStringList(task?.dependsOn || [])
+  return {
+    taskId,
+    prompt,
+    subagentType: task?.subagentType ? String(task.subagentType) : undefined,
+    category: task?.category ? String(task.category) : undefined,
+    plannedFiles: normalizeFileList(task?.plannedFiles),
+    acceptance: normalizeStringList(task?.acceptance),
+    dependsOn,
+    complexity,
+    timeoutMs: Number.isFinite(timeoutMs) && timeoutMs >= 1000 ? timeoutMs : 600000,
+    maxRetries: Number.isFinite(maxRetries) && maxRetries >= 0 ? maxRetries : 2
+  }
+}
+
+function normalizeStage(stage, defaults = {}, idx = 0) {
+  const stageId = String(stage?.stageId || stage?.id || `stage_${idx + 1}`).trim() || `stage_${idx + 1}`
+  const name = String(stage?.name || `Stage ${idx + 1}`).trim() || `Stage ${idx + 1}`
+  const tasks = Array.isArray(stage?.tasks)
+    ? stage.tasks.map((t) => normalizeTask(t, stageId, defaults)).filter(Boolean)
+    : []
+  return {
+    stageId,
+    name,
+    passRule: "all_success",
+    tasks
+  }
+}
+
+export function defaultStagePlan(objective, defaults = {}) {
+  const stageId = "stage_1"
+  return {
+    planId: newId("plan"),
+    objective: String(objective || "").trim(),
+    stages: [
+      {
+        stageId,
+        name: "Execution",
+        passRule: "all_success",
+        tasks: [
+          {
+            taskId: `${stageId}_task_1`,
+            prompt: String(objective || "").trim(),
+            plannedFiles: [],
+            acceptance: ["Task objective is fully usable"],
+            timeoutMs: Number(defaults.timeoutMs || 600000),
+            maxRetries: Number(defaults.maxRetries ?? 2)
+          }
+        ]
+      }
+    ]
+  }
+}
+
+export function validateAndNormalizeStagePlan(input, { objective = "", defaults = {} } = {}) {
+  if (!input || typeof input !== "object") {
+    return { plan: defaultStagePlan(objective, defaults), errors: ["plan is not object"] }
+  }
+
+  const plan = {
+    planId: String(input.planId || newId("plan")),
+    objective: String(input.objective || objective || "").trim(),
+    stages: Array.isArray(input.stages)
+      ? input.stages.map((s, idx) => normalizeStage(s, defaults, idx))
+      : []
+  }
+
+  const errors = []
+  if (!plan.objective) errors.push("objective is empty")
+  if (!plan.stages.length) errors.push("no stages")
+  for (const stage of plan.stages) {
+    if (!stage.tasks.length) errors.push(`stage "${stage.stageId}" has no tasks`)
+    // Early file isolation check — detect overlapping file ownership at plan time
+    const ownership = new Map()
+    for (const task of stage.tasks) {
+      for (const file of task.plannedFiles || []) {
+        if (ownership.has(file)) {
+          errors.push(`stage "${stage.stageId}": file "${file}" claimed by "${ownership.get(file)}" and "${task.taskId}"`)
+        } else {
+          ownership.set(file, task.taskId)
+        }
+      }
+    }
+  }
+
+  // Cross-stage dependency check: later stages should not own files already owned by earlier stages
+  // Downgraded to warnings — cross-stage overlap is common (e.g. shared config) and should not block the plan
+  const warnings = []
+  const globalOwnership = new Map()
+  for (const stage of plan.stages) {
+    for (const task of stage.tasks) {
+      for (const file of task.plannedFiles || []) {
+        if (globalOwnership.has(file)) {
+          const prev = globalOwnership.get(file)
+          warnings.push(`file "${file}" appears in stage "${prev}" and "${stage.stageId}"`)
+        } else {
+          globalOwnership.set(file, stage.stageId)
+        }
+      }
+    }
+  }
+
+  // Quality score: penalize tasks with no files or no acceptance criteria
+  let qualityScore = 100
+  let totalTasks = 0
+  for (const stage of plan.stages) {
+    for (const task of stage.tasks) {
+      totalTasks += 1
+      if (!task.plannedFiles.length) qualityScore -= 15
+      if (!task.acceptance.length) qualityScore -= 10
+    }
+  }
+  // Cross-stage file overlap: penalize quality but don't block
+  qualityScore -= warnings.length * 5
+  qualityScore = Math.max(0, Math.min(100, qualityScore))
+
+  if (errors.length) {
+    // If the parsed plan has at least one valid stage with tasks, keep it despite warnings
+    const hasValidStages = plan.stages.some(s => s.tasks.length > 0)
+    if (hasValidStages) {
+      // Filter out empty stages but preserve valid ones
+      plan.stages = plan.stages.filter(s => s.tasks.length > 0)
+      return { plan, errors, warnings, qualityScore: Math.max(0, qualityScore - errors.length * 10) }
+    }
+    return {
+      plan: defaultStagePlan(objective || plan.objective, defaults),
+      errors, warnings,
+      qualityScore: 0
+    }
+  }
+  return { plan, errors: [], warnings, qualityScore }
+}
+
+export async function runIntakeDialogue({
+  objective,
+  model,
+  providerType,
+  sessionId,
+  configState,
+  baseUrl = null,
+  apiKeyEnv = null,
+  agent = null,
+  signal = null,
+  maxRounds = 6
+}) {
+  const rounds = Math.max(1, Number(maxRounds || 6))
+  const transcript = []
+  let summary = ""
+
+  for (let i = 1; i <= rounds; i++) {
+    const prompt = [
+      "You are the INTAKE ANALYST for a production-grade parallel coding pipeline.",
+      "Multiple independent sub-agents will execute tasks IN PARALLEL — they cannot communicate with each other during execution.",
+      "Your job: resolve ALL ambiguities NOW so that parallel agents produce compatible, integrable code.",
+      "",
+      "## Analysis Categories (address ALL that apply)",
+      "",
+      "### 1. SCOPE & BOUNDARIES",
+      "- What files/modules/directories are IN scope for modification?",
+      "- What existing code must NOT be changed (public APIs, shared interfaces, config schemas)?",
+      "- Are there related features that should be explicitly excluded to prevent scope creep?",
+      "",
+      "### 2. TECHNOLOGY & PATTERNS",
+      "- What language version, runtime, and framework constraints apply?",
+      "- What existing patterns in the codebase MUST be followed (error handling, logging, naming, async style)?",
+      "- What existing utilities/helpers MUST be reused instead of reimplemented?",
+      "- Are there specific libraries to use or avoid?",
+      "",
+      "### 3. INTERFACE CONTRACTS",
+      "- What are the exact function signatures, parameter types, and return types?",
+      "- What data schemas are involved (DB models, API payloads, config shapes)?",
+      "- What error types should be thrown/returned and how should callers handle them?",
+      "- What events/hooks/callbacks are part of the contract?",
+      "",
+      "### 4. QUALITY CONSTRAINTS",
+      "- What test coverage is expected (unit, integration, e2e)?",
+      "- What backward compatibility guarantees must be maintained?",
+      "- Are there performance budgets (latency, memory, bundle size)?",
+      "- What security considerations apply (input validation, auth, secrets)?",
+      "",
+      "### 5. DEPENDENCY ORDER",
+      "- Which components must exist before others can be built?",
+      "- Which components are independent and can be built in parallel?",
+      "- Are there shared types/interfaces that must be defined first?",
+      "",
+      "## Output Rules",
+      "- For each question, provide your BEST ASSUMPTION as the answer based on the objective and codebase context.",
+      "- Be specific: 'use HS256 JWT with 24h expiry' not 'implement auth'.",
+      "- Set enough=true ONLY when you have sufficient clarity to generate a concrete file-level execution plan.",
+      "",
+      "Return STRICT JSON (no markdown wrapping):",
+      `{"enough":boolean,"summary":"technical summary with ALL resolved assumptions and concrete decisions","qa":[{"q":"specific question","a":"concrete answer with implementation detail"}]}`,
+      "",
+      `Round: ${i}/${rounds}`,
+      `Objective: ${objective}`,
+      summary ? `Previous summary: ${summary}` : ""
+    ].filter(Boolean).join("\n")
+
+    const out = await processTurnLoop({
+      prompt,
+      mode: "assistant",
+      model,
+      providerType,
+      sessionId,
+      configState,
+      baseUrl,
+      apiKeyEnv,
+      agent,
+      signal,
+      output: { write: () => {} },
+      allowQuestion: true
+    })
+
+    const parsed = parseJsonLoose(out.reply)
+    if (parsed && Array.isArray(parsed.qa)) {
+      for (const item of parsed.qa.slice(0, 10)) {
+        const q = String(item?.q || "").trim()
+        const a = String(item?.a || "").trim()
+        if (q || a) transcript.push({ q, a })
+      }
+      summary = String(parsed.summary || "").trim() || summary
+      const enough = Boolean(parsed.enough)
+      if (enough && i >= 2) break
+      continue
+    }
+
+    const fallbackLine = String(out.reply || "").trim()
+    if (fallbackLine) {
+      transcript.push({ q: `Round ${i} synthesis`, a: fallbackLine })
+      summary = fallbackLine
+    }
+    if (i >= 2) break
+  }
+
+  return {
+    transcript,
+    summary: summary || String(objective || "").trim()
+  }
+}
+
+export async function buildStagePlan({
+  objective,
+  intakeSummary = "",
+  model,
+  providerType,
+  sessionId,
+  configState,
+  baseUrl = null,
+  apiKeyEnv = null,
+  agent = null,
+  signal = null,
+  defaults = {}
+}) {
+  const plannerPrompt = [
+    "You are the EXECUTION PLANNER for a production-grade parallel coding pipeline.",
+    "Generate a stage plan that will be executed by independent sub-agents running in parallel.",
+    "",
+    "Return STRICT JSON ONLY (no markdown wrapping, no explanation) with this schema:",
+    '{"planId":"...","objective":"...","stages":[{"stageId":"...","name":"...","passRule":"all_success","tasks":[{"taskId":"...","prompt":"...","plannedFiles":["..."],"acceptance":["..."],"timeoutMs":600000,"maxRetries":2,"complexity":"low|medium|high"}]}]}',
+    "",
+    "## Execution Model (understand this before planning)",
+    "",
+    "1. An ARCHITECT agent creates ALL plannedFiles as scaffolds with detailed inline comments (signatures + logic descriptions, no implementation)",
+    "2. Each task is assigned to an INDEPENDENT sub-agent that reads the scaffold and implements real code",
+    "3. Sub-agents within the same stage run IN PARALLEL — they CANNOT see each other's work or communicate",
+    "4. Stages run SEQUENTIALLY — stage N+1 starts only after ALL tasks in stage N succeed",
+    "5. If a task fails, it is retried up to maxRetries times. If all retries fail, the stage fails.",
+    "",
+    "## File Assignment Rules (violations cause runtime failures)",
+    "",
+    "- Files that import from each other MUST be in the SAME task (parallel agents cannot resolve cross-task imports)",
+    "- A module and its test file MUST be in the same task (tests need the implementation to exist)",
+    "- A component and its type definitions MUST be in the same task",
+    "- Each task should own 2-8 files. 1 file → merge with a related task. >10 files → split into smaller tasks",
+    "- NO file may appear in multiple tasks within a stage or across stages",
+    "- Shared types/interfaces that multiple tasks depend on → put in stage 1 as a dedicated 'shared types' task",
+    "",
+    "## Stage Ordering Strategy",
+    "",
+    "- Stage 1: Shared types, interfaces, config schemas, utility functions (things others import from)",
+    "- Stage 2: Core business logic modules (independent of each other, depend on stage 1)",
+    "- Stage 3: Integration layer (routes, controllers, orchestrators that wire stage 2 modules together)",
+    "- Stage 4: Tests, validation, documentation (if not already co-located with their modules)",
+    "- Minimize stage count — only create a new stage when there is a REAL import dependency from a previous stage",
+    "- If all tasks are independent, use a SINGLE stage with multiple parallel tasks",
+    "",
+    "## Task Prompt Requirements (this is what the sub-agent sees)",
+    "",
+    "The sub-agent's ONLY context is: (1) your task prompt, (2) the scaffold file with inline comments, (3) the file ownership list.",
+    "The sub-agent has NO access to the original user objective, blueprint, or other tasks' prompts.",
+    "",
+    "Each task prompt MUST include:",
+    "1. WHAT to implement — specific behavior with concrete details, not vague goals",
+    "2. HOW to implement — key algorithms, data structures, patterns to use",
+    "3. INTEGRATION — what this task exports (function names, signatures), what format other tasks expect",
+    "4. ERROR HANDLING — what errors to throw/catch, how to handle edge cases (null, empty, invalid input)",
+    "5. TESTING — if test files are in this task, what test cases to write (happy path, error cases, edge cases)",
+    "",
+    "BAD prompt: 'implement the auth module'",
+    "GOOD prompt: 'Implement src/auth/jwt.mjs: generateToken(userId, role) returns signed JWT string using HS256 with 24h expiry, reading secret from process.env.JWT_SECRET. Throw AuthError if secret is missing. verifyToken(token) returns decoded payload or throws TokenExpiredError/InvalidTokenError. Export both functions as named exports. Test file: test happy path (valid token roundtrip), expired token, invalid signature, missing secret.'",
+    "",
+    "## Acceptance Criteria Rules",
+    "",
+    "- MUST be machine-verifiable commands or assertions:",
+    "  - 'node --check src/auth/jwt.mjs' (syntax check)",
+    "  - 'node --test test/auth.test.mjs' (test execution)",
+    "  - 'grep -q \"export function generateToken\" src/auth/jwt.mjs' (API existence)",
+    "- NEVER subjective: 'code is clean', 'implementation is correct', 'works as expected'",
+    "- FINAL stage MUST include: 'all modified files parse without syntax errors' AND 'test suite passes'",
+    "",
+    "## Complexity Rating",
+    "",
+    "- low: simple CRUD, config changes, straightforward utility functions (< 100 lines total)",
+    "- medium: business logic with multiple code paths, moderate error handling (100-500 lines)",
+    "- high: complex algorithms, concurrent operations, extensive edge cases (> 500 lines)",
+    "",
+    `## Objective`,
+    objective,
+    intakeSummary ? `\n## Intake Summary\n${intakeSummary}` : ""
+  ].filter(Boolean).join("\n")
+
+  const out = await processTurnLoop({
+    prompt: plannerPrompt,
+    mode: "plan",
+    model,
+    providerType,
+    sessionId,
+    configState,
+    baseUrl,
+    apiKeyEnv,
+    agent,
+    signal,
+    output: { write: () => {} },
+    allowQuestion: true
+  })
+
+  const parsed = parseJsonLoose(out.reply)
+  if (!parsed) {
+    return {
+      plan: defaultStagePlan(objective, defaults),
+      errors: ["planner returned unparseable response — falling back to single-stage plan"],
+      qualityScore: 0,
+      rawReply: out.reply
+    }
+  }
+  const { plan, errors, qualityScore } = validateAndNormalizeStagePlan(parsed, { objective, defaults })
+  return { plan, errors, qualityScore, rawReply: out.reply }
+}