gsd-pi 2.68.1-dev.c1497ab → 2.68.1

package/src/resources/extensions/gsd/guided-flow.ts

@@ -53,8 +53,25 @@ import {
   runPreparation,
   formatCodebaseBrief,
   formatPriorContextBrief,
+  formatEcosystemBrief,
+  type PreparationResult,
 } from "./preparation.js";
 
+// ─── Preparation result storage ─────────────────────────────────────────────
+// Stores the most recent preparation result for injection into discuss prompts.
+// S02 will consume this when building the prepared discussion prompt.
+let lastPreparationResult: PreparationResult | null = null;
+
+/** Get the most recent preparation result (for S02 prompt building). */
+export function getLastPreparationResult(): PreparationResult | null {
+  return lastPreparationResult;
+}
+
+/** Clear the preparation result (called after discussion completes). */
+export function clearPreparationResult(): void {
+  lastPreparationResult = null;
+}
+
 // ─── Re-exports (preserve public API for existing importers) ────────────────
 export {
   MILESTONE_ID_RE, generateMilestoneSuffix, nextMilestoneId,
@@ -410,7 +427,7 @@ function resolveAvailableModel<T extends { id: string; provider: string }>(
  * Build the discuss-and-plan prompt for a new milestone.
  * Used by all three "new milestone" paths (first ever, no active, all complete).
  */
-function buildDiscussPrompt(nextId: string, preamble: string, _basePath: string, preparationContext?: string): string {
+function buildDiscussPrompt(nextId: string, preamble: string, _basePath: string): string {
   const milestoneRel = `.gsd/milestones/${nextId}`;
   const inlinedTemplates = [
     inlineTemplate("project", "Project"),
@@ -422,7 +439,6 @@ function buildDiscussPrompt(nextId: string, preamble: string, _basePath: string,
   return loadPrompt("discuss", {
     milestoneId: nextId,
     preamble,
-    preparationContext: preparationContext ?? "",
     contextPath: `${milestoneRel}/${nextId}-CONTEXT.md`,
     roadmapPath: `${milestoneRel}/${nextId}-ROADMAP.md`,
     inlinedTemplates,
@@ -455,12 +471,59 @@ function buildHeadlessDiscussPrompt(nextId: string, seedContext: string, _basePa
   });
 }
 
+/**
+ * Build the prepared discuss prompt with brief injection.
+ * Uses the discuss-prepared template which encodes the 4-layer discussion protocol.
+ *
+ * @param nextId - The milestone ID being discussed
+ * @param preamble - Preamble text for the discuss prompt
+ * @param _basePath - Root directory of the project (unused, kept for signature consistency)
+ * @param prepResult - Preparation result containing briefs to inject
+ * @returns The prepared discuss prompt string
+ */
+function buildPreparedPrompt(
+  nextId: string,
+  preamble: string,
+  _basePath: string,
+  prepResult: PreparationResult,
+): string {
+  const milestoneRel = `.gsd/milestones/${nextId}`;
+
+  // Use context-enhanced instead of context for prepared discussions
+  const inlinedTemplates = [
+    inlineTemplate("project", "Project"),
+    inlineTemplate("requirements", "Requirements"),
+    inlineTemplate("context-enhanced", "Context Enhanced"),
+    inlineTemplate("roadmap", "Roadmap"),
+    inlineTemplate("decisions", "Decisions"),
+  ].join("\n\n---\n\n");
+
+  // Format the briefs from the preparation result
+  const codebaseBrief = prepResult.codebaseBrief || formatCodebaseBrief(prepResult.codebase);
+  const priorContextBrief = prepResult.priorContextBrief || formatPriorContextBrief(prepResult.priorContext);
+  const ecosystemBrief = prepResult.ecosystemBrief || formatEcosystemBrief(prepResult.ecosystem);
+
+  return loadPrompt("discuss-prepared", {
+    milestoneId: nextId,
+    preamble,
+    codebaseBrief,
+    priorContextBrief,
+    ecosystemBrief,
+    contextPath: `${milestoneRel}/${nextId}-CONTEXT.md`,
+    roadmapPath: `${milestoneRel}/${nextId}-ROADMAP.md`,
+    inlinedTemplates,
+    commitInstruction: buildDocsCommitInstruction(`docs(${nextId}): context, requirements, and roadmap`),
+    multiMilestoneCommitInstruction: buildDocsCommitInstruction("docs: project plan — N milestones"),
+  });
+}
+
 /**
  * Run preparation phase if enabled, then build the discuss prompt.
- * Preparation analyzes the codebase and prior context, injecting the results
- * as supplementary context into the standard discuss template. The discuss
- * template drives the conversation (asks "What's the vision?" first), while
- * the preparation briefs give the agent grounding in the existing codebase.
+ * This is the main entry point for new milestone discussions with preparation.
+ * Stores the preparation result for S02 to inject into the discuss prompt.
+ *
+ * When preparation succeeds, uses the discuss-prepared template with brief injection.
+ * Falls back to the standard discuss template when preparation is disabled or fails.
  *
  * @param ctx - Extension command context with UI for progress notifications
  * @param nextId - The milestone ID being discussed
@@ -474,13 +537,14 @@ async function prepareAndBuildDiscussPrompt(
   preamble: string,
   basePath: string,
 ): Promise<string> {
+  // Clear stale preparation result immediately to prevent cross-session/project
+  // state leaks. This ensures data from a prior milestone/project never leaks
+  // into subsequent discussions (adversarial review fix #3602).
+  lastPreparationResult = null;
+
   const prefs = loadEffectiveGSDPreferences()?.preferences ?? {};
 
-  // Run preparation if enabled (default: true) — results are injected as
-  // supplementary context into the standard discuss prompt, NOT as a
-  // replacement template. The discuss prompt always leads with "What's the
-  // vision?" so the user defines the scope, not the codebase analysis.
-  let preparationContext = "";
+  // Run preparation if enabled (default: true)
   if (prefs.discuss_preparation !== false) {
     try {
       const prepResult = await runPreparation(basePath, ctx.ui, {
@@ -488,23 +552,21 @@ async function prepareAndBuildDiscussPrompt(
         discuss_web_research: prefs.discuss_web_research,
         discuss_depth: prefs.discuss_depth,
       });
+      lastPreparationResult = prepResult;
 
+      // Use prepared prompt if preparation was enabled and produced results
       if (prepResult.enabled) {
-        const codebaseBrief = prepResult.codebaseBrief || formatCodebaseBrief(prepResult.codebase);
-        const priorContextBrief = prepResult.priorContextBrief || formatPriorContextBrief(prepResult.priorContext);
-        const parts: string[] = [];
-        if (codebaseBrief) parts.push(`### Codebase Brief\n\n${codebaseBrief}`);
-        if (priorContextBrief) parts.push(`### Prior Context Brief\n\n${priorContextBrief}`);
-        if (parts.length > 0) {
-          preparationContext = `\n\n## Preparation Context\n\nThe system analyzed the codebase before this discussion. Use these findings as background context — they describe what already exists, NOT what the user wants to build. Always ask the user what they want to build first.\n\n${parts.join("\n\n")}`;
-        }
+        return buildPreparedPrompt(nextId, preamble, basePath, prepResult);
       }
-    } catch (err) {
-      logWarning("guided", `preparation failed, proceeding without context: ${(err as Error).message}`);
+    } catch {
+      // If preparation throws, ensure stale data doesn't persist
+      lastPreparationResult = null;
     }
   }
 
-  return buildDiscussPrompt(nextId, preamble, basePath, preparationContext);
+  // Fall back to standard discuss prompt for backward compatibility
+  // lastPreparationResult is already null (cleared at entry or on error)
+  return buildDiscussPrompt(nextId, preamble, basePath);
 }
 
 /**

package/src/resources/extensions/gsd/model-router.ts

@@ -58,7 +58,7 @@ export interface ModelCapabilities {
 // Maps known model IDs to their capability tier. Used when tier_models is not
 // explicitly configured to pick the best available model for each tier.
 
-export const MODEL_CAPABILITY_TIER: Record<string, ComplexityTier> = {
+const MODEL_CAPABILITY_TIER: Record<string, ComplexityTier> = {
   // Light-tier models (cheapest)
   "claude-haiku-4-5": "light",
   "claude-3-5-haiku-latest": "light",
@@ -139,49 +139,15 @@ const MODEL_COST_PER_1K_INPUT: Record<string, number> = {
 // model selection within an eligible tier set.
 
 export const MODEL_CAPABILITY_PROFILES: Record<string, ModelCapabilities> = {
-  // ── Anthropic ──────────────────────────────────────────────────────────────
-  "claude-opus-4-6":              { coding: 95, debugging: 90, research: 85, reasoning: 95, speed: 30, longContext: 80, instruction: 90 },
-  "claude-sonnet-4-6":            { coding: 85, debugging: 80, research: 75, reasoning: 80, speed: 60, longContext: 75, instruction: 85 },
-  "claude-sonnet-4-5-20250514":   { coding: 85, debugging: 80, research: 75, reasoning: 80, speed: 60, longContext: 75, instruction: 85 },
-  "claude-3-5-sonnet-latest":     { coding: 82, debugging: 78, research: 72, reasoning: 78, speed: 62, longContext: 70, instruction: 82 },
-  "claude-haiku-4-5":             { coding: 60, debugging: 50, research: 45, reasoning: 50, speed: 95, longContext: 50, instruction: 75 },
-  "claude-3-5-haiku-latest":      { coding: 60, debugging: 50, research: 45, reasoning: 50, speed: 95, longContext: 50, instruction: 75 },
-  "claude-3-haiku-20240307":      { coding: 50, debugging: 40, research: 35, reasoning: 40, speed: 95, longContext: 40, instruction: 65 },
-  "claude-3-opus-latest":         { coding: 90, debugging: 85, research: 82, reasoning: 90, speed: 35, longContext: 75, instruction: 88 },
-
-  // ── OpenAI GPT ─────────────────────────────────────────────────────────────
-  "gpt-4o":                       { coding: 80, debugging: 75, research: 70, reasoning: 75, speed: 65, longContext: 70, instruction: 80 },
-  "gpt-4o-mini":                  { coding: 55, debugging: 45, research: 40, reasoning: 45, speed: 90, longContext: 45, instruction: 70 },
-  "gpt-4-turbo":                  { coding: 78, debugging: 72, research: 68, reasoning: 72, speed: 50, longContext: 65, instruction: 78 },
-  "gpt-4.1":                      { coding: 82, debugging: 78, research: 72, reasoning: 78, speed: 62, longContext: 72, instruction: 82 },
-  "gpt-4.1-mini":                 { coding: 58, debugging: 48, research: 42, reasoning: 48, speed: 88, longContext: 48, instruction: 72 },
-  "gpt-4.1-nano":                 { coding: 40, debugging: 30, research: 25, reasoning: 30, speed: 95, longContext: 30, instruction: 60 },
-  "gpt-5":                        { coding: 92, debugging: 88, research: 85, reasoning: 92, speed: 40, longContext: 85, instruction: 90 },
-  "gpt-5-mini":                   { coding: 62, debugging: 52, research: 48, reasoning: 52, speed: 88, longContext: 52, instruction: 74 },
-  "gpt-5-nano":                   { coding: 42, debugging: 32, research: 28, reasoning: 32, speed: 95, longContext: 32, instruction: 62 },
-  "gpt-5-pro":                    { coding: 94, debugging: 90, research: 88, reasoning: 94, speed: 35, longContext: 88, instruction: 92 },
-  "gpt-5.1":                      { coding: 93, debugging: 89, research: 86, reasoning: 93, speed: 42, longContext: 86, instruction: 91 },
-  "gpt-5.1-codex-max":            { coding: 90, debugging: 85, research: 70, reasoning: 85, speed: 55, longContext: 75, instruction: 85 },
-  "gpt-5.1-codex-mini":           { coding: 65, debugging: 55, research: 40, reasoning: 50, speed: 88, longContext: 48, instruction: 72 },
-  "gpt-5.2":                      { coding: 93, debugging: 90, research: 87, reasoning: 93, speed: 42, longContext: 87, instruction: 91 },
-  "gpt-5.2-codex":                { coding: 93, debugging: 90, research: 72, reasoning: 88, speed: 50, longContext: 78, instruction: 88 },
-  "gpt-5.3-codex":                { coding: 94, debugging: 91, research: 74, reasoning: 89, speed: 50, longContext: 80, instruction: 89 },
-  "gpt-5.3-codex-spark":          { coding: 68, debugging: 58, research: 42, reasoning: 52, speed: 90, longContext: 50, instruction: 74 },
-  "gpt-5.4":                      { coding: 95, debugging: 92, research: 88, reasoning: 94, speed: 42, longContext: 88, instruction: 92 },
-
-  // ── OpenAI o-series (reasoning-first) ──────────────────────────────────────
-  "o1":                           { coding: 78, debugging: 82, research: 78, reasoning: 90, speed: 20, longContext: 65, instruction: 82 },
-  "o3":                           { coding: 80, debugging: 85, research: 80, reasoning: 92, speed: 25, longContext: 70, instruction: 85 },
-  "o4-mini":                      { coding: 75, debugging: 80, research: 72, reasoning: 88, speed: 60, longContext: 65, instruction: 80 },
-  "o4-mini-deep-research":        { coding: 75, debugging: 80, research: 85, reasoning: 88, speed: 30, longContext: 80, instruction: 80 },
-
-  // ── Google ─────────────────────────────────────────────────────────────────
-  "gemini-2.5-pro":               { coding: 75, debugging: 70, research: 85, reasoning: 75, speed: 55, longContext: 90, instruction: 75 },
-  "gemini-2.0-flash":             { coding: 50, debugging: 40, research: 50, reasoning: 40, speed: 95, longContext: 60, instruction: 65 },
-  "gemini-flash-2.0":             { coding: 50, debugging: 40, research: 50, reasoning: 40, speed: 95, longContext: 60, instruction: 65 },
-
-  // ── DeepSeek ───────────────────────────────────────────────────────────────
-  "deepseek-chat":                { coding: 75, debugging: 65, research: 55, reasoning: 70, speed: 70, longContext: 55, instruction: 65 },
+  "claude-opus-4-6":   { coding: 95, debugging: 90, research: 85, reasoning: 95, speed: 30, longContext: 80, instruction: 90 },
+  "claude-sonnet-4-6": { coding: 85, debugging: 80, research: 75, reasoning: 80, speed: 60, longContext: 75, instruction: 85 },
+  "claude-haiku-4-5":  { coding: 60, debugging: 50, research: 45, reasoning: 50, speed: 95, longContext: 50, instruction: 75 },
+  "gpt-4o":            { coding: 80, debugging: 75, research: 70, reasoning: 75, speed: 65, longContext: 70, instruction: 80 },
+  "gpt-4o-mini":       { coding: 55, debugging: 45, research: 40, reasoning: 45, speed: 90, longContext: 45, instruction: 70 },
+  "gemini-2.5-pro":    { coding: 75, debugging: 70, research: 85, reasoning: 75, speed: 55, longContext: 90, instruction: 75 },
+  "gemini-2.0-flash":  { coding: 50, debugging: 40, research: 50, reasoning: 40, speed: 95, longContext: 60, instruction: 65 },
+  "deepseek-chat":     { coding: 75, debugging: 65, research: 55, reasoning: 70, speed: 70, longContext: 55, instruction: 65 },
+  "o3":                { coding: 80, debugging: 85, research: 80, reasoning: 92, speed: 25, longContext: 70, instruction: 85 },
 };
 
 // ─── Base Task Requirements Data Table ───────────────────────────────────────

package/src/resources/extensions/gsd/preferences-types.ts

@@ -20,7 +20,7 @@ import type {
   ReactiveExecutionConfig,
   GateEvaluationConfig,
 } from "./types.js";
-import type { DynamicRoutingConfig, ModelCapabilities } from "./model-router.js";
+import type { DynamicRoutingConfig } from "./model-router.js";
 
 export interface ContextManagementConfig {
   observation_masking?: boolean;          // default: true
@@ -255,8 +255,6 @@ export interface GSDPreferences {
   post_unit_hooks?: PostUnitHookConfig[];
   pre_dispatch_hooks?: PreDispatchHookConfig[];
   dynamic_routing?: DynamicRoutingConfig;
-  /** Per-model capability overrides. Deep-merged with built-in profiles for capability-aware routing (ADR-004). */
-  modelOverrides?: Record<string, { capabilities?: Partial<ModelCapabilities> }>;
   context_management?: ContextManagementConfig;
   token_profile?: TokenProfile;
   phases?: PhaseSkipPreferences;

package/src/resources/extensions/gsd/prompt-validation.ts

@@ -0,0 +1,88 @@
+/**
+ * GSD Prompt Validation — Validates enhanced context output before writing.
+ *
+ * Implements R109 validation requirement: CONTEXT.md must have required sections
+ * before being written to disk.
+ */
+
+/**
+ * Result of validating enhanced context output.
+ */
+export interface ValidationResult {
+  /** Whether all required sections are present. */
+  valid: boolean;
+  /** List of missing required sections. */
+  missing: string[];
+}
+
+/**
+ * Validate that enhanced context content has all required sections.
+ *
+ * Required sections per R109:
+ * - Scope section (## Scope, ## Milestone Scope, or ## Why This Milestone)
+ * - Architectural Decisions section (## Architectural Decisions)
+ * - Acceptance Criteria section (## Acceptance Criteria or ## Final Integrated Acceptance)
+ *
+ * Additionally validates that the Architectural Decisions section contains
+ * at least one decision entry (### heading or **Decision marker).
+ *
+ * @param content - The enhanced context markdown content
+ * @returns ValidationResult with valid flag and list of missing sections
+ */
+export function validateEnhancedContext(content: string): ValidationResult {
+  const missing: string[] = [];
+
+  // Required section 1: Scope (multiple acceptable header variants)
+  const hasScopeSection =
+    /^## Scope\b/m.test(content) ||
+    /^## Milestone Scope\b/m.test(content) ||
+    /^## Why This Milestone\b/m.test(content);
+
+  if (!hasScopeSection) {
+    missing.push("Milestone Scope or Why This Milestone");
+  }
+
+  // Required section 2: Architectural Decisions
+  const hasArchitecturalDecisions = /^## Architectural Decisions\b/m.test(content);
+  if (!hasArchitecturalDecisions) {
+    missing.push("Architectural Decisions");
+  }
+
+  // Required section 3: Acceptance Criteria (multiple acceptable header variants)
+  const hasAcceptanceCriteria =
+    /^## Acceptance Criteria\b/m.test(content) ||
+    /^## Final Integrated Acceptance\b/m.test(content);
+
+  if (!hasAcceptanceCriteria) {
+    missing.push("Acceptance Criteria");
+  }
+
+  // Additional validation: Architectural Decisions must have at least one entry
+  if (hasArchitecturalDecisions) {
+    // Extract the section content between ## Architectural Decisions and the next ## heading.
+    // Uses indexOf-based extraction instead of regex with \z (which is invalid in JavaScript
+    // regex — it's PCRE/Ruby syntax and JS treats it as literal 'z').
+    const sectionStart = content.indexOf("## Architectural Decisions");
+    if (sectionStart === -1) {
+      missing.push("Architectural Decisions");
+    } else {
+      const afterHeading = content.slice(sectionStart + "## Architectural Decisions".length);
+      const nextSection = afterHeading.search(/^## /m);
+      const sectionContent = nextSection === -1 ? afterHeading : afterHeading.slice(0, nextSection);
+
+      // Check for actual decision entries:
+      // - ### heading (subsection per decision)
+      // - **Decision marker (inline decision format)
+      const hasDecisionEntry = /^### /m.test(sectionContent) || /^\*\*Decision/m.test(sectionContent);
+
+      if (!hasDecisionEntry) {
+        missing.push("At least one architectural decision entry");
+      }
+    }
+  }
+
+  return {
+    valid: missing.length === 0,
+    missing,
+  };
+}