@opengsd/gsd-pi 1.9.0 → 1.10.0

package/src/resources/extensions/gsd/md-importer.ts

@@ -31,6 +31,7 @@ import {
   legacyMilestonesDir,
   gsdRoot,
   resolveTaskFiles,
+  resolveTaskFile,
   resolveMilestonePath,
 } from './paths.js';
 import { findMilestoneIds } from './guided-flow.js';
@@ -925,8 +926,7 @@ export function migrateHierarchyToDb(
           legacyTasksDir = isLegacySlice
             ? resolveTasksDir(basePath, milestoneId, sliceEntry.id)
             : null;
-          const summaryDir = legacyTasksDir ?? slicePath;
-          taskSummaryFile = join(summaryDir, `${taskEntry.id}-SUMMARY.md`);
+          taskSummaryFile = resolveTaskFile(basePath, milestoneId, sliceEntry.id, taskEntry.id, 'SUMMARY');
         }
         const hasTaskSummary = taskSummaryFile !== null && existsSync(taskSummaryFile);
 
@@ -950,9 +950,9 @@ export function migrateHierarchyToDb(
         // first execute-task. The surviving SUMMARY wins over a stale checkbox.
         // reopen-task removes the SUMMARY, so this never re-completes an
         // intentionally reopened task.
-        // Flat-phase summaries are phase-scoped (T01-SUMMARY.md); when a milestone
-        // has multiple slices, require the summary parent frontmatter to match so
-        // one slice's completion does not attest a sibling's same-named task.
+        // Flat-phase summaries are phase-scoped; when a milestone has multiple
+        // slices, require the summary parent frontmatter to match so one slice's
+        // completion does not attest a sibling's same-named task.
         let summaryAttestsCompletion = hasTaskSummary;
         if (summaryAttestsCompletion && !isLegacySlice && roadmap.slices.length > 1 && taskSummaryFile) {
           try {

package/src/resources/extensions/gsd/model-cost-table.ts

@@ -62,6 +62,8 @@ export const BUNDLED_COST_TABLE: ModelCostEntry[] = [
   // GPT-5.5 API list price, also used for live Codex OAuth routing.
   // Source: https://openai.com/api/pricing/
   { id: "gpt-5.5", inputPer1k: 0.005, outputPer1k: 0.03, updatedAt: "2026-04-23" },
+  // Mirrors gpt-5.5 list pricing until OpenAI publishes official gpt-5.6 pricing.
+  { id: "gpt-5.6", inputPer1k: 0.005, outputPer1k: 0.03, updatedAt: "2026-07-08" },
 
   // Google
   { id: "gemini-2.0-flash", inputPer1k: 0.0001, outputPer1k: 0.0004, updatedAt: "2025-03-15" },

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

@@ -111,10 +111,12 @@ export const MODEL_CAPABILITY_TIER: Record<string, ComplexityTier> = {
   "gpt-5.3-codex": "heavy",
   "gpt-5.4": "heavy",
   "gpt-5.5": "heavy",
+  "gpt-5.6": "heavy",
   "o1": "heavy",
   "o3": "heavy",
   "o4-mini": "heavy",
   "o4-mini-deep-research": "heavy",
+  "grok-4.5": "heavy",
 };
 
 // ─── Cost Table (per 1K input tokens, approximate USD) ───────────────────────
@@ -149,11 +151,13 @@ const MODEL_COST_PER_1K_INPUT: Record<string, number> = {
   "gpt-5.3-codex-spark": 0.0003,
   "gpt-5.4": 0.005,
   "gpt-5.5": 0.005,
+  "gpt-5.6": 0.005,
   "o4-mini": 0.005,
   "o4-mini-deep-research": 0.005,
   "gemini-2.0-flash": 0.0001,
   "gemini-2.5-pro": 0.00125,
   "deepseek-chat": 0.00014,
+  "grok-4.5": 0.002,
 };
 
 // ─── Capability Profiles Data Table ──────────────────────────────────────────
@@ -198,6 +202,9 @@ export const MODEL_CAPABILITY_PROFILES: Record<string, ModelCapabilities> = {
   // OpenAI's 2026-04-23 published eval deltas across coding, tool use, and long context.
   // Source: https://openai.com/index/introducing-gpt-5-5/
   "gpt-5.5":                      { coding: 96, debugging: 93, research: 89, reasoning: 95, speed: 42, longContext: 90, instruction: 93 },
+  // GPT-5.6 profile is provisional: a small uniform step over gpt-5.5, to be
+  // recalibrated once OpenAI publishes official eval results.
+  "gpt-5.6":                      { coding: 97, debugging: 94, research: 90, reasoning: 96, speed: 42, longContext: 91, instruction: 94 },
 
   // ── OpenAI o-series (reasoning-first) ──────────────────────────────────────
   "o1":                           { coding: 78, debugging: 82, research: 78, reasoning: 90, speed: 20, longContext: 65, instruction: 82 },
@@ -212,6 +219,11 @@ export const MODEL_CAPABILITY_PROFILES: Record<string, ModelCapabilities> = {
 
   // ── DeepSeek ───────────────────────────────────────────────────────────────
   "deepseek-chat":                { coding: 75, debugging: 65, research: 55, reasoning: 70, speed: 70, longContext: 55, instruction: 65 },
+
+  // ── xAI ────────────────────────────────────────────────────────────────────
+  // Grok 4.5 (2026-07): Opus-class frontier model tuned for coding and agentic
+  // work; notably faster and more token-efficient than peer heavy models.
+  "grok-4.5":                     { coding: 95, debugging: 90, research: 82, reasoning: 93, speed: 55, longContext: 80, instruction: 90 },
 };
 
 // ─── Base Task Requirements Data Table ───────────────────────────────────────

package/src/resources/extensions/gsd/paths.ts

@@ -136,6 +136,14 @@ function cachedReaddir(dirPath: string): string[] {
   return entries;
 }
 
+function isExistingFile(path: string): boolean {
+  try {
+    return statSync(path).isFile();
+  } catch {
+    return false;
+  }
+}
+
 /**
  * Clear the volatile directory listing caches.
  * Call after milestone transitions, file creation in planning directories,
@@ -223,6 +231,26 @@ export function buildTaskFileName(taskId: string, suffix: string): string {
   return `${taskId}-${suffix}.md`;
 }
 
+/**
+ * Build a flat-phase task artifact file name.
+ * ("S06", "T03", "SUMMARY") → "S06-T03-SUMMARY.md"
+ */
+export function buildFlatTaskFileName(sliceId: string, taskId: string, suffix: string): string {
+  return `${sliceId}-${taskId}-${suffix}.md`;
+}
+
+/**
+ * Extract the task ID from a task artifact filename.
+ * Supports flat-phase S##-T##-SUFFIX.md plus legacy T##-SUFFIX.md and
+ * T##-DESCRIPTOR-SUFFIX.md names.
+ */
+export function taskIdFromTaskFileName(fileName: string, suffix: string): string | null {
+  const flat = new RegExp(`^S\\d+-(T\\d+)-${suffix}\\.md$`, "i").exec(fileName);
+  if (flat?.[1]) return flat[1].toUpperCase();
+  const legacy = new RegExp(`^(T\\d+)(?:-.*)?-${suffix}\\.md$`, "i").exec(fileName);
+  return legacy?.[1]?.toUpperCase() ?? null;
+}
+
 // ─── Resolvers ─────────────────────────────────────────────────────────────
 
 /**
@@ -264,7 +292,7 @@ export function resolveFile(dir: string, idPrefix: string, suffix: string): stri
   if (!existsSync(dir)) return null;
   const target = `${idPrefix}-${suffix}.md`.toUpperCase();
   try {
-    const entries = cachedReaddir(dir);
+    const entries = cachedReaddirWithTypes(dir).filter(e => e.isFile()).map(e => e.name);
     // Direct match: ID-SUFFIX.md
     const direct = entries.find(e => e.toUpperCase() === target);
     if (direct) return direct;
@@ -284,18 +312,17 @@ export function resolveFile(dir: string, idPrefix: string, suffix: string): stri
 }
 
 /**
- * Find all task files matching a pattern in a tasks directory.
- * Returns sorted file names matching T##-SUFFIX.md or legacy T##-*-SUFFIX.md
+ * Find all task files matching a pattern in a tasks directory or flat phase dir.
+ * Returns sorted file names matching S##-T##-SUFFIX.md, T##-SUFFIX.md,
+ * or legacy T##-*-SUFFIX.md.
  */
 export function resolveTaskFiles(tasksDir: string, suffix: string): string[] {
   if (!existsSync(tasksDir)) return [];
   try {
-    // Current convention: T01-PLAN.md
-    const currentPattern = new RegExp(`^T\\d+-${suffix}\\.md$`, "i");
-    // Legacy convention: T01-INSTALL-PACKAGES-PLAN.md
-    const legacyPattern = new RegExp(`^T\\d+-.*-${suffix}\\.md$`, "i");
-    return cachedReaddir(tasksDir)
-      .filter(f => currentPattern.test(f) || legacyPattern.test(f))
+    return cachedReaddirWithTypes(tasksDir)
+      .filter(e => e.isFile())
+      .map(e => e.name)
+      .filter(f => taskIdFromTaskFileName(f, suffix) !== null)
       .sort();
   } catch {
     return [];
@@ -843,7 +870,7 @@ export function resolveMilestoneFile(
   const prefix = `${String(phaseNum).padStart(2, "0")}`;
   const flatName = `${prefix}-${suffix}.md`;
   const flatPath = join(mDir, flatName);
-  if (existsSync(flatPath)) return flatPath;
+  if (isExistingFile(flatPath)) return flatPath;
   // Legacy fallback: M001-SUFFIX.md
   const file = resolveFile(mDir, milestoneId, suffix);
   return file ? join(mDir, file) : null;
@@ -878,11 +905,11 @@ export function resolveSliceFile(
   const planNum = sliceIdToPlanNum(sliceId);
   const flatName = planFileName(phaseNum, planNum, suffix);
   const flatPath = join(phaseDir, flatName);
-  if (existsSync(flatPath)) return flatPath;
+  if (isExistingFile(flatPath)) return flatPath;
   // Also check plan-number-only format MM-SUFFIX.md (written by buildSliceFileName)
   const planOnlyName = `${String(planNum).padStart(2, "0")}-${suffix}.md`;
   const planOnlyPath = join(phaseDir, planOnlyName);
-  if (existsSync(planOnlyPath)) return planOnlyPath;
+  if (isExistingFile(planOnlyPath)) return planOnlyPath;
   // Try prefix match for the phase+plan number (handles suffix variations)
   const planPrefix = `${String(phaseNum).padStart(2, "0")}-${String(planNum).padStart(2, "0")}-`;
   try {
@@ -931,8 +958,10 @@ export function resolveTaskFile(
   const isLegacy = phaseDir.startsWith(legacyBase + "/") || phaseDir.startsWith(legacyBase + "\\");
 
   if (suffix !== "PLAN" && !isLegacy) {
-    const flatPath = join(phaseDir, buildTaskFileName(taskId, suffix));
-    return existsSync(flatPath) ? flatPath : null;
+    const flatPath = join(phaseDir, buildFlatTaskFileName(sliceId, taskId, suffix));
+    if (isExistingFile(flatPath)) return flatPath;
+    const legacyFlatPath = join(phaseDir, buildTaskFileName(taskId, suffix));
+    return isExistingFile(legacyFlatPath) ? legacyFlatPath : null;
   }
 
   const tDir = resolveTasksDir(basePath, milestoneId, sliceId);
@@ -1066,6 +1095,35 @@ export function targetSliceFile(
   return join(dir, fileName);
 }
 
+/**
+ * Build the canonical absolute write target for a task file.
+ * Readers that need legacy flat-phase fallback should call resolveTaskFile.
+ */
+export function targetTaskFile(
+  basePath: string, milestoneId: string, sliceId: string,
+  taskId: string, suffix: string, milestoneTitle?: string
+): string {
+  if (suffix === "PLAN") {
+    return targetSliceFile(basePath, milestoneId, sliceId, "PLAN", milestoneTitle);
+  }
+
+  const mDir = resolveMilestonePath(basePath, milestoneId);
+  const legacyBase = legacyMilestonesDir(basePath);
+  const isLegacy = mDir
+    ? mDir.startsWith(legacyBase + "/") || mDir.startsWith(legacyBase + "\\")
+    : isLegacyMilestonesLayout(basePath);
+  const milestoneDir = mDir
+    ?? dirname(targetMilestoneFile(basePath, milestoneId, "ROADMAP", milestoneTitle));
+
+  if (isLegacy) {
+    const slicesDir = join(milestoneDir, "slices");
+    const sliceDir = join(slicesDir, resolveDir(slicesDir, sliceId) ?? sliceId);
+    return join(sliceDir, "tasks", buildTaskFileName(taskId, suffix));
+  }
+
+  return join(milestoneDir, buildFlatTaskFileName(sliceId, taskId, suffix));
+}
+
 /**
  * Build relative .gsd/ path to a slice file.
  * Preserves an existing compatibility filename when present; otherwise falls
@@ -1105,5 +1163,5 @@ export function relTaskFile(
     return relSliceFile(basePath, milestoneId, sliceId, "PLAN");
   }
   const relS = relSlicePath(basePath, milestoneId, sliceId, milestoneTitle);
-  return `${relS}/${buildTaskFileName(taskId, suffix)}`;
+  return `${relS}/${buildFlatTaskFileName(sliceId, taskId, suffix)}`;
 }

package/src/resources/extensions/gsd/planning-subagent-policy.ts

@@ -0,0 +1,58 @@
+import { CONFIGURABLE_PLANNING_SUBAGENT_UNITS, type GSDPreferences } from "./preferences-types.js";
+import { loadEffectiveGSDPreferences } from "./preferences.js";
+import type { ToolsPolicy } from "./unit-context-manifest.js";
+
+const CONFIGURABLE_UNITS = new Set<string>(CONFIGURABLE_PLANNING_SUBAGENT_UNITS);
+
+function configuredAllowedSubagents(
+  unitType: string,
+  preferences: GSDPreferences | null | undefined,
+): readonly string[] {
+  if (!CONFIGURABLE_UNITS.has(unitType)) return [];
+  return preferences?.planning_subagents?.[
+    unitType as keyof NonNullable<GSDPreferences["planning_subagents"]>
+  ]?.allowed ?? [];
+}
+
+function mergeAllowedSubagents(
+  current: readonly string[],
+  configured: readonly string[],
+): readonly string[] {
+  return Array.from(new Set([...current, ...configured]));
+}
+
+export function applyPlanningSubagentPreferences(
+  unitType: string,
+  policy: ToolsPolicy | null | undefined,
+  preferences: GSDPreferences | null | undefined,
+): ToolsPolicy | null | undefined {
+  if (!policy) return policy;
+  const configured = configuredAllowedSubagents(unitType, preferences);
+  if (configured.length === 0) return policy;
+
+  if (policy.mode === "planning") {
+    return {
+      mode: "planning-dispatch",
+      allowedSubagents: mergeAllowedSubagents([], configured),
+    };
+  }
+
+  if (policy.mode === "planning-dispatch") {
+    return {
+      ...policy,
+      allowedSubagents: mergeAllowedSubagents(policy.allowedSubagents, configured),
+    };
+  }
+
+  return policy;
+}
+
+export function resolveEffectivePlanningToolsPolicy(
+  unitType: string,
+  policy: ToolsPolicy | null | undefined,
+  basePath?: string,
+): ToolsPolicy | null | undefined {
+  if (!basePath) return policy;
+  const preferences = loadEffectiveGSDPreferences(basePath)?.preferences;
+  return applyPlanningSubagentPreferences(unitType, policy, preferences);
+}

package/src/resources/extensions/gsd/planning-subagent-registry.ts

@@ -0,0 +1,31 @@
+/**
+ * Read-only planning specialists that controlled planning dispatch may use.
+ *
+ * Unit manifests and project preferences still declare per-unit subsets. This
+ * registry is the global safety classification checked by the write gate and
+ * preference validation before any configured planning allowlist takes effect.
+ */
+
+const PLANNING_DISPATCH_AGENT_REGISTRY = {
+  mnemo: { readOnlySpecialist: true },
+  scout: { readOnlySpecialist: true },
+  planner: { readOnlySpecialist: true },
+  reviewer: { readOnlySpecialist: true },
+  security: { readOnlySpecialist: true },
+  tester: { readOnlySpecialist: true },
+} as const satisfies Record<string, { readonly readOnlySpecialist: boolean }>;
+
+export const ALLOWED_PLANNING_DISPATCH_AGENTS = new Set<string>(
+  Object.entries(PLANNING_DISPATCH_AGENT_REGISTRY)
+    .filter(([, metadata]) => metadata.readOnlySpecialist)
+    .map(([agentId]) => agentId),
+);
+
+export function isReadOnlyPlanningDispatchAgent(agentId: string): boolean {
+  const metadata = PLANNING_DISPATCH_AGENT_REGISTRY[agentId as keyof typeof PLANNING_DISPATCH_AGENT_REGISTRY];
+  return metadata?.readOnlySpecialist === true;
+}
+
+export function allowedPlanningDispatchAgentsList(): string {
+  return [...ALLOWED_PLANNING_DISPATCH_AGENTS].join(", ");
+}

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

@@ -141,6 +141,7 @@ export const KNOWN_PREFERENCE_KEYS = new Set<string>([
   "git",
   "post_unit_hooks",
   "pre_dispatch_hooks",
+  "planning_subagents",
   "dynamic_routing",
   "disabled_model_providers",
   "uok",
@@ -217,6 +218,20 @@ export const KNOWN_UNIT_LABELS = [
 ] as const;
 export type UnitLabel = (typeof KNOWN_UNIT_LABELS)[number];
 
+export const CONFIGURABLE_PLANNING_SUBAGENT_UNITS = [
+  "plan-milestone",
+  "plan-slice",
+] as const;
+export type ConfigurablePlanningSubagentUnit = (typeof CONFIGURABLE_PLANNING_SUBAGENT_UNITS)[number];
+
+export interface PlanningSubagentUnitConfig {
+  allowed?: string[];
+}
+
+export type PlanningSubagentsConfig = Partial<
+  Record<ConfigurablePlanningSubagentUnit, PlanningSubagentUnitConfig>
+>;
+
 
 export const SKILL_ACTIONS = new Set(["use", "prefer", "avoid"]);
 
@@ -476,6 +491,7 @@ export interface GSDPreferences {
   git?: GitPreferences;
   post_unit_hooks?: PostUnitHookConfig[];
   pre_dispatch_hooks?: PreDispatchHookConfig[];
+  planning_subagents?: PlanningSubagentsConfig;
   dynamic_routing?: DynamicRoutingConfig;
   /** Provider IDs to exclude from /model and automatic model routing while leaving tool auth intact. */
   disabled_model_providers?: string[];

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

@@ -17,6 +17,7 @@ import { getGateIdsForTurn } from "./gate-registry.js";
 import {
   KNOWN_PREFERENCE_KEYS,
   KNOWN_UNIT_LABELS,
+  CONFIGURABLE_PLANNING_SUBAGENT_UNITS,
   GSD_MODEL_PHASE_KEYS,
 
   SKILL_ACTIONS,
@@ -26,6 +27,10 @@ import {
   type GSDSkillRule,
   type GSDThinkingLevel,
 } from "./preferences-types.js";
+import {
+  ALLOWED_PLANNING_DISPATCH_AGENTS,
+  allowedPlanningDispatchAgentsList,
+} from "./planning-subagent-registry.js";
 
 const VALID_TOKEN_PROFILES = new Set<TokenProfile>(["budget", "balanced", "quality", "burn-max"]);
 const VALID_THINKING_LEVELS = new Set<GSDThinkingLevel>([
@@ -747,6 +752,63 @@ export function validatePreferences(preferences: GSDPreferences): {
     }
   }
 
+  // ─── Planning Subagents ─────────────────────────────────────────────────
+  if (preferences.planning_subagents !== undefined) {
+    if (
+      !preferences.planning_subagents ||
+      typeof preferences.planning_subagents !== "object" ||
+      Array.isArray(preferences.planning_subagents)
+    ) {
+      errors.push("planning_subagents must be an object keyed by planning unit type");
+    } else {
+      const validUnits = new Set<string>(CONFIGURABLE_PLANNING_SUBAGENT_UNITS);
+      const validPlanningSubagents: NonNullable<GSDPreferences["planning_subagents"]> = {};
+
+      for (const [unitType, rawConfig] of Object.entries(preferences.planning_subagents as Record<string, unknown>)) {
+        if (!validUnits.has(unitType)) {
+          errors.push(
+            `planning_subagents unknown or unsupported unit "${unitType}" — supported units: ` +
+            `${CONFIGURABLE_PLANNING_SUBAGENT_UNITS.join(", ")}`,
+          );
+          continue;
+        }
+        if (!rawConfig || typeof rawConfig !== "object" || Array.isArray(rawConfig)) {
+          errors.push(`planning_subagents.${unitType} must be an object with an allowed array`);
+          continue;
+        }
+
+        const allowed = normalizeStringArray((rawConfig as Record<string, unknown>).allowed);
+        if (allowed.length === 0) {
+          errors.push(`planning_subagents.${unitType}.allowed must include at least one read-only planning specialist`);
+          continue;
+        }
+
+        const sanitizedAllowed: string[] = [];
+        for (const agentId of allowed) {
+          if (!ALLOWED_PLANNING_DISPATCH_AGENTS.has(agentId)) {
+            errors.push(
+              `planning_subagents.${unitType}.allowed includes "${agentId}", which is not classified as a ` +
+              `read-only planning specialist. Allowed agents: ${allowedPlanningDispatchAgentsList()}`,
+            );
+            continue;
+          }
+          sanitizedAllowed.push(agentId);
+        }
+
+        const dedupedAllowed = Array.from(new Set(sanitizedAllowed));
+        if (dedupedAllowed.length > 0) {
+          validPlanningSubagents[unitType as keyof NonNullable<GSDPreferences["planning_subagents"]>] = {
+            allowed: dedupedAllowed,
+          };
+        }
+      }
+
+      if (Object.keys(validPlanningSubagents).length > 0) {
+        validated.planning_subagents = validPlanningSubagents;
+      }
+    }
+  }
+
   // ─── Dynamic Routing ─────────────────────────────────────────────────
   if (preferences.dynamic_routing !== undefined) {
     if (typeof preferences.dynamic_routing === "object" && preferences.dynamic_routing !== null) {

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

@@ -817,6 +817,7 @@ function mergePreferences(base: GSDPreferences, override: GSDPreferences): GSDPr
       : undefined,
     post_unit_hooks: mergePostUnitHooks(base.post_unit_hooks, override.post_unit_hooks),
     pre_dispatch_hooks: mergePreDispatchHooks(base.pre_dispatch_hooks, override.pre_dispatch_hooks),
+    planning_subagents: mergePlanningSubagents(base.planning_subagents, override.planning_subagents),
     dynamic_routing: (base.dynamic_routing || override.dynamic_routing)
       ? { ...(base.dynamic_routing ?? {}), ...(override.dynamic_routing ?? {}) } as DynamicRoutingConfig
       : undefined,
@@ -943,6 +944,30 @@ function mergePreDispatchHooks(
   return merged.length > 0 ? merged : undefined;
 }
 
+function mergePlanningSubagents(
+  base?: GSDPreferences["planning_subagents"],
+  override?: GSDPreferences["planning_subagents"],
+): GSDPreferences["planning_subagents"] | undefined {
+  if (!base && !override) return undefined;
+  const merged: NonNullable<GSDPreferences["planning_subagents"]> = {};
+  const unitTypes = new Set([
+    ...Object.keys(base ?? {}),
+    ...Object.keys(override ?? {}),
+  ]);
+
+  for (const unitType of unitTypes) {
+    const allowed = mergeStringLists(
+      base?.[unitType as keyof NonNullable<GSDPreferences["planning_subagents"]>]?.allowed,
+      override?.[unitType as keyof NonNullable<GSDPreferences["planning_subagents"]>]?.allowed,
+    );
+    if (allowed?.length) {
+      merged[unitType as keyof NonNullable<GSDPreferences["planning_subagents"]>] = { allowed };
+    }
+  }
+
+  return Object.keys(merged).length > 0 ? merged : undefined;
+}
+
 // ─── System Prompt Rendering ──────────────────────────────────────────────────
 
 export function renderPreferencesForSystemPrompt(

package/src/resources/extensions/gsd/prompts/validate-milestone.md

@@ -24,18 +24,18 @@ Roadmap, slice summaries, assessments, requirements, decisions, and project cont
 
 ### Step 1 - Dispatch Parallel Reviewers
 
-The Inlined Context above already preloads the evidence reviewers need — the roadmap, per-slice SUMMARY/ASSESSMENT excerpts, requirements, and verification classes. **Embed the relevant preloaded evidence directly into each reviewer's task prompt** so reviewers work from it instead of re-reading the same artifacts from disk. Each reviewer should read a full file only when its excerpt is missing, truncated, or internally inconsistent — never as a routine first step. This avoids three reviewers independently re-surveying artifacts the orchestrator already holds.
+The Inlined Context above already preloads the evidence reviewers need — the roadmap, per-slice SUMMARY/ASSESSMENT excerpts, requirements, and verification classes. **Embed the relevant preloaded evidence directly into each reviewer's task prompt** so reviewers work from it instead of re-reading the same artifacts from disk. Each reviewer should read a full file only when its excerpt is missing, truncated, or internally inconsistent — never as a routine first step. When a reviewer needs a full artifact, use only explicit file paths from the On-demand Validation Artifacts / On-demand Milestone Context blocks or discover files with `find .gsd -type f`; never pass a phase, slice, `tasks/`, or `slices/` directory to `read` or `readFileSync`. This avoids three reviewers independently re-surveying artifacts the orchestrator already holds.
 
 Call `subagent` with `tasks: [...]` containing ALL THREE reviewers simultaneously:
 
 **Reviewer A - Requirements Coverage**
-Prompt: "Review milestone {{milestoneId}} requirements coverage. Working directory: {{workingDirectory}}. Use the preloaded requirements and slice SUMMARY evidence embedded in this task — do not re-read them from disk. For each requirement, mark COVERED, PARTIAL, or MISSING against that evidence. Read a full SUMMARY under `.gsd/milestones/{{milestoneId}}/slices/` only if its preloaded excerpt is missing, truncated, or inconsistent. Output table: Requirement | Status | Evidence. End with one-line verdict: PASS if all covered, NEEDS-ATTENTION if partials exist, FAIL if any missing."
+Prompt: "Review milestone {{milestoneId}} requirements coverage. Working directory: {{workingDirectory}}. Use the preloaded requirements and slice SUMMARY evidence embedded in this task — do not re-read them from disk. For each requirement, mark COVERED, PARTIAL, or MISSING against that evidence. Read a full SUMMARY only from an explicit file path in the On-demand Validation Artifacts list, or discover candidates with `find .gsd -type f -name '*-SUMMARY.md'`; never read a directory path. Output table: Requirement | Status | Evidence. End with one-line verdict: PASS if all covered, NEEDS-ATTENTION if partials exist, FAIL if any missing."
 
 **Reviewer B - Cross-Slice Integration**
-Prompt: "Review milestone {{milestoneId}} cross-slice integration. Working directory: {{workingDirectory}}. Use the preloaded roadmap and slice SUMMARY evidence embedded in this task — do not re-read them from disk. Find the boundary map (produces/consumes contracts) in the roadmap. For each boundary, confirm producer SUMMARY produced the artifact and consumer SUMMARY consumed it. Read `{{roadmapPath}}` or a full SUMMARY only if the preloaded evidence is missing, truncated, or inconsistent. Output table: Boundary | Producer Summary | Consumer Summary | Status. End with one-line verdict: PASS if all boundaries honored, NEEDS-ATTENTION if any gaps."
+Prompt: "Review milestone {{milestoneId}} cross-slice integration. Working directory: {{workingDirectory}}. Use the preloaded roadmap and slice SUMMARY evidence embedded in this task — do not re-read them from disk. Find the boundary map (produces/consumes contracts) in the roadmap. For each boundary, confirm producer SUMMARY produced the artifact and consumer SUMMARY consumed it. Read `{{roadmapPath}}` or a full SUMMARY from the On-demand Validation Artifacts list only if the preloaded evidence is missing, truncated, or inconsistent; never read a directory path. Output table: Boundary | Producer Summary | Consumer Summary | Status. End with one-line verdict: PASS if all boundaries honored, NEEDS-ATTENTION if any gaps."
 
 **Reviewer C - Assessment & Acceptance Criteria**
-Prompt: "Review milestone {{milestoneId}} assessment evidence and acceptance criteria. Working directory: {{workingDirectory}}. Use the preloaded milestone context, slice SUMMARY, and ASSESSMENT evidence embedded in this task — do not re-read them from disk. Read `.gsd/milestones/{{milestoneId}}/{{milestoneId}}-CONTEXT.md`, a full SUMMARY, or a full ASSESSMENT only if the preloaded excerpt is missing, truncated, or inconsistent. UAT files are specs, not evidence. Verify each criterion maps to passing evidence. Then review the inlined `Verification Classes (from planning)` table. For every planned row in that table, output a `Verification Classes` table with columns `Class | Planned Check | Evidence | Verdict`. Preserve every planned non-empty class row; do not summarize, rename, combine, or omit planned classes. The first cell of each row must be exactly `Contract`, `Integration`, `Operational`, or `UAT` when that class is present in planning. If a planned class lacks evidence, still include its canonical row and mark the verdict NEEDS-ATTENTION or FAIL. If a planned browser/UAT class has no ASSESSMENT with browser/runtime actions and assertions, return NEEDS-ATTENTION. If no verification classes were planned, say that explicitly. Output sections `Acceptance Criteria` with checklist `[ ] Criterion | Evidence`, and `Verification Classes` with the table. End with one-line verdict: PASS if all criteria and classes are covered by evidence, NEEDS-ATTENTION if gaps exist."
+Prompt: "Review milestone {{milestoneId}} assessment evidence and acceptance criteria. Working directory: {{workingDirectory}}. Use the preloaded milestone context, slice SUMMARY, and ASSESSMENT evidence embedded in this task — do not re-read them from disk. Read the milestone context from the On-demand Milestone Context file path, and read full SUMMARY or ASSESSMENT files only from explicit paths in the On-demand Validation Artifacts list, if the preloaded excerpt is missing, truncated, or inconsistent. UAT files are specs, not evidence. Verify each criterion maps to passing evidence. Then review the inlined `Verification Classes (from planning)` table. For every planned row in that table, output a `Verification Classes` table with columns `Class | Planned Check | Evidence | Verdict`. Preserve every planned non-empty class row; do not summarize, rename, combine, or omit planned classes. The first cell of each row must be exactly `Contract`, `Integration`, `Operational`, or `UAT` when that class is present in planning. If a planned class lacks evidence, still include its canonical row and mark the verdict NEEDS-ATTENTION or FAIL. If a planned browser/UAT class has no ASSESSMENT with browser/runtime actions and assertions, return NEEDS-ATTENTION. If no verification classes were planned, say that explicitly. Output sections `Acceptance Criteria` with checklist `[ ] Criterion | Evidence`, and `Verification Classes` with the table. End with one-line verdict: PASS if all criteria and classes are covered by evidence, NEEDS-ATTENTION if gaps exist."
 
 ### Step 2 - Synthesize Findings
 
@@ -85,6 +85,6 @@ If verdict is `needs-remediation`:
 
 **You MUST call `gsd_validate_milestone` before finishing. Do not manually write `{{validationPath}}`.**
 
-**File system safety:** When scanning milestone directories, use `ls` or `find` first. Never pass a directory path such as `tasks/` or `slices/` directly to `read`; it only accepts files.
+**File system safety:** When scanning milestone directories, use `find .gsd -type f` first and read only files. In flat-phase projects, `.gsd/phases/<NN>-<slug>/` is a directory, not an artifact. Never pass a directory path such as a phase directory, `tasks/`, or `slices/` directly to `read`, `readFile`, or `readFileSync`.
 
 When done, say: "Milestone {{milestoneId}} validation complete — verdict: <verdict>." Say this exactly once — if you already said it in a prior message, do not repeat it.

package/src/resources/extensions/gsd/provider-payload-policy.ts

@@ -2,10 +2,12 @@
  * Provider Payload Policy - ordered shaping of provider request payloads.
  *
  * The order is intentional:
- * 1. observation budgeting masks old tool results in auto-mode,
- * 2. display truncation caps tool-result text for every mode,
- * 3. the protected Source Context Block is appended after truncation,
- * 4. supported models receive the configured service tier.
+ * 1. superseded GSD context injections (memory/guided/forensics) are removed,
+ *    keeping only the latest, for every mode,
+ * 2. observation budgeting masks old tool results in auto-mode,
+ * 3. display truncation caps tool-result text for every mode,
+ * 4. the protected Source Context Block is appended after truncation,
+ * 5. supported models receive the configured service tier.
  */
 
 import type { ContextManagementConfig } from "./preferences-types.js";
@@ -14,6 +16,8 @@ import type { ServiceTierSetting } from "./service-tier.js";
 import {
   createObservationMask,
   createResponsesInputObservationMask,
+  filterSupersededContextInjections,
+  filterSupersededResponsesContextInjections,
   truncateContextResultMessages,
   truncateResponsesInputResultItems,
 } from "./context-masker.js";
@@ -80,10 +84,20 @@ function applyContextManagement(
   deps: ProviderPayloadPolicyDeps,
 ): void {
   const config = deps.loadContextManagementConfig();
+  applyContextInjectionFilter(payload);
   applyObservationBudget(payload, config, deps.isAutoActive());
   applyDisplayTruncation(payload, config);
 }
 
+function applyContextInjectionFilter(payload: Record<string, unknown>): void {
+  if (Array.isArray(payload.messages)) {
+    payload.messages = filterSupersededContextInjections(payload.messages as MessagePayload);
+  }
+  if (Array.isArray(payload.input)) {
+    payload.input = filterSupersededResponsesContextInjections(payload.input as ResponsesInputPayload);
+  }
+}
+
 function applyObservationBudget(
   payload: Record<string, unknown>,
   config: ContextManagementConfig | undefined,

package/src/resources/extensions/gsd/safety/destructive-guard.ts

@@ -14,7 +14,7 @@ interface DestructivePattern {
 }
 
 const DESTRUCTIVE_PATTERNS: readonly DestructivePattern[] = [
-  { pattern: /\brm\s+(-[^\s]*[rfRF][^\s]*\s+|.*\s+-[^\s]*[rfRF])/, label: "recursive delete" },
+  { pattern: /\brm\s+((?:-(?!-)[^\s]*[rR][^\s]*|--recursive)\s+|.*\s+(?:-(?!-)[^\s]*[rR][^\s]*|--recursive)(?:\s|$))/, label: "recursive delete" },
   { pattern: /\bgit\s+push\s+.*--force/, label: "force push" },
   { pattern: /\bgit\s+push\s+-f\b/, label: "force push" },
   { pattern: /\bgit\s+reset\s+--hard/, label: "hard reset" },

package/src/resources/extensions/gsd/skill-activation.ts

@@ -131,10 +131,23 @@ function resolvePreferredSkillNames(
  *  to prevent prompt injection via crafted directory names. */
 const SAFE_SKILL_NAME = /^[a-z0-9][a-z0-9-]*$/;
 
-function formatSkillActivationBlock(skillNames: string[]): string {
+function escapeXml(str: string): string {
+  return str
+    .replace(/&/g, "&")
+    .replace(/</g, "&lt;")
+    .replace(/>/g, "&gt;")
+    .replace(/"/g, "&quot;")
+    .replace(/'/g, "&apos;");
+}
+
+function formatSkillActivationBlock(skillNames: string[], skillPaths: Map<string, string>): string {
   const safe = skillNames.filter(name => SAFE_SKILL_NAME.test(name));
   if (safe.length === 0) return "";
-  const reads = safe.map(name => `Read the installed '${name}' skill file from <available_skills>`).join(". ");
+  const reads = safe.map(name => {
+    const path = skillPaths.get(name);
+    if (path) return `Read the skill file at \`${escapeXml(path)}\``;
+    return `Find '${name}' in <available_skills>, copy its <location> value, and pass that exact path to read`;
+  }).join(". ");
   return `<skill_activation>${reads}.</skill_activation>`;
 }
 
@@ -203,6 +216,7 @@ export function buildSkillActivationBlock(params: {
   const visibleSkills = loaded;
   const manifestScopedSkills = filterSkillsByManifest(visibleSkills, params.unitType);
   const installedNames = new Set(visibleSkills.map(skill => normalizeSkillReference(skill.name)));
+  const installedSkillPaths = new Map(visibleSkills.map(skill => [normalizeSkillReference(skill.name), skill.filePath]));
   warnIfManifestHasMissingSkills(params.unitType, installedNames);
   const avoided = new Set(resolvePreferenceSkillNames(prefs?.avoid_skills ?? [], params.base));
   const matched = new Set<string>();
@@ -252,7 +266,7 @@ export function buildSkillActivationBlock(params: {
   const ordered = [...matched]
     .filter(name => installedNames.has(name) && !avoided.has(name))
     .sort();
-  const activationBlock = formatSkillActivationBlock(ordered);
+  const activationBlock = formatSkillActivationBlock(ordered, installedSkillPaths);
 
   // Omit recommendations when the system catalog is manifest-scoped for this
   // unit — skill names are already listed in <available_skills>.