gsd-pi 2.77.0-dev.58d3d4d6c → 2.77.0-dev.cfd69e714

package/src/resources/extensions/gsd/git-service.ts

@@ -8,7 +8,7 @@
  * paths, commit type inference, and the runGit shell helper.
  */
 
-import { execFileSync, execSync } from "node:child_process";
+import { execFileSync } from "node:child_process";
 import { existsSync, mkdirSync, readFileSync, readdirSync, writeFileSync } from "node:fs";
 import { join } from "node:path";
 import { gsdRoot } from "./paths.js";
@@ -409,6 +409,111 @@ export function resolveMilestoneIntegrationBranch(
   };
 }
 
+// ─── Pre-Merge Command Tokenizer ──────────────────────────────────────────
+
+/**
+ * Tokenize a user-supplied pre-merge command string into argv form, with
+ * minimal support for double- and single-quoted strings. Designed to be
+ * sufficient for typical commands ("npm test", `npm run lint:ci`,
+ * `pnpm run tsc --noEmit`) without spawning a shell.
+ *
+ * Returns [] when the input is empty or whitespace-only.
+ * Throws when quoting is malformed.
+ *
+ * Used by GitServiceImpl.runPreMergeCheck to eliminate the shell-injection
+ * surface that running an arbitrary user string through a shell would create.
+ * (Issue #4980 HIGH-2)
+ */
+export function tokenizePreMergeCommand(input: string): string[] {
+  const tokens: string[] = [];
+  let current = "";
+  let i = 0;
+  let quote: "" | "'" | '"' = "";
+  let hasContent = false;
+
+  while (i < input.length) {
+    const ch = input[i]!;
+    if (quote) {
+      if (ch === quote) {
+        quote = "";
+      } else if (ch === "\\" && quote === '"' && i + 1 < input.length) {
+        current += input[i + 1];
+        i += 2;
+        continue;
+      } else {
+        current += ch;
+      }
+      i++;
+      continue;
+    }
+    if (ch === '"' || ch === "'") {
+      quote = ch;
+      hasContent = true;
+      i++;
+      continue;
+    }
+    if (ch === " " || ch === "\t") {
+      if (hasContent) {
+        tokens.push(current);
+        current = "";
+        hasContent = false;
+      }
+      i++;
+      continue;
+    }
+    if (ch === "\\" && i + 1 < input.length) {
+      current += input[i + 1];
+      i += 2;
+      hasContent = true;
+      continue;
+    }
+    current += ch;
+    hasContent = true;
+    i++;
+  }
+
+  if (quote) {
+    throw new Error(`Unterminated ${quote === '"' ? "double" : "single"} quote in pre-merge command`);
+  }
+  if (hasContent) tokens.push(current);
+  return tokens;
+}
+
+function containsUnquotedShellControl(input: string): boolean {
+  let i = 0;
+  let quote: "" | "'" | '"' = "";
+
+  while (i < input.length) {
+    const ch = input[i]!;
+    if (quote) {
+      if (ch === quote) {
+        quote = "";
+      } else if (ch === "\\" && quote === '"' && i + 1 < input.length) {
+        i += 2;
+        continue;
+      }
+      i++;
+      continue;
+    }
+
+    if (ch === '"' || ch === "'") {
+      quote = ch;
+      i++;
+      continue;
+    }
+    if (ch === "\\" && i + 1 < input.length) {
+      i += 2;
+      continue;
+    }
+    if (ch === ";" || ch === "&" || ch === "|" || ch === "`" || ch === "$" || ch === "<" || ch === ">") {
+      return true;
+    }
+    i++;
+  }
+
+  return false;
+}
+
 // ─── Git Helper ────────────────────────────────────────────────────────────
 
 
@@ -802,8 +907,34 @@ export class GitServiceImpl {
       }
     }
 
+    // Tokenize and run via execFileSync (no shell). Shell metacharacters in
+    // user-supplied prefs.pre_merge_check would otherwise be interpreted as
+    // chaining/redirection (e.g. `;`, `&&`, `|`, backticks) — a privesc
+    // surface in repos with a checked-in `.gsd/PREFERENCES.md`.
+    // (Issue #4980 HIGH-2)
+    if (containsUnquotedShellControl(command)) {
+      return {
+        passed: false,
+        skipped: false,
+        command,
+        error:
+          "pre_merge_check contains shell metacharacters (;, &&, |, $, backticks, redirects). " +
+          "Put complex commands in a script file (e.g. './scripts/pre-merge.sh') and reference the script path instead.",
+      };
+    }
+
+    const tokens = tokenizePreMergeCommand(command);
+    if (tokens.length === 0) {
+      return { passed: true, skipped: true };
+    }
+
     try {
-      execSync(command, { cwd: this.basePath, stdio: "pipe", encoding: "utf-8" });
+      execFileSync(tokens[0]!, tokens.slice(1), {
+        cwd: this.basePath,
+        stdio: "pipe",
+        encoding: "utf-8",
+        env: GIT_NO_PROMPT_ENV,
+      });
       return { passed: true, skipped: false, command };
     } catch (err) {
       const msg = getErrorMessage(err);

package/src/resources/extensions/gsd/gsd-db.ts

@@ -180,7 +180,7 @@ function openRawDb(path: string): unknown {
   return new Database(path);
 }
 
-const SCHEMA_VERSION = 22;
+export const SCHEMA_VERSION = 22;
 
 function indexExists(db: DbAdapter, name: string): boolean {
   return !!db.prepare(
@@ -2946,6 +2946,9 @@ export function insertAssessment(entry: {
   fullContent: string;
 }): void {
   if (!currentDb) throw new GSDError(GSD_STALE_STATE, "gsd-db: No database open");
+  // Idempotent: PRIMARY KEY is `path`, which is deterministic given (milestone_id, scope) per
+  // the artifact-path resolver. Retrying the same reassess-roadmap silently overwrites the row
+  // instead of accumulating duplicates.
   currentDb.prepare(
     `INSERT OR REPLACE INTO assessments (path, milestone_id, slice_id, task_id, status, scope, full_content, created_at)
      VALUES (:path, :milestone_id, :slice_id, :task_id, :status, :scope, :full_content, :created_at)`,
@@ -3100,7 +3103,7 @@ function rowToGate(row: Record<string, unknown>): GateRow {
     scope: row["scope"] as GateScope,
     task_id: (row["task_id"] as string) ?? "",
     status: row["status"] as GateStatus,
-    verdict: (row["verdict"] as GateVerdict) || "",
+    verdict: row["status"] === "pending" ? null : (row["verdict"] as GateVerdict),
     rationale: (row["rationale"] as string) || "",
     findings: (row["findings"] as string) || "",
     evaluated_at: (row["evaluated_at"] as string) ?? null,
@@ -3204,7 +3207,7 @@ export function getGateResults(milestoneId: string, sliceId: string, scope?: Gat
 export function markAllGatesOmitted(milestoneId: string, sliceId: string): void {
   if (!currentDb) return;
   currentDb.prepare(
-    `UPDATE quality_gates SET status = 'omitted', verdict = 'omitted', evaluated_at = :now
+    `UPDATE quality_gates SET status = 'complete', verdict = 'omitted', evaluated_at = :now
      WHERE milestone_id = :mid AND slice_id = :sid AND status = 'pending'`,
   ).run({
     ":mid": milestoneId,

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

@@ -39,7 +39,7 @@ import { ensureGitignore, ensurePreferences, untrackRuntimeFiles } from "./gitig
 import { loadEffectiveGSDPreferences } from "./preferences.js";
 import { resolveUokFlags } from "./uok/flags.js";
 import { ensurePlanV2Graph, isMissingFinalizedContextResult } from "./uok/plan-v2.js";
-import { detectProjectState } from "./detection.js";
+import { detectProjectState, hasGsdBootstrapArtifacts } from "./detection.js";
 import { showProjectInit, offerMigration } from "./init-wizard.js";
 import { validateDirectory } from "./validate-directory.js";
 import { showConfirm } from "../shared/tui.js";
@@ -315,12 +315,29 @@ function extractAssistantText(msg: any): string {
 
 /**
  * Return true if the assistant message contains any tool-use block.
+ *
+ * The canonical pi-ai `AssistantMessage.content` (see packages/pi-ai/src/types.ts)
+ * uses `type: "toolCall"` and `type: "serverToolUse"` for tool invocations —
+ * every provider (anthropic-direct, claude-code-cli, openai, etc.) normalizes
+ * incoming tool blocks into these two shapes before they reach guided-flow.
+ *
+ * The Anthropic API wire shape `"tool_use"` / `"server_tool_use"` does NOT appear
+ * in the internal AssistantMessage — those literals are only used when sending
+ * messages back out to the Anthropic API. Matching them here was a latent bug:
+ * `hasToolUse` returned `false` for every real tool call, which let the
+ * empty-turn nudge fire and pre-empt MCP tools that block on the user
+ * (e.g. `ask_user_questions`). See investigation in PR for #4658.
  */
 function hasToolUse(msg: any): boolean {
   if (!msg) return false;
   const content = msg.content;
   if (!Array.isArray(content)) return false;
-  return content.some((b: any) => b && typeof b === "object" && (b.type === "tool_use" || b.type === "tool-use"));
+  return content.some(
+    (b: any) =>
+      b &&
+      typeof b === "object" &&
+      (b.type === "toolCall" || b.type === "serverToolUse"),
+  );
 }
 
 /**
@@ -1491,9 +1508,7 @@ export async function showSmartEntry(
   // A zombie .gsd/ state (symlink exists but missing PREFERENCES.md and
   // milestones/) must trigger the init wizard, not skip it (#2942).
   const gsdPath = gsdRoot(basePath);
-  const hasBootstrapArtifacts = existsSync(gsdPath)
-    && (existsSync(join(gsdPath, "PREFERENCES.md"))
-        || existsSync(join(gsdPath, "milestones")));
+  const hasBootstrapArtifacts = hasGsdBootstrapArtifacts(gsdPath);
 
   if (!hasBootstrapArtifacts) {
     const detection = detectProjectState(basePath);

package/src/resources/extensions/gsd/memory-extractor.ts

@@ -18,7 +18,10 @@ import type { MemoryAction } from './memory-store.js';
 
 // ─── Types ──────────────────────────────────────────────────────────────────
 
-export type LLMCallFn = (system: string, user: string) => Promise<string>;
+export type LLMCallFn = ((system: string, user: string) => Promise<string>) & {
+  /** Promise resolving once the provider API key has been fetched (for tests). */
+  apiKeyReady?: Promise<string | undefined>;
+};
 
 // ─── Concurrency Guard ──────────────────────────────────────────────────────
 
@@ -92,8 +95,9 @@ export function buildMemoryLLMCall(ctx: ExtensionContext): LLMCallFn | null {
     // which returns undefined for OAuth users (Claude Max / Claude Pro).
     // See: https://github.com/gsd-build/gsd-2/issues/2959
     const resolvedKeyPromise = ctx.modelRegistry.getApiKey(selectedModel).catch(() => undefined);
-
-    return async (system: string, user: string): Promise<string> => {
+    // Expose on the returned fn so tests can await resolution deterministically
+    // (avoids arbitrary setTimeout polling for an internal microtask).
+    const llmCall = async (system: string, user: string): Promise<string> => {
       const { completeSimple } = await import('@gsd/pi-ai');
       const resolvedApiKey = await resolvedKeyPromise;
       const result: AssistantMessage = await completeSimple(selectedModel, {
@@ -111,6 +115,10 @@ export function buildMemoryLLMCall(ctx: ExtensionContext): LLMCallFn | null {
         .map(c => c.text);
       return textParts.join('');
     };
+    // Attach the in-flight API-key resolution so tests (and callers) can
+    // `await llmCall.apiKeyReady` rather than relying on setTimeout polling.
+    (llmCall as LLMCallFn & { apiKeyReady?: Promise<string | undefined> }).apiKeyReady = resolvedKeyPromise;
+    return llmCall;
   } catch {
     return null;
   }

package/src/resources/extensions/gsd/milestone-scope-classifier.ts

@@ -0,0 +1,366 @@
+// GSD-2 — Milestone scope classifier (#4781 / ADR-003 companion).
+//
+// Pure heuristics over milestone planning fields. Produces a PipelineVariant
+// that downstream dispatch logic can use to shape the auto-mode sequence.
+// No LLM calls, no file I/O, sub-millisecond.
+//
+// Distinct from `complexity-classifier.ts`, which decides *model tier*
+// (light/standard/heavy) for an individual unit. This module decides
+// *pipeline topology* for an entire milestone at plan-milestone time.
+//
+// This file ships the classifier in isolation. Dispatch-side wiring
+// lands in follow-up PRs so the classification contract can be reviewed
+// and tested before any behavior change reaches users.
+
+export type PipelineVariant = "trivial" | "standard" | "complex";
+
+export interface MilestoneScopeInput {
+  /** Milestone vision / elevator pitch. Free-form prose. */
+  vision?: string;
+  /** Success criteria, one per array entry. */
+  successCriteria?: string[];
+  /** Milestone title. */
+  title?: string;
+  /** Slice risks declared at plan-milestone time. */
+  keyRisks?: Array<{ risk?: string; whyItMatters?: string }>;
+  /** Definition-of-done lines. */
+  definitionOfDone?: string[];
+  /** Freeform "requirement coverage" marker. */
+  requirementCoverage?: string;
+  /** Verification hints (contract/integration/operational/uat). */
+  verificationContract?: string;
+  verificationIntegration?: string;
+  verificationOperational?: string;
+  verificationUat?: string;
+}
+
+export interface ScopeClassificationResult {
+  variant: PipelineVariant;
+  /** Short human-readable reasons, one per triggered signal. */
+  reasons: string[];
+  /** Sub-signals for telemetry / debugging. Stable across releases. */
+  signals: {
+    triggeredOverride: boolean;
+    complexCount: number;
+    trivialCount: number;
+    fileCountHint: number | null;
+  };
+}
+
+// ─── Keyword sets ─────────────────────────────────────────────────────────
+
+/**
+ * Override keywords that force `standard` (at minimum) regardless of
+ * apparent triviality. Presence of any of these signals work that is
+ * either security-sensitive, irreversible, or requires runtime verification
+ * a "trivial" pipeline would skip.
+ *
+ * Matched as case-insensitive word-boundary substrings. Conservative — err
+ * on the side of including a keyword; over-classifying to `standard` costs
+ * units, under-classifying could ship broken auth/security/migration work.
+ */
+const OVERRIDE_KEYWORDS: ReadonlyArray<string> = [
+  // Security-sensitive
+  "security", "auth", "authn", "authz", "authentication", "authorization",
+  "credential", "secret", "password", "token", "oauth", "encrypt", "decrypt",
+  "vulnerability", "exploit", "permission", "rbac", "acl",
+  // Data-migration / irreversible
+  "migration", "migrate", "schema change", "data migration",
+  "backfill", "drop column", "drop table",
+  // Compliance / regulatory
+  "compliance", "gdpr", "hipaa", "soc2", "pci",
+  // Infra / deploy — runtime verification needed
+  "deploy", "rollout", "canary", "production database",
+];
+
+/**
+ * Keywords that contribute to `complex` classification on their own.
+ * Different from OVERRIDE_KEYWORDS in that a single match bumps to
+ * complex, not just to standard.
+ */
+const COMPLEX_KEYWORDS: ReadonlyArray<string> = [
+  "multi-service", "distributed", "consensus", "saga", "eventual consistency",
+  "breaking change", "api contract change", "schema redesign",
+  "architect", "architecture", "refactor core",
+];
+
+/**
+ * Trivial-signal keywords: presence strongly suggests a simple, contained
+ * deliverable. Only effective when combined with low file count / no tests
+ * / no override keywords.
+ */
+const TRIVIAL_KEYWORDS: ReadonlyArray<string> = [
+  "single file", "one file", "static html", "static page",
+  "one-page", "landing page", "readme", "docs only", "typo", "rename",
+  "spelling", "comment", "changelog",
+  // Browser-only / no-build deliverable shapes (b23 forensic case).
+  "pure html", "browser-based", "no build step", "no build tooling",
+  "localstorage", "client-only", "no backend", "no server", "no backend.",
+];
+
+// ─── Heuristics ───────────────────────────────────────────────────────────
+
+/**
+ * Estimate how many distinct files the milestone will touch, based on
+ * explicit mentions in the input text. Returns `null` when no hint is
+ * discoverable — callers should treat that as "unknown, no signal."
+ */
+function extractFileCountHint(text: string): number | null {
+  // Explicit phrasing: "a single file", "two files", "3 files"
+  const singleFileMatch = /\b(a|one|single)\s+(file|page)\b/i.test(text);
+  if (singleFileMatch) return 1;
+
+  const digitMatch = text.match(/\b(\d+)\s+files?\b/i);
+  if (digitMatch) {
+    const n = parseInt(digitMatch[1], 10);
+    if (!Number.isNaN(n)) return n;
+  }
+
+  const wordMatch = text.match(/\b(two|three|four|five|six|seven|eight|nine|ten)\s+files?\b/i);
+  if (wordMatch) {
+    const wordMap: Record<string, number> = {
+      two: 2, three: 3, four: 4, five: 5,
+      six: 6, seven: 7, eight: 8, nine: 9, ten: 10,
+    };
+    return wordMap[wordMatch[1].toLowerCase()] ?? null;
+  }
+
+  return null;
+}
+
+function escapeRegExp(value: string): string {
+  return value.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+}
+
+function containsAnyKeyword(haystack: string, keywords: ReadonlyArray<string>): string[] {
+  const lower = haystack.toLowerCase();
+  const hits: string[] = [];
+  for (const kw of keywords) {
+    // Word-boundary match to prevent substring collisions (e.g. "auth"
+    // must not match "author", "api" must not match "capital"). Phrases
+    // containing non-word characters (hyphens, slashes) still work because
+    // `\b` sits at the word-char / non-word-char transition, so
+    // `\bbrowser-based\b` matches "browser-based" bounded by whitespace
+    // or punctuation on either side.
+    const pattern = new RegExp(String.raw`\b${escapeRegExp(kw)}\b`, "i");
+    if (pattern.test(lower)) hits.push(kw);
+  }
+  return hits;
+}
+
+/**
+ * True when `term` appears in the text without an immediately preceding
+ * negator (no / without / not / zero / skip) in the same clause. Used to
+ * keep phrases like "no backend" or "no tests" from flipping a trivial-
+ * class milestone to standard. Best-effort; imperfect English parsing,
+ * biased toward false negatives (if unsure, treats term as present —
+ * which routes to standard, the safe pipeline).
+ */
+function mentionsWithoutNegation(text: string, term: string): boolean {
+  const lower = text.toLowerCase();
+  const termPattern = new RegExp(String.raw`\b${term}\b`, "gi");
+  const matches = Array.from(lower.matchAll(termPattern));
+  for (const m of matches) {
+    const start = m.index ?? 0;
+    const windowStart = Math.max(0, start - 30);
+    const window = lower.slice(windowStart, start);
+    // Negator anywhere in the 30-char lookback window counts as negation —
+    // covers "no backend", "without a server", "not using api", "zero
+    // dependencies on an api". If a sentence break intervenes between the
+    // negator and the term, treat as a different clause (positive mention).
+    const hasNegator = /(^|[^a-z0-9])(no|without|not|zero|skip(s|ping)?|drops?)\b/i.test(window);
+    const hasSentenceBreak = /[.;!?]/.test(window);
+    if (hasNegator && !hasSentenceBreak) continue;
+    return true;
+  }
+  return false;
+}
+
+function mentionsTests(haystack: string): boolean {
+  return mentionsWithoutNegation(haystack, "test")
+      || mentionsWithoutNegation(haystack, "tests")
+      || mentionsWithoutNegation(haystack, "testing")
+      || mentionsWithoutNegation(haystack, "spec")
+      || mentionsWithoutNegation(haystack, "unit test")
+      || mentionsWithoutNegation(haystack, "integration test");
+}
+
+function mentionsBackend(haystack: string): boolean {
+  return mentionsWithoutNegation(haystack, "api")
+      || mentionsWithoutNegation(haystack, "backend")
+      || mentionsWithoutNegation(haystack, "server")
+      || mentionsWithoutNegation(haystack, "database")
+      || mentionsWithoutNegation(haystack, "endpoint");
+}
+
+// ─── DB adapter ───────────────────────────────────────────────────────────
+
+/**
+ * Shape adapter: convert a milestone DB row into the classifier input
+ * object. Keeps `milestone-scope-classifier.ts` free of DB types at the
+ * module boundary — callers can either use this helper or hand-build the
+ * input themselves (e.g. from plan-milestone tool params before insert).
+ */
+export function milestoneRowToScopeInput(row: {
+  title?: string;
+  vision?: string;
+  success_criteria?: string[];
+  key_risks?: Array<{ risk?: string; whyItMatters?: string }>;
+  definition_of_done?: string[];
+  requirement_coverage?: string;
+  verification_contract?: string;
+  verification_integration?: string;
+  verification_operational?: string;
+  verification_uat?: string;
+}): MilestoneScopeInput {
+  return {
+    title: row.title,
+    vision: row.vision,
+    successCriteria: row.success_criteria,
+    keyRisks: row.key_risks,
+    definitionOfDone: row.definition_of_done,
+    requirementCoverage: row.requirement_coverage,
+    verificationContract: row.verification_contract,
+    verificationIntegration: row.verification_integration,
+    verificationOperational: row.verification_operational,
+    verificationUat: row.verification_uat,
+  };
+}
+
+/**
+ * Compute the pipeline variant for a milestone by reading its planning
+ * fields from the DB and running the classifier. Returns `null` when
+ * classification is unavailable (DB closed, milestone missing, unexpected
+ * error) — callers MUST treat null as "run the full pipeline" so a
+ * classification failure never silently downshifts dispatch.
+ */
+export async function getMilestonePipelineVariant(mid: string): Promise<PipelineVariant | null> {
+  try {
+    const { isDbAvailable, getMilestone } = await import("./gsd-db.js");
+    if (!isDbAvailable()) return null;
+    const row = getMilestone(mid);
+    if (!row) return null;
+    return classifyMilestoneScope(milestoneRowToScopeInput(row)).variant;
+  } catch {
+    return null;
+  }
+}
+
+// ─── Public API ───────────────────────────────────────────────────────────
+
+/**
+ * Classify a milestone's pipeline variant based on its planning inputs.
+ *
+ * Precedence (matches implementation order — complex-first so that
+ * security-sensitive architecture refactors correctly route to complex
+ * rather than standard; the override hit is still recorded in
+ * `signals.triggeredOverride` for telemetry):
+ *  1. Complex-signal keyword OR ≥ 8 file hint OR architecture/refactor-core
+ *     language → `complex`.
+ *  2. Override keyword → `standard` (at minimum). Prevents trivial
+ *     misclassification of security / auth / migration work.
+ *  3. Trivial-signal keyword AND ≤ 2 file hint AND no tests mentioned AND
+ *     no backend mentioned → `trivial`.
+ *  4. Otherwise → `standard`.
+ *
+ * Ambiguity → `standard` (today's default). Safe to run the full pipeline.
+ */
+export function classifyMilestoneScope(input: MilestoneScopeInput): ScopeClassificationResult {
+  const haystack = [
+    input.title ?? "",
+    input.vision ?? "",
+    (input.successCriteria ?? []).join("\n"),
+    (input.keyRisks ?? []).map(r => `${r.risk ?? ""} ${r.whyItMatters ?? ""}`).join("\n"),
+    (input.definitionOfDone ?? []).join("\n"),
+    input.requirementCoverage ?? "",
+    input.verificationContract ?? "",
+    input.verificationIntegration ?? "",
+    input.verificationOperational ?? "",
+    input.verificationUat ?? "",
+  ].join("\n");
+
+  const overrideHits = containsAnyKeyword(haystack, OVERRIDE_KEYWORDS);
+  const complexHits = containsAnyKeyword(haystack, COMPLEX_KEYWORDS);
+  const trivialHits = containsAnyKeyword(haystack, TRIVIAL_KEYWORDS);
+  const fileCountHint = extractFileCountHint(haystack);
+  const hasTests = mentionsTests(haystack);
+  const hasBackend = mentionsBackend(haystack);
+
+  const reasons: string[] = [];
+
+  // Rule 2: complex-class signals. Evaluated before override because a
+  // complex + override input should land in complex, not standard.
+  if (complexHits.length > 0) {
+    reasons.push(`complex keywords: ${complexHits.slice(0, 3).join(", ")}`);
+  }
+  if (fileCountHint !== null && fileCountHint >= 8) {
+    reasons.push(`file count hint: ${fileCountHint}`);
+  }
+
+  const isComplex = complexHits.length > 0 || (fileCountHint !== null && fileCountHint >= 8);
+
+  if (isComplex) {
+    return {
+      variant: "complex",
+      reasons,
+      signals: {
+        triggeredOverride: overrideHits.length > 0,
+        complexCount: complexHits.length,
+        trivialCount: trivialHits.length,
+        fileCountHint,
+      },
+    };
+  }
+
+  // Rule 1: override keywords force standard.
+  if (overrideHits.length > 0) {
+    return {
+      variant: "standard",
+      reasons: [`override keywords: ${overrideHits.slice(0, 3).join(", ")}`],
+      signals: {
+        triggeredOverride: true,
+        complexCount: complexHits.length,
+        trivialCount: trivialHits.length,
+        fileCountHint,
+      },
+    };
+  }
+
+  // Rule 3: trivial signals — require ALL of: trivial-keyword, low file
+  // hint (or nothing suggesting high count), no test mention, no backend
+  // mention.
+  const fileCountOk = fileCountHint === null || fileCountHint <= 2;
+  const trivial =
+    trivialHits.length > 0 &&
+    fileCountOk &&
+    !hasTests &&
+    !hasBackend;
+
+  if (trivial) {
+    reasons.push(`trivial keywords: ${trivialHits.slice(0, 3).join(", ")}`);
+    if (fileCountHint !== null) reasons.push(`file count hint: ${fileCountHint}`);
+    reasons.push("no tests mentioned", "no backend mentioned");
+    return {
+      variant: "trivial",
+      reasons,
+      signals: {
+        triggeredOverride: false,
+        complexCount: complexHits.length,
+        trivialCount: trivialHits.length,
+        fileCountHint,
+      },
+    };
+  }
+
+  // Rule 4: fallback.
+  return {
+    variant: "standard",
+    reasons: reasons.length > 0 ? reasons : ["no strong signals — default"],
+    signals: {
+      triggeredOverride: overrideHits.length > 0,
+      complexCount: complexHits.length,
+      trivialCount: trivialHits.length,
+      fileCountHint,
+    },
+  };
+}

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

@@ -57,6 +57,9 @@ export const BUNDLED_COST_TABLE: ModelCostEntry[] = [
   { id: "gpt-5.3-codex-spark", inputPer1k: 0.0003, outputPer1k: 0.0012, updatedAt: "2026-03-29" },
   { id: "gpt-5.4", inputPer1k: 0.005, outputPer1k: 0.02, updatedAt: "2026-03-29" },
   { id: "gpt-5.4-mini", inputPer1k: 0.00075, outputPer1k: 0.0045, updatedAt: "2026-04-18" },
+  // 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" },
 
   // Google
   { id: "gemini-2.0-flash", inputPer1k: 0.0001, outputPer1k: 0.0004, updatedAt: "2025-03-15" },

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

@@ -109,6 +109,7 @@ export const MODEL_CAPABILITY_TIER: Record<string, ComplexityTier> = {
   "gpt-5.2-codex": "heavy",
   "gpt-5.3-codex": "heavy",
   "gpt-5.4": "heavy",
+  "gpt-5.5": "heavy",
   "o1": "heavy",
   "o3": "heavy",
   "o4-mini": "heavy",
@@ -144,6 +145,7 @@ const MODEL_COST_PER_1K_INPUT: Record<string, number> = {
   "gpt-5.3-codex": 0.005,
   "gpt-5.3-codex-spark": 0.0003,
   "gpt-5.4": 0.005,
+  "gpt-5.5": 0.005,
   "o4-mini": 0.005,
   "o4-mini-deep-research": 0.005,
   "gemini-2.0-flash": 0.0001,
@@ -187,6 +189,10 @@ export const MODEL_CAPABILITY_PROFILES: Record<string, ModelCapabilities> = {
   "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 },
+  // GPT-5.5 scores are relative to the existing gpt-5.4 profile and backed by
+  // 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 },
 
   // ── OpenAI o-series (reasoning-first) ──────────────────────────────────────
   "o1":                           { coding: 78, debugging: 82, research: 78, reasoning: 90, speed: 20, longContext: 65, instruction: 82 },