@mmerterden/multi-agent-pipeline 8.6.1 → 10.0.6

package/pipeline/scripts/run-smokes.mjs

@@ -0,0 +1,76 @@
+#!/usr/bin/env node
+// Structured runner for pipeline/scripts/smoke-*.sh.
+//
+// Replaces the bare `for f in smoke-*.sh; do bash "$f" || exit 1; done`
+// loop with per-suite timeouts, timing, and a summary, so a hung smoke
+// (network wait, interactive prompt) can never block `npm test` or CI
+// indefinitely.
+//
+// Env:
+//   SMOKE_TIMEOUT_SECONDS  per-suite timeout (default 180)
+//   SMOKE_NO_BAIL=1        run every suite and summarize instead of
+//                          stopping at the first failure
+// Args:
+//   optional substring filters: `node run-smokes.mjs adapters keychain`
+//   runs only suites whose filename contains one of the terms.
+
+import { spawnSync } from "node:child_process";
+import { readdirSync } from "node:fs";
+import { dirname, join } from "node:path";
+import { fileURLToPath } from "node:url";
+
+const scriptsDir = dirname(fileURLToPath(import.meta.url));
+const timeoutSec = Number(process.env.SMOKE_TIMEOUT_SECONDS || 180);
+const bail = process.env.SMOKE_NO_BAIL !== "1";
+const filters = process.argv.slice(2);
+
+const suites = readdirSync(scriptsDir)
+  .filter((f) => f.startsWith("smoke-") && f.endsWith(".sh"))
+  .filter((f) => filters.length === 0 || filters.some((t) => f.includes(t)))
+  .sort();
+
+if (suites.length === 0) {
+  console.error("run-smokes: no smoke suites matched");
+  process.exit(1);
+}
+
+const failures = [];
+let ran = 0;
+const startedAt = Date.now();
+
+for (const suite of suites) {
+  ran++;
+  const t0 = Date.now();
+  const res = spawnSync("bash", [join(scriptsDir, suite)], {
+    encoding: "utf-8",
+    timeout: timeoutSec * 1000,
+    maxBuffer: 16 * 1024 * 1024,
+  });
+  const secs = ((Date.now() - t0) / 1000).toFixed(1);
+
+  const timedOut = res.error && res.error.code === "ETIMEDOUT";
+  const ok = !timedOut && res.status === 0;
+
+  if (ok) {
+    console.log(`ok   ${suite} (${secs}s)`);
+    continue;
+  }
+
+  const reason = timedOut ? `timeout after ${timeoutSec}s` : `exit ${res.status}`;
+  console.error(`FAIL ${suite} (${reason}, ${secs}s)`);
+  if (res.stdout) console.error(res.stdout.split("\n").slice(-30).join("\n"));
+  if (res.stderr) console.error(res.stderr.split("\n").slice(-30).join("\n"));
+  failures.push(suite);
+  if (bail) break;
+}
+
+const totalSecs = ((Date.now() - startedAt) / 1000).toFixed(1);
+const skipped = suites.length - ran;
+console.log(
+  `run-smokes: ${ran - failures.length}/${suites.length} passed in ${totalSecs}s` +
+    (skipped > 0 ? ` (${skipped} not run after bail)` : ""),
+);
+if (failures.length > 0) {
+  console.error(`run-smokes: failed -> ${failures.join(", ")}`);
+  process.exit(1);
+}

package/pipeline/scripts/scan-skills.sh

@@ -1,5 +1,5 @@
 #!/usr/bin/env bash
-# scan-skills.sh — skill security scanner (v5.1.0)
+# scan-skills.sh  -  skill security scanner (v5.1.0)
 #
 # Scans skill directories (pipeline/skills/ OR ~/.claude/skills/ OR ~/.copilot/skills/)
 # for known-bad patterns. Multi-tier severity:
@@ -51,7 +51,7 @@ if [ ! -d "$ROOT" ]; then
   exit 64
 fi
 
-# Severity rank — lower = more severe. Used for threshold comparison.
+# Severity rank  -  lower = more severe. Used for threshold comparison.
 sev_rank() {
   case "$1" in
     critical) echo 0 ;;
@@ -63,7 +63,7 @@ sev_rank() {
 }
 THRESHOLD_RANK=$(sev_rank "$THRESHOLD")
 
-# Network endpoint allow-list — domains we expect to see in skills.
+# Network endpoint allow-list  -  domains we expect to see in skills.
 # Anything else flagged as medium (unknown endpoint).
 ALLOW_DOMAINS='(github\.com|raw\.githubusercontent\.com|api\.github\.com|codeload\.github\.com|githubusercontent\.com|anthropic\.com|claude\.ai|api\.anthropic\.com|api\.figma\.com|figma\.com|jira\.example\.com|confluence\.example\.com|bitbucket\.example\.com|example\.com|mmerterden\.vercel\.app|localhost|127\.0\.0\.1|npmjs\.org|npmjs\.com|registry\.npmjs\.org|vercel\.com|api\.vercel\.com|pkg\.github\.com|googleapis\.com|firebase\.google\.com|apple\.com|developer\.apple\.com|google\.com|atlassian\.com|atlassian\.net|openai\.com|api\.openai\.com)'
 
@@ -121,9 +121,9 @@ scan_critical() {
     add_finding critical "$f" "$ln" "eval-of-network" "eval of network-fetched content"
   done < <(grep -nE 'eval[[:space:]]+(\$\([[:space:]]*(curl|wget|fetch)|`[[:space:]]*(curl|wget|fetch))' "$f" 2>/dev/null)
 
-  # Unicode bidi override characters — invisible injection attack
+  # Unicode bidi override characters  -  invisible injection attack
   if LC_ALL=C grep -l $'\xe2\x80\xad\|\xe2\x80\xae\|\xe2\x81\xa6\|\xe2\x81\xa7\|\xe2\x81\xa8\|\xe2\x81\xa9' "$f" 2>/dev/null >/dev/null; then
-    add_finding critical "$f" "0" "unicode-bidi" "bidirectional control chars (U+202D-202E, U+2066-2069) — trojan source risk"
+    add_finding critical "$f" "0" "unicode-bidi" "bidirectional control chars (U+202D-202E, U+2066-2069)  -  trojan source risk"
   fi
 }
 
@@ -140,7 +140,7 @@ scan_high() {
     done < <(grep -nE '\b(eval|new[[:space:]]+Function|Function)\s*\(' "$f" 2>/dev/null)
   fi
 
-  # Python: exec(, eval( on non-literal — exclude re.compile (regex) and subprocess
+  # Python: exec(, eval( on non-literal  -  exclude re.compile (regex) and subprocess
   if [ "$ext" = "py" ]; then
     while IFS= read -r hit; do
       [ -z "$hit" ] && continue
@@ -152,7 +152,7 @@ scan_high() {
     done < <(grep -nE '(^|[^a-zA-Z0-9_.])(exec|eval)[[:space:]]*\(' "$f" 2>/dev/null)
   fi
 
-  # Hardcoded API keys — AWS, OpenAI, GitHub, generic sk-* with length
+  # Hardcoded API keys  -  AWS, OpenAI, GitHub, generic sk-* with length
   # Skip lines in FORBIDDEN/NEVER/example blocks (false positives from docs)
   while IFS= read -r hit; do
     [ -z "$hit" ] && continue
@@ -187,8 +187,8 @@ scan_high() {
 scan_medium() {
   local f="$1"
 
-  # Long base64 blobs — obfuscation indicator.
-  # Skip .md files (documentation often shows example b64) — only scan executable content types.
+  # Long base64 blobs  -  obfuscation indicator.
+  # Skip .md files (documentation often shows example b64)  -  only scan executable content types.
   case "$f" in
     *.md) ;;
     *)
@@ -204,7 +204,7 @@ scan_medium() {
 scan_info() {
   local f="$1"
 
-  # Unknown network endpoints — LOW severity (informational). URLs in skills
+  # Unknown network endpoints  -  LOW severity (informational). URLs in skills
   # are usually doc references, not exfil channels. Only the critical
   # shell-pipe-exec patterns detect active network abuse.
   while IFS= read -r hit; do
@@ -314,7 +314,7 @@ else
   fi
 
   if [ "$STRICT" -eq 0 ]; then
-    printf '\n  %s(warn-only mode — exit 0 regardless of findings; use --strict to halt on findings)%s\n' "$C_DIM" "$C_RST"
+    printf '\n  %s(warn-only mode  -  exit 0 regardless of findings; use --strict to halt on findings)%s\n' "$C_DIM" "$C_RST"
   fi
 fi
 

package/pipeline/scripts/search-logs.sh

@@ -1,5 +1,5 @@
 #!/usr/bin/env bash
-# search-logs.sh — cross-project search across multi-agent agent-log.md files.
+# search-logs.sh  -  cross-project search across multi-agent agent-log.md files.
 #
 # Greps ~/.claude/logs/multi-agent/**/agent-log.md (or --root PATH) for a query,
 # groups results by task, and ranks by (match_count × recency_weight).
@@ -12,7 +12,7 @@
 # Flags:
 #   --root PATH           Override log root. Default: $HOME/.claude/logs/multi-agent
 #   --project NAME        Filter to tasks under this project dir
-#   --since SPEC          7d | 30d | 2026-04-01 — drop tasks older than this
+#   --since SPEC          7d | 30d | 2026-04-01  -  drop tasks older than this
 #   --phase N             Only match inside "## Phase N" sections
 #   --semantic            v8.3+: query the per-repo triage corpus instead of
 #                         agent-log.md grep. Returns past triage findings
@@ -73,7 +73,7 @@ if [ -z "$QUERY" ]; then
   exit 64
 fi
 
-# v8.3+ — semantic mode: dispatch to triage-memory.mjs corpus query and exit.
+# v8.3+  -  semantic mode: dispatch to triage-memory.mjs corpus query and exit.
 # Honors --project as the repo-slug, --json (default off → human table),
 # --limit (passed as --top N). --since/--phase have no meaning here.
 if [ "$SEMANTIC" -eq 1 ]; then
@@ -147,7 +147,7 @@ while IFS= read -r f; do
   fi
   # Since filter: drop files with mtime older than since_epoch
   if [ "$since_epoch" -gt 0 ]; then
-    m=$(stat -f %m "$f" 2>/dev/null || stat -c %Y "$f" 2>/dev/null || echo 0)
+    m=$(stat -c %Y "$f" 2>/dev/null || stat -f %m "$f" 2>/dev/null || echo 0)
     [ "$m" -lt "$since_epoch" ] && continue
   fi
   echo "$f" >> "$FILTERED"
@@ -200,7 +200,7 @@ while IFS= read -r f; do
   title=$(head -1 "$f" | sed 's/^# *//' | cut -c1-80)
   branch=$(grep -m1 -iE '^\*\*Branch\*\*' "$f" 2>/dev/null | sed -E 's/.*Branch\*\*: *([^ |]+).*/\1/' | head -c60)
   [ -z "$branch" ] && branch="(unknown)"
-  mtime=$(stat -f %m "$f" 2>/dev/null || stat -c %Y "$f" 2>/dev/null || echo "$NOW")
+  mtime=$(stat -c %Y "$f" 2>/dev/null || stat -f %m "$f" 2>/dev/null || echo "$NOW")
 
   # Recency weight: 1.0 for last 7d, 0.5 for 7-30d, 0.2 older
   age_days=$(( (NOW - mtime) / 86400 ))
@@ -241,7 +241,7 @@ if [ "$JSON" -eq 1 ]; then
     i=$((i+1))
     jt=$(printf '%s' "$title" | sed 's/\\/\\\\/g; s/"/\\"/g')
     js=$(printf '%s' "$snippet" | sed 's/\\/\\\\/g; s/"/\\"/g')
-    date=$(date -r "$mtime" "+%Y-%m-%d" 2>/dev/null || echo "")
+    date=$(date -r "$mtime" "+%Y-%m-%d" 2>/dev/null || date -d "@$mtime" "+%Y-%m-%d" 2>/dev/null || echo "")
     printf '    { "task_id": "%s", "project": "%s", "branch": "%s", "date": "%s", "matches": %d, "score": %d, "title": "%s", "snippet": "%s" }' \
       "$task_id" "$project" "$branch" "$date" "$count" "$score" "$jt" "$js"
     [ "$i" -lt "$TOTAL_TASKS" ] && printf ','
@@ -251,7 +251,7 @@ if [ "$JSON" -eq 1 ]; then
 elif [ "$TSV" -eq 1 ]; then
   printf 'score\ttask_id\tproject\tbranch\tdate\tmatches\ttitle\tsnippet\n'
   while IFS=$'\t' read -r score task_id project branch mtime count title snippet; do
-    date=$(date -r "$mtime" "+%Y-%m-%d" 2>/dev/null || echo "")
+    date=$(date -r "$mtime" "+%Y-%m-%d" 2>/dev/null || date -d "@$mtime" "+%Y-%m-%d" 2>/dev/null || echo "")
     printf '%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\n' "$score" "$task_id" "$project" "$branch" "$date" "$count" "$title" "$snippet"
   done < "$RESULTS"
 else
@@ -280,7 +280,7 @@ else
   printf '\n'
 
   while IFS=$'\t' read -r score task_id project branch mtime count title snippet; do
-    date=$(date -r "$mtime" "+%Y-%m-%d" 2>/dev/null || echo "")
+    date=$(date -r "$mtime" "+%Y-%m-%d" 2>/dev/null || date -d "@$mtime" "+%Y-%m-%d" 2>/dev/null || echo "")
     printf '%s%s%s · %s%s%s · %s%s%s · %s%d match%s%s\n' \
       "$C_YEL" "$task_id" "$C_RST" "$C_DIM" "$project" "$C_RST" "$C_DIM" "$branch" "$C_RST" "$C_GRN" "$count" "$([ "$count" -eq 1 ] || echo es)" "$C_RST"
     [ -n "$title" ] && printf '  %s%s%s\n' "$C_BLD" "$title" "$C_RST"

package/pipeline/scripts/sign-skills.sh

@@ -1,8 +1,8 @@
 #!/usr/bin/env bash
-# sign-skills.sh — v6.2.H
+# sign-skills.sh  -  v6.2.H
 #
 # Generates a SHA-256 manifest of every SKILL.md under pipeline/skills/.
-# Output: pipeline/skills/.skill-manifest.json (in the authored tree — becomes
+# Output: pipeline/skills/.skill-manifest.json (in the authored tree  -  becomes
 # part of what install.js copies, so the deployed dirs get the manifest too).
 #
 # The manifest is NOT a security boundary. It's a tamper-detection signal:

package/pipeline/scripts/smoke-adapters.sh

@@ -1,5 +1,5 @@
 #!/usr/bin/env bash
-# smoke-adapters.sh — third-party AI tool adapters (cursor + copilot-chat).
+# smoke-adapters.sh  -  third-party AI tool adapters (cursor + copilot-chat).
 #
 # As of v8.5.4 the supported set narrowed to the two adapters the pipeline
 # owner actually uses: Cursor (knowledge layer) and GitHub Copilot Chat
@@ -7,12 +7,12 @@
 #
 # Round-trip integrity:
 #   1. Install writes the expected layout (file counts, dir structure)
-#   2. Orchestration skills (multi-agent-*) are NOT exported — knowledge layer only
+#   2. Orchestration skills (multi-agent-*) are NOT exported  -  knowledge layer only
 #   3. Uninstall removes ONLY pipeline-managed artifacts
 #   4. Marker-wrapped block (copilot-chat) preserves user-authored content
-#   5. Idempotent install — running twice produces identical tree
+#   5. Idempotent install  -  running twice produces identical tree
 #   6. Platform filter (--platform=ios|android) narrows the skill set
-#   7. Default behavior — no flags installs Claude only (backward compat)
+#   7. Default behavior  -  no flags installs Claude only (backward compat)
 #
 # All work happens in mktemp dirs; the user's actual project root is never
 # touched.
@@ -30,13 +30,14 @@ fail() { FAIL=$((FAIL + 1)); echo "  ✗ $1"; }
 cleanup() {
   [ -n "${TMP_CURSOR:-}" ] && rm -rf "$TMP_CURSOR"
   [ -n "${TMP_COPILOTCHAT:-}" ] && rm -rf "$TMP_COPILOTCHAT"
+  [ -n "${TMP_ANTIGRAVITY:-}" ] && rm -rf "$TMP_ANTIGRAVITY"
   [ -n "${TMP_IDEMP:-}" ] && rm -rf "$TMP_IDEMP"
   [ -n "${TMP_FILTER:-}" ] && rm -rf "$TMP_FILTER"
 }
 trap cleanup EXIT
 
 # ──────────────────────────────────────────────────────────────────────────
-echo "→ 1. Cursor adapter — install + uninstall round-trip"
+echo "→ 1. Cursor adapter  -  install + uninstall round-trip"
 TMP_CURSOR=$(mktemp -d)
 node "$INSTALL" --cursor --target="$TMP_CURSOR" >/dev/null 2>&1 || fail "cursor install exited non-zero"
 
@@ -59,7 +60,7 @@ else
   fail "cursor missing .cursor/rules/ directory"
 fi
 
-# Orchestration filter — multi-agent-{dev,review,...} should NOT be exported
+# Orchestration filter  -  multi-agent-{dev,review,...} should NOT be exported
 orch_leaked=$(find "$TMP_CURSOR/.cursor/rules" -name "multi-agent-multi-agent*" -o -name "multi-agent-dev.mdc" -o -name "multi-agent-review.mdc" 2>/dev/null | wc -l | tr -d ' ')
 if [ "$orch_leaked" -eq 0 ]; then
   pass "cursor: orchestration commands correctly filtered out"
@@ -67,6 +68,24 @@ else
   fail "cursor: $orch_leaked orchestration command(s) leaked into knowledge layer"
 fi
 
+# Orchestration layer  -  Cursor subagents + /multi-agent command + MCP.
+ma_agents=$(find "$TMP_CURSOR/.cursor/agents" -name "ma-*.md" 2>/dev/null | wc -l | tr -d ' ')
+if [ "$ma_agents" -ge 8 ]; then
+  pass "cursor orchestration: $ma_agents ma-* subagents emitted"
+else
+  fail "cursor orchestration: too few ma-* subagents ($ma_agents, expected >= 8)"
+fi
+if [ -f "$TMP_CURSOR/.cursor/commands/multi-agent.md" ]; then
+  pass "cursor orchestration: /multi-agent command present"
+else
+  fail "cursor orchestration: .cursor/commands/multi-agent.md missing"
+fi
+if grep -q "dev-toolkit" "$TMP_CURSOR/.cursor/mcp.json" 2>/dev/null; then
+  pass "cursor orchestration: mcp.json registers dev-toolkit"
+else
+  fail "cursor orchestration: mcp.json missing dev-toolkit server"
+fi
+
 # Round-trip via direct adapter call (CLI uninstall not yet wired into smoke).
 node --input-type=module -e "
   import('$REPO_ROOT/pipeline/adapters/cursor.mjs').then(m => {
@@ -74,6 +93,13 @@ node --input-type=module -e "
   });
 " >/dev/null 2>&1
 
+ma_after=$(find "$TMP_CURSOR/.cursor/agents" -name "ma-*.md" 2>/dev/null | wc -l | tr -d ' ')
+if [ "$ma_after" -eq 0 ] && [ ! -f "$TMP_CURSOR/.cursor/commands/multi-agent.md" ]; then
+  pass "cursor orchestration uninstall: agents + command removed"
+else
+  fail "cursor orchestration uninstall: left $ma_after agent(s) / command behind"
+fi
+
 remaining=$(find "$TMP_CURSOR/.cursor/rules" -name "multi-agent-*" 2>/dev/null | wc -l | tr -d ' ')
 if [ "$remaining" -eq 0 ]; then
   pass "cursor uninstall: 0 files remain"
@@ -82,7 +108,7 @@ else
 fi
 
 # ──────────────────────────────────────────────────────────────────────────
-echo "→ 2. Copilot Chat adapter — marker-wrapped block preservation"
+echo "→ 2. Copilot Chat adapter  -  marker-wrapped block preservation"
 TMP_COPILOTCHAT=$(mktemp -d)
 mkdir -p "$TMP_COPILOTCHAT/.github"
 USER_CONTENT="# My personal Copilot Chat instructions"
@@ -121,6 +147,12 @@ else
   fail "copilot-chat: per-skill instruction count too low ($perskill_count)"
 fi
 
+# Orchestration layer  -  custom agents + /multi-agent prompt + VS Code MCP.
+cc_agents=$(find "$TMP_COPILOTCHAT/.github/agents" -name "ma-*.agent.md" 2>/dev/null | wc -l | tr -d ' ')
+[ "$cc_agents" -ge 8 ] && pass "copilot-chat orchestration: $cc_agents custom agents emitted" || fail "copilot-chat orchestration: too few agents ($cc_agents)"
+[ -f "$TMP_COPILOTCHAT/.github/prompts/multi-agent.prompt.md" ] && pass "copilot-chat orchestration: /multi-agent prompt present" || fail "copilot-chat orchestration: prompt missing"
+grep -q "dev-toolkit" "$TMP_COPILOTCHAT/.vscode/mcp.json" 2>/dev/null && pass "copilot-chat orchestration: .vscode/mcp.json registers dev-toolkit" || fail "copilot-chat orchestration: .vscode/mcp.json missing dev-toolkit"
+
 # Round-trip
 node --input-type=module -e "
   import('$REPO_ROOT/pipeline/adapters/copilot-chat.mjs').then(m => {
@@ -141,7 +173,44 @@ else
 fi
 
 # ──────────────────────────────────────────────────────────────────────────
-echo "→ 3. Idempotency — install twice produces identical tree"
+echo "→ 2c. Antigravity adapter  -  orchestration round-trip + AGENTS.md preservation"
+TMP_ANTIGRAVITY=$(mktemp -d)
+printf '# user agents\nKEEP_ANTI_LINE\n' > "$TMP_ANTIGRAVITY/AGENTS.md"
+node "$INSTALL" --antigravity --target="$TMP_ANTIGRAVITY" >/dev/null 2>&1 || fail "antigravity install exited non-zero"
+anti_rules=$(find "$TMP_ANTIGRAVITY/.agent/rules" -name "multi-agent-*.md" 2>/dev/null | wc -l | tr -d ' ')
+[ "$anti_rules" -ge 8 ] && pass "antigravity: $anti_rules rules in .agent/rules" || fail "antigravity: too few rules ($anti_rules)"
+[ -f "$TMP_ANTIGRAVITY/.agent/workflows/multi-agent.md" ] && pass "antigravity: /multi-agent workflow present" || fail "antigravity: workflow missing"
+grep -q "KEEP_ANTI_LINE" "$TMP_ANTIGRAVITY/AGENTS.md" && pass "antigravity: user AGENTS.md content preserved" || fail "antigravity: user content lost"
+grep -q "dev-toolkit" "$TMP_ANTIGRAVITY/.agent/mcp_config.json" 2>/dev/null && pass "antigravity: mcp_config registers dev-toolkit" || fail "antigravity: mcp_config missing dev-toolkit"
+node --input-type=module -e "import('$REPO_ROOT/pipeline/adapters/antigravity.mjs').then(m => m.default.uninstall({ target: '$TMP_ANTIGRAVITY' }));" >/dev/null 2>&1
+anti_after=$(find "$TMP_ANTIGRAVITY/.agent/rules" -name "multi-agent-*.md" 2>/dev/null | wc -l | tr -d ' ')
+if [ "$anti_after" -eq 0 ] && grep -q "KEEP_ANTI_LINE" "$TMP_ANTIGRAVITY/AGENTS.md" 2>/dev/null && ! grep -q "multi-agent-pipeline:begin" "$TMP_ANTIGRAVITY/AGENTS.md" 2>/dev/null; then
+  pass "antigravity uninstall: artifacts removed, user content survived"
+else
+  fail "antigravity uninstall: incomplete (rules=$anti_after)"
+fi
+
+# ──────────────────────────────────────────────────────────────────────────
+echo "→ 2d. Codex CLI adapter  -  global install round-trip (~/.codex)"
+TMP_CODEX_HOME=$(mktemp -d)
+printf '# user codex memory\nKEEP_CODEX_LINE\n' > "$TMP_CODEX_HOME/.codex_seed" # marker file, not AGENTS.md
+mkdir -p "$TMP_CODEX_HOME/.codex"
+printf 'model = "gpt-5.5-codex"\n' > "$TMP_CODEX_HOME/.codex/config.toml" # pre-existing user TOML
+HOME="$TMP_CODEX_HOME" node "$INSTALL" --codex >/dev/null 2>&1 || fail "codex install exited non-zero"
+[ -f "$TMP_CODEX_HOME/.codex/prompts/multi-agent.md" ] && pass "codex: /multi-agent prompt present" || fail "codex: prompt missing"
+[ -f "$TMP_CODEX_HOME/.codex/AGENTS.md" ] && pass "codex: AGENTS.md skill index present" || fail "codex: AGENTS.md missing"
+grep -q "dev-toolkit" "$TMP_CODEX_HOME/.codex/config.toml" 2>/dev/null && pass "codex: config.toml registers dev-toolkit mcp" || fail "codex: config.toml missing mcp block"
+grep -q 'model = "gpt-5.5-codex"' "$TMP_CODEX_HOME/.codex/config.toml" 2>/dev/null && pass "codex: pre-existing user TOML preserved" || fail "codex: user TOML clobbered"
+HOME="$TMP_CODEX_HOME" node --input-type=module -e "import('$REPO_ROOT/pipeline/adapters/codex.mjs').then(m => m.default.uninstall());" >/dev/null 2>&1
+if [ ! -f "$TMP_CODEX_HOME/.codex/prompts/multi-agent.md" ] && grep -q 'model = "gpt-5.5-codex"' "$TMP_CODEX_HOME/.codex/config.toml" 2>/dev/null && ! grep -q "dev-toolkit" "$TMP_CODEX_HOME/.codex/config.toml" 2>/dev/null; then
+  pass "codex uninstall: artifacts removed, user TOML survived"
+else
+  fail "codex uninstall: incomplete"
+fi
+rm -rf "$TMP_CODEX_HOME"
+
+# ──────────────────────────────────────────────────────────────────────────
+echo "→ 3. Idempotency  -  install twice produces identical tree"
 TMP_IDEMP=$(mktemp -d)
 node "$INSTALL" --cursor --target="$TMP_IDEMP" >/dev/null 2>&1
 HASH_ONE=$(find "$TMP_IDEMP" -type f -exec md5 -q {} \; 2>/dev/null | sort | md5 -q 2>/dev/null || \
@@ -156,7 +225,7 @@ else
 fi
 
 # ──────────────────────────────────────────────────────────────────────────
-echo "→ 4. Platform filter — ios vs android narrow the skill set"
+echo "→ 4. Platform filter  -  ios vs android narrow the skill set"
 TMP_FILTER=$(mktemp -d)
 node "$INSTALL" --cursor --platform=ios --target="$TMP_FILTER" >/dev/null 2>&1
 ios_count=$(find "$TMP_FILTER/.cursor/rules" -type f 2>/dev/null | wc -l | tr -d ' ')
@@ -180,7 +249,7 @@ else
 fi
 
 # ──────────────────────────────────────────────────────────────────────────
-echo "→ 5. Default behavior — no flags installs Claude (backward compat)"
+echo "→ 5. Default behavior  -  no flags installs Claude (backward compat)"
 # We don't actually install to $HOME here; just verify the flag math.
 RESULT=$(node -e "
   const flags = [];

package/pipeline/scripts/smoke-add-detail.sh

@@ -1,5 +1,5 @@
 #!/usr/bin/env bash
-# smoke-add-detail.sh — contract check for the `add-detail` command.
+# smoke-add-detail.sh  -  contract check for the `add-detail` command.
 #
 # Does NOT hit real Bitbucket / Jira / GitHub. Instead:
 #   1. Generates a fake PR diff and body fixture.
@@ -77,7 +77,7 @@ DESC="$(jq -r .description "$TMP/payload.json")"
 if [[ "$DESC" == *$'\n'* ]]; then
   pass "description contains real newlines"
 else
-  fail "description is single-line — newlines were escaped"
+  fail "description is single-line  -  newlines were escaped"
 fi
 
 # Contract 3: no literal backslash-n sequence survived.
@@ -106,7 +106,7 @@ for glyph in 'Ö' 'İ' 'ç'; do
   fi
 done
 
-# Contract 6: Bitbucket PUT contract — all of reviewers/fromRef/toRef/draft present.
+# Contract 6: Bitbucket PUT contract  -  all of reviewers/fromRef/toRef/draft present.
 for field in reviewers fromRef toRef draft; do
   if jq -e "has(\"$field\")" "$TMP/payload.json" >/dev/null; then
     pass "payload has .$field"
@@ -120,7 +120,7 @@ LEN="$(jq '.reviewers | length' "$TMP/payload.json")"
 if [ "$LEN" -gt 0 ]; then
   pass "reviewers array preserved ($LEN entries)"
 else
-  fail "reviewers array is empty — would wipe defaults"
+  fail "reviewers array is empty  -  would wipe defaults"
 fi
 
 # Contract 8: fromRef / toRef are distinct (PR to self is rejected by Bitbucket).
@@ -129,7 +129,7 @@ TO="$(jq -r .toRef.id   "$TMP/payload.json")"
 if [ "$FROM" != "$TO" ]; then
   pass "fromRef ($FROM) ≠ toRef ($TO)"
 else
-  fail "fromRef == toRef — invalid PR"
+  fail "fromRef == toRef  -  invalid PR"
 fi
 
 echo

package/pipeline/scripts/smoke-agent-log-cost.sh

@@ -1,5 +1,5 @@
 #!/usr/bin/env bash
-# smoke-agent-log-cost.sh — v8.3.0
+# smoke-agent-log-cost.sh  -  v8.3.0
 #
 # Verifies the agent-log Cost Breakdown pipeline:
 #   1. render-agent-log-cost.sh exits 2 when no tracker data available
@@ -7,7 +7,7 @@
 #      from a hand-built object-shaped phase-tracker.json
 #   3. Renders identical block from the array-shaped phase-tracker.json that
 #      phase-tracker.sh actually writes
-#   4. Unknown model degrades to "—" with footnote, never blocks
+#   4. Unknown model degrades to " - " with footnote, never blocks
 #   5. log-metric.sh forwards tokens + model to tracker when LOG_METRIC_FORWARD_TO_TRACKER=1
 #   6. log-format.md ref doc declares the Tokens column + Cost Breakdown section
 #
@@ -118,8 +118,8 @@ cat > .worktrees/UNK-1/phase-tracker.json <<'JSON'
 ]}
 JSON
 out=$(bash "$RENDERER" UNK-1 2>/dev/null)
-if grep -qF '| made-up-9000 |' <<< "$out" && grep -qF '| — |' <<< "$out"; then
-  record_pass "unknown model row renders with USD as —"
+if grep -qF '| made-up-9000 |' <<< "$out" && grep -qF '|  -  |' <<< "$out"; then
+  record_pass "unknown model row renders with USD as  - "
 else
   record_fail "unknown model handling broken"
 fi
@@ -137,7 +137,7 @@ export METRICS_FILE="$tmp/metrics.jsonl"
 export TRACKER_FILE="$tmp/tracker.json"
 # Bootstrap tracker with one phase
 bash "$TRACKER" init "FWD-1" >/dev/null 2>&1 || true
-# init writes to $HOME/.claude/logs/multi-agent/FWD-1/tracker-state.json — override via TRACKER_FILE
+# init writes to $HOME/.claude/logs/multi-agent/FWD-1/tracker-state.json  -  override via TRACKER_FILE
 cat > "$TRACKER_FILE" <<'JSON'
 {"task_id":"FWD-1","started_at":"2026-05-05T00:00:00Z","phases":[
   {"id":"4","name":"Review","status":"in_progress","subs":[]}
@@ -145,16 +145,95 @@ cat > "$TRACKER_FILE" <<'JSON'
 JSON
 LOG_METRIC_FORWARD_TO_TRACKER=1 TRACKER_QUIET=1 \
   bash "$LOG_METRIC" FWD-1 4 review.reviewer_call \
-  model=opus tokens_in=1234 tokens_out=567 duration_ms=1000 >/dev/null 2>&1 || true
+  model=opus tokens_in=1234 tokens_out=567 tokens_cached=400 duration_ms=1000 >/dev/null 2>&1 || true
 if jq -e '.phases[0].tokens_in == 1234 and .phases[0].tokens_out == 567 and .phases[0].model == "opus"' "$TRACKER_FILE" >/dev/null 2>&1; then
   record_pass "log-metric forwards tokens + model to tracker"
 else
   record_fail "log-metric forwarder did not update tracker"
   cat "$TRACKER_FILE" >&2
 fi
+# End-to-end: tokens_cached must flow log-metric -> phase-tracker -> tracker state
+if jq -e '.phases[0].tokens_cached == 400' "$TRACKER_FILE" >/dev/null 2>&1; then
+  record_pass "log-metric forwards tokens_cached end-to-end"
+else
+  record_fail "tokens_cached did not reach the tracker"
+  cat "$TRACKER_FILE" >&2
+fi
+# A malformed tokens_cached must NOT poison the call (in/out must still land).
+cat > "$TRACKER_FILE" <<'JSON'
+{"task_id":"FWD-2","phases":[{"id":"4","name":"Review","status":"in_progress","subs":[]}]}
+JSON
+LOG_METRIC_FORWARD_TO_TRACKER=1 TRACKER_QUIET=1 \
+  bash "$LOG_METRIC" FWD-2 4 review.reviewer_call \
+  model=opus tokens_in=999 tokens_out=111 tokens_cached=abc duration_ms=1 >/dev/null 2>&1 || true
+if jq -e '.phases[0].tokens_in == 999 and .phases[0].tokens_out == 111 and (.phases[0].tokens_cached // 0) == 0' "$TRACKER_FILE" >/dev/null 2>&1; then
+  record_pass "malformed tokens_cached sanitized to 0; in/out preserved"
+else
+  record_fail "malformed tokens_cached poisoned the tokens call"
+  cat "$TRACKER_FILE" >&2
+fi
 unset METRICS_FILE TRACKER_FILE
 rm -rf "$tmp"
 
+# --- 5b: cached tokens are priced at the cache-read rate + surfaced ---
+tmp=$(mktemp -d)
+pushd "$tmp" >/dev/null
+mkdir -p .worktrees/CACHE-1
+cat > .worktrees/CACHE-1/phase-tracker.json <<'JSON'
+{"task_id":"CACHE-1","phases":[
+  {"id":"1","name":"Analysis","model":"opus","tokens_in":20000,"tokens_out":3000,"tokens_cached":15000,"subs":[]},
+  {"id":"3","name":"Dev","model":"sonnet","tokens_in":60000,"tokens_out":10000,"subs":[]}
+]}
+JSON
+out=$(bash "$RENDERER" CACHE-1 2>/dev/null)
+# Cache-exclusive: tokens_in (20000) and tokens_cached (15000) are disjoint.
+# 20000 fresh @5 + 15000 cached @0.50 + 3000 out @25 = 0.1825 -> floor-to-cents $0.18.
+# Without the cache discount (cached billed @5) it would be 0.25 -> $0.25, so the
+# discounted rate is what drives the row down to $0.18.
+if grep -qE '\| Analysis \| opus \| 20,000 \| 3,000 \| \$0\.18 \|' <<< "$out"; then
+  record_pass "cached tokens priced at cache-read rate (opus row = \$0.18)"
+else
+  record_fail "cached-token pricing wrong"
+  printf '%s\n' "$out" | head -12 >&2
+fi
+if grep -q 'Cache reads: 15,000 tokens' <<< "$out"; then
+  record_pass "cache-read footer surfaced when tokens_cached present"
+else
+  record_fail "cache-read footer missing"
+fi
+if grep -qE 'Top cost driver:' <<< "$out"; then
+  record_pass "top cost driver line rendered"
+else
+  record_fail "top cost driver line missing"
+fi
+popd >/dev/null
+rm -rf "$tmp"
+
+# --- 5c: no cache footer when no cached tokens (object-shape regression) ---
+tmp=$(mktemp -d)
+pushd "$tmp" >/dev/null
+mkdir -p .worktrees/NOCACHE-1
+cat > .worktrees/NOCACHE-1/phase-tracker.json <<'JSON'
+{"task_id":"NOCACHE-1","phases":[
+  {"id":"3","name":"Dev","model":"sonnet","tokens_in":50000,"tokens_out":8000,"subs":[]}
+]}
+JSON
+out=$(bash "$RENDERER" NOCACHE-1 2>/dev/null)
+if grep -q 'Cache reads:' <<< "$out"; then
+  record_fail "cache-read footer should be absent when no cached tokens"
+else
+  record_pass "no cache-read footer when tokens_cached absent"
+fi
+# Sonnet 50000 @3 + 8000 @15 = 0.27 -> $0.27 (unchanged from pre-cache behavior)
+if grep -qE '\| Dev \| sonnet \| 50,000 \| 8,000 \| \$0\.27 \|' <<< "$out"; then
+  record_pass "no-cache USD unchanged (backward-compatible)"
+else
+  record_fail "no-cache USD changed unexpectedly"
+  printf '%s\n' "$out" | head -12 >&2
+fi
+popd >/dev/null
+rm -rf "$tmp"
+
 # --- 6: log-format.md declares Tokens + Cost Breakdown ---
 if grep -q 'Tokens (in/out)' "$LOG_FORMAT_DOC"; then
   record_pass "log-format.md declares Tokens (in/out) column"

package/pipeline/scripts/smoke-agent-model-routing.sh

@@ -1,9 +1,9 @@
 #!/usr/bin/env bash
-# smoke-agent-model-routing.sh — v6.1.D
+# smoke-agent-model-routing.sh  -  v6.1.D
 #
 # Asserts per-persona model routing contract:
 #   1. Every agent file in pipeline/agents/ has a `preferredModel` frontmatter field.
-#   2. `preferredModel` is one of: opus | sonnet | haiku.
+#   2. `preferredModel` is one of: fable | opus | sonnet | haiku.
 #   3. `model:` (Claude-Code-native) matches `preferredModel` (single source of truth).
 #   4. Orchestrator docs mention `CLAUDE_CODE_SUBAGENT_MODEL` and `preferredModel`.
 #   5. Phase 4 review doc references the per-dispatch override mechanism.
@@ -38,7 +38,7 @@ for name in "${expected[@]}"; do
   pref=$(awk '/^preferredModel:/ {print $2; exit}' "$f" | tr -d '"'"'")
   mdl=$(awk '/^model:/ {print $2; exit}' "$f" | tr -d '"'"'")
   case "$pref" in
-    opus|sonnet|haiku)
+    fable|opus|sonnet|haiku)
       record_pass "$name has preferredModel=$pref"
       ;;
     *)

package/pipeline/scripts/smoke-ask-choice.sh

@@ -0,0 +1,42 @@
+#!/usr/bin/env bash
+# smoke-ask-choice.sh  -  the portable picker fallback used on platforms without
+# a native AskUserQuestion. Exercises default-by-index, default-by-label,
+# non-interactive first-option, label-vs-index resolution, and usage errors.
+
+set -uo pipefail
+
+REPO_ROOT="$(cd "$(dirname "$0")/../.." && pwd)"
+ASK="$REPO_ROOT/pipeline/lib/ask-choice.sh"
+
+PASS=0
+FAIL=0
+pass() { PASS=$((PASS + 1)); echo "  ✓ $1"; }
+fail() { FAIL=$((FAIL + 1)); echo "  ✗ $1"; }
+
+echo "→ 1. usage error with <3 args (exit 64)"
+bash "$ASK" "only a question" >/dev/null 2>&1
+[ "$?" -eq 64 ] && pass "too-few-args exits 64" || fail "too-few-args should exit 64"
+
+echo "→ 2. ASK_CHOICE_DEFAULT by 1-based index"
+out=$(ASK_CHOICE_DEFAULT=2 bash "$ASK" "Approve this plan?" "Approve" "Cancel" "Edit" 2>/dev/null)
+[ "$out" = "Cancel" ] && pass "index default 2 -> 'Cancel'" || fail "index default gave '$out' (expected Cancel)"
+
+echo "→ 3. ASK_CHOICE_DEFAULT by label (case-insensitive)"
+out=$(ASK_CHOICE_DEFAULT=approve bash "$ASK" "Approve this plan?" "Approve" "Cancel" 2>/dev/null)
+[ "$out" = "Approve" ] && pass "label default 'approve' -> 'Approve'" || fail "label default gave '$out'"
+
+echo "→ 4. out-of-range / unmatched default falls through to first option (non-TTY)"
+out=$(ASK_CHOICE_DEFAULT=99 bash "$ASK" "Pick" "A" "B" </dev/null 2>/dev/null)
+[ "$out" = "A" ] && pass "bad default + no TTY -> first option 'A'" || fail "bad default gave '$out'"
+
+echo "→ 5. non-interactive (no TTY, no default) selects first option, never blocks"
+out=$(bash "$ASK" "Continue?" "Yes" "No" </dev/null 2>/dev/null)
+[ "$out" = "Yes" ] && pass "no TTY -> first option 'Yes'" || fail "no TTY gave '$out'"
+
+echo "→ 6. stdout carries ONLY the label (menu/prompt go to stderr)"
+out=$(ASK_CHOICE_DEFAULT=1 bash "$ASK" "Q" "Only" "Other" 2>/dev/null)
+[ "$out" = "Only" ] && pass "stdout is clean single label" || fail "stdout not clean: '$out'"
+
+echo ""
+echo "══ ask-choice smoke: $PASS passed, $FAIL failed ══"
+[ "$FAIL" -eq 0 ]