pantheon-opencode 1.5.0-beta.1 → 1.5.0-beta.10

package/scripts/test-opencode-v1-v2-sandbox.sh

@@ -13,6 +13,11 @@
 #                    (config is regenerated for that version first).
 #   --prompts        Prompt battery via `opencode run --format json` for every
 #                    requested version (with --run vX) or both versions alone.
+#   --cost           Offline pantheon_cost probe for the requested
+#                    version (with --run vX) or both versions alone.
+#                    The probe's synthetic checks run in explicit --fixture mode.
+#   --rehydrate      Offline context_rehydrate + context_session_summary
+#                    probe for the requested version (with --run vX) or both.
 #   --reset          Wipe the sandbox root.
 #   --help           Usage.
 #
@@ -25,11 +30,8 @@
 #   PANTHEON_SANDBOX_MODEL   Model used by init/prompts
 #                            (default: opencode-go/mimo-v2.5)
 #   PANTHEON_PROMPT_TIMEOUT  Per-prompt timeout in seconds (default: 300)
-#   PANTHEON_RETRY_COOLDOWN  Cooldown seconds before the single FAIL retry
-#                            (default: 60; retry only for real FAIL, never AMBIENTAL)
-#
-# Exit codes: 0 = no real failures (AMBIENTAL issues do not fail the run),
-#             1 = real FAIL (see prompts-report.md in the sandbox root),
+# Exit codes: 0 = every required check returned explicit PASS,
+#             1 = any failure or untested check (see reports in the sandbox root),
 #             2 = usage error, 3 = sandbox not prepared.
 set -euo pipefail
 
@@ -49,11 +51,15 @@ V1_BIN="${OPENCODE_V1_BIN:-opencode}"
 V2_BIN="${OPENCODE_V2_BIN:-opencode2}"
 
 REPORT_FILE="$SANDBOX_ROOT/prompts-report.md"
+COST_REPORT_FILE="$SANDBOX_ROOT/pantheon-cost-report.md"
+REHYDRATE_REPORT_FILE="$SANDBOX_ROOT/context-rehydrate-report.md"
 EXTRACT_PY="$SANDBOX_ROOT/.prompt-extract-json.py"
 
 MODE_PREPARE=0
 MODE_RESET=0
 MODE_PROMPTS=0
+MODE_COST=0
+MODE_REHYDRATE=0
 RUN_VERSION=""
 
 usage() {
@@ -105,8 +111,8 @@ write_run_test_sh() {
   cat > "$SANDBOX_ROOT/run-test.sh" <<'RUNTEST'
 #!/usr/bin/env bash
 # Generated by pantheon-opencode scripts/test-opencode-v1-v2-sandbox.sh --prepare
-# Validates the global install: binaries, 5/5 MCP connectivity, doctor 0 errors,
-# AND the CONTENT of the pantheon://agents resource (zeus/hermes present).
+# Validates the global install: binaries, exactly 5 connected MCPs per version,
+# doctor 0 errors, AND the CONTENT of pantheon://agents (zeus/hermes present).
 set -euo pipefail
 
 SANDBOX_DIR="$(cd "$(dirname "$0")" && pwd)"
@@ -124,9 +130,68 @@ export HOME="$SANDBOX_HOME"
 export PATH="$NPM_PREFIX/bin:$SANDBOX_VENV/bin:$PATH"
 export npm_config_prefix="$NPM_PREFIX"
 
+STATUS_FAIL=0
+STATUS_PASS=0
+
+record_status() { # label status detail
+  local label="$1" status="$2" detail="$3"
+  case "$status" in
+    PASS) STATUS_PASS=$((STATUS_PASS + 1)) ;;
+    FAIL) STATUS_FAIL=$((STATUS_FAIL + 1)) ;;
+    *) STATUS_FAIL=$((STATUS_FAIL + 1)); status="FAIL" ;;
+  esac
+  printf '%s: %s — %s\n' "$status" "$label" "$detail"
+}
+
+check_binary() { # label binary
+  local label="$1" bin="$2" out rc=0
+  if ! command -v "$bin" >/dev/null 2>&1; then
+    record_status "$label" "FAIL" "binary '$bin' is not installed in the sandbox prefix"
+    return 0
+  fi
+  set +e
+  out="$("$bin" --version 2>&1)"
+  rc=$?
+  set -e
+  if [ "$rc" -eq 0 ]; then
+    record_status "$label" "PASS" "${out:-version command exited 0}"
+  else
+    record_status "$label" "FAIL" "version check exited $rc"
+  fi
+}
+
+check_mcp() { # label binary project
+  local label="$1" bin="$2" project="$3" out connected_count rc=0
+  if ! command -v "$bin" >/dev/null 2>&1; then
+    record_status "$label MCP" "FAIL" "binary '$bin' is not installed in the sandbox prefix"
+    return 0
+  fi
+  if [ ! -d "$project" ]; then
+    record_status "$label MCP" "FAIL" "project directory is missing: $project"
+    return 0
+  fi
+  set +e
+  out="$(cd "$project" && "$bin" mcp list 2>&1)"
+  rc=$?
+  set -e
+  out="$(printf '%s' "$out" | sed -E $'s/\x1B\\[[0-?]*[ -/]*[@-~]//g')"
+  printf '%s\n' "$out"
+  if [ "$rc" -ne 0 ]; then
+    record_status "$label MCP" "FAIL" "mcp list exited $rc"
+    return 0
+  fi
+  connected_count="$(printf '%s' "$out" | grep -Eic '(^|[^[:alnum:]_])connected([^[:alnum:]_]|$)')" || connected_count=0
+  if [ "$connected_count" -eq 5 ] \
+    && ! printf '%s' "$out" | grep -Eiq 'failed|✘'; then
+    record_status "$label MCP" "PASS" "exactly 5 connected MCPs"
+  else
+    record_status "$label MCP" "FAIL" "expected exactly 5 connected MCPs; found $connected_count or failures reported"
+  fi
+}
+
 echo "--- Binaries ---"
-opencode --version
-opencode2 --version
+check_binary "V1 binary" opencode
+check_binary "V2 binary" opencode2
 
 echo "--- Building tarball ---"
 cd "$REPO_DIR"
@@ -134,42 +199,70 @@ npm pack 2>/dev/null || { echo "ERROR: npm pack failed"; exit 1; }
 TGZ=$(ls -t pantheon-opencode-*.tgz | head -1)
 npm rm -g pantheon-opencode 2>/dev/null || true
 npm install -g "$TGZ"
+if ! git -C "$REPO_DIR" ls-files --error-unmatch -- "$TGZ" >/dev/null 2>&1; then rm -f -- "$TGZ"; fi
 
 echo "--- Init (headless) ---"
 pantheon-opencode init --headless -y
 
-echo "--- MCP List (V1) ---"
-opencode mcp list
-
-echo "--- MCP List (V2) ---"
-opencode2 mcp list
+echo "--- MCP Validation ---"
+check_mcp "V1" opencode "$SANDBOX_DIR/project-v1"
+check_mcp "V2" opencode2 "$SANDBOX_DIR/project-v2"
 
 echo "--- Doctor ---"
-pantheon-opencode doctor --target "$SANDBOX_DIR"
+if command -v pantheon-opencode >/dev/null 2>&1; then
+  set +e
+  DOCTOR_OUTPUT="$(pantheon-opencode doctor --target "$SANDBOX_DIR" 2>&1)"
+  DOCTOR_RC=$?
+  set -e
+  if [ "$DOCTOR_RC" -eq 0 ]; then
+    record_status "doctor" "PASS" "exit 0"
+  else
+    record_status "doctor" "FAIL" "doctor exited $DOCTOR_RC"
+  fi
+else
+  record_status "doctor" "FAIL" "pantheon-opencode is not installed in the sandbox prefix"
+fi
 
 echo "--- Gate (b): pantheon://agents CONTENT check ---"
 RESOURCES_SRV="$PANTHEON_GLOBAL/scripts/mcp_resources_server.py"
 if [ ! -f "$RESOURCES_SRV" ]; then
-  echo "ERROR: resources server not found at $RESOURCES_SRV"; exit 1
-fi
-AGENTS_CONTENT="$("$SANDBOX_VENV/bin/python" - "$RESOURCES_SRV" <<'PYEOF'
+  record_status "pantheon://agents resource" "FAIL" "resources server not found at $RESOURCES_SRV"
+else
+  set +e
+  AGENTS_CONTENT="$("$SANDBOX_VENV/bin/python" - "$RESOURCES_SRV" <<'PYEOF'
 import asyncio, importlib.util, sys
+from pathlib import Path
+resource_path = Path(sys.argv[1]).resolve()
+sys.path.insert(0, str(resource_path.parent))
 spec = importlib.util.spec_from_file_location("mrs", sys.argv[1])
 mod = importlib.util.module_from_spec(spec)
 spec.loader.exec_module(mod)
 print(asyncio.run(mod.list_agents()))
 PYEOF
 )"
-printf '%s\n' "$AGENTS_CONTENT" | head -8
-if printf '%s\n' "$AGENTS_CONTENT" | grep -qi 'zeus' \
-  && printf '%s\n' "$AGENTS_CONTENT" | grep -qi 'hermes'; then
-  echo "OK: pantheon://agents content includes zeus and hermes"
-else
-  echo "ERROR: pantheon://agents content missing zeus/hermes (connectivity alone is not enough)"; exit 1
+  AGENTS_RC=$?
+  set -e
+  printf '%s\n' "$AGENTS_CONTENT" | head -8
+  if [ "$AGENTS_RC" -ne 0 ]; then
+    record_status "pantheon://agents resource" "FAIL" "resource probe exited $AGENTS_RC"
+  elif printf '%s\n' "$AGENTS_CONTENT" | grep -qi 'zeus' \
+    && printf '%s\n' "$AGENTS_CONTENT" | grep -qi 'hermes'; then
+    record_status "pantheon://agents resource" "PASS" "pantheon://agents content includes zeus and hermes"
+  else
+    record_status "pantheon://agents resource" "FAIL" "content missing zeus/hermes (connectivity alone is not enough)"
+  fi
 fi
-
 echo ""
 echo "=== VALIDATION COMPLETE ==="
+if [ "$STATUS_FAIL" -gt 0 ] || [ "$STATUS_PASS" -eq 0 ]; then
+  FINAL_VERDICT="FAIL"
+else
+  FINAL_VERDICT="PASS"
+fi
+printf 'Final verdict: %s\n' "$FINAL_VERDICT"
+if [ "$FINAL_VERDICT" != "PASS" ]; then
+  exit 1
+fi
 echo "Next: open the isolated TUI with: $SANDBOX_DIR/start-pantheon.sh"
 RUNTEST
   chmod +x "$SANDBOX_ROOT/run-test.sh"
@@ -195,12 +288,11 @@ install_binaries() {
   tgz=$(ls -t pantheon-opencode-*.tgz | head -1)
   npm rm -g pantheon-opencode 2>/dev/null || true
   npm install -g "$REPO_DIR/$tgz" || die "install of $tgz failed"
+  if ! git -C "$REPO_DIR" ls-files --error-unmatch -- "$tgz" >/dev/null 2>&1; then rm -f -- "$tgz"; fi
   log "--- Installing OpenCode V1 ($OPENCODE_V1_SPEC) ---"
-  npm install -g "$OPENCODE_V1_SPEC" \
-    || warn "V1 binary install failed ($OPENCODE_V1_SPEC) — v1 prompts will be AMBIENTAL"
+  npm install -g "$OPENCODE_V1_SPEC" || die "install of $OPENCODE_V1_SPEC failed"
   log "--- Installing OpenCode V2 ($OPENCODE_V2_SPEC) ---"
-  npm install -g "$OPENCODE_V2_SPEC" \
-    || warn "V2 binary install failed ($OPENCODE_V2_SPEC) — v2 prompts will be AMBIENTAL"
+  npm install -g "$OPENCODE_V2_SPEC" || die "install of $OPENCODE_V2_SPEC failed"
 }
 
 prepare_project() {
@@ -271,19 +363,14 @@ PYEOF
 
 # ── Prompt battery ────────────────────────────────────────────────────────────
 
-# Known environmental failure signatures (never fail the run).
-AMBIENTAL_RE='bifrost|unauthorized|authentication failure|api key|credentials|401|docker|econnrefused|enotfound|etimedout|econnreset|connection refused|rate.?limit|429|quota|overloaded|edgesout|npm err'
-
 PROMPT_IDS=()
 PROMPT_TEXTS=()
 PROMPT_MARKERS=()
-PROMPT_AMB_MARKERS=()
 PROMPT_VERIFY=()
 
-add_prompt() { # id marker ambiental_marker verify text
+add_prompt() { # id marker ignored verify text
   PROMPT_IDS+=("$1")
   PROMPT_MARKERS+=("$2")
-  PROMPT_AMB_MARKERS+=("$3")
   PROMPT_VERIFY+=("$4")
   PROMPT_TEXTS+=("$5")
 }
@@ -293,9 +380,8 @@ build_battery() {
   PROMPT_IDS=()
   PROMPT_TEXTS=()
   PROMPT_MARKERS=()
-  PROMPT_AMB_MARKERS=()
   PROMPT_VERIFY=()
-  add_prompt "agents-resource" "AGENTS-OK" "AGENTS-MISSING" "" \
+  add_prompt "agents-resource" "AGENTS-OK" "" "" \
     "List your MCP resources and read the resource pantheon://agents. If the returned agent list includes an agent named 'zeus' and one named 'hermes', reply with exactly: AGENTS-OK. Otherwise reply with exactly: AGENTS-MISSING."
   add_prompt "memory-store" "MEMORY-STORE-OK" "" "" \
     "Use the pantheon-memory memory_store tool to store this exact entry: namespace 'default', key 'pantheon-sandbox-probe-$v', value 'alive'. Then reply with exactly: MEMORY-STORE-OK"
@@ -304,7 +390,7 @@ build_battery() {
   add_prompt "tmp-write-read" "TMP-OK" "" \
     "${TMPDIR:-/tmp}/pantheon-sandbox-probe-$v.txt::TMP-OK" \
     "Using the bash tool, run a python3 -c command that writes the text TMP-OK to the file ${TMPDIR:-/tmp}/pantheon-sandbox-probe-$v.txt (overwrite it), then run another python3 -c command to read that file back. Do not use any other tool or command to write or read the file — only python3 -c. If what you read back is exactly TMP-OK, reply with exactly: TMP-OK"
-  add_prompt "delegation" "DELEGATION-OK" "DELEGATION-UNAVAILABLE" "" \
+  add_prompt "delegation" "DELEGATION-OK" "" "" \
     "Delegate a tiny task to the @talos agent via the task tool: ask it to reply with the single word PONG. If the delegated response contains PONG, reply with exactly: DELEGATION-OK. If you cannot delegate to agents in this context, reply with exactly: DELEGATION-UNAVAILABLE"
 }
 
@@ -322,10 +408,9 @@ ATTEMPT_RESULT=""
 ATTEMPT_DETAIL=""
 
 run_prompt_attempt() { # v idx current_bin — sets ATTEMPT_RESULT / ATTEMPT_DETAIL
-  local v="$1" idx="$2" bin="$3" id marker amb_marker verify text
+  local v="$1" idx="$2" bin="$3" id marker verify text
   id="${PROMPT_IDS[$idx]}"
   marker="${PROMPT_MARKERS[$idx]}"
-  amb_marker="${PROMPT_AMB_MARKERS[$idx]}"
   verify="${PROMPT_VERIFY[$idx]}"
   text="${PROMPT_TEXTS[$idx]}"
 
@@ -352,75 +437,59 @@ run_prompt_attempt() { # v idx current_bin — sets ATTEMPT_RESULT / ATTEMPT_DET
   rm -f "$out_file" "$err_file"
 
   if [ "$rc" -eq 124 ]; then
-    ATTEMPT_RESULT="AMBIENTAL"
-    ATTEMPT_DETAIL="timeout after ${PANTHEON_PROMPT_TIMEOUT}s (provider latency)"
+    ATTEMPT_RESULT="FAIL"
+    ATTEMPT_DETAIL="timeout after ${PANTHEON_PROMPT_TIMEOUT}s"
+  elif [ "$rc" -ne 0 ]; then
+    ATTEMPT_RESULT="FAIL"
+    ATTEMPT_DETAIL="marker $marker absent; exit $rc; snippet: $(printf '%s' "$raw" | cut -c1-160)"
   elif printf '%s' "$extracted" | grep -q "$marker"; then
     ATTEMPT_RESULT="PASS"
     ATTEMPT_DETAIL="marker $marker found"
-  elif [ -n "$amb_marker" ] && printf '%s' "$extracted" | grep -q "$amb_marker"; then
-    ATTEMPT_RESULT="AMBIENTAL"
-    ATTEMPT_DETAIL="explicit $amb_marker"
   elif [ -n "$verify" ] \
     && [ "$(cat "${verify%%::*}" 2>/dev/null)" = "${verify##*::}" ]; then
     ATTEMPT_RESULT="PASS"
     ATTEMPT_DETAIL="probe file verified: ${verify%%::*} contains ${verify##*::}"
-  elif printf '%s' "$raw" | grep -Eiq "$AMBIENTAL_RE"; then
-    local sig
-    sig="$(printf '%s' "$raw" | grep -Eio "$AMBIENTAL_RE" | head -1)"
-    ATTEMPT_RESULT="AMBIENTAL"
-    ATTEMPT_DETAIL="environmental signature: $sig"
   else
     ATTEMPT_RESULT="FAIL"
     ATTEMPT_DETAIL="marker $marker absent; rc=$rc; snippet: $(printf '%s' "$raw" | cut -c1-160)"
   fi
 }
 
-RETRY_COOLDOWN="${PANTHEON_RETRY_COOLDOWN:-60}"
-RETRIES_USED=0
-
-run_prompt() { # v idx current_bin — ONE retry after cooldown for real FAIL only
-  local v="$1" idx="$2" bin="$3" id first_detail
+run_prompt() { # v idx current_bin — exactly one attempt
+  local v="$1" idx="$2" bin="$3" id
   id="${PROMPT_IDS[$idx]}"
   run_prompt_attempt "$v" "$idx" "$bin"
-  if [ "$ATTEMPT_RESULT" = "FAIL" ]; then
-    first_detail="$ATTEMPT_DETAIL"
-    log "[$v] prompt $id FAIL on first attempt — retrying once after ${RETRY_COOLDOWN}s cooldown"
-    sleep "$RETRY_COOLDOWN"
-    run_prompt_attempt "$v" "$idx" "$bin"
-    RETRIES_USED=$((RETRIES_USED + 1))
-    if [ "$ATTEMPT_RESULT" != "FAIL" ]; then
-      ATTEMPT_DETAIL="$ATTEMPT_DETAIL [retry used — first attempt: $first_detail]"
-    else
-      ATTEMPT_DETAIL="$ATTEMPT_DETAIL [FAIL persisted after retry — first attempt: $first_detail]"
-    fi
-  fi
   record "$v" "$id" "$ATTEMPT_RESULT" "$ATTEMPT_DETAIL"
 }
 
 check_mcp_list() { # v
-  local v="$1" bin out rc=0
+  local v="$1" bin out connected_count rc=0
   if ! bin="$(sandbox_bin_v "$v")"; then
-    record "$v" "mcp-list" "AMBIENTAL" "binary not installed in sandbox prefix"
+    record "$v" "mcp-list" "FAIL" "binary not installed in sandbox prefix"
     return 0
   fi
   set +e
   out="$(cd "$(project_for "$v")" && "$bin" mcp list 2>&1)"
   rc=$?
   set -e
+  out="$(printf '%s' "$out" | sed -E $'s/\x1B\\[[0-?]*[ -/]*[@-~]//g')"
   if [ "$rc" -ne 0 ]; then
     record "$v" "mcp-list" "FAIL" "opencode mcp list exited $rc"
-  elif printf '%s' "$out" | grep -qi "connected" \
+    return 0
+  fi
+  connected_count="$(printf '%s' "$out" | grep -Eic '(^|[^[:alnum:]_])connected([^[:alnum:]_]|$)')" || connected_count=0
+  if [ "$connected_count" -eq 5 ] \
     && ! printf '%s' "$out" | grep -Eiq "failed|✘"; then
-    record "$v" "mcp-list" "PASS" "$(printf '%s' "$out" | grep -ci 'connected') connected"
+    record "$v" "mcp-list" "PASS" "$connected_count connected (expected exactly 5)"
   else
-    record "$v" "mcp-list" "FAIL" "no connected MCPs or failures reported"
+    record "$v" "mcp-list" "FAIL" "expected exactly 5 connected MCPs; found $connected_count or failures reported"
   fi
 }
 
 check_doctor() { # v
   local v="$1" pan rc=0
   if ! pan="$(sandbox_bin_for "pantheon-opencode")"; then
-    record "$v" "doctor" "AMBIENTAL" "pantheon-opencode not installed in sandbox prefix"
+    record "$v" "doctor" "FAIL" "pantheon-opencode not installed in sandbox prefix"
     return 0
   fi
   set +e
@@ -434,17 +503,101 @@ check_doctor() { # v
   fi
 }
 
+installed_package_root() {
+  local pan resolved
+  if ! pan="$(sandbox_bin_for "pantheon-opencode")"; then
+    return 1
+  fi
+  resolved="$(readlink -f "$pan" 2>/dev/null || true)"
+  [ -n "$resolved" ] || return 1
+  dirname "$(dirname "$resolved")"
+}
+
+run_cost_probe() { # v
+  local v="$1" package_root probe raw status detail
+  if ! package_root="$(installed_package_root)"; then
+    record "$v" "pantheon-cost" "FAIL" "sandbox package is not installed"
+    return 0
+  fi
+  probe="$package_root/scripts/probe-pantheon-cost.mjs"
+  if [ ! -f "$probe" ]; then
+    record "$v" "pantheon-cost" "FAIL" "installed package has no offline cost probe"
+    return 0
+  fi
+  if ! command -v node >/dev/null 2>&1; then
+    record "$v" "pantheon-cost" "FAIL" "node runtime is not available"
+    return 0
+  fi
+  set +e
+  raw="$(node "$probe" --version "$v" --fixture --json 2>&1)"
+  local rc=$?
+  set -e
+  case "$raw" in
+    *'"status":"PASS"'*) status="PASS" ;;
+    *'"status":"NOT_TESTED"'*) status="FAIL" ;;
+    *'"status":"AMBIENTAL"'*) status="FAIL" ;;
+    *) status="FAIL" ;;
+  esac
+  detail="$(printf '%s' "$raw" | tr '\n' ' ' | cut -c1-400)"
+  if [ "$rc" -ne 0 ] && [ "$status" = "PASS" ]; then
+    status="FAIL"
+    detail="$detail (probe exit $rc)"
+  elif [ "$rc" -ne 0 ] && [ "$status" != "FAIL" ]; then
+    detail="$detail (probe exit $rc)"
+  fi
+  record "$v" "pantheon-cost" "$status" "$detail"
+}
+
+run_rehydrate_probe() { # v
+  local v="$1" package_root probe raw status detail
+  if ! package_root="$(installed_package_root)"; then
+    record "$v" "context-rehydrate" "FAIL" "sandbox package is not installed"
+    return 0
+  fi
+  probe="$package_root/scripts/probe-context-rehydrate.mjs"
+  if [ ! -f "$probe" ]; then
+    record "$v" "context-rehydrate" "FAIL" "installed package has no offline context probe"
+    return 0
+  fi
+  if ! command -v node >/dev/null 2>&1; then
+    record "$v" "context-rehydrate" "FAIL" "node runtime is not available"
+    return 0
+  fi
+  set +e
+  raw="$(timeout 30 node "$probe" --version "$v" --json 2>&1)"
+  local rc=$?
+  set -e
+  case "$raw" in
+    *'"status":"PASS"'*) status="PASS" ;;
+    *'"status":"NOT_TESTED"'*) status="FAIL" ;;
+    *'"status":"AMBIENTAL"'*) status="FAIL" ;;
+    *) status="FAIL" ;;
+  esac
+  detail="$(printf '%s' "$raw" | tr '\n' ' ' | cut -c1-400)"
+  if [ "$rc" -eq 124 ]; then
+    status="FAIL"
+    detail="context probe timed out after 30s"
+  elif [ "$rc" -ne 0 ] && [ "$status" = "PASS" ]; then
+    # A non-zero probe exit must never be reported as PASS.
+    status="FAIL"
+    detail="$detail (probe exit $rc)"
+  elif [ "$rc" -ne 0 ] && [ "$status" != "FAIL" ]; then
+    detail="$detail (probe exit $rc)"
+  fi
+  record "$v" "context-rehydrate" "$status" "$detail"
+}
+
 run_version_prompts() { # v
   local v="$1" bin
   if ! bin="$(sandbox_bin_v "$v")"; then
-    log "Binary '$(bin_for "$v")' not in sandbox prefix — marking all $v prompts AMBIENTAL."
+    log "Binary '$(bin_for "$v")' not in sandbox prefix — failing all $v checks."
     build_battery "$v"
     local i
     for i in "${!PROMPT_IDS[@]}"; do
-      record "$v" "${PROMPT_IDS[$i]}" "AMBIENTAL" "binary $(bin_for "$v") not installed in sandbox"
+      record "$v" "${PROMPT_IDS[$i]}" "FAIL" "binary $(bin_for "$v") not installed in sandbox"
     done
-    record "$v" "mcp-list" "AMBIENTAL" "binary $(bin_for "$v") not installed in sandbox"
-    record "$v" "doctor" "AMBIENTAL" "binary $(bin_for "$v") not installed in sandbox"
+    record "$v" "mcp-list" "FAIL" "binary $(bin_for "$v") not installed in sandbox"
+    record "$v" "doctor" "FAIL" "binary $(bin_for "$v") not installed in sandbox"
     return 0
   fi
   prepare_project "$v"
@@ -459,8 +612,17 @@ run_version_prompts() { # v
   check_doctor "$v"
 }
 
+probe_verdict() { # pass total → PASS when every required check was explicitly PASS
+  local pass="$1" total="$2"
+  if [ "$pass" -eq "$total" ] && [ "$pass" -gt 0 ]; then
+    printf 'PASS'
+  else
+    printf 'FAIL'
+  fi
+}
+
 write_report() { # versions...
-  local versions=("$@") total_fail=0 total_pass=0 total_amb=0
+  local versions=("$@") total_fail=0 total_pass=0
   {
     printf '# Pantheon Sandbox Prompts Report\n\n'
     printf -- '- **Date:** %s\n' "$(date -u '+%Y-%m-%d %H:%M UTC')"
@@ -475,26 +637,108 @@ write_report() { # versions...
       for id in "${PROMPT_IDS[@]}" mcp-list doctor; do
         res="${RESULTS["$v:$id"]:-SKIP}"
         case "$res" in
-          PASS) total_pass=$((total_pass + 1)) ;;
-          AMBIENTAL) total_amb=$((total_amb + 1)) ;;
-          FAIL) total_fail=$((total_fail + 1)) ;;
+        PASS) total_pass=$((total_pass + 1)) ;;
+        *) total_fail=$((total_fail + 1)) ;;
         esac
         printf '| %s | %s | %s |\n' "$id" "$res" "${DETAILS["$v:$id"]:-—}"
       done
       printf '\n'
     done
     printf '## Summary\n\n'
-    printf -- '- PASS: %d\n- FAIL (real): %d\n- AMBIENTAL: %d\n- Retries used: %d\n\n' \
-      "$total_pass" "$total_fail" "$total_amb" "$RETRIES_USED"
-    if [ "$total_fail" -eq 0 ]; then
-      printf '**Verdict: PASS** — no real failures (AMBIENTAL issues do not block).\n'
-    else
-      printf '**Verdict: FAIL** — real failures detected (see table above).\n'
-    fi
+    printf -- '- PASS: %d\n- Not PASS (blocking): %d\n\n' \
+      "$total_pass" "$total_fail"
+    printf '**Verdict: %s**\n' "$(probe_verdict "$total_pass" "$((total_pass + total_fail))")"
   } > "$REPORT_FILE"
   log "Report written to $REPORT_FILE"
 }
 
+write_cost_report() { # versions...
+  local versions=("$@") total_fail=0 total_pass=0
+  {
+    printf '# Pantheon Cost Probe Report\n\n'
+    printf -- '- **Date:** %s\n' "$(date -u '+%Y-%m-%d %H:%M UTC')"
+    printf -- '- **Sandbox:** %s\n\n' "$SANDBOX_ROOT"
+    printf '| Version | Result | Detail |\n|---|---|---|\n'
+    local v result
+    for v in "${versions[@]}"; do
+      result="${RESULTS["$v:pantheon-cost"]:-NOT_TESTED}"
+      case "$result" in
+        PASS) total_pass=$((total_pass + 1)) ;;
+        *) total_fail=$((total_fail + 1)) ;;
+      esac
+      printf '| %s | %s | %s |\n' "$v" "$result" "${DETAILS["$v:pantheon-cost"]:-—}"
+    done
+    printf '\n## Summary\n\n'
+    printf -- '- PASS: %d\n- Not PASS (blocking): %d\n\n' \
+      "$total_pass" "$total_fail"
+    printf '**Verdict: %s**\n' "$(probe_verdict "$total_pass" "$((total_pass + total_fail))")"
+  } > "$COST_REPORT_FILE"
+  log "Cost probe report written to $COST_REPORT_FILE"
+}
+
+write_rehydrate_report() { # versions...
+  local versions=("$@") total_fail=0 total_pass=0
+  {
+    printf '# Pantheon Context Rehydration Probe Report\n\n'
+    printf -- '- **Date:** %s\n' "$(date -u '+%Y-%m-%d %H:%M UTC')"
+    printf -- '- **Sandbox:** %s\n\n' "$SANDBOX_ROOT"
+    printf '| Version | Probe | Result | Detail |\n|---|---|---|---|\n'
+    local v result
+    for v in "${versions[@]}"; do
+      result="${RESULTS["$v:context-rehydrate"]:-NOT_TESTED}"
+      case "$result" in
+        PASS) total_pass=$((total_pass + 1)) ;;
+        *) total_fail=$((total_fail + 1)) ;;
+      esac
+      printf '| %s | context_rehydrate + context_session_summary | %s | %s |\n' \
+        "$v" "$result" "${DETAILS["$v:context-rehydrate"]:-—}"
+    done
+    printf '\n## Summary\n\n'
+    printf -- '- PASS: %d\n- Not PASS (blocking): %d\n\n' \
+      "$total_pass" "$total_fail"
+    printf '**Verdict: %s**\n' "$(probe_verdict "$total_pass" "$((total_pass + total_fail))")"
+  } > "$REHYDRATE_REPORT_FILE"
+  log "Context probe report written to $REHYDRATE_REPORT_FILE"
+}
+
+run_cost() { # versions...
+  local versions=("$@") v fail=0
+  mkdir -p "$SANDBOX_ROOT"
+  for v in "${versions[@]}"; do
+    log "[$v] pantheon_cost offline probe ..."
+    run_cost_probe "$v"
+    log "[$v] pantheon-cost → ${RESULTS["$v:pantheon-cost"]}"
+  done
+  write_cost_report "${versions[@]}"
+  for v in "${versions[@]}"; do
+    [ "${RESULTS["$v:pantheon-cost"]:-}" = "PASS" ] || fail=$((fail + 1))
+  done
+  if [ "$fail" -gt 0 ]; then
+    log "Cost probe: $fail check(s) without explicit PASS."
+    exit 1
+  fi
+  log "Cost probe: PASS."
+}
+
+run_rehydrate() { # versions...
+  local versions=("$@") v fail=0
+  mkdir -p "$SANDBOX_ROOT"
+  for v in "${versions[@]}"; do
+    log "[$v] context_rehydrate/session_summary offline probe ..."
+    run_rehydrate_probe "$v"
+    log "[$v] context probe → ${RESULTS["$v:context-rehydrate"]}"
+  done
+  write_rehydrate_report "${versions[@]}"
+  for v in "${versions[@]}"; do
+    [ "${RESULTS["$v:context-rehydrate"]:-}" = "PASS" ] || fail=$((fail + 1))
+  done
+  if [ "$fail" -gt 0 ]; then
+    log "Context probe: $fail check(s) without explicit PASS."
+    exit 1
+  fi
+  log "Context probe: PASS."
+}
+
 run_battery() { # versions...
   local versions=("$@")
   mkdir -p "$SANDBOX_ROOT"
@@ -507,14 +751,14 @@ run_battery() { # versions...
   local v fail=0 id
   for v in "${versions[@]}"; do
     for id in "${PROMPT_IDS[@]}" mcp-list doctor; do
-      [ "${RESULTS["$v:$id"]:-}" = "FAIL" ] && fail=$((fail + 1)) || true
+      [ "${RESULTS["$v:$id"]:-}" = "PASS" ] || fail=$((fail + 1))
     done
   done
   if [ "$fail" -gt 0 ]; then
-    log "Prompts battery: $fail real FAIL(s)."
+    log "Prompts battery: $fail check(s) without explicit PASS."
     exit 1
   fi
-  log "Prompts battery: no real failures."
+  log "Prompts battery: PASS."
 }
 
 # ── Base validation (no prompts) ──────────────────────────────────────────────
@@ -535,13 +779,13 @@ run_base() { # v
   check_doctor "$v"
   log "[$v] doctor → ${RESULTS["$v:doctor"]}"
   local fail=0
-  [ "${RESULTS["$v:mcp-list"]}" = "FAIL" ] && fail=$((fail + 1)) || true
-  [ "${RESULTS["$v:doctor"]}" = "FAIL" ] && fail=$((fail + 1)) || true
+  [ "${RESULTS["$v:mcp-list"]}" = "PASS" ] || fail=$((fail + 1))
+  [ "${RESULTS["$v:doctor"]}" = "PASS" ] || fail=$((fail + 1))
   if [ "$fail" -gt 0 ]; then
-    log "Base validation $v: $fail real FAIL(s)."
+    log "Base validation $v: $fail check(s) without explicit PASS."
     exit 1
   fi
-  log "Base validation $v: OK."
+  log "Base validation $v: PASS."
 }
 
 # ── CLI ───────────────────────────────────────────────────────────────────────
@@ -551,6 +795,8 @@ while [ $# -gt 0 ]; do
     --prepare) MODE_PREPARE=1 ;;
     --reset) MODE_RESET=1 ;;
     --prompts) MODE_PROMPTS=1 ;;
+    --cost) MODE_COST=1 ;;
+    --rehydrate) MODE_REHYDRATE=1 ;;
     --run)
       shift
       case "${1:-}" in
@@ -569,7 +815,9 @@ if [ "$MODE_RESET" -eq 1 ]; then
   exit 0
 fi
 
-if [ "$MODE_PREPARE" -eq 0 ] && [ -z "$RUN_VERSION" ] && [ "$MODE_PROMPTS" -eq 0 ]; then
+if [ "$MODE_PREPARE" -eq 0 ] && [ -z "$RUN_VERSION" ] \
+  && [ "$MODE_PROMPTS" -eq 0 ] && [ "$MODE_COST" -eq 0 ] \
+  && [ "$MODE_REHYDRATE" -eq 0 ]; then
   usage
 fi
 
@@ -585,7 +833,30 @@ if [ -n "$RUN_VERSION" ]; then
     sandbox_env
     run_base "$RUN_VERSION"
   fi
+  if [ "$MODE_COST" -eq 1 ]; then
+    sandbox_env
+    run_cost "$RUN_VERSION"
+  fi
+  if [ "$MODE_REHYDRATE" -eq 1 ]; then
+    sandbox_env
+    run_rehydrate "$RUN_VERSION"
+  fi
 elif [ "$MODE_PROMPTS" -eq 1 ]; then
   sandbox_env
   run_battery v1 v2
+  if [ "$MODE_COST" -eq 1 ]; then
+    run_cost v1 v2
+  fi
+  if [ "$MODE_REHYDRATE" -eq 1 ]; then
+    run_rehydrate v1 v2
+  fi
+elif [ "$MODE_COST" -eq 1 ]; then
+  sandbox_env
+  run_cost v1 v2
+  if [ "$MODE_REHYDRATE" -eq 1 ]; then
+    run_rehydrate v1 v2
+  fi
+elif [ "$MODE_REHYDRATE" -eq 1 ]; then
+  sandbox_env
+  run_rehydrate v1 v2
 fi