pantheon-opencode 1.5.0-beta.1 → 1.5.0-beta.2

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/ambiental 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/ambiental context_rehydrate + context_session_summary
+#                    probe for the requested version (with --run vX) or both.
 #   --reset          Wipe the sandbox root.
 #   --help           Usage.
 #
@@ -49,11 +54,15 @@ 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 +114,8 @@ 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 "$0")" && pwd)"
@@ -124,9 +133,81 @@ export HOME="$SANDBOX_HOME"
 export PATH="$NPM_PREFIX/bin:$SANDBOX_VENV/bin:$PATH"
 export npm_config_prefix="$NPM_PREFIX"
 
+STATUS_FAIL=0
+STATUS_AMBIENTAL=0
+STATUS_NOT_TESTED=0
+STATUS_PASS=0
+MCP_AMBIENTAL_RE='needs authentication|unauthorized|authentication failure|api key|credentials|401|docker|econnrefused|enotfound|etimedout|econnreset|connection refused|rate.?limit|429|quota|overloaded|edgesout|npm err'
+
+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)) ;;
+    AMBIENTAL) STATUS_AMBIENTAL=$((STATUS_AMBIENTAL + 1)) ;;
+    NOT_TESTED) STATUS_NOT_TESTED=$((STATUS_NOT_TESTED + 1)) ;;
+  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 "$label" "NOT_TESTED" "binary '$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}"
+  elif printf '%s' "$out" | grep -Eiq "$MCP_AMBIENTAL_RE"; then
+    record_status "$label" "AMBIENTAL" "version check unavailable (exit $rc)"
+  else
+    record_status "$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 "$label MCP" "NOT_TESTED" "binary '$bin' is not installed in the sandbox prefix"
+    return 0
+  fi
+  if [ ! -d "$project" ]; then
+    record_status "$label MCP" "NOT_TESTED" "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 [ "$rc" -ne 0 ]; then
+    if printf '%s' "$out" | grep -Eiq "$MCP_AMBIENTAL_RE"; then
+      record_status "$label MCP" "AMBIENTAL" "mcp list environmental failure (exit $rc)"
+    else
+      record_status "$label MCP" "FAIL" "mcp list exited $rc"
+    fi
+    return 0
+  fi
+  connected_count="$(printf '%s' "$out" | grep -Eic '(^|[^[:alnum:]_])connected([^[:alnum:]_]|$)')" || connected_count=0
+  if [ "$connected_count" -eq 5 ] \
+    && ! printf '%s' "$out" | grep -Eiq 'failed|✘' \
+    && ! printf '%s' "$out" | grep -Eiq "$MCP_AMBIENTAL_RE"; then
+    record_status "$label MCP" "PASS" "exactly 5 connected MCPs"
+  elif printf '%s' "$out" | grep -Eiq "$MCP_AMBIENTAL_RE"; then
+    record_status "$label MCP" "AMBIENTAL" "environmental MCP status; found $connected_count connected"
+  else
+    record_status "$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 "$REPO_DIR"
@@ -134,42 +215,80 @@ npm pack 2>/dev/null || { echo "ERROR: npm pack failed"; exit 1; }
 TGZ=$(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 [ "$DOCTOR_RC" -eq 0 ]; then
+    record_status "doctor" "PASS" "exit 0"
+  elif printf '%s' "$DOCTOR_OUTPUT" | grep -Eiq "$MCP_AMBIENTAL_RE"; then
+    record_status "doctor" "AMBIENTAL" "environmental doctor failure (exit $DOCTOR_RC)"
+  else
+    record_status "doctor" "FAIL" "doctor exited $DOCTOR_RC"
+  fi
+else
+  record_status "doctor" "NOT_TESTED" "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" "NOT_TESTED" "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
+    if printf '%s' "$AGENTS_CONTENT" | grep -Eiq "$MCP_AMBIENTAL_RE"; then
+      record_status "pantheon://agents resource" "AMBIENTAL" "resource probe environmental failure (exit $AGENTS_RC)"
+    else
+      record_status "pantheon://agents resource" "FAIL" "resource probe exited $AGENTS_RC"
+    fi
+  elif printf '%s\n' "$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 ]; then
+  FINAL_VERDICT="FAIL"
+elif [ "$STATUS_NOT_TESTED" -gt 0 ]; then
+  FINAL_VERDICT="NOT_TESTED"
+elif [ "$STATUS_AMBIENTAL" -gt 0 ]; then
+  FINAL_VERDICT="AMBIENTAL"
+else
+  FINAL_VERDICT="PASS"
+fi
+printf 'Final verdict: %s\n' "$FINAL_VERDICT"
+if [ "$FINAL_VERDICT" = "FAIL" ]; then
+  exit 1
+fi
 echo "Next: open the isolated TUI with: $SANDBOX_DIR/start-pantheon.sh"
 RUNTEST
   chmod +x "$SANDBOX_ROOT/run-test.sh"
@@ -195,6 +314,7 @@ 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"
@@ -273,6 +393,7 @@ PYEOF
 
 # 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'
+MCP_AMBIENTAL_RE='needs authentication|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=()
@@ -354,12 +475,20 @@ run_prompt_attempt() { # v idx current_bin — sets ATTEMPT_RESULT / ATTEMPT_DET
   if [ "$rc" -eq 124 ]; then
     ATTEMPT_RESULT="AMBIENTAL"
     ATTEMPT_DETAIL="timeout after ${PANTHEON_PROMPT_TIMEOUT}s (provider latency)"
+  elif [ "$rc" -ne 0 ]; then
+    if [ -n "$amb_marker" ] && printf '%s' "$extracted" | grep -q "$amb_marker"; then
+      ATTEMPT_RESULT="AMBIENTAL"
+      ATTEMPT_DETAIL="explicit $amb_marker (exit $rc)"
+    elif printf '%s' "$raw" | grep -Eiq "$AMBIENTAL_RE"; then
+      ATTEMPT_RESULT="AMBIENTAL"
+      ATTEMPT_DETAIL="environmental signature with exit $rc"
+    else
+      ATTEMPT_RESULT="FAIL"
+      ATTEMPT_DETAIL="marker $marker absent; exit $rc; snippet: $(printf '%s' "$raw" | cut -c1-160)"
+    fi
   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"
@@ -398,7 +527,7 @@ run_prompt() { # v idx current_bin — ONE retry after cooldown for real FAIL on
 }
 
 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"
     return 0
@@ -407,13 +536,24 @@ check_mcp_list() { # v
   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" \
-    && ! printf '%s' "$out" | grep -Eiq "failed|✘"; then
-    record "$v" "mcp-list" "PASS" "$(printf '%s' "$out" | grep -ci 'connected') connected"
+    if printf '%s' "$out" | grep -Eiq "$MCP_AMBIENTAL_RE"; then
+      record "$v" "mcp-list" "AMBIENTAL" "mcp list environmental failure (exit $rc)"
+    else
+      record "$v" "mcp-list" "FAIL" "opencode mcp list exited $rc"
+    fi
+    return 0
+  fi
+  connected_count="$(printf '%s' "$out" | grep -Eic '(^|[^[:alnum:]_])connected([^[:alnum:]_]|$)')" || connected_count=0
+  if [ "$connected_count" -eq 5 ] \
+    && ! printf '%s' "$out" | grep -Eiq "failed|✘" \
+    && ! printf '%s' "$out" | grep -Eiq "$MCP_AMBIENTAL_RE"; then
+    record "$v" "mcp-list" "PASS" "$connected_count connected (expected exactly 5)"
+  elif printf '%s' "$out" | grep -Eiq "$MCP_AMBIENTAL_RE"; then
+    record "$v" "mcp-list" "AMBIENTAL" "environmental MCP status; found $connected_count connected"
   else
-    record "$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
 }
 
@@ -434,6 +574,90 @@ 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 "$v" "pantheon-cost" "NOT_TESTED" "sandbox package is not installed"
+    return 0
+  fi
+  probe="$package_root/scripts/probe-pantheon-cost.mjs"
+  if [ ! -f "$probe" ]; then
+    record "$v" "pantheon-cost" "NOT_TESTED" "installed package has no offline cost probe"
+    return 0
+  fi
+  if ! command -v node >/dev/null 2>&1; then
+    record "$v" "pantheon-cost" "AMBIENTAL" "node runtime is not available"
+    return 0
+  fi
+  set +e
+  raw="$(node "$probe" --version "$v" --fixture --json 2>&1)"
+  local rc=$?
+  set -e
+  case "$raw" in
+    *'"status":"PASS"'*) status="PASS" ;;
+    *'"status":"NOT_TESTED"'*) status="NOT_TESTED" ;;
+    *'"status":"AMBIENTAL"'*) status="AMBIENTAL" ;;
+    *) 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 "$v" "context-rehydrate" "NOT_TESTED" "sandbox package is not installed"
+    return 0
+  fi
+  probe="$package_root/scripts/probe-context-rehydrate.mjs"
+  if [ ! -f "$probe" ]; then
+    record "$v" "context-rehydrate" "NOT_TESTED" "installed package has no offline context probe"
+    return 0
+  fi
+  if ! command -v node >/dev/null 2>&1; then
+    record "$v" "context-rehydrate" "AMBIENTAL" "node runtime is not available"
+    return 0
+  fi
+  set +e
+  raw="$(timeout 30 node "$probe" --version "$v" --json 2>&1)"
+  local rc=$?
+  set -e
+  case "$raw" in
+    *'"status":"PASS"'*) status="PASS" ;;
+    *'"status":"NOT_TESTED"'*) status="NOT_TESTED" ;;
+    *'"status":"AMBIENTAL"'*) status="AMBIENTAL" ;;
+    *) status="FAIL" ;;
+  esac
+  detail="$(printf '%s' "$raw" | tr '\n' ' ' | cut -c1-400)"
+  if [ "$rc" -eq 124 ]; then
+    status="AMBIENTAL"
+    detail="context probe timed out after 30s"
+  elif [ "$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 "$v")"; then
@@ -459,8 +683,22 @@ run_version_prompts() { # v
   check_doctor "$v"
 }
 
+probe_verdict() { # fail pass ambiental not_tested → one final status
+  # Verdict: PASS; Verdict: FAIL; Verdict: AMBIENTAL; Verdict: NOT_TESTED remain distinct.
+  local fail="$1" pass="$2" ambiental="$3" not_tested="$4"
+  if [ "$fail" -gt 0 ]; then
+    printf 'FAIL'
+  elif [ "$not_tested" -gt 0 ]; then
+    printf 'NOT_TESTED'
+  elif [ "$ambiental" -gt 0 ]; then
+    printf 'AMBIENTAL'
+  else
+    printf 'PASS'
+  fi
+}
+
 write_report() { # versions...
-  local versions=("$@") total_fail=0 total_pass=0 total_amb=0
+  local versions=("$@") total_fail=0 total_pass=0 total_amb=0 total_not_tested=0
   {
     printf '# Pantheon Sandbox Prompts Report\n\n'
     printf -- '- **Date:** %s\n' "$(date -u '+%Y-%m-%d %H:%M UTC')"
@@ -475,26 +713,114 @@ write_report() { # versions...
       for id in "${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)) ;;
+        AMBIENTAL) total_amb=$((total_amb + 1)) ;;
+        FAIL) total_fail=$((total_fail + 1)) ;;
+        NOT_TESTED) total_not_tested=$((total_not_tested + 1)) ;;
         esac
         printf '| %s | %s | %s |\n' "$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 [ "$total_fail" -eq 0 ]; then
-      printf '**Verdict: PASS** — no real failures (AMBIENTAL issues do not block).\n'
-    else
-      printf '**Verdict: FAIL** — real failures detected (see table above).\n'
-    fi
+    printf -- '- PASS: %d\n- FAIL (real): %d\n- AMBIENTAL: %d\n- NOT_TESTED: %d\n- Retries used: %d\n\n' \
+      "$total_pass" "$total_fail" "$total_amb" "$total_not_tested" "$RETRIES_USED"
+    printf '**Verdict: %s**\n' "$(probe_verdict "$total_fail" "$total_pass" "$total_amb" "$total_not_tested")"
   } > "$REPORT_FILE"
   log "Report written to $REPORT_FILE"
 }
 
+write_cost_report() { # versions...
+  local versions=("$@") total_fail=0 total_pass=0 total_ambient=0 total_not_tested=0
+  {
+    printf '# Pantheon Cost Probe Report\n\n'
+    printf -- '- **Date:** %s\n' "$(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)) ;;
+        FAIL) total_fail=$((total_fail + 1)) ;;
+        AMBIENTAL) total_ambient=$((total_ambient + 1)) ;;
+        NOT_TESTED) total_not_tested=$((total_not_tested + 1)) ;;
+      esac
+      printf '| %s | %s | %s |\n' "$v" "$result" "${DETAILS["$v:pantheon-cost"]:-—}"
+    done
+    printf '\n## Summary\n\n'
+    printf -- '- PASS: %d\n- FAIL (real): %d\n- AMBIENTAL: %d\n- NOT_TESTED: %d\n\n' \
+      "$total_pass" "$total_fail" "$total_ambient" "$total_not_tested"
+    printf '**Verdict: %s**\n' "$(probe_verdict "$total_fail" "$total_pass" "$total_ambient" "$total_not_tested")"
+  } > "$COST_REPORT_FILE"
+  log "Cost probe report written to $COST_REPORT_FILE"
+}
+
+write_rehydrate_report() { # versions...
+  local versions=("$@") total_fail=0 total_pass=0 total_ambient=0 total_not_tested=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=$((total_pass + 1)) ;;
+        FAIL) total_fail=$((total_fail + 1)) ;;
+        AMBIENTAL) total_ambient=$((total_ambient + 1)) ;;
+        NOT_TESTED) total_not_tested=$((total_not_tested + 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- FAIL (real): %d\n- AMBIENTAL: %d\n- NOT_TESTED: %d\n\n' \
+      "$total_pass" "$total_fail" "$total_ambient" "$total_not_tested"
+    printf '**Verdict: %s**\n' "$(probe_verdict "$total_fail" "$total_pass" "$total_ambient" "$total_not_tested")"
+  } > "$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"]:-}" = "FAIL" ] && fail=$((fail + 1)) || true
+  done
+  if [ "$fail" -gt 0 ]; then
+    log "Cost probe: $fail real FAIL(s)."
+    exit 1
+  fi
+  log "Cost probe: no real failures (NOT_TESTED/AMBIENTAL is non-blocking)."
+}
+
+run_rehydrate() { # versions...
+  local versions=("$@") 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"]:-}" = "FAIL" ] && fail=$((fail + 1)) || true
+  done
+  if [ "$fail" -gt 0 ]; then
+    log "Context probe: $fail real FAIL(s)."
+    exit 1
+  fi
+  log "Context probe: no real failures (NOT_TESTED/AMBIENTAL is non-blocking)."
+}
+
 run_battery() { # versions...
   local versions=("$@")
   mkdir -p "$SANDBOX_ROOT"
@@ -551,6 +877,8 @@ 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 +897,9 @@ 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 +915,30 @@ 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

package/src/mcp/mcp_persistence_server.py

@@ -566,7 +566,7 @@ async def kv_delete_namespace(
 # ── Context Checkpoint Tools ────────────────────────────────────────────────────
 # Session-scoped context storage for deepwork checkpoints, phase state,
 # tool call history, and reasoning chains. All entries use namespace
-# "checkpoint:{slug}" with auto-TTL of 4 hours (session duration).
+# "checkpoint:{slug}:{session}" with auto-TTL of 4 hours (session duration).
 
 DEFAULT_CONTEXT_TTL: int = 14400  # 4 hours
 
@@ -1023,6 +1023,31 @@ def _load_checkpoint_for_rehydrate(
     return None
 
 
+def _checkpoint_namespaces(
+    conn: sqlite3.Connection, slug: str, session_id: str | None
+) -> list[str]:
+    """Resolve checkpoint namespaces while preserving session isolation.
+
+    ``context_save`` always creates a session-qualified namespace, including
+    when the caller omits ``session_id`` (it generates one and returns it).
+    Runtime callers may not retain that return value after compaction, so an
+    unqualified rehydrate/summary must discover the newest matching session.
+    The prefix comparison is parameterized and avoids LIKE wildcard leakage.
+    """
+    if session_id:
+        return [f"checkpoint:{slug}:{session_id}"]
+    prefix = f"checkpoint:{slug}:"
+    rows = conn.execute(
+        "SELECT namespace FROM kv_store "
+        "WHERE substr(namespace, 1, ?) = ? "
+        "AND (expires_at IS NULL OR datetime(expires_at) > datetime('now')) "
+        "AND deleted_at IS NULL "
+        "GROUP BY namespace ORDER BY MAX(updated_at) DESC LIMIT 20",
+        (len(prefix), prefix),
+    ).fetchall()
+    return [str(row[0]) for row in rows if row and row[0]]
+
+
 def _refresh_heartbeat_ttl(conn: sqlite3.Connection, namespace: str) -> None:
     """Extend an unexpired heartbeat TTL; never shorten, never resurrect."""
     row = conn.execute(
@@ -1055,7 +1080,8 @@ def _refresh_heartbeat_ttl(conn: sqlite3.Connection, namespace: str) -> None:
 @mcp.tool(
     name="context_rehydrate",
     description="Rehydrate critical context after native compaction. "
-    "Reads 'latest' + serialized 'tail' from checkpoint:<slug> and rebuilds "
+    "Reads 'latest' + serialized 'tail' from a session-qualified checkpoint "
+    "namespace and rebuilds "
     "goal/phase/delegation blocks deterministically (no LLM). Refreshes the "
     "heartbeat TTL. Idempotent. Honors PANTHEON_COMPACTION=off.",
 )
@@ -1068,14 +1094,21 @@ async def context_rehydrate(
     if _compaction_disabled():
         return None
     conn = _db(scope)
-    ns = f"checkpoint:{slug}:{session_id}" if session_id else f"checkpoint:{slug}"
-    checkpoint = _load_checkpoint_for_rehydrate(conn, ns)
+    namespaces = _checkpoint_namespaces(conn, slug, session_id)
+    checkpoint = None
+    selected_namespace = None
+    for namespace in namespaces:
+        checkpoint = _load_checkpoint_for_rehydrate(conn, namespace)
+        if checkpoint is not None:
+            selected_namespace = namespace
+            break
     if checkpoint is None:
         return None
     blocks = build_rehydration_blocks(checkpoint)
     if blocks is None:
         return None
-    _refresh_heartbeat_ttl(conn, ns)
+    if selected_namespace is not None:
+        _refresh_heartbeat_ttl(conn, selected_namespace)
     return blocks
 
 
@@ -1095,8 +1128,12 @@ async def context_session_summary(
     if os.environ.get(SESSION_END_SUMMARY_ENV, "").lower() == "off":
         return None
     conn = _db(scope)
-    ns = f"checkpoint:{slug}:{session_id}" if session_id else f"checkpoint:{slug}"
-    checkpoint = _load_checkpoint_for_rehydrate(conn, ns)
+    namespaces = _checkpoint_namespaces(conn, slug, session_id)
+    checkpoint = None
+    for namespace in namespaces:
+        checkpoint = _load_checkpoint_for_rehydrate(conn, namespace)
+        if checkpoint is not None:
+            break
     if checkpoint is None:
         return None
     return build_session_end_summary(checkpoint)