@ai-dev-methodologies/rlp-desk 0.18.1 → 0.18.2

package/CHANGELOG.md

@@ -11,6 +11,30 @@ For pre-v0.15.4 versions, refer to `git log` and individual GitHub release notes
 - Post-v0.15.6: remove `RLP_LIFECYCLE_METRICS` flag entirely (per plan v3 ADR follow-ups).
 - Phase D.1 (handoff documents) + Phase D.2 (per-stage agent role specialization) — both deferred per `docs/plans/v0.15.4-release-runbook.md` §7.6.
 
+## [0.18.2] — 2026-06-25
+
+Five leader-resilience fixes (per-us, consensus, and config-validation paths), each
+validated by a real-LLM dogfood. Consensus is opt-in; default per-us campaigns benefit
+from the final-verify and config-validation fixes.
+
+### Fixed
+- Final verification is resilient to verifier non-determinism: a user story that already
+  passed its per-US check is re-verified (up to a small bound) on a fail verdict, so a
+  flaky false-fail must reproduce before it charges a fix-loop failure. A single
+  non-deterministic false-fail can no longer block a complete, correct campaign.
+- The `claude` rate-limit banner ("API Error: … temporarily limiting requests … · Rate
+  limited") is now recognized as a transient API condition and routed to the bounded
+  backoff, instead of being misclassified as a frozen-pane deadlock.
+- Numeric configuration knobs (e.g. `--max-iter`, `--cb-threshold`, timeouts) are now
+  validated: a non-integer or out-of-range value falls back to its default with a warning,
+  instead of silently mis-evaluating (a bad `--max-iter` previously could make a campaign
+  run zero iterations).
+- Leader auto-commit recovery no longer blocks a campaign whose work the worker already
+  committed (a commit-timing race previously surfaced as "nothing to commit" → blocked).
+- `--consensus final-only` now actually runs at the final verification in the default
+  per-us verify mode (it was previously bypassed by the sequential final-verify path, so
+  the recommended consensus configuration was a silent no-op in the default mode).
+
 ## [0.18.1] — 2026-06-25
 
 ### Fixed

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ai-dev-methodologies/rlp-desk",
-  "version": "0.18.1",
+  "version": "0.18.2",
   "description": "Fresh-context iterative loops for Claude Code — autonomous task completion with independent verification",
   "scripts": {
     "postinstall": "node scripts/postinstall.js",

package/src/scripts/run_ralph_desk.zsh

@@ -3,6 +3,31 @@ set -uo pipefail
 # NOTE: We use set -u (undefined var check) and pipefail, but NOT set -e
 # because the main loop uses explicit error checks throughout.
 
+# D-19: validate an env-overridable INTEGER knob. A non-integer value (operator
+# typo, or a bad CLI arg like `--max-iter abc` threaded into the env) otherwise
+# mis-evaluates under `set -u` inside (( )) arithmetic — e.g. a non-integer
+# MAX_ITER makes the main-loop bound error so the campaign silently runs ZERO
+# iterations; a non-integer CB_THRESHOLD breaks the circuit breaker. The `<->`
+# integer glob is checked FIRST (and short-circuits) so the arithmetic never runs
+# on a non-integer. A malformed / below-min / above-max value → the default.
+_validate_int_knob() {
+  local _name="$1" _default="$2" _min="${3:-0}" _max="${4:-0}"
+  local _val="${(P)_name}"
+  local _bad=0
+  if ! [[ "$_val" == <-> ]]; then
+    _bad=1
+  elif (( _val < _min )); then
+    _bad=1
+  elif (( _max > 0 && _val > _max )); then
+    _bad=1
+  fi
+  if (( _bad )); then
+    local _range="min=$_min"; (( _max > 0 )) && _range="$_range, max=$_max"
+    print -r -- "WARNING: $_name='$_val' is not a valid integer ($_range) — using default $_default" >&2
+    eval "$_name=$_default"
+  fi
+}
+
 # =============================================================================
 # Ralph Desk Tmux Runner
 #
@@ -56,6 +81,14 @@ HEARTBEAT_STALE_THRESHOLD="${HEARTBEAT_STALE_THRESHOLD:-120}"
 MAX_RESTARTS="${MAX_RESTARTS:-3}"
 IDLE_NUDGE_THRESHOLD="${IDLE_NUDGE_THRESHOLD:-30}"
 MAX_NUDGES="${MAX_NUDGES:-3}"
+# D-19: validate the numeric knobs above (set -u + (( )) arithmetic safety).
+_validate_int_knob MAX_ITER 20 1
+_validate_int_knob POLL_INTERVAL 5 1
+_validate_int_knob ITER_TIMEOUT 600 1
+_validate_int_knob HEARTBEAT_STALE_THRESHOLD 120 1
+_validate_int_knob MAX_RESTARTS 3 0
+_validate_int_knob IDLE_NUDGE_THRESHOLD 30 1
+_validate_int_knob MAX_NUDGES 3 0
 WITH_SELF_VERIFICATION="${WITH_SELF_VERIFICATION:-0}"
 WITH_SELF_VERIFICATION_REQUESTED="$WITH_SELF_VERIFICATION"  # preserves original user intent for traceability (governance §1f)
 SV_SKIPPED_REASON=""                                         # set when SV is disabled despite user request
@@ -88,6 +121,7 @@ TEST_DENSITY_MODE="${TEST_DENSITY_MODE:-warn}"
 # .sisyphus/mission-abort.json and exits non-zero so contract defects don't
 # silently loop. infra_failure category and the very first iteration are exempt.
 BLOCK_CB_THRESHOLD="${BLOCK_CB_THRESHOLD:-3}"
+_validate_int_knob BLOCK_CB_THRESHOLD 3 1   # D-19
 CONSECUTIVE_BLOCKS=0
 LAST_BLOCK_REASON=""
 
@@ -229,6 +263,18 @@ FINAL_VERIFIER_ENGINE="${FINAL_VERIFIER_ENGINE:-claude}"
 WORKER_EFFORT="${WORKER_EFFORT:-}"
 VERIFIER_EFFORT="${VERIFIER_EFFORT:-}"
 FINAL_VERIFIER_EFFORT="${FINAL_VERIFIER_EFFORT:-}"
+# D-18: max final-verify attempts for a US that ALREADY passed per-US. A verifier
+# fail verdict on already-per-US-passed work must REPRODUCE across all attempts
+# (first pass wins) before it charges a fix-loop failure — guards against verifier
+# non-determinism defeating a complete, correct campaign. A genuinely-regressed US
+# (or one never per-US-passed) still fails on the first attempt.
+FINAL_VERIFY_MAX_ATTEMPTS="${FINAL_VERIFY_MAX_ATTEMPTS:-3}"
+# D-18/D-19: a non-integer value ("abc") would mis-evaluate under set -u in the
+# (( )) retry-loop arithmetic — skipping the loop and silently FALSE-FAILING a US
+# — and an unbounded value would be ruinously expensive. Validate to an integer
+# in 1..10 via the shared _validate_int_knob helper (D-19 generalized this
+# per-knob fix into one validator used by every numeric knob).
+_validate_int_knob FINAL_VERIFY_MAX_ATTEMPTS 3 1 10
 
 # Auto-detect engine from model format for env var path (CLI path uses parse_model_flag)
 _auto_detect_engine WORKER_MODEL WORKER_ENGINE WORKER_CODEX_MODEL WORKER_CODEX_REASONING WORKER_EFFORT
@@ -260,6 +306,7 @@ elif [[ "${FINAL_CONSENSUS:-0}" = "1" ]]; then
 fi
 CONSENSUS_SCOPE="${CONSENSUS_SCOPE:-${CONSENSUS_MODE}}"
 CB_THRESHOLD="${CB_THRESHOLD:-6}"           # consecutive failures before BLOCKED (default: 6)
+_validate_int_knob CB_THRESHOLD 6 1   # D-19: must be valid before the (( *2 )) below
 # Effective CB threshold: doubled when consensus mode active
 if [[ "$CONSENSUS_MODE" != "off" ]]; then
   EFFECTIVE_CB_THRESHOLD=$(( CB_THRESHOLD * 2 ))
@@ -268,6 +315,8 @@ else
 fi
 _API_MAX_RETRIES="${_API_MAX_RETRIES:-5}"
 _API_RETRY_INTERVAL_S="${_API_RETRY_INTERVAL_S:-30}"
+_validate_int_knob _API_MAX_RETRIES 5 1        # D-19
+_validate_int_knob _API_RETRY_INTERVAL_S 30 1  # D-19
 
 # --- Derived Paths ---
 DESK="$ROOT/${RLP_DESK_RUNTIME_DIR:-.rlp-desk}"
@@ -847,7 +896,25 @@ _bug8_check_synth_allowed() {
       log "  Bug #8 F-8 recovery: done-claim + Worker's uncommitted tracked changes — auto-committing $us_id work (files: $_bug8_first5)."
       log_debug "[GOV] iter=$iter bug8=recover_autocommit us_id=$us_id files='$_bug8_first5'"
       local -a _bug8_add=("${(@f)_bug8_worker_files}")
-      if git -C "$ROOT" add -- "${_bug8_add[@]}" && git -C "$ROOT" commit -q -m "chore(leader-recovery): commit Worker's uncommitted $us_id changes (Bug #8 F-8)"; then
+      # D-20 (codex LOW): fail-safe on an empty file list. The upstream
+      # `[[ -z "$_bug8_worker_files" ]]` guard already makes this unreachable, but
+      # never let an empty array turn `git diff --quiet HEAD --` into a whole-tree
+      # check (which could falsely read "already committed" → false PASS). BLOCK.
+      if (( ${#_bug8_add} == 0 )); then
+        log_error "  Bug #8: empty worker-file list at auto-commit (unexpected) — refusing synthesis."
+        write_blocked_sentinel "worker_incomplete_uncommitted: empty file list at auto-commit" "$us_id" "metric_failure"
+        return 1
+      fi
+      if git -C "$ROOT" diff --quiet HEAD -- "${_bug8_add[@]}" 2>/dev/null; then
+        # D-20: the Worker committed these files itself in the window between the
+        # dirty-detection above and now (a reap/commit race) — the working tree is
+        # already clean vs HEAD for them, i.e. the work IS committed. The old code
+        # ran `git add … && git commit`, which exited non-zero ("nothing to commit")
+        # and BLOCKED a correct, fully-committed campaign. Treat "already committed"
+        # as success: proceed to synthesis (the Verifier still gates correctness).
+        log "  Bug #8 F-8 (D-20): Worker files already committed (commit race) — nothing to auto-commit; proceeding."
+        log_debug "[GOV] iter=$iter bug8=autocommit_noop_already_committed us_id=$us_id files='$_bug8_first5'"
+      elif git -C "$ROOT" add -- "${_bug8_add[@]}" && git -C "$ROOT" commit -q -m "chore(leader-recovery): commit Worker's uncommitted $us_id changes (Bug #8 F-8)"; then
         log "  Leader-recovery auto-commit OK (Worker files only) — Verifier will gate correctness."
       else
         log_error "  Bug #8: leader-recovery auto-commit failed. Refusing synthesis. files: $_bug8_first5"
@@ -1434,6 +1501,8 @@ typeset -gA PANE_PROMPT_STUCK_SINCE
 typeset -gA PANE_DISMISS_FAILED_COUNT
 PROMPT_STALL_TIMEOUT="${PROMPT_STALL_TIMEOUT:-300}"  # 5 min default
 PROMPT_DISMISS_FAIL_LIMIT="${PROMPT_DISMISS_FAIL_LIMIT:-20}"  # ~100s of fruitless dismiss attempts
+_validate_int_knob PROMPT_STALL_TIMEOUT 300 1      # D-19
+_validate_int_knob PROMPT_DISMISS_FAIL_LIMIT 20 1  # D-19
 
 # v5.7 §4.17: generic no-progress timeout (codex Critic HIGH — closes the gap
 # where an undetected prompt or alive-but-frozen Worker bypasses Layer 4).
@@ -1441,6 +1510,7 @@ PROMPT_DISMISS_FAIL_LIMIT="${PROMPT_DISMISS_FAIL_LIMIT:-20}"  # ~100s of fruitle
 # seconds AND signal file still missing, write BLOCKED `infra_failure` reason
 # `worker_no_progress` so silent infinite-wait is impossible.
 PROGRESS_NO_CHANGE_TIMEOUT="${PROGRESS_NO_CHANGE_TIMEOUT:-600}"  # 10 min default
+_validate_int_knob PROGRESS_NO_CHANGE_TIMEOUT 600 1  # D-19
 typeset -gA PANE_LAST_CHANGE_TS  # epoch when content last changed
 typeset -gA PANE_LAST_CONTENT_FOR_PROGRESS  # captured content for diff
 
@@ -1449,6 +1519,7 @@ typeset -gA PANE_LAST_CONTENT_FOR_PROGRESS  # captured content for diff
 # CODEX_IDLE_GRACE_S (default 120s) before BLOCK. Per-pane bookkeeping to
 # avoid granting it repeatedly. Bug Report #3 (BOS 2026-05-04).
 CODEX_IDLE_GRACE_S="${CODEX_IDLE_GRACE_S:-120}"
+_validate_int_knob CODEX_IDLE_GRACE_S 120 1  # D-19
 typeset -gA PANE_CODEX_IDLE_GRACED
 # v0.14.2: per-verifier-pane trace flag — log the verdict-lookup outcome
 # exactly once per byte-stasis transition. Bug Report #4 (BOS 2026-05-05).
@@ -2520,9 +2591,22 @@ poll_for_signal() {
     if [[ -n "$pane_output_for_retry" ]] &&
        ( echo "$pane_output_for_retry" | grep -qiE '(^|[^[:digit:]])500([^[:digit:]]|$)' \
       || echo "$pane_output_for_retry" | grep -qiE '(^|[^[:digit:]])529([^[:digit:]]|$)' \
+      || echo "$pane_output_for_retry" | grep -qiE '(^|[^[:digit:]])429([^[:digit:]]|$)' \
       || echo "$pane_output_for_retry" | grep -qi 'overloaded' \
       || echo "$pane_output_for_retry" | grep -qi 'too many requests' \
-      || echo "$pane_output_for_retry" | grep -qi 'service unavailable' ); then
+      || echo "$pane_output_for_retry" | grep -qi 'service unavailable' \
+      || echo "$pane_output_for_retry" | grep -qiE 'api error.*temporarily limiting requests' ); then
+      # D-17a: the last pattern catches the claude TUI rate-limit banner
+      # ("API Error: Server is temporarily limiting requests (not your usage
+      # limit) · Rate limited") that previously fell through to the 600s
+      # frozen-pane BLOCK with a misleading "deadlock" reason. It requires BOTH
+      # the "API Error" banner prefix AND the distinctive multi-word phrase
+      # "temporarily limiting requests" on the SAME line (codex MEDIUM): a Worker
+      # implementing a rate-limiter feature, quoting the phrase, or merely
+      # discussing API rate-limit handling does NOT false-trigger backoff — only
+      # the actual error banner does. Routes to the bounded API backoff below
+      # (5×30s) → recovers a transient limit, else BLOCKs as infra (recoverable),
+      # not as a misleading frozen-pane deadlock.
       is_api_text_retry=1
     fi
 
@@ -2832,6 +2916,94 @@ _all_us_verified() {
   return 0
 }
 
+# D-18 helper: one final-verify pass for a single US. Returns 0=pass verdict,
+# 1=fail verdict, 2=infra-terminal (launch/poll hard fail — sentinel handling
+# already done by the poll). Reads/updates globals (VERIFIER_PANE, FINAL_VERIFIER_*,
+# SIGNAL_FILE, VERDICT_FILE, SESSION_CONFIG). Extracted from run_sequential_final_verify
+# so the caller can re-verify a per-US-passed US on a flake without duplicating the
+# dispatch/poll logic. Does NOT set FAILED_US (the caller owns that).
+_final_verify_one_us() {
+  local us="$1" iter="$2"
+
+  # Temporarily override signal file to scope verifier to this US
+  local orig_signal
+  orig_signal=$(cat "$SIGNAL_FILE" 2>/dev/null)
+  echo "{\"status\":\"verify\",\"us_id\":\"$us\",\"summary\":\"sequential final verify\"}" | atomic_write "$SIGNAL_FILE"
+
+  # Write scoped verifier trigger
+  write_verifier_trigger "$iter"
+  local verifier_prompt="$LOGS_DIR/iter-$(printf '%03d' $iter).verifier-prompt.md"
+
+  # Clean verifier pane
+  local verifier_cmd
+  verifier_cmd=$(tmux display-message -p -t "$VERIFIER_PANE" '#{pane_current_command}' 2>/dev/null)
+  if [[ "$verifier_cmd" == "node" || "$verifier_cmd" == "claude" || "$verifier_cmd" == "codex" ]]; then
+    tmux send-keys -t "$VERIFIER_PANE" C-c 2>/dev/null; sleep 0.5
+    tmux send-keys -t "$VERIFIER_PANE" "/exit" C-m 2>/dev/null; sleep 2
+  fi
+  wait_for_pane_ready "$VERIFIER_PANE" 10 2>/dev/null || true
+
+  # Launch verifier. D-1: the FINAL (ALL) verify uses FINAL_VERIFIER_* (the
+  # "final 엄격" knob — a configured stronger model, e.g. opus, for the final
+  # gate), NOT the lighter per-US VERIFIER_*. This is the configured-final-model
+  # distinction, distinct from the removed per-iteration verifier auto-upgrade.
+  local verifier_launch
+  if [[ "$FINAL_VERIFIER_ENGINE" = "codex" ]]; then
+    verifier_launch="${CODEX_BIN:-codex} -m $FINAL_VERIFIER_CODEX_MODEL -c model_reasoning_effort=\"$FINAL_VERIFIER_CODEX_REASONING\" -c mcp_servers='{}' --disable plugins --dangerously-bypass-approvals-and-sandbox"
+    launch_verifier_codex "$VERIFIER_PANE" "$verifier_prompt" "$iter" "$verifier_launch"
+  else
+    verifier_launch="$(build_claude_cmd tui "$FINAL_VERIFIER_MODEL" "" "" "$FINAL_VERIFIER_EFFORT")"
+    launch_verifier_claude "$VERIFIER_PANE" "$verifier_prompt" "$iter" "$verifier_launch" || {
+      log_error "Failed to launch final verifier for $us"
+      return 2
+    }
+  fi
+
+  # Poll for verdict. D-4: distinguish rc==2 (hard-fail, sentinel already
+  # written → terminal) from rc==1 (transient pane race/timeout) and give ONE
+  # replace-pane + re-dispatch retry before failing the US — the F-10 retry
+  # parity the per-US main verifier site has but this final-verify path lacked
+  # (a single transient poll miss falsely failed a US at the most expensive
+  # end-of-campaign moment, charging a bogus consecutive failure).
+  rm -f "$VERDICT_FILE"
+  local poll_rc=0
+  poll_for_signal "$VERDICT_FILE" "$VERIFIER_HEARTBEAT" "$VERIFIER_PANE" "$verifier_launch" "Verifier-final" || poll_rc=$?
+  if (( poll_rc == 2 )); then
+    log_error "Verifier hard-fail (rc=2, sentinel written) for $us in final verify"
+    return 2
+  fi
+  if (( poll_rc == 1 )); then
+    log "  Verifier-final transient poll fail for $us — replacing pane + retrying once (D-4)"
+    replace_worker_pane "$VERIFIER_PANE" "verifier"
+    VERIFIER_PANE=$(jq -r '.panes.verifier' "$SESSION_CONFIG")
+    if [[ "$FINAL_VERIFIER_ENGINE" = "codex" ]]; then
+      launch_verifier_codex "$VERIFIER_PANE" "$verifier_prompt" "$iter" "$verifier_launch"
+    else
+      launch_verifier_claude "$VERIFIER_PANE" "$verifier_prompt" "$iter" "$verifier_launch" || return 2
+    fi
+    rm -f "$VERDICT_FILE"; poll_rc=0
+    poll_for_signal "$VERDICT_FILE" "$VERIFIER_HEARTBEAT" "$VERIFIER_PANE" "$verifier_launch" "Verifier-final" || poll_rc=$?
+    if (( poll_rc != 0 )); then
+      log_error "Verifier poll failed for $us after replace+retry (rc=$poll_rc)"
+      return 2
+    fi
+  fi
+
+  # Bug #7 Fix-Q/R: reap verifier pane between per-US final verifications so
+  # the previous codex/claude TUI cannot continue running while the next per-
+  # US verifier dispatch reuses the same pane.
+  _kill_pane_process "$VERIFIER_PANE" "verifier-final"
+  _lock_sentinel "$VERDICT_FILE"
+  # PR-0b-narrow: stamp leader handshake ack on the verdict (audit-only).
+  _stamp_ack_field "$VERDICT_FILE"
+
+  # Read verdict
+  local verdict
+  verdict=$(jq -r '.verdict' "$VERDICT_FILE" 2>/dev/null)
+  [[ "$verdict" == "pass" ]] && return 0
+  return 1
+}
+
 run_sequential_final_verify() {
   local iter="$1"
   FAILED_US=""
@@ -2842,91 +3014,38 @@ run_sequential_final_verify() {
   for us in $(echo "$US_LIST" | tr ',' ' '); do
     log "  Final verify: checking $us..."
 
-    # Temporarily override signal file to scope verifier to this US
-    local orig_signal
-    orig_signal=$(cat "$SIGNAL_FILE" 2>/dev/null)
-    echo "{\"status\":\"verify\",\"us_id\":\"$us\",\"summary\":\"sequential final verify\"}" | atomic_write "$SIGNAL_FILE"
-
-    # Write scoped verifier trigger
-    write_verifier_trigger "$iter"
-    local verifier_prompt="$LOGS_DIR/iter-$(printf '%03d' $iter).verifier-prompt.md"
-
-    # Clean verifier pane
-    local verifier_cmd
-    verifier_cmd=$(tmux display-message -p -t "$VERIFIER_PANE" '#{pane_current_command}' 2>/dev/null)
-    if [[ "$verifier_cmd" == "node" || "$verifier_cmd" == "claude" || "$verifier_cmd" == "codex" ]]; then
-      tmux send-keys -t "$VERIFIER_PANE" C-c 2>/dev/null; sleep 0.5
-      tmux send-keys -t "$VERIFIER_PANE" "/exit" C-m 2>/dev/null; sleep 2
-    fi
-    wait_for_pane_ready "$VERIFIER_PANE" 10 2>/dev/null || true
-
-    # Launch verifier. D-1: the FINAL (ALL) verify uses FINAL_VERIFIER_* (the
-    # "final 엄격" knob — a configured stronger model, e.g. opus, for the final
-    # gate), NOT the lighter per-US VERIFIER_*. This is the configured-final-model
-    # distinction, distinct from the removed per-iteration verifier auto-upgrade.
-    local verifier_launch
-    if [[ "$FINAL_VERIFIER_ENGINE" = "codex" ]]; then
-      verifier_launch="${CODEX_BIN:-codex} -m $FINAL_VERIFIER_CODEX_MODEL -c model_reasoning_effort=\"$FINAL_VERIFIER_CODEX_REASONING\" -c mcp_servers='{}' --disable plugins --dangerously-bypass-approvals-and-sandbox"
-      launch_verifier_codex "$VERIFIER_PANE" "$verifier_prompt" "$iter" "$verifier_launch"
-    else
-      verifier_launch="$(build_claude_cmd tui "$FINAL_VERIFIER_MODEL" "" "" "$FINAL_VERIFIER_EFFORT")"
-      launch_verifier_claude "$VERIFIER_PANE" "$verifier_prompt" "$iter" "$verifier_launch" || {
-        log_error "Failed to launch final verifier for $us"
+    # D-18: a US that already passed per-US gets up to FINAL_VERIFY_MAX_ATTEMPTS
+    # final-verify attempts on a FAIL verdict (first pass wins). A verifier
+    # false-fail (non-determinism) on already-correct, per-US-passed work must
+    # REPRODUCE across all attempts before it charges a fix-loop failure — else a
+    # single flake defeats a complete, correct campaign (D-16 dogfood: codex
+    # false-failed pytest-36/36 work → fix-loop churn → stale BLOCK). A US that
+    # never passed per-US (or a genuine regression) fails on the first attempt.
+    local _fv_max=1
+    # FINAL_VERIFY_MAX_ATTEMPTS is validated to 1..10 at declaration, so no clamp here.
+    if echo ",$VERIFIED_US," | grep -q ",$us,"; then _fv_max=$FINAL_VERIFY_MAX_ATTEMPTS; fi
+    local _fv_attempt=0 _fv_rc=1
+    while (( _fv_attempt < _fv_max )); do
+      (( _fv_attempt++ ))
+      _final_verify_one_us "$us" "$iter"; _fv_rc=$?
+      if (( _fv_rc == 2 )); then   # infra-terminal (launch/poll hard fail) — no retry
         FAILED_US="$us"
+        log "  Sequential final verify FAILED at $us (infra)"
+        log_debug "[FLOW] iter=$iter phase=sequential_final_verify failed_us=$us reason=infra attempts=$_fv_attempt"
         return 1
-      }
-    fi
-
-    # Poll for verdict. D-4: distinguish rc==2 (hard-fail, sentinel already
-    # written → terminal) from rc==1 (transient pane race/timeout) and give ONE
-    # replace-pane + re-dispatch retry before failing the US — the F-10 retry
-    # parity the per-US main verifier site has but this final-verify path lacked
-    # (a single transient poll miss falsely failed a US at the most expensive
-    # end-of-campaign moment, charging a bogus consecutive failure).
-    rm -f "$VERDICT_FILE"
-    local poll_rc=0
-    poll_for_signal "$VERDICT_FILE" "$VERIFIER_HEARTBEAT" "$VERIFIER_PANE" "$verifier_launch" "Verifier-final" || poll_rc=$?
-    if (( poll_rc == 2 )); then
-      log_error "Verifier hard-fail (rc=2, sentinel written) for $us in final verify"
-      FAILED_US="$us"
-      return 1
-    fi
-    if (( poll_rc == 1 )); then
-      log "  Verifier-final transient poll fail for $us — replacing pane + retrying once (D-4)"
-      replace_worker_pane "$VERIFIER_PANE" "verifier"
-      VERIFIER_PANE=$(jq -r '.panes.verifier' "$SESSION_CONFIG")
-      if [[ "$FINAL_VERIFIER_ENGINE" = "codex" ]]; then
-        launch_verifier_codex "$VERIFIER_PANE" "$verifier_prompt" "$iter" "$verifier_launch"
-      else
-        launch_verifier_claude "$VERIFIER_PANE" "$verifier_prompt" "$iter" "$verifier_launch" || { FAILED_US="$us"; return 1; }
       fi
-      rm -f "$VERDICT_FILE"; poll_rc=0
-      poll_for_signal "$VERDICT_FILE" "$VERIFIER_HEARTBEAT" "$VERIFIER_PANE" "$verifier_launch" "Verifier-final" || poll_rc=$?
-      if (( poll_rc != 0 )); then
-        log_error "Verifier poll failed for $us after replace+retry (rc=$poll_rc)"
-        FAILED_US="$us"
-        return 1
-      fi
-    fi
-
-    # Bug #7 Fix-Q/R: reap verifier pane between per-US final verifications so
-    # the previous codex/claude TUI cannot continue running while the next per-
-    # US verifier dispatch reuses the same pane.
-    _kill_pane_process "$VERIFIER_PANE" "verifier-final"
-    _lock_sentinel "$VERDICT_FILE"
-    # PR-0b-narrow: stamp leader handshake ack on the verdict (audit-only).
-    _stamp_ack_field "$VERDICT_FILE"
-
-    # Check verdict
-    local verdict
-    verdict=$(jq -r '.verdict' "$VERDICT_FILE" 2>/dev/null)
-    if [[ "$verdict" != "pass" ]]; then
+      (( _fv_rc == 0 )) && break   # pass verdict
+      log "  Sequential final verify: $us verdict=fail (attempt $_fv_attempt/$_fv_max)"
+      log_debug "[FLOW] iter=$iter phase=sequential_final_verify us=$us attempt=$_fv_attempt/$_fv_max verdict=fail"
+      (( _fv_attempt < _fv_max )) && log "  D-18: re-verifying $us — a per-US-passed US's fail must reproduce to count (verifier flake guard)."
+    done
+    if (( _fv_rc != 0 )); then
       FAILED_US="$us"
-      log "  Sequential final verify FAILED at $us"
-      log_debug "[FLOW] iter=$iter phase=sequential_final_verify failed_us=$us verdict=$verdict"
+      log "  Sequential final verify FAILED at $us (failed all $_fv_attempt/$_fv_max attempt(s))"
+      log_debug "[FLOW] iter=$iter phase=sequential_final_verify failed_us=$us verdict=fail attempts=$_fv_attempt max=$_fv_max"
       return 1
     fi
-    log "  Sequential final verify: $us PASSED"
+    log "  Sequential final verify: $us PASSED$([[ $_fv_attempt -gt 1 ]] && echo " (after $_fv_attempt attempts — earlier verdict was a verifier flake)")"
 
     # Archive per-US final verdict
     cp "$VERDICT_FILE" "$LOGS_DIR/iter-$(printf '%03d' $iter).final-verdict-${us}.json" 2>/dev/null
@@ -3760,7 +3879,15 @@ main() {
         update_status "verifier" "running"
 
         # --- Sequential final verify: per-US scoped checks instead of one big ALL verify ---
-        if [[ "$signal_us_id" == "ALL" && "$VERIFY_MODE" == "per-us" && -n "$US_LIST" ]]; then
+        # D-21: do NOT take the sequential single-verifier path when consensus applies to
+        # the final verify — it would BYPASS consensus entirely (run_sequential_final_verify
+        # returns 0 before the consensus check below), so `--consensus final-only` was a
+        # silent no-op in the DEFAULT per-us mode (the recommended consensus config). When
+        # consensus is on, fall through to run_consensus_verification "ALL" (which uses the
+        # stricter FINAL_VERIFIER_MODEL + FINAL_CONSENSUS_MODEL and the designed
+        # claude+codex final path). The per-US timeout-prevention split is the off-consensus
+        # optimization only.
+        if [[ "$signal_us_id" == "ALL" && "$VERIFY_MODE" == "per-us" && -n "$US_LIST" ]] && ! _should_use_consensus "$signal_us_id"; then
           log "  Final ALL verify: using sequential per-US strategy (timeout prevention)"
           local seq_rc=0
           run_sequential_final_verify "$ITERATION" || seq_rc=$?