diff --git a/scripts/engine.sh b/scripts/engine.sh
index 77b302f4b..56fd97db6 100644
--- a/scripts/engine.sh
+++ b/scripts/engine.sh
@@ -496,7 +496,12 @@ run_triage() {
   # Capture output for token estimation when logging is enabled (tee is a no-op otherwise).
   local _tok_tmp=""
   if [ -n "${TOKEN_LOG_FILE:-}" ]; then
+    unset _ENGINE_USAGE_OUT
     _tok_tmp="$(mktemp 2>/dev/null || true)"
+    # Per-call usage-sidecar key: a unique mktemp path, exported so the engine's
+    # pipeline subshell and _record_engine_tokens agree on it. Unique per call →
+    # concurrent run_agentic/run_duck (review-one-pr.sh) never collide.
+    [[ -n "$_tok_tmp" ]] && local -x _ENGINE_USAGE_OUT="${_tok_tmp}.usage"
   fi
   while [ "$attempt" -le "$RETRY_MAX_ATTEMPTS" ]; do
     rc=0
@@ -567,7 +572,12 @@ run_agentic() {
   local tier="${3:-deep}"
   local _tok_tmp="" rc=0
   if [ -n "${TOKEN_LOG_FILE:-}" ]; then
+    unset _ENGINE_USAGE_OUT
     _tok_tmp="$(mktemp 2>/dev/null || true)"
+    # Per-call usage-sidecar key: a unique mktemp path, exported so the engine's
+    # pipeline subshell and _record_engine_tokens agree on it. Unique per call →
+    # concurrent run_agentic/run_duck (review-one-pr.sh) never collide.
+    [[ -n "$_tok_tmp" ]] && local -x _ENGINE_USAGE_OUT="${_tok_tmp}.usage"
   fi
   case "$REVIEW_ENGINE" in
     claude)
@@ -688,7 +698,12 @@ run_duck() {
   local model="$2"
   local _tok_tmp="" rc=0
   if [ -n "${TOKEN_LOG_FILE:-}" ]; then
+    unset _ENGINE_USAGE_OUT
     _tok_tmp="$(mktemp 2>/dev/null || true)"
+    # Per-call usage-sidecar key: a unique mktemp path, exported so the engine's
+    # pipeline subshell and _record_engine_tokens agree on it. Unique per call →
+    # concurrent run_agentic/run_duck (review-one-pr.sh) never collide.
+    [[ -n "$_tok_tmp" ]] && local -x _ENGINE_USAGE_OUT="${_tok_tmp}.usage"
   fi
   case "$DUCK_ENGINE" in
     claude)
@@ -842,7 +857,11 @@ run_writer() {
   # to stdout, not stderr), so is_rate_limited never fired and fallback engines
   # were never tried.
   local _tmp rc=0
+  unset _ENGINE_USAGE_OUT
   _tmp="$(mktemp 2>/dev/null || true)"
+  # Per-call usage-sidecar key (see run_triage); unique mktemp path shared with
+  # the engine subshell via export.
+  [[ -n "$_tmp" ]] && local -x _ENGINE_USAGE_OUT="${_tmp}.usage"
 
   case "$REVIEW_ENGINE" in
     claude)
diff --git a/scripts/lib/token-metrics.sh b/scripts/lib/token-metrics.sh
index ba22b9f1e..e29facda6 100644
--- a/scripts/lib/token-metrics.sh
+++ b/scripts/lib/token-metrics.sh
@@ -131,18 +131,26 @@ emit_token_record() {
 # ─────────────────────────────────────────────────────────────────────────────
 
 # _engine_usage_sidecar
-# Prints the path of the per-process usage sidecar (empty when TOKEN_LOG_FILE is
-# unset). Engine invocations often run inside a `cmd | tee` pipeline — i.e. a
-# SUBSHELL — so the LAST_* globals they set never reach the parent. Writing usage
-# to this file (the filesystem is shared) lets _record_engine_tokens read it back.
-# $$ (shell PID) is constant across pipeline subshells: both the writing left-hand
-# side and the reading parent see the same value. Separate background processes
-# (e.g. `run_agentic &` forking a new process vs `(run_duck) &` running as a
-# subshell) each have a distinct $$, so concurrent tier-2 calls never share a
-# sidecar file.
+# Prints the path of the per-CALL usage sidecar (empty when TOKEN_LOG_FILE is
+# unset). Engine invocations run inside a `cmd | tee` pipeline — i.e. a SUBSHELL —
+# so the LAST_* globals they set never reach the parent; writing usage to this file
+# (the filesystem is shared) lets _record_engine_tokens read it back.
+#
+# The key must be unique per concurrent engine call AND identical between the
+# writer (the pipeline subshell) and the reader (the calling shell). $$ fails the
+# first requirement: review-one-pr.sh backgrounds `run_agentic &` and `(run_duck) &`
+# at the same time, and $$ stays the parent PID inside both, so they would collide
+# on one sidecar. BASHPID fails the second: it differs between the pipe subshell
+# and the caller. So each run_* exports _ENGINE_USAGE_OUT, derived from its own
+# per-call mktemp file (unique by construction) and inherited by the pipeline
+# subshell. The $$ form is only a fallback for any non-concurrent direct caller.
 _engine_usage_sidecar() {
   [ -n "${TOKEN_LOG_FILE:-}" ] || return 0
-  printf '%s' "${TOKEN_LOG_FILE}.last-usage.$$"
+  if [[ -n "${_ENGINE_USAGE_OUT:-}" ]]; then
+    printf '%s' "$_ENGINE_USAGE_OUT"
+  else
+    printf '%s' "${TOKEN_LOG_FILE}.last-usage.$$"
+  fi
 }
 
 # reset_engine_usage
diff --git a/tests/dev-lead/unit/test_engine_writer.bats b/tests/dev-lead/unit/test_engine_writer.bats
index 76a5e33d5..e3e281bda 100644
--- a/tests/dev-lead/unit/test_engine_writer.bats
+++ b/tests/dev-lead/unit/test_engine_writer.bats
@@ -486,6 +486,37 @@ STUB
   [ "$(jq -r '.output_tokens'     < "$log")" = "60" ]
 }
 
+@test "usage: failed mktemp + stale _ENGINE_USAGE_OUT does NOT reuse the stale sidecar" {
+  # Regression for the per-call-key fix: run_* clears _ENGINE_USAGE_OUT before
+  # mktemp, so a failed mktemp falls back to estimation instead of reading a prior
+  # call's sidecar. Without it, the planted 999/9/9/9 below would be logged.
+  _source_engine "claude"
+  export STUB_ENGINE_EXIT=0
+  export STUB_ENGINE_RESPONSE="verdict"
+  local log; log=$(mktemp)
+  export TOKEN_LOG_FILE="$log"
+  export TEST_OWNED_TOKEN_LOG="$log"
+
+  # A leftover per-call key + sidecar content from a "previous" call.
+  local stale; stale="$(mktemp)"
+  printf '999\t9\t9\t9\n' > "$stale"
+  export _ENGINE_USAGE_OUT="$stale"
+
+  # Force mktemp to fail inside the engine via a PATH shim.
+  local shim; shim="$(mktemp -d)"
+  printf '#!/usr/bin/env bash\nexit 1\n' > "$shim/mktemp"
+  chmod +x "$shim/mktemp"
+
+  PATH="$shim:$PATH" run run_triage "$TEST_PROMPT"
+
+  rm -rf "$shim"; rm -f "$stale"; unset _ENGINE_USAGE_OUT
+  [ "$status" -eq 0 ]
+  # The stale cache figures must NOT appear in the record.
+  [ "$(jq -r '.cache_read_tokens'     < "$log")" != "9" ]
+  [ "$(jq -r '.cache_creation_tokens' < "$log")" != "9" ]
+  [ "$(jq -r '.input_tokens'          < "$log")" != "999" ]
+}
+
 @test "token: run_writer_with_fallback logs the fallback engine (not rate-limited primary)" {
   _source_engine "claude"
   export DEV_LEAD_DRY_RUN=false
diff --git a/tests/dev-lead/unit/test_token_metrics.bats b/tests/dev-lead/unit/test_token_metrics.bats
index 485f2ff07..86123033d 100644
--- a/tests/dev-lead/unit/test_token_metrics.bats
+++ b/tests/dev-lead/unit/test_token_metrics.bats
@@ -246,10 +246,11 @@ teardown() {
   [ "$LAST_OUTPUT_TOKENS" = "85" ]   # 60 candidates + 25 thoughts
 }
 
-@test "_engine_usage_sidecar: path includes shell PID to isolate parallel invocations" {
-  # $$ is constant across pipeline subshells (left-hand side inherits parent PID)
-  # so both the writing subshell and the reading parent agree on the same path,
-  # while separate background processes each get their own distinct $$.
+@test "_engine_usage_sidecar: fallback path is \$\$-keyed when no per-call key set" {
+  # Without a per-call _ENGINE_USAGE_OUT the sidecar falls back to a $$-keyed path.
+  # NOTE: $$ does NOT isolate concurrent background jobs (they share the parent PID)
+  # — per-call isolation comes from _ENGINE_USAGE_OUT (see the concurrency test).
+  unset _ENGINE_USAGE_OUT
   local sidecar
   sidecar="$(_engine_usage_sidecar)"
   [ -n "$sidecar" ]
@@ -320,3 +321,53 @@ line two'
   run emit_token_record "pr-review" "triage" "claude" "haiku" 100 0 50 ""
   [ "$status" -eq 0 ]
 }
+
+# ── usage sidecar keying (concurrency safety) ─────────────────────────────────
+
+@test "_engine_usage_sidecar: uses the exported per-call key when set" {
+  export _ENGINE_USAGE_OUT="$TOKEN_LOG_FILE.callX.usage"
+  run _engine_usage_sidecar
+  unset _ENGINE_USAGE_OUT
+  [ "$output" = "$TOKEN_LOG_FILE.callX.usage" ]
+}
+
+@test "_engine_usage_sidecar: set-but-empty key falls back to \$\$ (mktemp-failure state)" {
+  # run_* clears _ENGINE_USAGE_OUT before mktemp; if mktemp fails the key is empty/
+  # unset and MUST fall back rather than reuse a prior/inherited key.
+  export _ENGINE_USAGE_OUT=""
+  run _engine_usage_sidecar
+  unset _ENGINE_USAGE_OUT
+  [ "$output" = "$TOKEN_LOG_FILE.last-usage.$$" ]
+}
+
+@test "_engine_usage_sidecar: empty when TOKEN_LOG_FILE is unset" {
+  unset TOKEN_LOG_FILE _ENGINE_USAGE_OUT
+  run _engine_usage_sidecar
+  [ -z "$output" ]
+}
+
+@test "usage sidecar: concurrent calls sharing \$\$ stay isolated (no cross-read)" {
+  # Reproduces the review-one-pr.sh pattern: two engine calls backgrounded at once.
+  # Background subshells share the parent's $$, so a $$-keyed sidecar would collide;
+  # the per-call _ENGINE_USAGE_OUT (unique mktemp path) must keep them isolated.
+  local dir; dir="$(mktemp -d)"
+  printf '%s' '{"result":"a","usage":{"input_tokens":111,"cache_read_input_tokens":1,"cache_creation_input_tokens":2,"output_tokens":11}}' > "$dir/a.json"
+  printf '%s' '{"result":"b","usage":{"input_tokens":222,"cache_read_input_tokens":3,"cache_creation_input_tokens":4,"output_tokens":22}}' > "$dir/b.json"
+  (
+    export _ENGINE_USAGE_OUT="$dir/A.usage"
+    parse_engine_usage claude "$dir/a.json"
+    sleep 0.3   # widen the race window for any colliding writer
+    cut -f1 < "$(_engine_usage_sidecar)" > "$dir/a.out"
+  ) &
+  (
+    export _ENGINE_USAGE_OUT="$dir/B.usage"
+    parse_engine_usage claude "$dir/b.json"
+    sleep 0.3
+    cut -f1 < "$(_engine_usage_sidecar)" > "$dir/b.out"
+  ) &
+  wait
+  local a b; a="$(cat "$dir/a.out")"; b="$(cat "$dir/b.out")"
+  rm -rf "$dir"
+  [ "$a" = "111" ]
+  [ "$b" = "222" ]
+}
diff --git a/tests/token_report.bats b/tests/token_report.bats
index 117447fb1..0421c3897 100644
--- a/tests/token_report.bats
+++ b/tests/token_report.bats
@@ -162,9 +162,11 @@ setup() {
   run bash -c "
     set -euo pipefail
     source '${BATS_TEST_DIRNAME}/../scripts/token_report.sh'
-    # Simulate the inner conversion loop for a binary-corrupted source file.
+    # Simulate the inner conversion loop for a corrupted source file. Use truncated
+    # JSON, which jq rejects deterministically across versions (jq 1.7 exits 0 on
+    # NUL bytes, so binary input made this test flaky).
     f='$tmpdir/bad.jsonl'
-    printf '\x00\x01\x02\x03' > \"\$f\"
+    printf '{\"a\":' > \"\$f\"
     dest='$jsonl_dir/999-bad.jsonl'
     if jq -c --arg repo 'r' 'select(type==\"object\") | . + {repo: \$repo}' \
         \"\$f\" > \"\$dest\" 2>/dev/null; then