diff --git a/README.md b/README.md
index 716ee2c..509dcb1 100644
--- a/README.md
+++ b/README.md
@@ -145,6 +145,8 @@ Considering Claude as your main and Codex as the copilot:
 Fable writes the plan, 5.6 Sol reviews it, Luna implements, back to Fable who reviews and fixes the diff, runs the testing gate, then a new Sol thread reviews again the code. All in one claude code session. Writer and reviewer are never the same thread.  
 As of mid july 2026, this Fable + GPT5.6 harness combo is absolute peak.
 
+One honest caveat: the implementer's `--notes` are injected into the reviewer's prompt, so a lazy (or scheming) writer agent can talk the reviewer out of findings. The skills mitigate this (reviews surfaced verbatim, capped rounds, push-back must be justified), but if a review converges suspiciously fast, read the notes.
+
 ## MCP Servers: Less Is More
 
 Last piece of advise before your new coding quest: Every MCP server you add is extra context, extra latency, and extra confusion. Keep it minimal. The one use case where MCP genuinely shines is **up-to-date documentation**, so your agent stops hallucinating deprecated APIs/whatever. Two servers cover it: [Context7](https://github.com/upstash/context7) for current library & framework docs, and [Exa](https://github.com/exa-labs/exa-mcp-server) for web search when the answer isn't in any doc. No bloat beyond that.
diff --git a/skills/codex-implement/scripts/start.sh b/skills/codex-implement/scripts/start.sh
index 4e6f1af..bec63e8 100755
--- a/skills/codex-implement/scripts/start.sh
+++ b/skills/codex-implement/scripts/start.sh
@@ -21,6 +21,8 @@ export STATE_DIR
 # shellcheck source=../../codex-plan-review/scripts/_common.sh
 source "$SCRIPT_DIR/../../codex-plan-review/scripts/_common.sh"
 
+require_tools codex jq
+
 PROMPT_FILE=""
 while [ $# -gt 0 ]; do
     case "$1" in
diff --git a/skills/codex-implement/state/.gitignore b/skills/codex-implement/state/.gitignore
new file mode 100644
index 0000000..3601fc4
--- /dev/null
+++ b/skills/codex-implement/state/.gitignore
@@ -0,0 +1,5 @@
+# Per-target Codex implementation state (thread ids, last report text,
+# raw event logs). Local-only — these are session artifacts, not
+# source-of-truth artifacts.
+*
+!.gitignore
diff --git a/skills/codex-plan-review/scripts/_common.sh b/skills/codex-plan-review/scripts/_common.sh
index d59807e..e973b8c 100755
--- a/skills/codex-plan-review/scripts/_common.sh
+++ b/skills/codex-plan-review/scripts/_common.sh
@@ -16,30 +16,48 @@ mkdir -p "$STATE_DIR"
 # implementation runs Luna, reviews (plan + code) run Sol, effort xhigh.
 # Adjust these defaults to your preferred models.
 # CODEX_MODEL / CODEX_EFFORT act as per-run overrides.
-case "$STATE_DIR" in
-    *codex-implement*) CODEX_MODEL="${CODEX_MODEL:-gpt-5.6-luna}" ;;
-    *)                 CODEX_MODEL="${CODEX_MODEL:-gpt-5.6-sol}" ;;
+# Match the trailing path components exactly — a repo path that merely
+# contains "codex-implement" must not flip reviews to the implement model.
+case "${STATE_DIR%/}" in
+    */codex-implement/state | codex-implement/state)
+        CODEX_MODEL="${CODEX_MODEL:-gpt-5.6-luna}" ;;
+    *)  CODEX_MODEL="${CODEX_MODEL:-gpt-5.6-sol}" ;;
 esac
 CODEX_EFFORT="${CODEX_EFFORT:-xhigh}"
 export CODEX_MODEL CODEX_EFFORT
 
+# Fail fast when a required external tool is missing. Without this,
+# `set -e` + suppressed stderr would kill the caller with no message.
+require_tools() {
+    local tool
+    for tool in "$@"; do
+        if ! command -v "$tool" >/dev/null 2>&1; then
+            echo "error: required tool not found on PATH: $tool" >&2
+            return 1
+        fi
+    done
+}
+
 # Derive a per-target key from a path-like string. For real paths we
 # resolve to absolute; for non-path targets (branch names, commit
-# ranges) we sanitize in place. Replace '/' with '__'; force any other
-# non-portable characters to '_'.
+# ranges) we use the string as-is. The key is a sanitized, readable
+# form plus a checksum of the resolved target, so distinct targets
+# that sanitize identically (e.g. "foo/bar" vs "foo__bar") never
+# share state files.
 target_key() {
-    local target="$1"
+    local target="$1" resolved sanitized sum
     if [ -e "$target" ]; then
-        local abs
-        abs="$(realpath -- "$target" 2>/dev/null || readlink -f -- "$target")"
-        if [ -z "$abs" ]; then
+        resolved="$(realpath -- "$target" 2>/dev/null || readlink -f -- "$target")"
+        if [ -z "$resolved" ]; then
             echo "error: cannot resolve target path: $target" >&2
             return 1
         fi
-        printf '%s' "$abs" | sed 's|^/||; s|/|__|g'
     else
-        printf '%s' "$target" | sed 's|^/||; s|/|__|g; s|[^A-Za-z0-9._-]|_|g'
+        resolved="$target"
     fi
+    sanitized="$(printf '%s' "$resolved" | sed 's|^/||; s|/|__|g; s|[^A-Za-z0-9._-]|_|g')"
+    sum="$(printf '%s' "$resolved" | cksum | cut -d' ' -f1)"
+    printf '%s.%s' "$sanitized" "$sum"
 }
 
 # Backwards-compatible alias used by older script call sites.
@@ -57,23 +75,21 @@ events_file() {
     printf '%s/%s.events.ndjson' "$STATE_DIR" "$(target_key "$1")"
 }
 
-# Load a prompt template from $1 and substitute {{TARGET}} and
-# {{EXTRA_PROMPT}} placeholders with the values of the $TARGET and
-# $EXTRA_PROMPT environment variables. Other text passes through
-# verbatim — no surprise expansion of unrelated $VAR sequences.
-# Writes the substituted prompt to stdout.
+# Load a prompt template from $1 and substitute {{TARGET}},
+# {{EXTRA_PROMPT}} and {{IMPLEMENTER_NOTES}} placeholders with the
+# values of the corresponding environment variables. The replacement
+# expansions are double-quoted, which forces bash to treat them as
+# literal strings — without the quotes, bash >= 5.2 (patsub_replacement)
+# expands '&' to the matched text, exactly the mangling awk's gsub did.
 load_prompt() {
-    local tpl="$1"
+    local tpl="$1" content
     if [ ! -f "$tpl" ]; then
         echo "error: prompt template not found: $tpl" >&2
         return 1
     fi
-    awk -v target="${TARGET-}" -v extra="${EXTRA_PROMPT-}" -v notes="${IMPLEMENTER_NOTES-}" '
-        {
-            gsub(/\{\{TARGET\}\}/, target)
-            gsub(/\{\{EXTRA_PROMPT\}\}/, extra)
-            gsub(/\{\{IMPLEMENTER_NOTES\}\}/, notes)
-            print
-        }
-    ' "$tpl"
+    content="$(cat "$tpl")"
+    content="${content//'{{TARGET}}'/"${TARGET-}"}"
+    content="${content//'{{EXTRA_PROMPT}}'/"${EXTRA_PROMPT-}"}"
+    content="${content//'{{IMPLEMENTER_NOTES}}'/"${IMPLEMENTER_NOTES-}"}"
+    printf '%s\n' "$content"
 }
diff --git a/skills/codex-plan-review/scripts/resume.sh b/skills/codex-plan-review/scripts/resume.sh
index 0c69212..e7ec06c 100755
--- a/skills/codex-plan-review/scripts/resume.sh
+++ b/skills/codex-plan-review/scripts/resume.sh
@@ -11,6 +11,8 @@ SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 # shellcheck source=_common.sh
 source "$SCRIPT_DIR/_common.sh"
 
+require_tools codex
+
 PROMPT_FILE=""
 IMPLEMENTER_NOTES=""
 while [ $# -gt 0 ]; do
@@ -49,6 +51,11 @@ if [ ! -f "$THREAD_FILE" ]; then
     exit 2
 fi
 THREAD_ID="$(cat "$THREAD_FILE")"
+if [ -z "$THREAD_ID" ]; then
+    echo "error: thread file for $TARGET is empty: $THREAD_FILE" >&2
+    echo "       run reset.sh, then start.sh to begin a fresh session." >&2
+    exit 2
+fi
 
 PROMPT="$(load_prompt "$PROMPT_FILE")"
 
diff --git a/skills/codex-plan-review/scripts/start.sh b/skills/codex-plan-review/scripts/start.sh
index 6b245f4..ede9c19 100755
--- a/skills/codex-plan-review/scripts/start.sh
+++ b/skills/codex-plan-review/scripts/start.sh
@@ -12,6 +12,8 @@ SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 # shellcheck source=_common.sh
 source "$SCRIPT_DIR/_common.sh"
 
+require_tools codex jq
+
 PROMPT_FILE=""
 while [ $# -gt 0 ]; do
     case "$1" in