Check if this issue is valid — if so, understand the root cause and fix it. At skills/continuous-learning-v2/agents/session-guardian.sh, line 35:

<comment>`grep -F "$project_root"` does a substring match, so a project like `/home/dev/app` will also match `/home/dev/app-backend`. This causes both an incorrect cooldown read here and data loss on line 47 (`grep -vF` deletes both entries). Append a literal tab to the pattern so it matches only the exact key in the tab-delimited log.</comment>

<file context>
@@ -0,0 +1,56 @@
+  trap 'rm -rf "$_lock_dir"' EXIT INT TERM
+
+  last_spawn=0
+  last_spawn=$(grep -F "$project_root" "$LOG_PATH" 2>/dev/null | tail -n1 | awk '{print $NF}') || true
+  last_spawn="${last_spawn:-0}"
+
</file context>

Verify each finding against the current code and only fix it if needed.

In `@skills/continuous-learning-v2/agents/observer-loop.sh` around lines 36 - 40,
The session-guardian cooldown is recorded too early in observer-loop.sh (it runs
before the claude availability check), so move the call that records the
cooldown (the bash "$(dirname "$0")/session-guardian.sh" invocation) to occur
immediately after the claude launch attempt (the line that runs claude --model
haiku ...) or split session-guardian.sh into two functions (e.g.,
session_guardian_check and session_guardian_record) and call the "check" before
the claude availability test and the "record" only when you actually spawn a
Haiku session; ensure you update references to LOG_FILE and preserve the
existing skip log message ("Observer cycle skipped by session-guardian") so the
cooldown is only consumed when the spawn succeeds.

Verify each finding against the current code and only fix it if needed.

In `@skills/continuous-learning-v2/agents/session-guardian.sh` around lines 20 -
21, The cooldown key currently derives project_root/project_name by running git
and basename in session-guardian.sh (variables project_root and project_name);
change it to use the provided environment variable PROJECT_DIR (set by
start-observer.sh) as the canonical project directory/key. Replace uses of
project_root/project_name when building the cooldown identifier with PROJECT_DIR
(or basename "$PROJECT_DIR" if you only need the name) so cooldowns are tied to
the observer-provided PROJECT_DIR rather than the current working dir or git
root.

Verify each finding against the current code and only fix it if needed.

In `@skills/continuous-learning-v2/agents/session-guardian.sh` around lines 28 -
31, The lock contention path currently logs and lets the loser continue; change
it so a failed mkdir on _lock_dir causes the script to skip this cycle instead
of proceeding. In the session-guardian.sh block that checks "if ! mkdir
\"$_lock_dir\" ...", replace the current "proceed" behavior with an early
exit/return that stops further work (e.g., log the contention and exit 0 or
return from the function). Keep the successful-creation cleanup/trap logic for
removing _lock_dir intact so the winner still releases the lock.

Verify each finding against the current code and only fix it if needed.

In `@skills/continuous-learning-v2/agents/session-guardian.sh` around lines 35 -
50, The current use of grep -F with project_root matches prefixes and can steal
or remove other projects' entries; replace the grep invocations that compute
last_spawn and that filter out old entries (the lines using grep -F
"$project_root" and grep -vF "$project_root" writing to tmp_log) with exact-key
matching, e.g. use awk (or grep -Fx) to select/remove only lines whose first
field exactly equals project_root; ensure the last_spawn extraction uses the
same exact-key logic (the variable last_spawn calculation) and the tmp_log
rewrite (the removal of the old entry before appending the new
"${project_root}\t${now}") also uses exact-key matching so only the intended
project's timestamp is read/removed.