diff --git a/templates/agent/HEARTBEAT.md b/templates/agent/HEARTBEAT.md
index 2b447d006b..0fdaf70580 100644
--- a/templates/agent/HEARTBEAT.md
+++ b/templates/agent/HEARTBEAT.md
@@ -175,11 +175,27 @@ timeout and it still failed" reports tonight never actually tested a raised time
 **RUN EACH FILE AS A BACKGROUND / UNCAPPED CALL. NEVER GATE COMPLETION ON A TIMEOUT, EXIT CODE, OR ELAPSED TIME.**
 ```bash
 cortextos bus kb-ingest "$f" --org "$CTX_ORG" --agent "$CTX_AGENT_NAME" --scope private --force \
-  > "/tmp/kb-ingest-${CTX_AGENT_NAME}-$$-$(basename "$f").log" 2>&1 &
+  > "/tmp/kb-ingest-${CTX_AGENT_NAME}-$(date -u +%Y%m%d-%H%M%S)-$$-$(basename "$f").log" 2>&1 &
 ```
 `$$` alone collides across concurrent agent sessions on a shared host — `$CTX_AGENT_NAME` in the
 filename is required, not cosmetic (infra, 2026-08-25: a single heartbeat's log came back
-interleaved with 7+ other agents' ingest output under the un-namespaced path).
+interleaved with 7+ other agents' ingest output under the un-namespaced path). `$CTX_AGENT_NAME-$$`
+alone is still not sufficient: PIDs get reused across days on a long-running box, and time-of-day-only
+labels (`hb2342`, `hb0342`) collide identically — infra, 2026-09-05, caught its own prior-day 429 log
+being misread as current via an hb-time-only filename. The UTC date+time component kills both the
+cross-agent and cross-day/cross-hour collisions in one form.
+
+⚠️ **`${BASHPID}` was tried and reverted the same day (CodeRabbit's original suggestion on PR #178):
+it is a BASH-ONLY special variable and is UNDEFINED in zsh, which is the actual shell every agent's
+Bash tool (and this box's `$SHELL`) executes commands in — verified directly (`ZSH_VERSION` set,
+`BASH_VERSION` empty, `${BASHPID}` expands to nothing). Shipping it produced literal double-hyphen,
+empty-differentiator filenames in the real execution environment despite passing verification via an
+explicit `bash script.sh` subprocess — a different, non-representative shell context. `$$` remains
+the working differentiator here; it does not distinguish two `&`-launched siblings from the exact
+same loop iteration, but the Step 10 loop always launches DISTINCT files per iteration (basename
+already differentiates), so that theoretical gap does not occur in the documented usage.** PID
+remains only as the final tiebreaker
+for two calls landing in the same second.
 
 The only valid completion signal is the literal text `Ingest complete` appearing in that log — not `rc=0`,
 not a chunk count alone, not silence. If it hasn't appeared by the time you move on to other work, check