Summary: Editing a custom Action’s environment variable (e.g. via the Action Editor’s Save/“Update and Close”) persists the
new value in storage, but the action’s live execution runtime continues using the old value until something forces a fresh
deploy. There’s no UI indication that this is happening — a saved change looks successful and simply doesn’t apply.
Actions affected: propose_personas and confirm_personas (custom Python actions, both using JOBSGLIDER_BASE_URL and
PICKAXE_API_SECRET env vars).
Reproduction, with exact evidence:
- JOBSGLIDER_BASE_URL was originally misconfigured (pointed at the wrong host). Corrected via the Action Editor.
- Confirmed via my own server’s request logs that calls were still arriving at the old wrong host — every call returned an
identical, byte-for-byte 404 response (same HTML page), timestamped well after the save. - Re-saved the value three separate times across ~1 hour, using both “Save Action” and “Update and Close” — identical stale
result every time, confirmed via server logs each time (not just UI appearance). - Asked Wingman to introspect the actual stored value directly (bypassing the UI). It confirmed the value was correctly
updated — cited the masked-value suffix changing as proof of a real write. - Despite the confirmed-correct stored value, the action still executed against the old host — another identical stale 404 at
2026-09-25T16:32:25Z. - Wingman’s diagnosis: most serverless Python runtimes read os.environ once at container warm-start, not per-invocation — a
warm instance can hold a stale value indefinitely regardless of how many times the stored config is updated. - Wingman performed a no-op manifest PUT (same code, no logic changes) to force a fresh deploy/cold start for both actions,
redeployed at 2026-09-25T16:46:32Z / 16:46:33Z. - Immediately after, the action executed correctly against the new value — real success at 16:46:40Z, four minutes after the
forced redeploy, roughly an hour after the underlying config was first corrected.
Impact: Any environment variable change to a custom Action is unreliable without a manual, undocumented workaround (a no-op
manifest redeploy). A builder has no way to know their save didn’t actually apply — the UI gives no warning, and the only
symptom is the action continuing to fail or misbehave exactly as before the “fix.”
Requested fix: Either (a) read environment variables per-invocation rather than caching them at container start, or (b)
automatically trigger a redeploy/cache invalidation whenever a variable is saved, so a save is never silently a no-op. At
minimum, (c) surface in the UI when a variable change requires a manual redeploy to take effect.

