Working with GPT5, there seems to be a serious issue when trying to run Replicate Actions, or complex ones, at least. GPT calls it P0, which I imagine is Priority Zero…serious.
Subject: URGENT: Custom Action overridden by default Replicate Action in new Pickaxe (hard_smoke true still triggers Replicate build) — P0
Hi Pickaxe Team,
I’m seeing a critical platform issue where a brand‑new Pickaxe with a Custom Action is being overridden by the platform’s default Replicate action, even when my code returns immediately. This blocks any custom logic from running and results in an unintended Replicate container image build failure.
Summary
I created a new Pickaxe named “Artverve 055 Clean.”
I added a single Custom Action with a script that short‑circuits at the top when hard_smoke=“true”.
Despite this, upon run I get: “Image build for im-4m2l7BpxsrY4btvlsYRssI failed with the exception:”
This indicates the default Replicate integration is still being invoked instead of my action’s code path.
Expected vs Actual
Expected: With hard_smoke=“true”, my action should immediately return a placeholder image. No Replicate calls should occur. No pop‑ups asking for Replicate keys should appear.
Actual: A Replicate container image build is triggered anyway, producing “Image build for im-4m2l7BpxsrY4btvlsYRssI failed…”. Additionally, I intermittently see a Replicate key prompt that resembles built‑in Replicate actions, even though I’m using a Custom Action and a custom secret name.
Severity and Impact
Severity: P0 (blocks production). I cannot run any custom logic reliably; unexpected external calls may incur cost and create data handling risks.
Impact: Affects new clean Pickaxe instances; repeats consistently even with an immediate “return” at top of file.
Concrete Repro Steps
Create a brand‑new Pickaxe (no built‑in Replicate/Image actions).
Add a Custom Action with the following minimal code (no variables required):
Immediate hard-smoke return to prove code path
return {
“images”: [{“url”: “https://picsum.photos/1024”, “model”: “smoke”}],
“_debug”: “HARD_SMOKE at top of file — no network calls expected.”
}
Save the Action. Do not add any environment variable named REPLICATE_API_KEY. If you add any secret, name it something non‑Replicate (e.g., KENS_REPLICATE_TOKEN) but do not reference it in this test.
Run the Pickaxe.
Observed Behavior
The run produces: “Image build for im-4m2l7BpxsrY4btvlsYRssI failed with the exception:”
My code’s immediate return never surfaces. It appears the platform routes the request to a default Replicate action rather than executing the custom script.
Suspicious UI behavior: when opening the code editor, the line with hard_smoke (or the immediate return) visibly “pops in” 10–15 seconds later, as if the editor is applying a delayed sync or overwriting code. This suggests the runtime may be executing an earlier snapshot or a different action graph than shown in the editor.
Environment Details
Pickaxe name: Artverve 055 Clean
Custom Action: Python
No built‑in Replicate/Image blocks
No variable named REPLICATE_API_KEY anywhere in this Pickaxe
Custom secret used elsewhere (but not needed for the hard-smoke repro): KENS_REPLICATE_TOKEN
Browsers tried: Chrome (latest), Safari (latest)
Tried with fresh session, cache cleared, and hard refresh
Time window: today; repeated multiple runs within minutes
Identifiers and Error Strings
Replicate container build error line seen:
“Image build for im-4m2l7BpxsrY4btvlsYRssI failed with the exception:”
I don’t have a visible prediction URL because my custom code never returns (the platform appears to bypass it), but you can search backend logs around these times under the Pickaxe name above.
Mitigations Attempted
Created a brand‑new Pickaxe from scratch; no packages, no variables, no actions except the single Custom Action.
Ensured no references to REPLICATE_API_KEY; used a differently named secret (KENS_REPLICATE_TOKEN) in other tests.
Forced an immediate return at the top of the file (hard-smoke) to guarantee zero network calls.
Re-saved the action multiple times, verified code content, published, then re-ran.
Isolated models and added alternative logic in other attempts; still saw the same container build error indicating a call outside my script.
Why this is high priority
The platform appears to auto-invoke a built‑in Replicate flow for a Custom Action, ignoring explicit code that should terminate execution. This can:
Trigger external API calls that the user did not intend.
Create unexpected costs and data handling concerns.
Make the Custom Action editor unreliable (code editor shows one thing while runtime executes another graph).
Requests
Immediate investigation and disablement of any implicit Replicate auto‑routing for this Pickaxe (and ideally a global toggle per Pickaxe to “Disable default provider actions”).
Validation that the Custom Action runtime executes exactly the code shown/saved in the editor at the time of run, with no delayed sync or shadow graph.
Review of routing/decision logs for this Pickaxe to confirm why a Replicate action was invoked.
Clarification on any heuristics that trigger built‑in Replicate flows (e.g., presence of image output schema, secret name scanning, or keywords) so we can avoid accidental triggers.
A short‑term patch (feature flag) to completely bypass built‑in Replicate for this Pickaxe until a permanent fix is deployed.
Optional Diagnostic Artifacts I can provide
Screenshots of the code editor where the hard_smoke/return line appears 10–15s after opening.
Video capture of: Save → immediate run → error → re-open editor → delayed code “pop in.”
Exact timestamps of runs and any prediction IDs visible in logs, if you surface them.
Temporary Workaround Request
Provide a way (toggle or support‑side flag) to ensure Custom Action runs are sandboxed with zero platform auto-actions, regardless of editor state or detected providers.
Contact
Happy to jump on a quick call or share a live repro.
Please assign an incident number and advise on ETA.
Thanks for prioritizing this. It’s a blocker for deploying a custom image generation workflow and appears to be a systemic routing/execution bug.