Tasks, references, and results

This page maintains common task and file rules. See CLI workflows and MCP workflows for execution steps.

Submit once and retain the task ID

Generation can return outputs directly or a task_id. Retain the ID, model, inputs, and reported status, then query the original task until completion or confirmed failure. Interpret actual returned states; a tool wait timeout is not generation failure.

Outcome	Action
A task_id exists	Query/wait for the original task; batch multiple IDs
Still queued or running	Continue querying; stopping a wait does not cancel
Completed	Save every original; report count, format, known costs, and expiry
Confirmed failed	Preserve safe errors and time; check billing without assuming a refund
Unknown submission without ID	Recover through the original route as described below

Released CLI and MCP currently have no usable public cancellation entry. Closing a terminal, conversation, or connection neither cancels the task nor establishes no charge.

Recover timeouts and unknown outcomes

Preserve the original identity and submission record

CLI retains quote_id and its recovery record; MCP retains client_request_id, original inputs, and errors. Prefer querying an existing task_id over generating again.

Recover through the original route

Follow CLI recovery or MCP recovery. Keep the original identity, inputs, and request ID. Do not change accounts, Keys, or CLI/MCP routes and resubmit.

Establish task and billing facts

Save completed outputs; inspect billing after confirmed failure. If still unknown, preserve the error/time and contact support. Unknown does not mean nothing generated or charged.

Make references readable by the model

Check accepted image/video/audio types, counts, sizes, and reference combinations.
Use a file readable in the execution environment or a URL accessible to the server. Chat attachments and desktop paths are not automatically remote-readable.
Confirm the upload receipt and file URL before adding it to model inputs and estimating again.
Preserve the original upload_id when the outcome is unknown. Request a replacement only after confirmed failure or expiry; receipt support depends on deployment.

Route	Input location and limits	Instructions
CLI local file	Readable by the CLI; 95 MiB maximum, models may be stricter	CLI references
Remote MCP URL/inline data	Server-readable URL; 1 MiB base64 limit	MCP upload tools
Remote MCP single-use upload	prepare_upload, upload from a file-capable environment, then get_upload; 95 MiB limit	MCP references
Local stdio	file_path on the MCP process machine; model/file-service limits apply	Local MCP

Expiry and storage

Item	Current default
Original generated URL	Usually 24 hours; save promptly
Uploaded reference	Usually 72 hours; re-upload after expiry
CLI local quote	Usually 15 minutes; re-estimate/approve unsubmitted expired quotes
Remote single-use upload authorization	Usually 15 minutes, one use
Remote upload status URL	Usually 1 hour; retain the original upload_id

Actual returned values apply. Receipt compatibility depends on file-service deployment. Downloaded originals remain in user storage; temporary URLs, documentation, and chat attachments are not permanent storage. Re-upload expired references and reconfirm changed inputs.

Download and preview troubleshooting

Query the original task and confirm completion. Use its original URL; headers cannot fix expired, private, or invalid links.
CLI users can use download. MCP users can open originals in a browser or save them from a file-capable environment. See CLI downloads and MCP delivery.
Check HTTP status, Content-Type, and actual format. Error HTML/JSON is not media. Match extensions to real formats rather than renaming every image PNG.
Open the original when a preview is gray or only a link appears. If it works, deliver/save it. Host previews must not trigger unapproved regeneration or editing.

Default Python requests rejected

When a valid original works in a browser/curl but default Python fails, controlled code can set a product User-Agent. This image example assumes result_url is already a valid original URL:

from urllib.request import Request, urlopen
from shutil import copyfileobj

request = Request(result_url, headers={"User-Agent": "EvoLinkClient/1.0"})
with urlopen(request, timeout=60) as response:
    if not response.headers.get_content_type().startswith("image/"):
        raise ValueError("Expected an image response")
    with open("result.png", "xb") as output:
        copyfileobj(response, output)

The example rejects non-image responses and avoids overwrites. Change the extension when the actual format is not PNG. Video/audio need matching MIME checks, filenames, and separate verification. This approach was verified only for particular metadata/image cases, not every client, file, or network. CLI downloads already set a product UA.

Information for support

Provide OS, agent/version, CLI/MCP version, route, failed stage, time, error code, expected/actual behavior, existing task_id, quote_id/client_request_id, upload_id, and redacted diagnostics. Mention SSH, containers, or proxies. Hide OAuth tokens, Keys, full authorization links, single-use upload URLs, signed private URLs, and private files. Do not share credential files or full environment dumps. Start with CLI troubleshooting or MCP troubleshooting to locate the stage.

Start here

CLI

MCP

Shared resources

Tasks, references, and results

Submit once and retain the task ID

Recover timeouts and unknown outcomes

Make references readable by the model

Expiry and storage

Download and preview troubleshooting

Default Python requests rejected

Information for support

​Submit once and retain the task ID

​Recover timeouts and unknown outcomes

​Make references readable by the model

​Expiry and storage

​Download and preview troubleshooting

​Default Python requests rejected

​Information for support

Submit once and retain the task ID

Recover timeouts and unknown outcomes

Make references readable by the model

Expiry and storage

Download and preview troubleshooting

Default Python requests rejected

Information for support