Limits & Known Behaviour

Some of what AETHER Coder does surprises people the first time, and a few of those surprises are deliberate design decisions rather than faults. This article collects them in one place, so you can tell "that is how it works" apart from "that is broken".

If something is genuinely failing rather than behaving as described here, go to Troubleshooting.

Four images per message

You may attach four images to a single message, and 20 images and files combined.

The cap exists because of how the extension works rather than as an arbitrary quota: the client owns the transcript and sends it back on every turn, and attached images are replayed along with it. Six screenshots on the first message are still being carried on the twentieth, and a request eventually grows past what the platform will accept.

In practice that makes attachments worth choosing rather than collecting:

• Attach the screenshot that shows the fault, not the four that surround it.
• Crop to the part you are asking about. A cropped screenshot is also easier for the agent to read.
• When the work moves on to a different screen, start a new task. The old images stop being carried, and the new task starts small.

The connect form does not validate

Nothing you type into the connect form is checked as you type it, and clicking Continue does not test the connection. A wrong URL, key or token is accepted quietly and fails on your first message instead.

This is why the first message of the first task is the moment to watch. If it fails immediately, go back to Settings → API Configuration and check all three values — see Connect to AETHER.

The cost badge is real, with a few gaps

The extension's Task view shows a running total of what the task was actually billed, at the same marked-up rate the Cost Report reads — not a local estimate. Each request freezes its own cost as it completes, so a later price or model change never rewrites an earlier line.

A few things it does not cover:

• It sums chat-completion requests only. Embedding calls the agent makes while searching your knowledge base during a task are billed but do not add to the badge.
• On a retried request, the platform bills every attempt while the badge keeps only the last one.
• A task resumed from history keeps the cost each of its messages was saved with; older messages saved before this behaviour shipped keep the $0.0000 they were saved with.
• A model with no configured price still shows $0.0000 — pricing has to exist for a request to be billed at all.

For reconciliation, the Cost Report remains the authoritative number, priced across all your other usage too. See Cost, Budgets & Auditing.

The model id must be aether

Model ID in Settings → API Configuration is not a model name and it does not choose a model. It is the fixed identifier the platform expects, and the real model is resolved from the agent's configuration in AETHER.

Change it and the agent does not fail cleanly — it falls back to a generic prompt and behaves noticeably worse: answers grow vaguer, it drifts out of the format it is meant to follow, and it describes a change rather than making it. If quality drops for no obvious reason, this field is the first thing to check.

Only the AETHER connection is supported

The settings screen exposes other provider options in its dropdown. They are not part of this product, they are not tested against your platform, and they are not supported.

Selecting one takes you outside everything AETHER provides: no agent knowledge base, no platform tools, no budget enforcement, and no audit trail attributing the session to you. Leave the connection as AETHER.

Long tasks degrade

The context window is finite and every turn carries the whole task with it, so a conversation that runs long enough will get slower, more expensive and less accurate — in that order.

Compaction buys you room but it is not free of loss, and no amount of it makes one enormous task as good as three scoped ones. Prefer /newtask at a natural boundary — see Context: Mentions, Images & Focus.

Approval is the safety model

There is no undo outside your workspace. What protects you is the review gate: diffs you approve, commands you approve, and checkpoints you can restore to.

Auto-approving edits and commands removes that gate by design. That is a reasonable trade in a scratch repository or on a throwaway branch, and a poor one on a working tree you care about or a machine that can reach production credentials. Make it a deliberate choice rather than something that accumulates — see Running Commands & Auto-Approve.

Note: the agent runs as you. It has your file permissions, your terminal and your credentials, and it does not commit or push on your behalf. Start tasks from a clean working tree so the diff at the end shows the agent's work and nothing else.

Next steps

• Context: Mentions, Images & Focus — keeping tasks small and well-fed.
• Troubleshooting — when it really is broken.
• Cost, Budgets & Auditing — where the authoritative numbers live.