Troubleshooting¶

Errors and connection notices appear in a slim strip above the composer. Open its chevron for the full diagnostic, or dismiss the notice. A new occurrence appears again. A model connection failure also offers Retry to test the connection and Open settings to review the model setup; Retry does not resend your chat request.

What Nawa says, what it means, and what to do. Every turn also has a trace id (shown with an error, and in the receipts); keep it if you ask for help, since it names the exact run in Nawa's log.

"Nawa couldn't start"¶

Nawa's engine stopped while Nawa was starting. The window shows the error code and the log that holds the engine's own account of it, %APPDATA%\Nawa\data\server.log. Close Nawa and open it again; if it fails again, keep that log for your report.

An answer or tool call ended early¶

If the provider interrupts an answer, Nawa keeps the available text as an unverified draft. The diagnostic includes the provider’s stop reason. If a batch of tool calls is incomplete, Nawa stops before dispatching any call in that batch. Earlier successful actions remain in the transcript.

"LLM unavailable"¶

Nawa could not reach the selected model, or its server remained unavailable or rate-limited. Nawa retries transient failures before stopping. During connection recovery, the SVG connection notice shows the current attempt with shimmering progress text. It clears when connected or when you stop the turn. Connection configuration problems, such as an untrusted certificate or an HTTPS address pointing to a plain HTTP server, are reported immediately.

Open the notice's chevron to read the diagnostic. Check that the model server is running, the model is loaded, and its address, port and protocol are correct. For 429, wait for the provider's rate limit to clear; for 5xx, check the provider's service or try another configured model.

Retry tests the connection; it does not resend your prompt. Use Open settings to review the model setup. Once the connection works, send your request again.

Nawa uses the Windows proxy configuration, including PAC scripts, and the Windows certificate trust store. Check proxy and certificate configuration if the diagnostic points there. SOCKS proxies and proxies requiring NTLM/Kerberos sign-in are not supported.

"Model request failed"¶

The model provider rejected a request or returned an unusable response. When it answered, the card says the status it returned and the message its answer states.

401 or 403 — the provider did not accept the authentication or access requested. Check that the selected provider, server address and key belong to the same service, then test the connection. The status alone does not establish that you copied the key incorrectly. If the same key and model work in another app, include that fact in your report; never send the key itself.
400 with a message about the request — the model rejected something Nawa sent. Two known cases are handled: a provider that accepts only auto, none or required as the tool choice, and a deployment that cannot force a tool call (it may answer 500 instead). Nawa asks again the accepted way. It remembers a refusal for that deployment for a day, unless you press Connect in Settings: then it checks again on the next request that needs it. A 500 is not remembered; Nawa checks again the next time it needs to force a call. Anything else is worth reporting with the trace id.

"Output limit reached"¶

The model wrote as much as it may write in one reply, and asking it to go on did not finish the answer. Ask it to continue; it receives what it had written. If Settings → Budgets → Output tokens per model call holds a number, a larger one, or blank for the model's own allowance, gives each reply more room.

A reply ends with an "Unverified draft" card¶

A reply that failed, or that you stopped, while the model was still writing its answer keeps what it had written in a closed card under the reply. Nawa did not check it the way it checks an answer, so it can be unfinished or wrong, and its links and citations are plain text. Open the card to read or copy it. Ask the assistant to continue and it starts from the same draft.

The model prints tool-call JSON or says it is done without delivering¶

Nawa attempts recovery through the provider’s native tool-calling interface. The coder is reminded to call finish_coding; repeated missing-tool replies produce a clear incomplete result. This does not guarantee that every model or custom server can recover. Keep the trace ID, confirm the server supports native tool calls, and check the selected Main and Coder models in this chat’s composer menu. A provider billing or credit rejection must be resolved with that provider.

The setup card says "Key missing"¶

No key is saved on that card. Paste one and test the connection. A card whose key was saved but whose model was never chosen says so instead; test the connection and choose the model.

"Context window full", or the provider says the context is too large¶

Nawa uses the model's reported context size. If a provider refuses an oversized request and states a smaller limit, Nawa records that limit, compacts the conversation and retries. A server that reports no size initially uses an explicitly labelled assumption of 256,000 tokens.

The chat holds more than the model can read at once. Nawa says Context window full when it measured that a request cannot fit, or when the model stopped its reply because its window was full; a provider that refuses the request itself says so in its own words ("context size has been exceeded", "maximum context length"). Nawa compacts long chats on its own; if this still appears, the model's window is smaller than Nawa measured, or one message (a large attachment, a huge tool result) is bigger than the window. Start a new chat for the next task, or attach a smaller file. The composer's context gauge shows how full the window is.

The connector says "Needs authentication"¶

The server lists its tools but refuses to run them until you sign in. Press Sign in on its entry. If the sign-in page asks for a client ID, the service does not register clients on its own; the entry explains where to create one.

The decision model says "No Laya worker is loaded"¶

The Laya server answered that no model is loaded. Start the worker on the server (from its dashboard), then press Connect again. "Decision request could not reach the provider: …" means nothing answered at the server URL: the URL is wrong or the server is down, and the rest of the line is what the connection reported.

Through OpenRouter, "Save a key on the OpenRouter card first." means the Jev card is set to use the OpenRouter card's key and that card has none.

A file tab says "… no longer exists in this workspace"¶

The file the tab shows has been deleted or rebuilt away, since an earlier session or while it was shown. The drawer shows the workspace's files under that line, so you can open what the workspace holds now; Refresh tries the file again, or close the tab.

A permission card keeps waiting¶

Nawa is paused until you choose. Cards never time out on their own; if you have left the chat, the card is still there, with the time it has waited. Choose, or press Stop generating (the send button while a turn runs) to end the turn.

The web is off¶

If Nawa answers from memory when you expected research, turn the web toggle in the composer on. If a search provider is not configured, open Settings → Web and add its key or choose Automatic or Web search (DDGS) to search without one. DDGS runs in the background; engine failures and rate limits are reported as incomplete. Retry or configure another provider as a fallback. Use Connect to check the selected provider.

An app does not open in the drawer¶

The drawer runs one app at a time and stops the others; a tab whose app was stopped shows a dot and starts again when you show it. If an app fails to start, the menu on its card offers Open in External Browser and Open in Windows Explorer so you can see the files it built. If the drawer opened but stayed empty after an update, restart Nawa: the running engine and the renderer must be the same version.

Report a bug¶

Click the bug icon at the bottom of the left pane, beside the gear, or choose Help → Report a Bug…. Describe what you expected and what happened instead; add your email if you would like a reply. Send sends your message and a diagnostic report to the Nawa team at nawaagent.com, and the dialog then shows the report's reference, such as R-1042. If it cannot be sent, the dialog says why and keeps your text for Try again.

The report covers the chat that is open, and no other chat:

From Nawa's logs, the lines that belong to that chat (the tools it ran, errors and timings, and short tool inputs such as file names and search terms) and the lines that belong to no chat at all, such as Nawa starting, updating and network or provider errors. Lines from other chats are left out. Earlier versions of Nawa did not mark which chat a log line belonged to, so of the lines they wrote, only those naming this chat are included.
The chat's records in Nawa's database, such as its title, a reply still in progress, builds, and automations with their recent runs and results. These diagnostic records are included even when recent messages are off. The report states the number of database rows omitted by its limits.
With Include recent messages from this chat on, the chat's last messages and the results of their tools. Images, attachments, workspace files and the database file itself are never included.
Your model connections: each one's provider, server address and model, and whether a key is saved for that address. A saved key appears only as its length, its public prefix such as sk-or-v1-, and whether it contains spaces; the key itself, the names of your setups and your other account settings are not included.
App and system facts, and your settings without passwords or keys. The system facts include your graphics adapters and drivers, whether Nawa is drawn by the graphics card or in software, whether Nawa runs translated on an ARM processor, each screen's size, work area and scale, and the size and zoom of Nawa's windows, but not their titles. Your Windows user folder appears as %USERPROFILE%, and your computer's name is not included. Tailscale addresses and device names appear as numbered placeholders, such as [tailscale-ip-1] and [tailnet-host-1].ts.net, so the report can still tell your devices apart.

Reported from a page without a chat, such as Plugins, the report holds only Nawa's own records.

Help → Nawa Help opens this documentation on nawaagent.com.

Where to look further¶

The receipts in the transcript: every step's input, output and timing, including the model's own requests.
%APPDATA%\Nawa\data\server.log: the engine's log, one line per tool call and per turn, with [turn], [tool], [decisions] and [system-error] entries keyed by trace id.
%APPDATA%\Nawa\data\runtime\trace-index.jsonl: the timing of every phase of every turn, as one JSON line each.

Each log keeps its newest 5 MB. The 5 MB before them are in the same file with .1 added, such as server.log.1; older entries are deleted.