Debugging

margin logs every AI request and response locally so you can inspect what's happening under the hood.

AI Logs

Every time you send a request to the AI (in edit mode or chat mode), margin saves a log entry that includes:

• Your input prompt / instruction
• The complete context sent to the writer agent (characters, styles, additional context, etc.)
• The full AI response
• Token usage information
• Session ID and timestamp

These logs are stored as JSON files in the outputs/ai_logs/ folder inside your workspace.

Viewing Logs

You can view logs directly in the app through the AI Assist panel -- look for the log/book icon that opens the request inspector. This shows the last API request, the input and output context, the total session context, and remaining context window.

Editing or Clearing Logs

• Logs are plain JSON files -- you can open, edit, or delete them with any text editor.
• To clear all logs for a session, use the delete option in the log viewer.

Logs are purely local. They never leave your machine.

Inspecting Prompts

The prompt templates used by the Planner and Writer agents are stored as markdown files in the prompts/ folder at the project root:

File	Purpose
prompts/simple-planner.md	Instructions for the Planner agent
prompts/simple-writer.md	Instructions for the Writer agent
prompts/simple-chat.md	Instructions for Chat mode

You can read these files to understand exactly what instructions are being sent to the AI. Edits take effect immediately -- no restart needed.

Editing prompts changes how the AI behaves. If something breaks, you can restore the original from the project's git history.

Telemetry

margin has no external telemetry. There are no analytics, no tracking scripts, and no data sent to external servers (except the AI provider you explicitly configure).

The only "telemetry" is local:

• AI logs stored in outputs/ai_logs/ for debugging your own requests
• Image logs stored in outputs/image_logs/images.json — every generation's prompt, seed, provider, and asset path; viewable in-app under History → Images
• Settings saved in your platform's config directory — ~/.config/slm-writing-engine/settings.json on Linux, ~/Library/Application Support/slm-writing-engine/settings.json on macOS, %APPDATA%\slm-writing-engine\settings.json on Windows
• No crash reports, no usage stats, no pings home

Everything stays on your machine.

Image Issues

Images are plain Markdown references (![alt](assets/foo.png)), so most problems come down to a mismatch between the text and the file. Your Markdown is never auto-deleted or rewritten — fix the underlying cause and the image comes back.

Broken image placeholder

The referenced asset file is missing — it was renamed, moved, or deleted. Check that the file exists under your workspace's assets/ folder and that the path in the Markdown matches exactly (including case). Editing the source line to the correct path restores the image.

Pasted image doesn't appear

• If an upload toast appeared, the file wasn't a supported image (PNG, JPG, WebP, GIF) — the pasted text itself is untouched.
• If a pasted image URL produced no image, the download failed (bad link, offline host, or a non-image response). Your pasted URL text stays in the document; try the URL in a browser to check it.

Wrong size after editing dimensions

Only positive integers count: ![alt|0](x.png) or ![alt|-5](x.png) fall back to natural size. A | followed by non-digits (e.g. ![a|b](x.png)) is treated as literal alt text, not a size. An {align=...} value other than left, center, or right is left as visible text — correct the spelling to apply it.

Remote image never becomes local

Only explicitly pasted/imported image URLs are downloaded into assets/. An image URL you typed by hand into Markdown stays remote by design — paste the URL as document content if you want margin to import a local copy.

Image looks too large after import

New images render at natural size bounded by the editor width. Select the image and drag a handle to persist a smaller width.

Image generation fails

• "Prompt is required" / unknown style / missing reference — your request or style selection is wrong. Check the prompt and Settings → Images → Styles.
• "API key rejected", "not configured", stale mapping warnings — your provider configuration is wrong. Re-check Settings → Images and use Test provider; for ComfyUI, re-import the workflow if you changed it (mappings go stale when nodes are renamed).
• Anything else (timeouts, execution failures, bad responses) — the provider itself failed. Check the provider is reachable (local ComfyUI running? service status page?) and try again.
• Imagine again changed more than expected — only your typed change description plus the style is submitted alongside the reference image. A bare instruction ("make it blue") gives the model little to anchor on, so results vary more — describe the change relative to the current image.
• Every Imagine again looks oddly similar — your workflow's saved seed is being reused. Register a seed mapping in Settings → Images so each run gets a fresh random seed; the submitted seed is recorded in History → Images.

Generation history

Image runs are logged to outputs/image_logs/images.json in your workspace (prompt, submitted text, seed, asset path) — separate from chat ai_logs/. Delete entries from History → Images in the sidebar; the asset files themselves are kept.