v0.1.9 — Know the Cost, Cancel Anytime & Exclude the Noise

Author

fariazinhu

🗓️ July 2026

This release puts you in control of both spend and scope. Every /genesis and /sync now shows an up-front cost estimate you confirm before anything runs, you can cancel mid-run with a keypress, and the new /exclude keeps large, irrelevant paths out of the scan entirely. Anthropic also graduates to its own native provider.

✨ Features

• Cost estimate + confirmation gate. Before generating, /genesis and /sync show a breakdown — number of model calls, input tokens (measured from the real prompts), an estimated output range, and a projected cost range — then ask to proceed. It appears even on free models (where it reads Free · $0.00), so you always know what a run costs before it starts. Skip the prompt with --yes / -y for automation
• Live, universal pricing. Prices come from OpenRouter's public model catalog — which also lists OpenAI, Anthropic, and Google models — so cost estimates work even when you call those providers directly. :free models are recognized as zero-cost; unknown models fall back gracefully to a small built-in table (or just show token counts)
• Cancel with ESC. Press ESC (or q) during a run to stop it — genesis and sync both abort the in-flight model calls immediately, even while waiting out a rate-limit backoff. Docs already written are kept
• /exclude — keep big paths out of the scan. Point Aether away from folders that don't need documenting (generated code, vendored deps, legacy modules) to shrink the context and lower cost. Excluded paths never reach the AI — not the file contents, the directory tree, or the change-detection fingerprint
  • /exclude <path> — exclude a file or folder
  • /exclude @ — inline path picker: type to filter, ↑↓ to move, ⏎ to fill in the path
  • /exclude remove @ — pick from what's already excluded to re-include it
  • /exclude — list what's excluded
  • Stored in .aether/settings/exclude.json — commit it to share the list with your team
• Native Anthropic provider. Anthropic now uses its real Messages API (/v1/messages, x-api-key, streaming token usage) instead of being routed through the OpenAI-compatible client. Extended thinking stays off by default

🏗️ Under the hood

• Dedicated OpenRouter provider — split out from the generic OpenAI-compatible client
• Reasoning turned off for docs. Reasoning tokens are billed even though Aether strips the <think> block before saving — pure waste for doc generation. Aether now disables reasoning per provider (OpenRouter's unified reasoning flag; reasoning_effort: low on OpenAI's reasoning models): cheaper and faster, with no change to the docs
• Real token usage in the stream — requests now ask providers to include usage in the final chunk, so estimates can be checked against actuals
• Estimates use real sizes — /sync measures each doc being refreshed from disk (the patch prompt re-sends it), so the input/output estimate tracks reality instead of a flat guess

🐛 Fixes (0.1.5 – 0.1.8)

• Clearer connection errors — a failed provider check now says whether it timed out, was rejected (bad key), or couldn't connect, instead of one generic "service down" message
• Install script and version-reporting fixes
• Interactive prompt polish