Ante Files

v0.2.9 - 2026-10-05

Added

• s in the /fork picker opens the forked child in a terminal split beside the parent instead of switching to it, with the cut prompt ready to edit; uses the same split launchers as /term (tmux, iTerm2, WezTerm, Kitty, Ghostty)
• Opt-in context-tokens status line item shows the context window as used/limit ctx (for example 13.6k/400k ctx), the same figure /context reports; turn it on in /statusline
• A provider's http_headers value in catalog.json may contain {session_id}, filled with the conversation's id on each request (a subagent sends its parent's id), for gateways that need a per-session header

Changed

• /rewind and /fork reach the prompts before a /resume and the prompts a fork copied from its parent, back to the last compaction or provider switch, instead of only the current run's
• Every provider call runs under one timeout owned by the transport: headers, error body, OAuth refresh, retry backoff and the WebSocket-to-HTTP fallback share one budget
• Catalog pruned: GPT-5.4 Pro on direct OpenAI (GPT-5.5 Pro stays)
• llama.cpp engine bumped to b11379 (Metal memory-leak fix, Vulkan cache-stride fix, GGUF overflow checks)
• Compaction logs the fold's configuration and the summary's thinking/text split, to diagnose cut-off summaries on local servers

Fixed

• Responses API "input exceeds the context window" rejections are treated as context overflows and recovered by compaction