1.0.44

• CLAUDE.md and AGENTS.md are loaded in full: if they do not fit, Swival stops with a message naming the files instead of trimming them.
  --instructions-full loads them regardless, an unreadable file is reported and skipped, and loaded rules now survive context compaction.
• /remember and /init now take effect in the running session, not just the next one.
  If the change cannot be applied, Swival prints the exact line and file and holds back model turns until you fix it.
• Without tokenizer data, context sizes are now treated as estimates: Swival warns once, and no longer compacts or clamps output on a guess.
• Added /reasoning to inspect or change the reasoning effort during a REPL session.
  /reasoning LEVEL applies to the next turn and subsequently spawned subagents, /reasoning - restores the previous level, and selecting a profile restores that profile's setting.
• If a model rejects a prompt containing an image, Swival now removes the image and retries the request as text-only.
  A model-wide vision rejection also withdraws view_image for that endpoint and model for the rest of the process, while a rejection specific to one image leaves the tool available.
• The generic provider now sends stable x-opencode-session and x-opencode-client headers to opencode.ai, allowing the OpenCode Go service to keep related turns, retries, compactions, and worktree calls in one session.
  /clear, /new, and subagents start independent sessions.
• Automatic Git metadata collection and /audit now ignore repository-configured hooks, filesystem monitors, filters, diff helpers, text conversion, and alternate attribute sources.
  This prevents a received directory with a preserved malicious .git/config from running code before a prompt or during an audit.
• grep include globs now honor their directory components, and recursive searches skip FIFOs and other special files instead of risking a hang.
  Token counting also accepts text containing reserved-token spellings such as <|endoftext|> and falls back to conservative byte counting when the tokenizer data is unavailable offline.
• Regenerating an existing config with --init-config now preserves all recognized settings, inline nested tables, quoted keys, and escaped control characters instead of silently dropping or corrupting values that are absent from the current template.
  Relative command-provider model paths inside profiles are now resolved relative to the config file, and switching profiles correctly replaces a profile-specific user_agent.
• A2A request bodies are now size-limited while they stream, so clients cannot bypass the configured limit by omitting or understating Content-Length.
• Automatic correction for a single-model server is now scoped to the endpoint as well as the requested model, preventing a correction learned from one server from changing requests sent to another.
• --no-subagents is now honored even when the context window would otherwise auto-enable subagents, and an outer Ctrl-C is reported to lifecycle exit hooks as an interruption with status 130.
• /audit now preserves tracked Git pathnames containing newlines and treats unreadable files as failed evidence instead of silently omitting them from deep review.