v0.5.1

Community feedback release: 18 fixes and features sourced from GitHub issues and the Discord server — every report from the last month triaged, fixed, and independently verified.

Reasoning & providers

• No more chain-of-thought leaking into replies — unparseable reasoning is never "salvaged" as the agent's answer; prompts are now transport-aware, so native function-calling providers stop being ordered to emit text-JSON tool calls (#286, fixes #285)
• Think-tag prefill no longer sent to native-tools chat providers — stops Ollama Cloud silently corrupting literal <think> strings; live reasoning display survives grammar fallover (#287, fixes #283)
• Live reasoning no longer freezes at the first { in the model's thoughts — the stream parser only hands off to tool-call parsing on a real {"tool": " payload (#292)
• Gemini RetryInfo.retryDelay honored on 429/503 — cooldowns from the error body instead of burning retries on fixed backoff (#271, fixes #106)
• Final step reserved for a real answer — the agent replies instead of stalling at max_steps; user cancellation is no longer mislabelled as max_steps (#272, fixes #107 — based on @mikemikimike's #269)

Security & correctness

• MCP tools with trust: approval_gated actually route through the approval gate — strict readOnlyHint === true exempts; everything else fails closed (#273, fixes #132)
• Search cache + provider cooldown persist across restarts — per-task processes stop re-spending quota (#274, fixes #256)
• skills.catalogTokenBudget is honored — the skill catalog is no longer hard-capped at 4096 chars (#288)
• Warn on repeated test runs against an unchanged workspace (#276, fixes #118)

Local models

• CPU llama.cpp fallback — iGPU-only Windows boxes get an automatic, visible fallback to the CPU build instead of a silent model-load failure; backendVariant config for manual control (#293)
• Multi-GPU tensor split — localModels.managed.tensorSplit passes --split-mode layer --tensor-split to the managed llama-server (#298)
• Ollama steer — pointing the External llama.cpp URL at an Ollama server now offers a one-key jump into the OpenAI-compatible provider wizard, prefilled (#290)

TUI

• Clicking a slash-command completion runs the clicked command, not the half-typed buffer (#289)
• Free/paid price filter on both cloud model pickers — press p (#295)
• Mouse-report hardening over ssh — coordinate spam can't leak into the composer; a tripped session auto-disables mouse tracking with a /mouse hint (#297)
• Legacy conhost guard — Windows 10 cmd/PowerShell (outside Windows Terminal) reserves the bottom row to stop shaking/duplicated last-row rendering, with a one-time Windows Terminal hint; opt out with ATOMIC_AGENT_CONHOST_GUARD=0 (#291)

Memory

• memory export --vault <path> — one-way, idempotent Obsidian export of notes/lessons/procedures with frontmatter and wikilinks (#296)
• MEMORY_GUIDE.md — the memory system end to end, with worked examples (#294)

Thanks to everyone who reported on Discord and GitHub: l.hk, thegreatteacher, detraxsenpai, adithr, johnfasterlane_83003, magicw0rm5, gokcanakalin, yablokolabs, Ramyazzam, sosidudku1, and @mikemikimike for #269.