Loading documentation...
Loading documentation...
Release history for TachiBot MCP — new models, tools, and orchestration capabilities.
TACHI_PLANNER_TIER=standard for the cheaper lineup.gpt-6-luna default. Override with TACHI_JURY_OPENAI_MODEL.code_focus includes planner_maker but not qwen_reason, so two steps pointed at a missing tool. They now fall back to deepseek_reason, and the 50% gate sends the right params.foo:5 and . could slip past it before.TACHI_GEMINI_PRO_MODEL=gemini-3.1-pro-preview; any Argon failure falls back to 3.1 Pro, then Flash.openai_search moved to GPT-6.1 sol. Cheaper than terra ($2/$10 vs $2/$12), and on a fresh-news query it found the actual launch where terra returned a minor update. GPT-6.1 sol is also Astra’s first fallback.grok_search moved to grok-4.7. This reverses the Sep 29 call: across three queries 4.7 returned 19 / 79 / 2 unique citations against 4.6’s 2 / 4 / 2.stepfun_reason runs step-5-preview. 600B / 27B-active MoE, 1M context, $1/$2.70, falling back to step-3.7-flash.jev_decide is new. TypeSafe Jev answers typed questions about a state — a yes/no probability, an option pick, or a score — as JSON instead of prose. Under a second and about $0.00002 a call, so it fits routing, groundedness checks and risk scoring.condition.if: "gate.answers.grounded.noul < 0.6" runs only when Jev doubts the answer; the new jev-gated-answer workflow researches further only then. /council, /write, /review and /prompt use Jev as a quick cross-check.ernie_reason is removed. Baidu’s model expired from OpenRouter on Oct 8 and every call would 404. Still 67 tools: jev_decide in, ernie_reason out.files only inside the working directory (no .env, keys or credentials), and cannot call the workflow tools themselves.[degraded: X unavailable … answered by Y]. Set TACHI_MODEL_FALLBACK=off to forbid swaps entirely; billing failures never fall back.length stop with empty content now retries once with a larger budget, then returns an explicit error. Long grok_search runs heartbeat instead of hitting the client’s 60s timeout.openai_reason, openai_brainstorm, openai_code_review, spec_writer, tachi and the GPT reviewers, falling back to sol, then terra.qwq_reason is back. Now on qwen3.8-27b after OpenRouter retired qwq-32b. Same name and parameters.storm, re3 and de_generic fill the structure, draft and revise stages. 85 → 88 techniques.technique="auto"; blog_writer’s limit raised to 16000 tokens.