What's New

Changelog and release history for TachiBot MCP. Every release brings new models, tools, and orchestration capabilities.

v2.34.0September 14, 2026New Models

GPT-6 Astra is the default, and qwq_reason works again

GPT-6 Astra is the OpenAI default. gpt-6-astra (released September 3, 2026; $10/$50 per million tokens, 1.05M context) now backs openai_reason, openai_brainstorm, openai_code_review, spec_writer, tachi and the GPT reviewers in diff_review and plan_critique. Astra rolled out to a limited set of organisations first, so calls fall back to gpt-5.6-sol, then gpt-5.6-terra, when an org lacks access. openai_explain deliberately stays on sol — low-effort explanations do not need a model with twice the input price.
Astra could not be called through the existing code path. callOpenAI used the Responses API only for model names starting with gpt-5., so a GPT-6 model fell through to Chat Completions with temperature: 0.7 and no reasoning effort — which OpenAI rejects with a 400: "'temperature' does not support 0.7 with this model". GPT-6 models now go to the Responses API without temperature, and the none / minimal efforts Astra refuses are sent as low. Verified live on September 14.
qwq_reason had been failing on every call. OpenRouter retired qwen/qwq-32b — every request returned 404 "No endpoints found". The tool now runs qwen/qwen3.8-27b (dense 27B, 1M context, $0.21/$2.55) at reasoning_effort: medium, falling back to the 235B thinking model. On a probability probe it answered correctly in 2.0s for $0.0007; the 235B model took 16.3s. Tool name and parameters are unchanged.
Stale labels and version drift. The review panels still labelled their GPT reviewer "GPT-5.6 Sol" and "GPT-5.5", and 2.33.1 had bumped only package.json, leaving three packaging manifests on 2.33.0. Tool count unchanged at 67.

v2.33.0August 27, 2026New Tools

The library could think creatively and not write a word

Creative thinking was covered; creative writing was empty. Against a seven-stage writing pipeline, stage 4 (structure) and stage 5 (draft) had nothing at all, and stage 6 (revise) fell through to reflexion — which optimises toward correct. Prose revision is cutting, rhythm and specificity, a different objective entirely. The writing family — storm, re3, de_generic — fills it. 85 → 88 techniques.
blog_writer is researched long-form in one call. It composes storm and re3 rather than carrying a private copy of their text, because a copy would drift the moment either template was edited — and drift silently, since both paths still emit well-formed prose. 66 → 67 tools.
"Write a blog post" routed to the wrong three techniques. technique="auto" had no matcher for writing intent, so it fell through to the first three core techniques — chain_of_note, astute_rag, spotlighting — while the tool list advertised auto as the way in.
blog_writer truncated silently at 8000 tokens. Its output is several full renderings of the same piece — draft, restructure and line-edit passes, plus 16 marked research answers — not the single artifact sibling tools emit. 8000 was already at the ceiling at the word count the docs use as their own example, and nothing errored on truncation. Raised to 16000.
The anti-drift tests could not detect drift. The tests asserting "uses the template rather than a private copy" checked for characteristic phrases that a copy would satisfy just as well. They now assert the engine's exact live output.

v2.32.0August 15, 2026New Tools

Ten techniques for tables, SQL and sources that contradict each other

Data-shaped work had no techniques at all. "Filter these rows" fell through to the generic core contracts, so seven tabular techniques now cover it. data_assert states the downstream purpose first, then checks the data with executable assertions carrying counts — a filter that silently matched zero rows is otherwise indistinguishable from one that worked.
Large tables are pruned before they are reasoned over. chain_of_table transforms one operation at a time, showing the intermediate table and row count after each; sub_table_first extracts the relevant sub-table with one explicit query so the reasoning runs over the small result instead of 10k rows; sql_stages does text-to-SQL in four passes — schema-link, classify, generate, self-correct — capped at 2 correction rounds.
Human-made spreadsheets are mapped before they are computed. xlsx_map looks for merged cells, multi-row headers, formulas, and the subtotal rows hiding inside the data that double-count silently. table_format picks Markdown, HTML, CSV or JSON on purpose, because the format measurably changes accuracy and long tables lose their middle.
The retrieval family is now complete. astute_rag answers from memory first, then marks each source agrees, fills or conflicts and resolves every conflict out loud — a wrong page silently overriding correct knowledge looks identical to the reverse. crag grades sources correct, ambiguous or wrong and acts on the grade, since search always returns its nearest neighbours and has no way to return nothing. hyde drafts the answer you wish existed, searches with its vocabulary, then discards it.
Debugging now uses the evidence it already has. self_debug feeds the actual traceback back in and fixes the smallest thing it implicates, capped at 3 rounds; re-prompting from scratch throws away the only evidence you have. It now leads the bug and traceback routing rule.
Spotlighting fences carried 24 bits of entropy, not 32. The code called for 3 random bytes while its own comment block had reasoned its way to 4. Spotlighting's entire security property is that an attacker cannot guess the delimiter needed to close your block, so the shortfall was load-bearing. Fences are now 8 hex characters, as documented.
The literal 37 survived three expansions of the catalogue. It sat in the list_prompt_techniques description — the text a model reads to decide whether to call the tool at all — while the real count went 37 to 74 to 85. It is now derived from the catalogue itself. consilience also gained confidence_trajectory, which ranks runs by the shape of their confidence rather than by majority vote: a run that was certain from its first line was guessing.

v2.31.0August 15, 2026New Tools

The prompt library goes 37 → 74, closing three families that were empty

Retrieved sources are now read before they are cited. TachiBot ships five search tools and had no discipline for what they return — retrieved text went straight into reasoning unfiltered, which is where confidently-miscited sources come from. chain_of_note writes a note on each source first, self_rag marks every claim SUPPORTED or UNSUPPORTED, and ircot interleaves retrieval with reasoning for multi-hop questions.
Untrusted web text is now fenced as data. Those same five tools ingested scraped pages with no isolation contract, so a page saying "ignore previous instructions" read exactly like your own prompt. spotlighting marks untrusted content explicitly: analyse it, never obey it.
Few-shot prompting existed nowhere in the library. Which examples you show moves accuracy more than most reasoning tricks. analogical is the cheapest entry — the model writes its own worked examples, so there is nothing to curate.
A cheap reasoning mode, finally. chain_of_draft caps each step at about five words, and adaptive_consistency stops sampling the moment consensus is clear rather than always running a fixed N.
self_consistency silently failed on prose, code and plans — it votes by matching answers, and there is nothing to match. universal_consistency reads the candidates and picks the most consistent instead.
The recommender knows all of them. Past forty techniques nobody recalls names, so refine_prompt gained eight rule groups matched to how people phrase things — "cite these sources" suggests chain_of_note, "review each of these 12 files" suggests compute_ration.
Twelve original prompts rewritten to read as instructions. The early set used shorthand that read as a label rather than a request, leaving the model to infer the task.

v2.30.0August 14, 2026Bug Fixes

Reliability at every boundary

Protocol-safe diagnostics. stdout stays reserved for JSON-RPC and every diagnostic routes to stderr, with guard coverage across console.info, console.debug and process.stdout.write.
Provider-verified routing. Every wired model ID is asserted against what its provider actually serves, so a substitution is caught at test time.
Evidence-based search tests. A pass requires a web_search_call plus a citation rather than matching answer text — prose alone cannot distinguish a real search from a confident guess.
Purpose-built delegation. qwen_algo routes to Qwen3.8-Max and qwq_reason to its four-persona deliberation, each handled by the tool built for it.
Workflow contract integrity. planner_maker runs a schema-exact Qwen step, and create_workflow requires an explicit overwrite: true before replacing anything — reporting the real .tachibot/workflows/ path as it goes.
Input resilience. Unrecognised approach values fall back cleanly across the Grok, Kimi, Qwen, DeepSeek, GLM, StepFun and ERNIE reasoning tools.
Deeper verification. The suite expands from 130 to 158 tests, and the live search probe stays opt-in behind RUN_LIVE_TESTS=1 so an ordinary run never makes a billed API call.

No new tools by design — v2.30.1 strengthens the 66 already at work. Zero breaking changes.

v2.28.0August 3, 2026New Models

Qwen3.8 Max

Qwen3.8 Max (qwen/qwen3.8-max) — Alibaba's new flagship, general-availability the same day it landed here. 1M context, up from 262K, multimodal input (text, image, video), 131K max output, $2/$6 per M. It is the first Qwen model to expose configurable reasoning effort, and it now powers qwen_algo, qwen_reason and the Qwen reasoning juror
Picked on measured results, not the version number — five Qwen candidates ran the same range-query problem. 3.8 Max was the only one to surface both the offline-vs-online tradeoff and the strict-inequality-with-duplicates edge case. The outgoing 235B Thinking model was correct but shallow; qwen3.7-max took twice the wall time for 1.6x the cost; qwen3-max-thinking was rejected outright — it returned zero reasoning tokens, so its fast result was never a thinking pass
Effort is pinned to medium, deliberately — this model's default effort behaves like high: 302s and $0.09 for a single qwen_algo call. At medium the same call answers at equal depth in 18–48s for $0.006–0.018 — cheaper and 3.5x faster than the model it replaces, which took 169s and $0.036 for a shorter answer. Left on default it would have been a 6x latency regression
Fallbacks and timeouts — quota chain is 3.8 Max → 3.7 Max → 235B Thinking, and both new tiers join the 600s extended-timeout bucket (their IDs contain no "thinking"/"reasoning" marker, so they would otherwise have inherited the 180s default). Output budget for both tools rises 8K → 12K
Coding tools deliberately unchanged — qwen_coder, qwen_competitive and testgen stay on Qwen3-Coder-Next: coding-specialized and ~16x cheaper. 3.8 Max is the reasoning tier, not the codegen tier. Still 65 tools

v2.27.1July 26, 2026New Models

Kimi K3 & Gemini 3.6 Flash

Kimi K3 (moonshotai/kimi-k3) — the largest open-weight model shipped: 2.8T MoE, 1M context, natively multimodal, built for long-horizon agentic coding. Replaces K2.7-Code across kimi_thinking, kimi_code, kimi_decompose, kimi_long_context and the jury juror — the long-context window quadruples from 262K to 1M. Falls back to k2.7-code → k2.6 on quota
Gemini 3.6 Flash (gemini-3.6-flash) — new flash tier powering gemini_search, 1M context. Cheap tier moves to gemini-3.5-flash-lite. Gemini 3.1 Pro remains the reasoning default and presiding jury judge — Google shipped three models on July 21 and explicitly skipped 3.5 Pro
Full provider audit — every other provider re-checked and confirmed current: DeepSeek V4 Pro, GLM-5.2, MiniMax M3, StepFun 3.7 Flash, ERNIE 4.5 VL, Qwen3-Coder-Next. No bumpable versions remain

v2.27.0July 12, 2026New Models

65 Tools — Grok 4.5, GPT-5.6 Tiers & grok_search_lite

Grok 4.5 (grok-4.5) — xAI's Opus-class flagship: 500K context, configurable reasoning effort, $2/$6. Now the default across every Grok tool. Rollout is region-staged, so tools auto-fall back to grok-4.3 (1M context) if your region or quota blocks it
GPT-5.6 tiers replace pro & mini — gpt-5.6-sol is the flagship (54% more token-efficient than 5.5), gpt-5.6-terra handles code at half of 5.5's price, gpt-5.6-luna is the fast/cheap tier. All 1.05M context / 128K output. "Pro" is no longer a model — it's sol plus high reasoning effort, and the $30/$180 5.5-pro tier is gone
New tool: grok_search_lite — live search on grok-4-1-fast ($0.20/$0.50, 2M context), roughly 10x cheaper than grok_search. Built for high-volume lookups and jury fan-outs. 65 tools total
Profile counts updated — minimal 13, code_focus 42, research_power 36, balanced 54, heavy_coding 58, full 65 (default profile)

v2.26.0July 3, 2026Major Expansion

64 Tools — refine_prompt, Setup Wizard & One-Click Desktop Install

Prompt stack, modernized — refine_prompt (new tool, cheap/fast model): raw query → goal-first brief + what-changed + open questions, never auto-fires or executes. list_prompt_techniques now defaults to the ~9 core techniques (all=true for the full 31); preview_prompt_technique gained technique="auto" recommendations and its tool param is now optional. In Claude Code, /prompt refine presents open questions as clickable choices
Setup, de-mystified — tachibot init (new CLI wizard): detects your API keys and clients, prints the exact config for Claude Code and Claude Desktop, never writes or echoes keys. One-click Claude Desktop install via the .mcpb release asset — no JSON editing
3 new skills — /setup (guided config via doctor), /spec (request → spec), /triage (ranked bug triage). 17 skills total
Fixes — focus orchestration output cut from 37 lines of scaffolding to 10 focused lines; GPT-5.5 high-effort reasoning no longer cut off at 3 minutes (timeout 180s → 600s); npm test exits 0 again
Profile counts updated — minimal 13, code_focus 42, research_power 35, balanced 53, heavy_coding 57, full 64 (default profile)

v2.25.0July 2, 2026New Tools

63 Tools — debug_triage & spec_writer

debug_triage (Grok 4.3) — ranked root-cause hypotheses with likelihoods, the cheapest discriminating check for each, and the minimal fix for the top candidate. Narrows a hypothesis space instead of jumping to one answer
spec_writer (GPT-5.5) — turns a loose feature request into a reviewable spec: user stories, Given/When/Then acceptance criteria, out-of-scope, open questions. For sign-off before planning — feed the approved spec to planner_maker
tachibot init CLI lands (setup wizard: key/client detection, per-client config emission) — expanded further in v2.26.0

v2.24.0July 1, 2026Major Expansion

61 Tools — Security Review, Test Gen, Diff & Plan Red-Team

Four new gap tools — testgen (runnable tests via Qwen3-Coder-Next, edge cases enumerated first), security_review (OWASP/CWE audit via DeepSeek V4 Pro — taint/data-flow analysis, severity + fix per finding), diff_review (multi-model diff-aware review — Kimi K2.7-Code + DeepSeek V4 Pro + GPT-5.5, deduplicated and severity-ranked by a Gemini judge), plan_critique (adversarial plan red-team — pre-mortem, hidden assumptions, ranked risks, Gemini-judged verdict)
doctor now documented — zero-cost setup diagnostic added in an earlier release, only now reflected here: detected API keys, visible vs hidden tools (and why), active profile + source, a concrete first command to try
2 new skills — /review (wraps diff_review on your current git diff), /redteam (wraps plan_critique on any plan). 14 skills total
Profile counts updated — minimal 13, code_focus 39, research_power 35, balanced 50, heavy_coding 54, full 61 (default profile)

v2.23.3June 17, 2026Model Bump

GLM-5.2 — Long-Horizon Coding, 1M Context

GLM-5.1 → GLM-5.2 — z-ai/glm-5.2 (released June 13). Zhipu's open-weights flagship for long-horizon tasks — usable 1M-token context (up from agentic-tier 5.1), two thinking-effort levels (High/Max)
Benchmarks up — SWE-Bench Pro 62.1 (from 58.4), Terminal-Bench 2.1 81.0 (from 62.0); beats GPT-5.5 on several long-horizon coding benchmarks at ~1/6 the cost. List price $1.40/$4.40 per M
Powers glm_reason and the glm juror. GLM-5.1 kept as quota fallback (GLM-5.2 → GLM-5.1 → GLM-5)

v2.23.2June 15, 2026Model Bump

Kimi K2.7-Code — Coding-Specialized, 262K Context

Kimi K2.6 → K2.7-Code — moonshotai/kimi-k2.7-code (released June 12). Coding-specialized variant built on K2.6 — +21.8% on Kimi Code Bench v2 with lower token use, 262K context, native multimodal, always-thinking mode
Now powers all Kimi tools — kimi_thinking, kimi_code, kimi_decompose, kimi_long_context and the kimi juror. List price $0.75/$3.50 per M
Quota fallback — K2.6 retained (K2.7-Code → K2.6 → K2-Thinking). No general-purpose K2.7 exists yet — the coding-tuned model serves reasoning + long-context too

v2.23.1June 11, 2026Model Bump

MiniMax M3 — 1M Context, Sparse Attention

MiniMax M2.7 → M3 — minimax/minimax-m3 (released May 31). 1M-token context (up from 200K), native multimodal, tuned for long-horizon agentic work
MSA sparse attention — ~1/20 the compute of the previous generation at 1M context, with faster prefill and decode
Same promo pricing — $0.30/$1.20 per M tokens, unchanged from M2.7. Powers minimax_code, minimax_agent, and the minimax juror
Quota fallback — M2.7 retained as automatic fallback if M3 hits provider limits

v2.23.0June 11, 2026Refactor

Central Tool Registry + One Honest Local Juror

Central scan registry — all provider tools now register through registry.ts + a defineModelTool factory, with golden wire-contract tests locking all 57 tool schemas. Zero behavior change
Removed the hermes juror — it was a persona prompt on the same local weights as local. Jury independence comes from different model weights, not different system prompts. hermes stays as a deduped legacy alias; 12 jurors total

v2.22.0June 10, 2026Major Expansion

57 Tools — DeepSeek, GLM, StepFun, ERNIE + Free Local Models

Four new reasoning providers — deepseek_reason + deepseek_algo (DeepSeek V4 Pro, top AIME/CodeElo), glm_reason (GLM-5.1, SWE-Bench Pro leader), stepfun_reason (Step 3.7 Flash), ernie_reason (ERNIE 4.5 VL)
Free local models — local_query talks to Ollama, LM Studio, llama.cpp, or vLLM. Zero-cost, offline, private; Ollama uses the native API so long contexts work
12-juror panel — jury adds deepseek, glm, stepfun, ernie, and a free offline local juror (runs whatever LOCAL_LLM_MODEL points at, e.g. a Nous Hermes build via Ollama). New lab-diverse default: grok,deepseek,kimi,openai
3 new skills — /lens (256K long-context analysis), /reflect (grounded reflexion), /tot (Tree-of-Thought with jury pruning). 12 skills total
Grok 4.3 default (v2.21.4) — 1M context, configurable reasoning effort, cheaper than 4.20. Kimi K2.6 hotfix (v2.21.5) — all call sites off the retired K2.5

v2.21.3May 29, 2026Model Bumps

Gemini 3.5 Flash — New Flash + Search Tier

Gemini 3.5 Flash is GA — gemini-3.5-flash shipped at Google I/O (May 19). Agentic/coding focus, 1M context, $1.50/$9 per M. SWE-bench Verified 78.8%
Now the Flash + search tier — gemini_search grounding and the flash variant now route to 3.5 Flash
Reasoning default unchanged — stays on gemini-3.1-pro-preview. Gemini 3.5 Pro is not out yet (expected June 2026)

v2.21.1April 26, 2026Model Bumps

GPT-5.5 + Kimi K2.6 — Frontier Refresh

OpenAI gpt-5.4 → gpt-5.5 — agentic-focused, 1.1M context, omnimodal. Released April 23. gpt-5.5-pro for premium tier ($30/$180)
Kimi K2.5 → K2.6 — 1T MoE, leads SWE-bench Pro for long-horizon coding. Released April 20. Auto-fallback to K2.5 on quota errors
Qwen3.6-Plus available — qwen/qwen3.6-plus registered ($0.325/$1.95). Coder stays on qwen3-coder-next until 3.6-coder ships
API-verified — all model IDs confirmed live via OpenAI /v1/models, OpenRouter, and xAI before release
Holding — Grok 5 unreleased (Q2 expected); Gemini 3.5 in preview (May GA at I/O); gpt-5.4-mini retained for code/explain (no 5.5-mini yet)

v2.21.0April 13, 2026UX Fix

Auto-Alias Param Names — LLMs Stop Failing on Wrong Keys

Transparent remap — query ↔ problem ↔ prompt ↔ question ↔ topic auto-mapped before Zod validation
Zero per-tool changes — one z.preprocess() hook in safeAddTool. Every tool benefits automatically
Eliminates -32602 errors — LLMs that reach for the wrong synonym now succeed instead of hard-failing on schema validation
11 unit tests covering directional aliasing, primary-wins, and missing-key behavior

v2.20.0April 10, 2026Model Bump

Grok 4.20 Flagship — Lower Hallucination, 16-Agent Swarm

Grok 4 → 4.20 — all defaults moved to flagship: grok-4.20-0309-reasoning for reason/search (low hallucination, 2M ctx)
Multi-agent architect — grok_architect now runs grok-4.20-multi-agent-0309 (4–16 parallel agents)
Code/debug/brainstorm — non-reasoning variant for fast turn-around on lighter tasks
Timeout fixes — AbortController on OpenAI (90s default, 180s high-reasoning) + Grok (60–180s by mode). Stops hung calls dead
Cleanup — killed stale grok-4-0709 + gpt-4-mini refs across 6 scattered files

v2.19.1March 21, 2026

MiniMax M2.7 — Self-Evolving AI

MiniMax M2.5 → M2.7 — 2,300B MoE (100B active), 200K context. #1 on Artificial Analysis Intelligence Index
SWE-Pro 56.22% — matches GPT-5.3-Codex. Multi-SWE-Bench 52.7% (#1, beats Opus 4.6 and GPT-5.4)
Same pricing — $0.30/$1.20 per M tokens. Massive quality leap at zero extra cost

v2.18.0March 21, 2026Major Release

AI That Proves Its Work

Stop babysitting LLMs. Deploy a pipeline that reads actual files, cross-examines across five models, and demands passing tests before moving forward.

Absolute goal alignment — define success criteria once. The engine verifies every step against your exact goals — drift gets caught at step 1, not step 50
No blind spots reach production — 5-model rotation cross-examines code: Gemini deduces, Grok detects drift, GPT validates strategy, Qwen cross-checks, Kimi decomposes
Hard evidence, not hallucinated progress — checkpoints demand raw git diffs, passing test results, and modified file lists. Zero reliance on paraphrased summaries
Never hit a dead end — structured amendment protocol detects drift, proposes revisions with evidence and impact analysis. You approve before it pivots
39 tools operate in reality — every analysis tool reads actual source code from disk via the files parameter. Models judge implementations, not stories about them
Your project gets smarter every run — post-completion reflexion saves architectural lessons to your devlog. Knowledge compounds across sessions

v2.17.2March 21, 2026

Files Parameter Rollout + Smart File Reader

files on 8 more tools — grok_architect, grok_brainstorm, openai_explain, openai_search, kimi_code, kimi_long_context, gemini_judge, gemini_brainstorm
Directory expansion — pass src/tools/ to read all code files in a directory
Smart char budget — multi-file reads distribute tokens across files to prevent context overflow
23 of 37 tools now support the files parameter

v2.17.1March 21, 2026

Smart Task Decomposition

kimi_decompose readability overhaul — output now uses OVERVIEW / STRUCTURE / DETAILS / RISKS sections
Smart decomposition — infers context, constraints, risks, and measurable criteria automatically
Reasoning leak fixed — strips Kimi K2.5 chain-of-thought from output
Tuned for format adherence — temp 0.3, 4500 tokens, 360s timeout

v2.17.0March 21, 2026

GPT-5.4-mini + Model Cleanup

GPT-5.4-mini — new fast coding model (400k context, $0.75/$4.50 per 1M tokens, SWE-Bench 54.4%)
Code tasks upgraded — openai_code_review now uses gpt-5.4-mini — 94% of flagship quality, 70% cheaper
GPT-5.3 series retired — gpt-5.3-codex and gpt-5.3 removed; coding capabilities absorbed into gpt-5.4
Simplified lineup — gpt-5.4 (flagship), gpt-5.4-mini (coding/fast), gpt-5.4-pro (expert)

v2.16.1March 6, 2026

Gemini 3.1 Pro Migration

Gemini 3.1 Pro — migrated from gemini-3-pro-preview to gemini-3.1-pro-preview before March 9 retirement
1M context window — enhanced reasoning capabilities with Gemini 3.1 Pro
Stale entries removed — cleaned up old display names and pricing for retired model

v2.16.0March 6, 2026

GPT-5.4 Upgrade + Brainstorm Fix

GPT-5.4 default — most capable model (Mar 2026), $2.50/$15 per 1M tokens
GPT-5.4-pro — expert model with higher compute ($30/$180 per 1M tokens)
GPT-5.3-codex — new agentic coding model for code review tasks
Gemini 3.1 Flash-Lite — added as fastest/cheapest option in 3.1 series
openai_brainstorm fixed — eliminated fragile duplicate API function; now uses shared retry/fallback logic
Token limits bumped — GPT-5.4 reasoning tokens eat into output limit; all OpenAI tools now have higher defaults

v2.15.6February 26, 2026

Full Audit: 6 More Fixes + Cost Optimization

6 tools had required enum anti-pattern — usage_stats, openrouter_multi, gemini_judge, planner_maker, planner_runner, create_workflow. All fixed
gemini_judge — had zero required params. Made perspectives required as primary content param
perplexity_reason downgraded — sonar-pro ($3/$15/M) → sonar-reasoning ($1/$5/M), 3x cheaper
perplexity_research removed — sonar-deep-research ($5/$25/M) was burning $12 in 3 days
All 51 tools audited — zero remaining required enum violations

v2.15.5February 26, 2026

Tool Parameter Fixes + Gemini Stability

Fixed parameter validation on qwen_coder, kimi_code, minimax_code — AI clients were misusing required enum task param. Added query as required primary param, made task optional with defaults
kimi_long_context — task enum now optional (default: analyze)
Gemini 3.1 → 3.0 rollback — Reverted to stable gemini-3-pro-preview (3.1 had timeout/503 issues)
Gemini timeout 30s → 90s — Pro models need longer than Flash

v2.15.0February 12, 2026

31 Prompt Techniques + /blueprint Skill + MiniMax M2.5

9 new prompt techniques — reflexion (Shinn 2023), react (Yao 2022), scot (Li 2025, +13.79% HumanEval), pre_mortem, rubber_duck, test_driven, pre_post, bdd_spec, least_to_most. Total: 31 techniques
/blueprint skill — Multi-model council → bite-sized TDD implementation plans. 7-step pipeline: Grok search → Qwen+Kimi analysis → GPT pre-mortem → Gemini final TDD output
MiniMax M2.5 — SWE-Bench 80.2% (was 72.5%). Embedded SCoT, reflexion, rubber_duck techniques. Per-task temperatures
Planner → writing-plans bridge — planner_maker now outputs bite-sized TDD steps (exact files, test-first, commit points)
Enhanced skills — /breakdown uses least_to_most + pre_mortem, /judge adds pre-mortem to critique, /decompose adds contracts, /prompt auto-recommends from 30 intents
51 tools across 7 providers, 9 skills for Claude Code

v2.14.7February 5, 2026

Gemini Judge + Multi-Model Jury

gemini_judge — Science-backed LLM-as-a-Judge evaluation (arXiv:2411.15594). 4 modes: synthesize, evaluate, rank, resolve
jury — Multi-model jury panel. Configurable jurors (grok, openai, qwen, kimi, perplexity, minimax) run in parallel, Gemini synthesizes verdict. Based on "Replacing Judges with Juries" (Cohere, arXiv:2404.18796)
Perplexity fix — sonar-pro model ID corrected (was using lightweight sonar by mistake)
perplexity_research — Removed in v2.15.6 (cost too high)
51 tools across 7 providers in the full profile

v2.14.6February 5, 2026

Qwen3-Coder-Next

qwen_coder upgraded — Qwen3-Coder-Next (80B/3B MoE, 262K context, SWE-Bench >70%)
3x cheaper — $0.07/$0.30 per M tokens (was $0.22/$0.88)
2x context — 262K tokens (was 131K)
Auto-fallback — Falls back to legacy 480B coder on provider failure

v2.14.5February 2, 2026

Claude Code Integration

Tool annotations — All 51 tools now have MCP-standard annotations for better discovery
Token overhead reduced — Stripped ANSI formatting, clean plain text output
25K character safety net — Smart truncation prevents Claude Code context overflow

v2.10January 28, 2026

Multi-Model Planner

Multi-model council creates verified implementation plans
Model roles — Grok searches ground truth, Qwen analyzes feasibility, GPT-5.2 critiques, Gemini scores quality
New models — Kimi K2.5 (multimodal + agent swarm), MiniMax M2.5 (SWE-Bench 80.2%)
New tools — qwen_reason, minimax_code, minimax_agent, gemini_search
Smart routing — Tool routing based on availability, cost, and quality

v2.8January 20, 2026

Prompt Techniques

FocusExecutionService for clean mode orchestration
22 research-backed techniques — first_principles, tree_of_thoughts, council_of_experts, and more
Preview before execute — See enhanced prompts before running them
Heartbeat support for long-running operations

v2.7.9January 2, 2026

Search Grounding

qwen_algo — O(1)-first algorithm analysis with Qwen3-235B-Thinking (235B MoE, LiveCodeBench 91.4)
gemini_search — Google Search grounding with dynamic retrieval
Format utilities for consistent output across tools

v2.3December 28, 2025

Enhanced Thinking

nextThought with finalJudge — Auto-call judge model when session completes
Context aliases — Use "none", "recent", "all" instead of magic numbers
Context distillation — Compress 8000+ tokens to ~500 (5x savings)
usage_stats tool for tracking tool usage and costs

v2.1November 25, 2025

Gateway Mode

OpenRouter Gateway — One API key for all models
Unified billing through OpenRouter

v2.0October 15, 2025

Major Rewrite

Multi-model orchestration rebuilt from scratch
Tool profiles for context control
YAML workflow engine with variable interpolation
6 AI providers, 31+ tools (now 51 tools across 7 providers)

Explore More

Dive deeper into TachiBot's capabilities

Documentation

Back to the main docs

Tools Overview

See all 51 available tools

Examples

Practical usage patterns and workflows

Loading documentation...