What's New

v2.33.0

• The writing family (storm, re3, de_generic). The library had creative thinking (what_if, innovate, alt_view, creative_use) and zero creative writing: against a 7-stage pipeline, stage 4 (structure) and stage 5 (draft) were empty, and stage 6 (revise) fell through to reflexion — which optimises toward correct, while prose revision is cutting, rhythm and specificity. 85 → 88 techniques.
• blog_writer — researched long-form in one call, composing storm and re3 rather than carrying a private copy of their text. A copy would drift the moment either template was edited, and drift silently, because both paths still emit well-formed prose. research: true runs the persona interview first and feeds its 3 surprising answers into the four-pass write. 66 → 67 tools; enabled in full, balanced and research_power.
• preview_prompt_technique(technique="auto") had no route for writing intent. RECOMMEND_RULES carried no matcher, so "write a blog post" fell through to CORE_TECHNIQUES.slice(0, 3) — chain_of_note / astute_rag / spotlighting — while list_prompt_techniques advertised auto as the way in. Writing verbs now route to storm, re3, de_generic.
• blog_writer truncated silently at maxTokens: 8000. Its output is several full renderings of the same piece (re3's draft, restructure and line-edit passes, plus storm's 16 marked answers), unlike sibling tools that emit one artifact. 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.
• Anti-drift tests could not detect drift. The "uses the template rather than a private copy" tests asserted characteristic phrases that a copy would also satisfy; they now assert the engine's exact live output.

v2.32.0

• Tabular family (7 new techniques) — data-shaped work had no coverage at all, so "filter these rows" fell through to the generic core contracts. data_assert states the downstream purpose first, then checks the data with executable assertions carrying counts, because a filter that silently matched zero rows is indistinguishable from one that worked. Plus chain_of_table, sub_table_first, sql_stages, xlsx_map, table_format, self_debug.
• Retrieval family completed (3 new) — astute_rag answers from memory first, then marks each source agrees/fills/conflicts and resolves every conflict out loud; crag grades sources correct/ambiguous/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.
• Spotlighting fences carried 24 bits of entropy, not 32. untrustedFence() called randomBytes(3) 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. Now 8 hex characters, as documented.
• The literal 37 survived three expansions of the catalogue (37 → 74 → 85) inside the list_prompt_techniques description — the text a model reads to decide whether to call the tool at all. It is now derived from FLAT_TECHNIQUES.length.
• 74 → 85 techniques. Tool count unchanged at 66.

v2.31.0

• Prompt technique library goes 37 → 74, closing three families that were at zero coverage. Retrieval (4): TachiBot ships five search tools and had no discipline for what to do with what they return — retrieved text went straight into reasoning unfiltered, which is where confidently-miscited sources come from. Prompt-injection defence (1): those same tools ingest untrusted web text with no isolation contract, so a page saying "ignore previous instructions" was read exactly like our own prompt; spotlighting fences it and states the rule. Few-shot (5): nothing touched example selection or ordering, despite this moving accuracy more than most reasoning tricks.
• Eight trace-management techniques from Jun–Aug 2026 research, including hourglass (compress to a rule, discard the exploration, regenerate from the rule instead of appending errors) and backtrack (rewind to the last validated checkpoint rather than restarting or patching).
• Twelve original prompts rewritten to read as instructions. The early set used telegraphic shorthand that reads as a label rather than a request, leaving the model to infer the task.
• Recommendation rules for all of it — 74 techniques is past the point where anyone recalls names, so RECOMMEND_RULES gained eight groups matched to how people actually phrase the situation. Without these the new techniques were reachable only by knowing their name in advance.

v2.30.1

• Tool descriptions no longer name-drop model versions. 19 descriptions were rewritten to name the provider and the task instead of pinned strings like "GPT-5.2", "Kimi K3" or "424B MoE" — those silently rot on every model bump and then actively mislead the routing model. Two were outright wrong: openai_reason and openai_search advertised "GPT-5.2" while the code calls the current OpenAI flagship tier.
• grok_search finally says why to pick it. Its entire description was "Web search" — no way to choose it over the four other search tools. It now states its live X/news grounding edge and cross-references grok_search_lite as the cheaper path. The five OpenRouter reasoners (deepseek_reason, glm_reason, stepfun_reason, ernie_reason, qwen_reason) likewise differed only by vendor trivia; each now carries an actual routing rule for when to pick it.
• Claude Desktop manifest corrected — it advertised 65 tools (actual: 66), omitted the heavy_coding profile from the TACHIBOT_PROFILE help, and its OPENAI_API_KEY hint claimed that key powers the Qwen and QwQ tools (those route via OpenRouter).
• Packaging no longer breaks the working tree. scripts/package-extension.sh ran npm install --production, pruning devDependencies and leaving tsc unable to build afterward. It now restores the full dependency tree when it finishes.
• The server introduced itself as v2.0.0. serverInfo.version was hardcoded from a 12-tool era ~28 minor versions ago, so Claude Desktop's connector panel misidentified every install. It now reads the real version from package.json. The /setup wizard's profile sizes were stale in the same way and now match the six real profiles.
• Correction to v2.30.0: grok_search_lite is not broken, and is staying. That release claimed grok-4.3 "does not invoke web search either" — which would have made the cheap search tier pointless. Re-probed Aug 15 asking today's date: grok-4.3 ran 2 web searches and returned a citation in 10.4s; grok-4.6 did the same in 36.4s. Lite is grounded, cheaper, and 3–4x faster — prefer it for high-volume lookups. grok-4.5 is still the one that doesn't ground. The cheap tier now has its own grounding test, since nothing previously covered it.
• No tools added or removed — the count stays 66.