v0.13.0

[0.13.0] — 2026-07-02

Receipts & first-touch. Cross-tier native thinking, cost-aware model routing, an overhauled first-ten-minutes developer experience (typed tool authoring, fail-fast build(), ReactiveAgents.quick()), honest abstention as a first-class terminal, a new efficiency-first Blueprint strategy, two token-waste guards (StallPolicy, FabricationGuard), and a broad sweep of correctness fixes across providers and the kernel.

Added

• .withThinking(...) — native thinking on every provider, opt-in. One builder switch (true or { effort, budgetTokens }) enables native reasoning across Anthropic, OpenAI, Gemini, and local models. Off by default everywhere: undefined never auto-enables. Budgets are bounded and reserved on top of the answer budget so thinking can never starve the visible answer.
• .withModelRouting(options?) — cost-aware model routing, opt-in. Routes each run to the cheapest capable model of the configured provider by task complexity, on both the inline and reasoning paths. Provider-agnostic, capability-gated (never routes below a model whose context window fits), and advisory — it degrades to the configured model on any error.
• Tool authoring v2 — schema in, typed args out. defineTool({ name, description, input, handler }) accepts a Standard Schema input (Zod, Effect, Valibot) and gives the plain-async handler inferred argument types — no more Record<string, unknown> or as never casts. defineTool also validates its options and rejects wrong field names (e.g. parameters/execute) with a message naming the correct field instead of crashing with TypeError: schema.ast. The Effect-handler form remains as the advanced path.
• Fail-fast build() via .withStrictValidation(). Opt in to catch a missing API key or unknown model at build time with a typed error and fix instructions, instead of a raw 401/404 on the first call.
• ReactiveAgents.quick() — a two-line agent. Resolves provider, model, and iteration defaults from the environment: const agent = await ReactiveAgents.quick(); await agent.run("...").
• .withLlmTimeout(ms) — per-LLM-call timeout for local models. Complements the run-level .withTimeout(); configures the local/Ollama per-call timeout that was previously hardcoded at 120 seconds (hosted providers keep their own limits). Timeout errors now name the model, elapsed time, and a cold-load/GPU-contention hint, and the in-flight local request is aborted server-side so the GPU stops burning after the client gives up.
• Harness-forced abstention — a run that cannot succeed says so. When a task is structurally impossible (a required tool is unavailable, or synthesis is repeatedly ungrounded), the run terminates with terminatedBy: "abstained" and a typed result.abstention { reason, missing } instead of fabricating an answer.
• Blueprint strategy — plan once, execute in parallel, zero extra LLM calls. For static, decomposable tasks the whole plan is knowable up front: Blueprint generates a plan, verifies it, executes independent steps in parallel with no per-step LLM call, then solves. Adaptive routing sends static-decomposable tasks to Blueprint automatically. This is the 7th reasoning strategy.
• .withStallPolicy(...) — fail fast on ignored nudges. When the model ignores required-tool nudges and makes no progress across consecutive iterations, the harness escalates and delivers accumulated artifacts (or fails) instead of looping to the full nudge cap — bounding wasted iterations/tokens on stuck runs while leaving progressing runs untouched. Sensible defaults apply when unset.
• .withFabricationGuard(mode) — reject invented measurements. An always-on verifier check (default "block") that rejects empirical performance numbers (benchmark timings, % speed-ups) absent from the tool-observation corpus; high-precision (only perf measurements are policed). Softenable to "warn" or "off", also via the RA_FABRICATION_GUARD env var.
• Evaluation gate CLI. rax eval gate runs the project lift rule over a benchmark report (default-on | opt-in | reject); --ledger appends a weakness→hypothesis→verdict chain and rax eval ledger reads it. Benchmark runs capture a per-run RunDiagnosis (honesty label, failure modes) when a trace dir is set.
• Scaffold templates. create-reactive-agent adds with-structured-output, with-approval-gates, and with-memory templates on the current API.
• ToolBuilder.create() static factory. A convenience entry point for the tool builder.
• Model coverage. Static capabilities for qwen3:4b and cogito:8b; suffix-less claude-sonnet-4-5 alias; o5-reasoning capability entry.
• Packaging & discoverability. npm keywords + descriptions added to 31 published packages; docs SEO/AEO foundation (JSON-LD, sitemap, AI-crawler robots); docs site gains stability tags, last-updated stamps, and change indicators.

Changed

• Gemini no longer thinks by default. Thinking is uniformly opt-in via .withThinking(); Gemini's former thinks-by-default behavior is off unless enabled. When enabled, its previously unbounded thinking budget is bounded so the answer is never truncated by hidden reasoning.

Fixed

• gpt-5.x models work for normal (non-thinking) calls. The OpenAI adapter chose the token-limit field by thinking state, so gpt-5.x default calls sent max_tokens — which those models reject with a 400. The field is now capability-driven: gpt-5.x / o-series always get max_completion_tokens; the gpt-4o family is unchanged.
• Thinking request shapes verified against live provider APIs. Anthropic uses the adaptive shape on current-generation models and the legacy budget_tokens shape on older ones; temperature is dropped when thinking is enabled (both Anthropic and OpenAI reject it otherwise).
• Provider errors are readable. A model typo now produces one clean error line with a suggestion instead of duplicated raw JSON and an internal stack; the run() boundary maps errors to a plain Error with cause (full internals behind a debug flag).
• withRetryPolicy now retries the path the kernel actually uses. It previously wrapped only complete(); the reactive kernel runs through stream() (and structured output through completeStructured()), so a transient failure during a normal run was never retried despite the configured policy. All three call sites are now retried.
• withMinIterations(N) enforces the full floor. It previously forced only a single extra pass regardless of N (a lone if, not a loop). It now loops to the configured minimum.
• Cross-provider tool-call arguments are never dropped. Tool-call arguments arriving as a JSON string (some Ollama models) were silently dropped to {} at the resolver; they are now coerced, so every adapter delivers parsed arguments.
• withVerificationStep({ mode: "loop" }) removed. The "loop" mode was documented but unimplemented (it skipped verification with a log-only warning). "reflect" is the only supported mode; the false option is gone.
• Thinking-mode models no longer starve their own answer. Gemini (formerly thinking-on by default) consumed the entire output budget with hidden reasoning, truncating the answer; thinking budget is now bounded and reserved on top of the answer budget. Cloud complete() timeout raised 30s→120s for thinking models. A non-OK empty-success guard was ported to Anthropic/OpenAI for parity.
• Blueprint plans over the full tool catalog with parameter semantics, instead of a truncated tool list.
• Configured-off phases stay off. runGuardedPhase now honors phase.skip, so direct callers no longer execute phase bodies that were disabled by configuration.
• Context & structured-output correctness sweep. JSON repair is string-safe ("True Story"/"NaN Industries" survive); the message window keeps mid-thread user instructions instead of silently dropping them over budget; field-provenance uses boundary matching and recurses into nested objects; duplicate tool names warn at sanitization instead of silently dispatching to the wrong tool; streaming partial-JSON reparse is gated (O(N²) removed) and walkback bounded.
• Output-ownership invariant. A run that did real work never ships an empty final answer (terminal synthesis assembles a deliverable when the output is empty but candidates exist).
• Tree-of-Thought degrades gracefully under a wall-clock budget instead of being killed mid-exploration on slow thinking models.
• Parallel-batch tool calls are healed with the same tier-parity pipeline as single calls; the healed flag is corrected.
• Negative cost on Anthropic prompt-cache hits; agent-config serialization drift closed with an AST-driven anti-drift guard; the severed experience-tips loop re-wired; the verifier rejects terminal continuation-intent as a final answer.