v0.6.0

0.6.0

Minor Changes

• #26 3c80dd7 - Added mode: 'auto': the question runs System 1 first and escalates exactly once to System 2 when the answer's confidence falls below autoModeThreshold (default 0.7; for noul() the confidence equivalent is max(noul, 1 − noul)). The escalated answer is returned as-is even at low confidence — one deliberate pass, never a loop. Confident System 1 answers complete without touching System 2, so the fast path stays in the milliseconds.

• #26 3c80dd7 - Added global debug logging: debug: true on any question (both modes) logs the internals to stderr — the prompt, the wire request body, the endpoint, raw responses with their HTTP status, every retry with its reason and backoff, and in System 2 the model's reasoning text and each structured-rejection reason.

• #26 3c80dd7 - Implement System 2 mode (mode: 'system2') for choice(), score() and noul(): the model deliberates and answers a minimal structured object — one integer 0..10 rating per candidate — and the same answer shapes as System 1 come out of it (winner = argmax, distribution = normalized ratings, same entropy confidence). Requests use response_format with a strict ratings schema, temperature: 0, no default token budget, and a parse-validate-retry fallback that feeds rejection reasons back to the model so engines with broken or missing structured-output support still work. Adds the thinking question option (valid in 'system2'/'auto' mode; false sends reasoning_effort: 'none') and the mode-partitioned question types: Question, ScoreQuestion and NoulQuestion switch from interfaces to unions over mode, so thinking set with mode: 'system1' no longer compiles — breaking for consumers extending those types with interface X extends ... (allowed by semver at 0.x).

v0.5.0

0.5.0

Minor Changes

• #17 8c84db1 - Add the noul() primitive: answers a yes/no question with the probability that the answer is "yes" (0..1).

v0.4.0

0.4.0

Minor Changes

• #15 d88ce52 Thanks @expilu! - feat: score() — rate a position on a spectrum

  New score() primitive for decisions that are a position on a scale of ordered,
  described levels: its answer is a score (the probability-weighted mean of the
  level numbers, so it can fall between two levels), the per-level probabilities
  and legend keyed by level number as string, and the usual 0..1 confidence.
  Accepts 2..10 level descriptions in criteria (an ordered array, low end of the
  scale first); one dimension per question is recommended. Not implemented yet in System 2 mode.

• #15 d88ce52 Thanks @expilu! - feat: shared System 1 core, shared mode router and the Mode type

  choice() System 1 answers are now produced through src/system1/system1.ts (the
  one-token logprobs engine any System 1 question type reuses) and src/utils/mode/route-mode.ts
  (the mode router: default to System 1, System 2 answers "Not implemented yet" from a single
  stub). New exported types: Mode (the shared 'system1' | 'system2' union) and BaseQuestion
  (the fields every question type carries — Question extends it). No behavior change:
  same prompts, same responses, same errors.

v0.3.0

0.3.0

Minor Changes

• #12 87127d5 Thanks @expilu! - Breaking (pre-1.0): the provider settings on Question moved into a new Model type.

  • apiBaseUrl, apiKey and model no longer sit flat on Question; they are now
    question.model.{apiBaseUrl, apiKey, model}. This groups everything about how the
    model is reached and served into one reusable object to pass to every choice().
  • Model gains extraBody: extra fields forwarded verbatim into the chat completions
    request body for engine- or model-specific settings (i.e. chat_template_kwargs
    thinking toggles on llama.cpp/vLLM/SGLang, reasoning_effort on OpenAI/OpenRouter/
    Ollama, Ollama's native think). Reserved request keys the library's single-token
    trick depends on (model, messages, stream, logprobs, top_logprobs,
    max_tokens, temperature) cannot be overridden through it; chat_template_kwargs
    merges one level deep with user keys winning per key.
  • System 1's request now asks for top_logprobs: 20 instead of 50: 20 is the highest
    portable window (OpenAI and OpenRouter cap it at 20, vLLM's server default
    --max-logprobs is 20). On llama.cpp (accepts up to 50) the smaller window is
    enough for realistic option counts.
  • New Model type exported from the package root; generateText accepts extra
    top-level body fields via ChatCompletionRequest's index signature.

  Hardening

  • The transport no longer follows redirects (redirect: 'error'): the Bearer token
    stays off unexpected paths, and a misconfigured base URL fails loudly.
  • Response bodies are read under a 10 MB safety cap instead of being buffered
    unconditionally; an over-cap body fails fast and is not retried. A declared
    Content-Length over the cap is refused before reading anything.
  • Non-finite maxRetries (i.e. NaN) no longer causes an infinite retry loop: it
    falls back to the default, negatives clamp to 0 and fractions to whole attempts.
  • extraBody ignores __proto__ and constructor keys, so prototype-smuggled
    config can never re-parent the request object.
  • Malformed entries inside top_logprobs (missing/null token) are skipped instead
    of crashing mid-read.
  • Error messages strip query strings from the request URL, so providers that take
    credentials as query parameters cannot leak them into logs.
  • The endpoint path is joined through the URL API, preserving a query string in
    apiBaseUrl instead of swallowing it, and an invalid base URL throws a clear
    Invalid model.apiBaseUrl error.
  • CI actions are pinned by commit SHA; the release script spawns every subprocess
    as an argv array (no shell), so interpolated values can never be re-parsed as
    shell syntax.

v0.2.0

0.2.0

Minor Changes

• #6 0cce138 Thanks @expilu! - Removed the openai SDK dependency: the library now talks to the OpenAI-compatible
  endpoint with an in-library HTTP transport.