Project and packaging updates

0.4.1 - 2026-09-25

Project and packaging updates; no change to the library's behavior.

Changed

• Package metadata: development status Beta, more keywords, and links to
  Discussions, the decision benchmark leaderboard, release notes and funding.
• The README links to good first issues and Discussions and shows the docs,
  leaderboard and OpenSSF Scorecard badges.
• The submission guide says submitters regenerate the leaderboard page in
  their pull request; the release guide explains that release tags are
  permanent.

Security

• Every GitHub Action in the workflows is pinned to a commit.
• An OpenSSF Scorecard workflow checks the repository's supply chain weekly.

Decision Benchmarks

0.4.0 - 2026-09-25

The decision benchmark.

Added

• The decision benchmark 1.0 (thinkless bench decisions): eight public
  tasks across the four question kinds (banking77, clinc150 with out of
  scope, MASSIVE in five languages, MultiWOZ 2.2 conversations, jailbreaks,
  Civil Comments toxicity, HelpSteer2 helpfulness ratings and WNUT 2017
  entities), with frozen rows and published hashes, calibration rows from a
  separate split, result files that hold every prediction, verify that
  recomputes every metric, a generated leaderboard and open submissions.
  Baselines for GLiNER 2.5, Laya, Qwen3-1.7B and Qwen 3.7 Flash.

Changed

• A list of chat messages now renders as role: text lines for providers
  that read text, instead of one key: value line per field.
• README links to the documentation point at the documentation site.

Shadow mode, framework adapters, limits, a decision server and MCP tools.

0.3.0 - 2026-09-25

Shadow mode, framework adapters, limits, a decision server and MCP tools.

Added

• Shadow mode (thinkless.shadow): run a candidate engine next to existing
  decision code without changing what it returns, log both answers, and
  report agreement with a 95% interval, the share settled without an LLM,
  cost and latency on both sides, thresholds fitted on production traffic,
  and a verdict per question (thinkless shadow report). Disagreements
  export as labeling rows (thinkless shadow export). The same class audits
  a live engine against an LLM.
• Framework adapters (thinkless.integrations): a LangGraph router, decision
  node and async decision node; OpenAI Agents SDK input guardrails, tool input
  guardrails and route_agent; and framework-neutral Router, route and
  gate, tested with LangChain tools and Pydantic AI. Extras langgraph and
  openai-agents.
• Engine deadlines (deadline_ms on the engine and per call), spend limits
  (SpendLimit, lifetime or rolling window) and per-run caps
  (engine.run(..., max_cost_usd=...)). Skipped providers are recorded with
  reason deadline or spend_limit.
• A decision server (thinkless serve, thinkless.server.app.create_app)
  with /v1/decide and a System One compatible /v1/systemone, so another
  engine's SystemOne provider, or TypeSafe's SDK, can use it. Extra
  server.
• MCP tools (thinkless mcp, thinkless.server.mcp.create_mcp_server): one
  decide_<question> tool per registered question. Extra mcp.
• thinkless trace export turns traced decisions into labeling rows, and
  thinkless trace drift compares two periods and exits with 1 when a
  question drifted.
• question_from_spec rebuilds a question from its spec.
• Docs: an integrations section, a migration guide, shadow mode, serving,
  a pilot playbook and a FAQ, published to GitHub Pages by a new workflow.

ThinkLess 0.2.0

0.2.0 - 2026-09-25

First release on PyPI.

Added

• OpenRouterLLM and the openrouter:<slug> spec: one key for models from
  Anthropic, OpenAI, Google, Qwen, DeepSeek and others.
• Billed cost: Completion.cost_usd and ProviderResult.cost_usd carry the
  cost a backend reports (OpenRouter does); the engine prefers it over the
  price table and records cost_source on spans. Benchmark reports add a
  billed cost row, and the intent benchmark reports cost per 1k per provider
  and per cascade threshold.
• --reasoning for every CLI command that builds an LLM, and
  from_spec(..., reasoning=...), mapped to OpenRouter's reasoning,
  Anthropic's effort and local thinking switches.
• HFClassifier: any Hugging Face text-classification model as a provider for
  the questions it was trained for.
• load_env(), and the CLI reads ./.env without overriding set variables.
• Escalation context: when a question escalates to the LLM decider, the
  engine includes the decisions already settled in the same batch. It removed
  false injection flags on requests for a human that appeared when an LLM saw
  the injection question alone. On by default (Engine(escalation_context=...)).
• The System One wire format is tested against TypeSafe's official SDK models.

Fixed

• LLMDecider detects replies cut off at the token limit, logs why, and
  retries once with a larger budget instead of abstaining silently.
• The OpenAI-compatible client retries once without response_format when an
  endpoint rejects JSON mode.
• A redaction test that failed about one run in fifty when a random span id
  contained its marker.

0.1.0 - 2026-09-25

First public release.

Added

• Typed questions: Choice, Score, YesNo and Extract, following the
  System One wire format.
• Engine with a confidence cascade across providers, batched provider calls,
  accepted, uncertain and abstained statuses, per-question and
  per-provider thresholds ("question@provider"), question-level provider
  allowlists, and async variants of every call.
• Normalized confidence, (k * p_max - 1) / (k - 1), applied uniformly across
  providers.
• Decision providers: Rules, GLiNER (GLiNER 2.5), Laya, SystemOne
  (TypeSafe Jev and compatible servers such as Kev and OpenJev) and
  LLMDecider.
• LLM backends: TransformersLLM, OpenAICompatibleLLM (OpenAI, Ollama, vLLM
  and others) and AnthropicLLM with server-side refusal fallbacks.
• Tracing: spans for runs, steps, decisions, provider attempts, generations,
  tools and rules; JSONL, memory, console and OpenTelemetry sinks; run
  summaries with decisions by plane, escalations, tokens, cost and time by
  plane; content capture switch; a self-contained HTML viewer.
• A price table with sources, overridable with THINKLESS_PRICING.
• The support demo: an agent, a mock store, 53 labeled tickets and a separate
  48-row calibration set.
• Benchmarks: the support benchmark across llm, hybrid and models modes,
  and intent benchmarks on Banking77, CLINC150 and Emotion with a simulated
  cascade.
• thinkless CLI: doctor, demo, bench support, bench intents,
  calibrate, trace ls, trace show and trace view.