mohdel v0.114.0

Added

• Anthropic per-TTL cache-creation accounting. The streaming adapter now reads
  usage.cache_creation.ephemeral_1h_input_tokens and exposes
  cacheWrite1hInputTokens on the done result. computeCost prices the 1h write
  tier via a new optional spec field cacheWrite1hPrice; the 5m portion
  (cacheWriteInputTokens − cacheWrite1hInputTokens) stays on cacheWritePrice.
  A request may mix TTLs, so both tiers can be non-zero.
• Gemini implicit-cache read accounting. The adapter now reads
  usageMetadata.cachedContentTokenCount and exposes cacheReadInputTokens on
  the done result, priced via the existing cacheReadPrice spec field.

Fixed

• Gemini cached input was billed at full inputPrice. Cached tokens are now
  subtracted from inputTokens (subset convention, matching the OpenAI adapter)
  and priced at cacheReadPrice, so reported cost reflects the provider's
  cached-token discount.
• Anthropic 1h cache writes were priced at the 5m rate. They now bill at
  cacheWrite1hPrice when the spec provides it.

Notes

• Additive and backward-compatible. When cacheWrite1hInputTokens is absent or a
  spec omits cacheWrite1hPrice, the whole write bills at cacheWritePrice —
  prior behaviour. Realizing the corrected rates requires catalog specs to carry
  cacheWrite1hPrice (Anthropic) and cacheReadPrice (Gemini); without them the
  new fields are observability-only and cost is unchanged.

mohdel v0.113.0

Added

• Qwen Cloud provider (qwen/*) — OpenAI-compatible chat completions
  against Alibaba's international DashScope endpoint
  (dashscope-intl.aliyuncs.com/compatible-mode/v1), key via
  QWEN_API_SK. Thinking is wired through the new qwen
  reasoning-field variant: outputEffort maps to enable_thinking
  plus a numeric thinking_budget from the spec's
  thinkingEffortLevels; Qwen hybrid models think by default, so
  effort none sends an explicit enable_thinking: false.

mohdel v0.112.0

Added

• POST /v1/transcription on thin-gate's data plane — closes the
  "factory path only" gap from 0.111.0; the cross-process client can
  now transcribe via callTranscription(envelope, { socketPath })
  from mohdel/client. Same one-shot shape as /v1/image: plain JSON
  response, no streaming, no enforcer, op: "transcription" driver
  tag. audio.fileUri is file:// (gate sessions read the path —
  requires a shared filesystem with the caller) or data: (inline,
  subject to the 16 MiB body cap). Rust protocol structs
  (TranscriptionEnvelope / TranscriptionResult), conformance
  fixtures, and an end-to-end session-dispatch test included; the
  image and transcription paths now share one one-shot pool-exchange
  helper in the gate.

mohdel v0.111.0

Added

• Speech-to-text via model.transcribe(audio, options?) — a third call
  primitive alongside answer() and image(). One shared adapter posts
  multipart audio to the OpenAI-compatible /audio/transcriptions
  endpoint; registered for groq, mistral, and openai. Audio comes
  from a file:// or data: URI; result is
  { status, text, language, durationSeconds, cost, timestamps }.
  Factory path only for now — thin-gate has no /v1/transcription route
  yet, so the cross-process client cannot transcribe.
• mo transcribe <model> <audio-file> CLI command — MIME type guessed
  from the extension (--mime to override), --language / --prompt
  hints, --json output, duration/cost summary on stderr.
• Catalog support for transcription entries: type: "transcription",
  transcriptionPrice (USD per audio minute), "audio" in
  inputFormat. Cost is duration × per-minute price when the provider
  reports duration, with a token-pricing fallback for OpenAI's
  gpt-4o-*-transcribe models (computeTranscriptionCost in
  _pricing.js).
• Live transcription smoke tests (test/live/transcription.live.test.js),
  key-gated per provider; the audio fixture is a generated sine WAV.

Changed

• README repositioned around the gateway wedge new "How it compares"
  section (LiteLLM, Vercel AI SDK, OpenRouter, raw SDKs).
• package.json description rewritten in searcher vocabulary:
  LiteLLM-style unified API, provider names spelled out, per-call USD
  cost tracking, speech-to-text.

mohdel v0.110.0

Fixed

• toEnvelopePrompt now accepts the tool role in addition to tool_result
  in the factory { system?, messages } prompt shape, carrying toolCallId
  and toolName onto the envelope.

mohdel v0.109.0

Changed

• fireworks provider definition moved into alphabetical position in
  providers.js (between cerebras and gemini). No behavioral change —
  the provider config is byte-for-byte identical, only its source order
  changed.
• @anthropic-ai/sdk ^0.98.0 → ^0.100.1
• @google/genai ^2.6.0 → ^2.8.0
• groq-sdk ^1.2.0 → ^1.2.1
• openai ^6.39.0 → ^6.42.0
• @clack/prompts ^1.4.0 → ^1.5.1 (optional)
• lint-staged ^17.0.5 → ^17.0.7 (dev)
• release-it ^20.0.1 → ^20.2.0 (dev)
• vitest ^4.1.7 → ^4.1.8 (dev)

mohdel v0.108.2

The release-it after:bump hook now refreshes both package-lock.json
and Cargo.lock after the version bump, so CI's npm ci succeeds and
the working tree stays clean through the cargo build. No runtime changes.

mohdel v0.108.1

No code changes from 0.108.0. The 0.108.0 release was tagged but
never landed on the npm registry — the publish workflow was migrated
to npm trusted publishing (OIDC) mid-release and the original tag was
cut against the legacy token-based workflow. 0.108.1 is the first
version of the 0.108 line actually available on npm.

mohdel v0.108.0

Changed

• Session log lines now carry trace context as a nested
  span: { traceId, spanId, traceFlags } object instead of flat
  traceId / spanId fields at the root, matching the standard
  OpenTelemetry pino convention. The public createLogger API and the
  wire/event protocol are unchanged.
• @anthropic-ai/sdk ^0.96.0 → ^0.98.0
• @google/genai ^2.3.0 → ^2.6.0
• openai ^6.37.0 → ^6.39.0
• lint-staged ^17.0.4 → ^17.0.5 (dev)
• vitest ^4.1.6 → ^4.1.7 (dev)

mohdel v0.107.0

Added

• mo doctor — single-command health check covering config dir, environment
  file, API keys per provider, curated.json parse + schema validation, and
  default-model resolution. Pretty output by default; --json for machine
  consumers. Exits 0 on success (warnings allowed) or 1 on errors.
• JSON Schema for curated.json at config/curated.schema.json. Editors that
  speak JSON Schema (VS Code, JetBrains, Helix, Neovim with coc-json) get
  autocomplete, inline type checking, and hover docs while editing the catalog.
  Two ways to wire it: an inline $schema pointer at the top of the file, or a
  .vscode/settings.json json.schemas mapping.
• config/curated.example.json — five worked entries (minimal, full-featured
  with thinking effort + cache pricing + leaderboard, deprecated stub, image
  generation, custom rate limits) with inline $schema pointer.
• docs/CATALOG.md — full reference for curated.json: required vs
  recommended vs capability fields, image entries, custom-field convention,
  editor support, and the mo curate / mo model add / mo check workflow.
• docs/GLOSSARY.md — short definitions for envelope, thin-gate, session,
  factory, creator vs provider, thinking effort, status, cooldown, and the
  rest of the recurring vocabulary.
• docs/COOKBOOK.md — six copy-paste recipes (summarize a file, stream to
  terminal, swap providers via env var, tool-use round trip, vision, batch
  with cost totals).
• --help for mo model add and mo curate (and the mo curate alias),
  each with concrete examples and pointers to the catalog docs.
• Friendly next-step hints in mo ask errors: model not in catalog →
  mo curate <provider> / mo model add <id>; API-key missing or rejected →
  mo setup <provider>; broken deprecation chain → mo check. Hints are
  pure pattern-match on err.message/err.detail, so the lib layer stays
  neutral.
• Top-level meta-key support in curated.json: keys starting with $ (e.g.
  $schema) or _ (e.g. _comment) are preserved on load/save but excluded
  from every iteration site (alias map, suggestion search, rank index,
  pickers, validation). Lets users embed a JSON Schema pointer or inline
  notes in the catalog without polluting downstream behaviour.
• New src/lib/common.js helpers: isMetaKey, catalogEntries,
  catalogKeys, catalogValues — used internally by every catalog
  iteration site so meta-key filtering stays consistent.

Changed

• README rewrite for clarity.
• Returning-user mo menu (when at least one provider is already
  configured) leads with mo ask <model> "...", surfaces mo doctor, and
  includes mo provider setup <p> for adding a second key.
• @anthropic-ai/sdk ^0.95.2 → ^0.96.0
• @google/genai ^2.2.0 → ^2.3.0