Lemon 2026.10.6 (stable)

Added

• Custom providers and models. A [providers.<id>] table with api,
  base_url, api_key_env/api_key/api_key_secret, optional
  auth = "none", headers, discovery = "openai" | "ollama" and
  [[providers.<id>.models]] defines an OpenAI-, Anthropic- or
  Google-compatible endpoint (omp models.yml parity). Its models resolve as
  <id>:<model> / <id>/<model> for the default model, /model, model roles,
  task/compaction selectors and fallback providers, appear in the /model
  picker and lemon providers status, and are checked by
  lemon config validate. On built-in providers, headers are sent with every
  request and models entries add or replace catalog models. See
  Custom Providers and Models.
• More provider sign-ins in lemon model. lemon model openai-codex --auth device signs in to ChatGPT with a device code (headless, also
  --non-interactive). openrouter adds browser sign-in that mints an API key,
  xai adds SuperGrok / X Premium+ device sign-in, kimi_coding (kimi-code)
  adds Kimi Code device sign-in, and zai adds GLM Coding Plan sign-in that
  provisions an API key. xAI and Kimi logins refresh automatically;
  --status/--logout cover all of them. See
  CLI.
• API-key sign-in for every catalog provider. lemon model now offers
  Google Gemini API, Groq, Cerebras, DeepSeek, Mistral, Hugging Face, Qwen
  (DashScope), Vercel AI Gateway, OpenCode Zen/Go, MiniMax China and every
  hosted OpenAI-compatible preset (model choice lists the endpoint's models);
  the key is stored encrypted and referenced as api_key_secret.
• OAuth logins can be force-refreshed after the provider rejects an unexpired
  token (Credentials.refresh_provider_api_key/4, used by 401 recovery).
• Local and open-model provider presets. Ollama, LM Studio, llama.cpp,
  vLLM and LiteLLM work without an API key (ollama:qwen3:32b), with omp's
  default endpoints and OLLAMA_HOST/*_BASE_URL overrides; hosted presets
  cover Ollama Cloud, Together, DeepInfra, NVIDIA, Novita, Baseten, Venice,
  Synthetic, NanoGPT, ZenMux, AIML API, SiliconFlow, GMI Cloud, Qianfan, ai&,
  Wafer and Kilo. lemon models --provider <id> --discover lists an endpoint's
  models, and the /model picker includes discovered models for configured
  presets. See OpenAI-Compatible Endpoints.

Fixed

• Session context queries fall back to saved accounting when the live owner
  retires during the read, instead of exiting on a stale registry PID. Timeouts
  and unexpected owner crashes still surface.
• Background process launches no longer bypass admission after queue failures.
  Fixed, owner-bound launch identities reconcile lost acknowledgements and
  reject conflicting retries instead of risking duplicate external effects.
• Background execution now holds a validated live-process capacity limit until
  confirmed termination, bounds pending admission and waiting, prevents canceled
  or expired queued commands from starting, and retains unknown capacity through
  restart. Yielded launches remain owner-scoped and inspectable; cancellation
  receipts no longer claim termination before it is confirmed.
• Local Unix background capacity now stays held when the root shell exits but
  descendants survive. Owned process groups are bound before command execution;
  live members are retained before forwarding stdin that can end the root, so
  verified cancellation still works after natural root exit. Groups reconcile
  until confirmed empty; stale ownership and unsupported workload witnesses
  remain unknown rather than freeing slots or signaling recycled PIDs.
  Repeated cancellation also requires the backend-aware workload witness:
  an empty host wrapper group cannot release SSH, container or PTY capacity.
  TUI tool cards and compact shelves also preserve queued and outcome-unknown
  receipts instead of presenting them as started or failed.
• Unknown Bash launch outcomes now preserve their structured process identity
  through the actual agent event and TUI render path. Late or reopened durable
  storage restores retained capacity and owner monitors before new admission.
  Pre-execution startup rejections retain the real failure reason.
• Unix process admission now uses a portable POSIX stdin gate, including dash
  on Linux, preserving subsequent command input and refusing an unreleased gate.
  Verified group signals use fixed argv with an OS-owned kill executable.
• JSONL storage rejects unreadable or corrupt journals rather than publishing
  empty state. Failed loads remain recoverable after repair; rejected writes
  return bounded errors before cache mutation without crashing the Store.
• Stored OAuth refreshes bind persistence to the selected login's opaque
  revision. Atomic saves cannot overwrite a concurrent login, logout, or
  delete/recreate; stale or failed saves cannot return the rotated credential.
  Persistence callbacks take (name, encoded, expected_revision) and require
  revision-bearing reloads and explicit success acknowledgements. Transient
  fallback reselects authoritative storage; non-persisting refresh stays ephemeral.
• Copilot token-endpoint HTTP 401/403 requires a fresh sign-in rather than
  returning a cached token, including near-expiry refreshes.
• Provider header defaults no longer override per-model headers on the wire.
  OpenAI Completions/Responses, Anthropic and Google compatible endpoints merge
  header names case-insensitively: provider defaults, then model overrides,
  then explicit caller overrides, without duplicate mixed-case headers.
• Provider CI checks use the existing OAuth decode branching convention without
  changing refresh locking, and Codex CLI import regressions check rejection
  and non-mutation rather than incidental error wording.
• OAuth callback parsing and xAI endpoint validation reject malformed URL
  authorities and ports before extracting codes or exchanging credentials.
• Provider sign-in failures no longer print credential payloads or arbitrary
  token-endpoint response strings. Token grants and model discovery do not
  follow redirects; Z.AI callbacks require the returned state, and device-code
  deadlines include callback and HTTP time.
• Copilot, Gemini CLI and Antigravity refreshes serialize by stored login,
  reload after waiting, and fail closed on deleted/replaced logins or expired
  access tokens. Google invalid_grant requires a fresh sign-in.
• Named lemon model --status/--logout --secret-name NAME targets that account
  without deleting another login or clearing its selected provider settings.
• Discovery cache metadata is isolated by endpoint, protocol, credential and
  headers, rather than the last listing for a provider id. Custom Google
  endpoints no longer inherit Google account options; explicitly keyless
  custom Anthropic and Google endpoints omit generated credential headers.
• Unknown provider ids in default_model/default_provider and custom
  provider option keys no longer create atoms from config input.
• lemon model github_copilot --status/--logout no longer crash once a
  Copilot login is stored.
• With auth_source = "oauth", a provider's API-key env var (for example
  GOOGLE_GEMINI_CLI_API_KEY) no longer overrides the stored sign-in.
• Concurrent ChatGPT (Codex) token refreshes no longer replay a rotated refresh
  token: refresh is single-flight per stored login and re-reads the secret.
• OpenAI-compatible providers no longer fall back to OPENAI_API_KEY: a
  DeepSeek, Together or other non-OpenAI request without its own key fails
  instead of sending the OpenAI key to that endpoint. Groq, xAI, Cerebras,
  DeepSeek, Mistral, Hugging Face, Qwen and Vercel AI Gateway now resolve their
  key environment variables and default llm_<id>_api_key secrets.

---