Oh My Pi v18.20.0 Release Notes — What's New

prepareCompaction() runs as the compaction.prepare loop phase, so a ui.loop-blocked stall from a cold token walk over a long branch names compaction instead of "unknown" (#15027, #18867 by @wolfiesch)

Added the google-decisions judgment API for Google's Gemini Decisions API (gemini-decision-1), with text and image input (#19060 by @kevcube)
Added apiKey to successful OAuthAccessResolution results from oauth.accessById() and oauth.accessAll(): the key normal requests send for that account, which differs from accessToken for providers with structured credentials (#19030 by @Xytronix)
Added an optional exemptionAlias to ToolCallLoopGuard so transport calls can match the dispatched tool's name in exemptTools (#18860)
Added the liquid-systemone judgment API for Liquid AI's decision models, which rejects image sets over Liquid's 10,000 32×32-patch request cap before sending (#18076, #18100 by @SuperUserA)
Added the nace-systemone judgment API for Nace's Drex decision models, which sends images as Drex's top-level images data URLs and replaces the base64 image, audio, video and PDF data URLs that Drex rejects in judged text with a description (#18950 by @alphastorm)

Fixed a spent Codex account retrying every minute for days when its siblings were only refused the requested model: their short model-policy blocks no longer count as a sibling becoming available, so the usage-limit wait follows the spent account's reset and model fallback can take over (#17637, #19019 by @wolfiesch)
Auth-gateway client tokens (serve --clients) can no longer set who a request is from, who pays, or how the gateway reaches the provider: chat routes drop their request headers, userProfileId, initiatorOverride, metadata, serviceTier, cacheRetention and the other gateway-only options, while the gateway token keeps them (#19067 by @alphastorm)
Fixed the auth gateway's /v1/pi/stream dropping taskBudget, include and anthropicPrefixMismatchBehavior, which its other chat routes accept (#19067 by @alphastorm)
Fixed requests to Vertex AI Mistral partner models (mistral-medium-3, mistral-small-2503, codestral-2) failing because they were sent in Anthropic's request format (#19010, #19022 by @yuzu-octopus)
Added Gemini's xhigh thinking level, sent on the Gemini API and Vertex for models whose ladder exposes it (Gemini 4 Hydrogen).
Fixed omp failing to start with ENOTDIR: not a directory, stat '~/.aws/credentials' when ~/.aws is a file instead of a directory (#19083 by @iamshatter)

Added the Gemini Decision 1 judge model (google/gemini-decision-1, GEMINI_API_KEY); assign it to the judge role to use it (#19060 by @kevcube)
Added Nace (nace) with the drex-max-v1.1, drex-max-finance and drex-v2.0 judge models and /login nace (NACE_API_KEY) (#18950 by @alphastorm)
Added Gemini 4 Hydrogen (gemini-4-hydrogen) on Google, Vertex, OpenCode Zen, OpenRouter, Vercel AI Gateway, and Kilo, with low, medium, high, and xhigh thinking levels on the native Google transports.

Factory Droid requests identify as droid 1.7.1 and report the interactive CLI as their surface, and prefix-bound Claude models (Opus, Sonnet and Haiku 5.5, Fable 5.1) on Factory's direct Anthropic route now ask Anthropic to drop stale signed thinking instead of rejecting a turn whose earlier context changed (#18771 by @will-bogusz)

Fixed alibaba-token-plan/qwen4-max-preview (including its qwen4.0-max-preview alias) showing no context or output limits, image input, or thinking levels (#18986).
Fixed Vertex AI Mistral partner models (mistral-medium-3, mistral-small-2503, codestral-2) misrouting to the first-party Gemini endpoint; they now ride the publishers/mistralai raw-predict endpoint with ADC auth and an OpenAI-shaped body (#19010, #19022 by @yuzu-octopus).
Added Liquid AI (liquid) with the d1.1 and d1-ultra judge models and /login liquid (LIQUID_API_KEY) (#18076, #18100 by @SuperUserA)

secrets.yml regex rules are now Rust regex patterns run natively, in linear time: lookarounds and backreferences are no longer supported, and a (?P<secret>…) group marks the part to redact (text around it is context that stays); \d, \w and \s are Unicode-aware. A regex rule that does not compile stops startup with an error naming the file and entry instead of being skipped. Migrate (?<=password=)\S+ to password=(?P<secret>\S+); see docs/secrets.md (#18965 by @H4vC)
Secret placeholders are now four camelCased words after the friendly name, e.g. GitHubToken_correctHorseStapleBattery, with no $$ delimiters, and the transcript shows the real value wherever the model wrote one. Placeholders in the old $$NAME_HASH:CASE$$ form, in sessions saved before this change, stay literal text, including in tool calls the model copies them into; start a new session to get working placeholders (#19014 by @H4vC)
What the model writes from placeholders (replies, thinking, tool calls, summaries) is stored as written and replayed byte-identical, keeping thinking signatures and prompt caches valid; model output that still holds a value (written while redaction was off, or before the secret was configured) is redacted on replay. Values a regex rule finds later no longer rewrite earlier history. A friendlyName

…