New features
- Claude Opus 5.5 is available, and OpenAI's newest models (GPT-6 Astra, GPT-5.6 Sol, Terra and Luna) show in the model list as soon as Codex offers them; Hydra's cookbook gains teams built on them.
- Droppy Code keeps Claude Code and Codex up to date by itself: both only update inside their own interactive sessions, so a Mac that used them through Droppy Code alone stayed on an old version and a new model refused to run. The app runs each CLI's own updater twice a day, and a message a model refuses for an outdated CLI is sent again by itself once the update is in.
- Hydra's panel shows recorded lead and head tokens by provider and model, with queue, startup, execution and landing timings where available.
- Custom MCP servers can be added in Settings using a local command or a remote HTTP address, with secret fields stored in the Keychain.
Bug fixes
- A finished Hydra chat now merges by itself every time and never says it is waiting for, or busy in, another chat: a job that finishes while another merges goes right after it, its merge request is built on the remote as it is now so a checkout that fell behind no longer opens a conflicting one, and each chat keeps one merge request that a retry updates instead of opening a new one every five minutes. A failure that clears by itself is retried quietly within a minute.
- Settings cards and the chat timeline lay out eagerly, so a row at the edge of the window can no longer freeze the app.
- Droppy Code writes far less to disk: the library is saved at most every four seconds instead of after every status change, long chats save less often while a turn runs, and web responses are no longer cached on disk. A busy team had written over two gigabytes in an hour and a quarter.
- Local file links in chat open again, Codex token spend is labelled with the turn's model, and the Codex lead no longer shows up among Hydra's worker events.
- The Hydra sidebar auto-hide setting persists and defaults to off.
- Token activity separates what it measured from what it inherited: daily totals recorded before spend kept its source are still counted, and the line under the grid names how much of the year's total those are instead of presenting them as measured.
Refinements
- Restored the previous Droppy Code design and removed the judgment service integration.
- Hydra parses streamed delegations incrementally, reuses incremental checkout snapshots and avoids repeated timeline scans while waiting to land edits.
- Native API providers run up to four independent automatic read tools concurrently and load bounded root project guidelines.
- A dedicated Appearance pane with app-wide zoom, sharper text and the terminal's font; OpenCode Zen and OpenCode Go are separate providers; model and provider settings collapse.
Thanks
- Nikita Tsyganov (@nik1tsyganov): the Hydra auto-hide setting (!358); Antigravity no longer gets Claude's MCP flag (!373); typed Hydra judgments (!374); a held-back delegation block is never silent (!375); local file links (!397); the Codex lead out of worker events (!398); Codex token spend per model (!400); eager layout against the edge-row freeze (!402).
- Bryan Galdámez (@JosueGalRe): custom MCP server definitions and settings (!305); collapsible model and provider settings (!376); main build fixes (!377).
- René (@diazdesandi): shared helpers and the AGPL string (!378); OpenCode Zen and Go split (!379); the Appearance pane, app-wide zoom and the terminal's font (!380); the timeline's lazy window anchor (!381); audit cleanups with five bug fixes (!401).