Gluon 1.2.0

Added

• ctrl+p in the agent choice sets who approves a build session's commands and edits, for an agent that asks before every
  one: its own behaviour (the default, unchanged), or one of its own levels: accept edits and auto (Claude Code), accept
  edits and never ask (Antigravity), never ask (Grok Build), auto and never ask (Kimi Code). The option's row shows the
  level, a resumed session gets it again, and gluon --launch takes it as --permissions. Codex and OpenCode have none:
  they don't ask before every command or edit.

Changed

• The intake agent on the ChatGPT plan keeps working when a codex update adds something Gluon hasn't checked (codex
  0.162 was refused for ultrafast_mode): Gluon turns off every Codex feature it hasn't checked, and runs with a
  feature codex keeps on or a new model-catalog field, which gluon doctor warns about, instead of refusing that codex.
  If codex fails to start, the error names the features Gluon turned off. When the model uses a tool of Codex's own,
  Gluon interrupts the turn and tells the model to use Gluon's tools, and the answer goes on, instead of stopping the
  intake agent and failing the message (the tool may already have run). gluon doctor shows these warnings.
• Claude Code's Haiku 5.5 runs at high effort by default (was medium), like GPT-6 Luna in the same slot.
• The routing reason names the effort it means: the model's default and what the session gets, e.g. effort +0 from haiku's default (high) → claude-code/haiku@high; when routing rounds up to a stronger level, it says the steps it
  applied there.
• The intake agent looks up what a quick check would settle instead of guessing: its spec's Context holds only what it saw,
  a bug's diagnosis comes with the evidence for it and without unsupported suspects, unrequested scope stays out of the
  Assumptions too, and a request that names only a symptom gets a question about the outcome wanted.
• Explore sessions aren't asked to run things: tests, builds and scripts that write may be refused or fail in a read-only
  session, so the intake agent keeps an explore session's goal to what reading the code can show, and the agent's brief
  says to read the code instead and say what it couldn't run.
• When the intake agent answers a question itself, it says that it only read the code and offers a session to verify it.

Removed

• Sessions no longer get a git worktree: Gluon says nothing about worktrees to the intake agent or the agent, and the agent
  works where Gluon was started, as each harness does on its own. A session saved with a worktree still resumes (the
  worktree is ignored), and gluon stats leaves the worktree columns empty.

Fixed

• A message sent while the intake agent works is queued instead of refused: it shows above the chat box, goes when the
  agent's turn ends, and esc cancels it. If the turn ends with a question or a proposal, it waits and enter sends it
  as your answer, so nothing becomes an answer unseen.
• Claude Code on Amazon Bedrock is priced at AWS's own prices, which Bedrock bills by, instead of Claude Code's list price
  (Sonnet 5.5's cache reads are $0.10 per million tokens on AWS's price list, not $0.20). gluon cost-report names Claude Code's
  list price as the cause where its own figure differs.
• Gluon's own updates: a newer version Gluon remembers from an earlier check is confirmed with GitHub before it is announced or
  downloaded (a remembered version that GitHub no longer lists is never acted on); while Gluon is up to date it still asks once a
  day. A version whose automatic install failed is not downloaded again for a day.
• On macOS, gluon stats sql keeps its 256 MB memory cap: macOS's SQLite ignores the heap limit Gluon sets, so the query's process
  now watches its own memory there and stops a query that passes it ("the query needs more than 256 MB of memory").
• Claude Code 2.1.296: Gluon reads its new slash menu, which marks the highlighted item with ❯, so Enter on a highlighted /clear
  (after typing /cl, say) asks first again instead of clearing at once, and the empty input box's Try "…" hint isn't read as
  typed text.
• Antigravity in a Gluon started inside tmux no longer shows Working for good: there agy re-sends two terminal mode settings every
  2 seconds while it waits, and Gluon took them for work. Output that only sets terminal modes isn't work now.

---

Draft built by the Release workflow from cf33c2a. SHA256SUMS signed with keyless cosign (SHA256SUMS.sigstore.json). Install and verify: see docs/getting-started/install.md.

Gluon 1.1.0

Added

• Gluon updates itself. At start, at most once a day, it checks for a newer release and by default installs it in the
  background for your next start; gluon update does it now and gluon update --check only says whether there is one.
  Nothing is installed unless the release's SHA256SUMS carries a valid Sigstore signature by the Release workflow on
  main (checked by Gluon itself, no cosign needed) and the binary matches it. The config key updates (auto,
  notify, off) and the variable GLUON_UPDATES choose what happens. This contacts github.com and Sigstore's
  tuf-repo-cdn.sigstore.dev. 1.0.0 has no updater: run the installer once more to get it (Updating).
• Pull requests from outside contributors need every commit signed off (git commit -s, the Developer Certificate of Origin); CI checks it. See CONTRIBUTING.md.

Changed

• On the ChatGPT plan the intake agent runs on GPT-6.1 Sol (gpt-6.1-sol) instead of GPT-6 Luna: step 2 of the default brain.order. An order you set yourself is kept.

Fixed

• The intake agent on the ChatGPT plan works with a codex that doesn't know every feature Gluon turns off: codex features list failed with exit code 1 (seen on Windows). Gluon now turns off only the features the installed codex lists, and a failure shows codex's own reason. Codex 0.161 is supported.
• On the ChatGPT plan, a tool of Codex's own that reaches the intake agent's turn (a command, a file change, a web search, a type Gluon doesn't know) stops codex app-server at once, and the turn fails saying so.
• The documented tarball install is bun add -g "$PWD/gluon-<version>.tgz": bun add -g ./gluon-<version>.tgz failed, because bun add -g resolves a relative path against Bun's global directory.

---

Draft built by the Release workflow from 7e86826. SHA256SUMS signed with keyless cosign (SHA256SUMS.sigstore.json). Install and verify: see docs/getting-started/install.md.

Gluon 1.0.0

First public release.

Gluon is a control platform for coding agents: a terminal app (gluon) that runs Claude Code, Codex,
Antigravity, Grok Build, OpenCode and Kimi Code side by side, each on your own API key or your own
subscription.

Added

• The intake agent: an LLM with read-only access to your repository asks a few questions, writes a spec and
  classifies the session.
• Routing: routing.yaml (yours, or the default) turns that classification into an agent, a model and an
  effort, by code and not by the LLM. Sessions start in build, explore or plan mode, and each session in a git
  repository can work in its own worktree and branch.
• Six harnesses, each started in its own UI inside Gluon's frame: start several sessions, switch between them
  and go home with one key. Connections: a subscription, an API key, Amazon Bedrock or OpenRouter, depending on
  the agent.
• Cost and context figures that Gluon calculates itself; what a harness reports about its own cost or window is
  only compared. Price and window tables are built on your machine, never shipped.
• Saved sessions (gluon sessions, gluon resume), a local history of the sessions you launched
  (gluon stats, off with analytics: off) and gluon doctor, gluon uninstall and the other commands in the
  CLI reference.
• Installers and executables for Linux x64 and arm64 (glibc and musl), macOS x64 and arm64 and Windows x64
  (install.sh, install.ps1), plus an npm-style tarball for machines with Bun. Every release carries
  SHA256SUMS, a CycloneDX SBOM and a keyless cosign signature of SHA256SUMS; the installers refuse a
  download that does not match.
• Documentation: getting started, guides, a
  generated reference and concepts.

---

Draft built by the Release workflow from f7181b7. SHA256SUMS signed with keyless cosign (SHA256SUMS.sigstore.json). Install and verify: see docs/getting-started/install.md.