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.