v0.1.3

radian 0.1.3

This release lets the file tools read a skill's own files and, in the full
tier, any file outside the workspace except a fixed deny set. It also adds the
/trust and radian trust commands that the startup hint already named.

File tools read more

• Skill directories, both tiers. Each discovered skill's directory (the
  one that holds its SKILL.md) is a read-only root for read, grep and
  glob. A skill that tells the model to read references/… next to its
  SKILL.md now works. Project skills count only when the project is
  trusted.
• Any path in the full tier. In the full tier (no sandbox, the
  default), read, grep and glob read any absolute or ~ path, except
  the sandbox's deny set: .ssh, .aws, .gnupg, .netrc, .npmrc,
  .config/gcloud, .docker/config.json and the browser profile and cookie
  directories under $HOME, and the sessions directory, memory/,
  credentials/, trust/ and access/ under ~/.radian. The check follows
  symbolic links. Before, the file tools refused every path outside the
  workspace in every tier, while the full tier's shell read everything.
• write and patch still change only the workspace, in every tier.
• The workspace tier refuses other outside paths as before; the refusal
  lists the readable roots, now including skill directories.
• ~/.radian/access/ (which projects run without a sandbox) is now denied to
  the radian exec sandbox as well; the full-screen interface already denied
  it.

docs/user/permissions.md lists every rule.

Trust commands

• /trust in the full-screen interface: /trust or /trust status shows
  this directory's decision and where it comes from; /trust on,
  /trust off and /trust forget save or remove it. A change takes effect
  at the next start.
• radian trust status|on|off|forget|list does the same from the command
  line; list prints every saved decision.
• Before, the startup hint said /trust to trust it, but the command did not
  exist, so a project with nothing to load had no way to save trust.

Smaller changes

• The untrusted project hint under the wordmark is drawn in its own theme
  role, trust_hint (yellow by default), so it stands out. A project you
  distrusted says distrusted project · /trust on to trust it.
• The startup tip about ! no longer says the command runs in the sandbox:
  it runs in the sandbox only in the workspace tier.

Install

brew upgrade radian (or brew install hisea/tap/radian), a release
tarball (check it against its .sha256), or cargo install from source.
The manual is at https://getrad.dev/docs/.

v0.1.2

radian 0.1.2

A small fix release.

Fixes

• A command whose output pipe closes early no longer panics. radian docs
  piped into head used to stop with failed printing to stdout: Broken pipe. Now radian docs, radian docs <topic>, radian footer export,
  radian theme import and radian exec --fork stop writing and exit 0 when
  the reader is gone, as head and git do; any other write error prints one
  line on stderr and exits 1. docs/user/cli.md states this.
• The radian docs unknown-topic message names the topic in matching
  backticks.

Tests

• The test that a committed IME string reaches the input in one frame no
  longer depends on machine load. It failed once on a heavily loaded Mac
  because it measured 150 ms of wall-clock time. A unit test now checks that
  a buffered burst drains without one poll wait per event, and the terminal
  test checks that no frame shows part of the string.

Install

brew upgrade radian (or brew install hisea/tap/radian), a release
tarball (check it against its .sha256), or cargo install from source.
The targets and packaging are the same as 0.1.1. The manual is at
https://getrad.dev/docs/.

v0.1.1

radian 0.1.1

This release ships the full user manual inside the binary, and fixes two
footer configurations that failed without a sign and one radian exec
hang on a second Ctrl-C.

The user manual

radian docs lists the manual's pages and radian docs <topic> prints one;
the agent reads the same pages through radian://docs/. The manual now
documents every input that changes radian's behavior, with its type, its
default, every allowed value, its limits and when it takes effect:

• cli.md: every command, subcommand and flag of radian and
  radian exec, with values, defaults, conflicts and exit codes.
• configuration.md: every configuration key in every section.
• environment.md: every environment variable radian reads, and the files
  whose presence or content changes behavior.
• footer.md (new): every footer item, its placeholders, its parameters,
  its icons per glyph set, thresholds, command items, and each built-in
  preset in full, exactly as radian footer export <preset> prints it.
• commands.md, keybindings.md (new), hooks.md (new),
  context-files.md (new), extensions.md (new), mcp.md, themes.md,
  skills.md, memory.md, subagent-and-jobs.md, quickstart.md and
  index.md.

Tests read the names, value sets and defaults from the code and fail when a
page falls behind.

Two pages were wrong in 0.1.0 and are corrected: a skill's frontmatter key
is disable-model-invocation (not disable_model_invocation), and the
memory_search tool's count field is limit (not max_results). A skill
file written from the old page had no effect.

--help text no longer carries internal history wording.

Fixes

• Footer command items: a command (or period) written inline in a
  [[tui.footer.rows]] item was marked as a command item but never ran.
  It is now not read, and a warning tells you to move it to
  [tui.footer.items.<id>]. A command under a built-in item id (for
  example [tui.footer.items.model]) started a process whose output was
  never shown. It now starts nothing, and a warning names the id.
• radian exec listens for SIGINT with one persistent subscription, so a
  second Ctrl-C while the run winds down can no longer hang the process.
• /memory import <path> in the full-screen interface now reads the file.
  Before, it passed the path to the importer as if it were the Markdown, and
  the import always failed. A relative path resolves against the workspace;
  a path with .. or through a symbolic link is refused.
• /memory add -, /memory edit <id> - and /memory import - in the
  full-screen interface write nothing and say that - reads stdin only in
  radian memory. Before, /memory add - stored a record whose text was
  -.

Install

brew upgrade radian (or brew install hisea/tap/radian), a release
tarball (check it against its .sha256), or cargo install from source.
The targets and packaging are the same as 0.1.0.

v0.1.0

radian 0.1.0

radian is a coding agent in one Rust binary, and the agent runtime it is
built on: radian opens a full-screen terminal UI, and radian exec runs
the same agent headless.

This first release ships the radian binary and its rad alias (two
separate binaries sharing one entry point), for four targets:

• macOS aarch64-apple-darwin and x86_64-apple-darwin, signed with a
  Developer ID certificate and notarized. Each is a
  radian-<version>-<target>.tar.gz holding radian and rad.
• Linux x86_64-unknown-linux-musl and aarch64-unknown-linux-musl,
  statically linked (musl), so the same tarball layout runs on any distro
  regardless of its glibc.

Install from Homebrew (brew install hisea/tap/radian, which links only
radian: homebrew-core already has a rad formula), from a release tarball
(check it against its .sha256, unpack radian and rad inside), or from
source with Cargo. On Linux the workspace profile runs the sandbox through
bubblewrap, which must be installed first.

The release is licensed under either MIT or Apache-2.0. New users start with
docs/user/quickstart.md;
the design is in
ARCHITECTURE.md,
and the decisions that constrain the code are in
docs/adr/.