v0.7.2

Changes:

• Added support for Amazon Bedrock, Azure OpenAI, Cloudflare Workers AI and
  Databricks, which build their address from a variable you export
  (AWS_REGION, AZURE_RESOURCE_NAME, CLOUDFLARE_ACCOUNT_ID,
  DATABRICKS_HOST) instead of a fixed one. These providers have no model
  list to fetch, so their models come from the built-in catalog and Test
  verifies your key with a single-token request.
• Execution blocks now explicitly displays used tools.
• Imza can now be driven from editors that support the Agent Client
  Protocol (ACP), running imza's sandbox and permission layers with the interface
  provided by the IDE or text editor of choice. Imza exposes many of it's
  features through ACP. However features of imza that are beyond ACP specification
  are not exposed. And they should be configured through imza itself or it's
  configuration file.

v0.7.1

Changes:

• Each agent message now shows the tokens it used and the date right of
  the model name, and the new /usage command opens a spending dashboard
  with per-model and per-day charts for this session and all saved
  sessions combined.
• /variant now takes the reasoning effort right on the input line
  (/variant high) with autocomplete suggestions, instead of opening a
  picker dialog.
• Plan documentation now loads on demand.
• Fixed imza.fs.ts_query ignoring its query predicates, so filtered
  searches returned unfiltered matches.
• Read-only shell commands now cover far more common forms.
• Secret-bearing files (.env, keys, credentials, .git contents) now
  always require explicit approval, even inside an allowed directory,
  and shell commands naming them are never auto-accepted.
• Fixed provider connections showing as failed.
• Fixed bugs regarding imza --update.
• The three loader tools are now consistently named load_skill,
  load_module, and load_mcp, and their chat cards and approval
  dialogs are labeled uniformly.

v0.7.0

Changes:

• Imza is now installable with Homebrew: brew install umutsevdi/imza/imza;
  every release updates the cask automatically.
• New /undo command reverts the last file change the agent made, one file
  at a time.
• API keys and subscription tokens now live in your operating system's
  keychain. Existing keys move over automatically on the first start.
• Added arm64 Linux packages.
• New /cd command switches the session to another directory without
  restarting.
• Slash commands now understand the names other coding agents use: /quit
  completes to /exit, /clear to /new, /agents to /subagents, /reasoning to
  /variant, and so on; a typo gets a suggestion.
• New /permissions dialog manages the persistent permission grants, so
  stopping the same prompts no longer requires editing config.json.
• Added a central in-app notification system.
• Increased subagent limit to 8.
• Fixed the chat input cursor disappearing or blinking.
• Improved performance during diff scans.
• Reduced CPU usage during streaming.
• Fixed a rare crash when closing Imza while a provider connection was
  still being tested.
• Fixed sessions opened by another Imza process not being recognized as
  in use on Windows, letting a switch overwrite them.
• Fixed child processes on Windows losing their environment.
• Fixed a bug that caused agent's mode to be displayed incorrectly when
  loaded from a session.

v0.6.4

Changes:

• Fixed formatting issues in one-shot answers (--ask/--exec) and --update.
• Imza now ships its own documentation as a bundled imza-docs skill; the
  agent can answer questions about Imza itself, and the skill is always
  allowed and cannot be removed.
• New /auto command lets you turn permission approvals on or off while
  the app is running; the input box changes color while approvals are
  skipped.
• Updated sidebar widgets.
• Subscription sign-in refresh can no longer freeze the app when the
  token service stalls, and the Review tab no longer keeps a background
  worker alive after it closes.
• Wrapping a command in bash -c (or sh, cmd /c, powershell -c, and
  similar) no longer sidesteps the shell rules.
• Searching with imza.fs.grep now skips build output, dependency trees,
  and version-control internals, and reads each file once — searching a
  whole project finishes quickly instead of timing out. A path you name
  explicitly is still searched as given.
• Two new bindings round out what used to need a shell: imza.json to read
  and write JSON in Lua, and imza.fs.copy/move to copy or move a file
  through the permission gate.
• Copying, moving, and deleting a file now show up in the diff view, so a
  deletion reads as a deletion and a new file as a new file instead of
  happening invisibly.

v0.6.3

Changes:

• Added bindings for imza.sleep and imza.bg.shell.
• Increased timeout limits for shell commands and lua bindings.
• Startup no longer waits on the first git scan, so the window appears
  sooner; the Changes widget fills in a moment later.
• The agent can now ask to connect an MCP server it discovers; approval
  is per connect, and --allow-mcp/--allow-skill pre-approve servers
  and skills for one-shot runs. Loading a skill also grants read access
  to its own folder, so skill scripts run without a second approval.

v0.6.2

Changes:

• The agent now records the decisions it makes while building for transparency.
• Imza can generate a handoff document based on what was planned, how it was
  implemented, and what changed along the way. Run it from Review. Each run
  keeps its own report, so a session that goes through several stints keeps
  one per stint.
• Discover and install skills without leaving Imza: /skills now has a
  Search tab backed by skills.sh, showing install counts and marking the
  ones you already have.
• The sidebar Documents list now labels plan revisions, build logs, and
  handoff reports consistently, numbered per kind.
• Any viewer can now be exported: press e to save its content under
  .imza/ and open it in your editor.
• On Windows the multi-line shortcut is now shown as Alt+Shift+Enter, since
  terminals reserve Alt+Enter for the full-screen toggle.
• Fixed a regression that caused OpenAI Subscription to fail.

v0.6.1

Changes:

• Imza opens faster: system detection and the provider catalog load in the
  background while the first frame renders.
• Delete and directory creation actions are now monitored by Imza's runtime.
• The sidebar widgets can be hidden individually through the new sidebar
  object in config.json.
• The startup banner now shows a rotating line.
• Ctrl+C and Ctrl+D now ask for confirmation before closing a session that
  has a conversation, so a mistyped key no longer ends your work.

v0.6.0

MCP Release

Changes:

• Full Model Context Protocol (MCP) support: remote MCP servers or
  local programs (stdio transport) connect through the new /mcp dialog,
  which offers a curated catalogue of 38 popular services (GitHub, GitLab,
  Linear, Slack, Notion, Figma, Stripe, and more) alongside a
  custom-endpoint option. Services that use browser sign-in (like
  Cloudflare and Linear) authorize with one press, and their access
  renews in the background. The agent discovers each server's tools on
  demand and calls them directly. Connected servers appear in the sidebar
  with live tool counts, tool rosters refresh automatically when servers
  change, failed connections recover with retries, and every server
  carries its own credential and timeout settings. Servers can be flagged
  to connect automatically at startup, the sidebar lists only the servers
  the agent can actually reach, and inactive servers are shown as such.
• Added the EVREN provider.
• The config file can now pre-approve directories and shell commands, and
  list extra instruction files to follow alongside AGENTS.md.
• Startup memory usage reduced by replacing the JSON library with one
  that does not leave its parsing scratch space in memory.
• Reduced idle CPU and memory usage of the side panel.
• Performance improvement on attachment encoding.
• Fixed sessions becoming unloadable when the transcript contained terminal
  escape codes (colored command output); affected saved sessions load
  again.
• Fixed MCP servers locking up after their session expired; they reconnect
  and keep taking calls.
• Keystrokes are no longer lost or garbled while the agent runs commands.
• Fixed sessions becoming unusable after a tool result contained invalid
  text; affected saved sessions load again.
• Fixed a rare crash when the task list updated while the sidebar was
  redrawing.

v0.5.2

Changes:

• Compaction now runs in the background after a turn finishes, so
  replies are no longer delayed by it.
• The Sidechat no longer re-summarizes the whole conversation.
• Thought blocks are no longer shown when the model produced no
  reasoning content.
• Simplified plan document generation.
• The loading animation now sits to the left of the hints line while
  the agent is working.
• Fixed a stuck state after interrupting a tool that blocked /compact
  and session switching.

v0.5.1

Changes:

• Added support for Google, DeepInfra, Fireworks AI, SiliconFlow and
  CoreWeave.
• Added a loading animation to the input bar while the agent is working.
• Fixed scroll behaviour of Plan Document pane.
• Bindings now aborts the script on error instead of being silently ignored.
• Fixed Sidechat pane layout issues on Build mode.
• Fixed Sidechat pane animation issues.