v0.0.1

The first e release brings a Rust coding agent to macOS and Linux. Use hosted or local models, resume and branch conversations, and add your own tools, commands, and themes.

Breaking changes

• ls, find, and the dedicated skill tool are removed. Use bash for directory listings and filename searches, and read to load a skill's SKILL.md.
• OpenCode Zen's provider ID is now opencode-zen. Existing opencode credentials still work; saved model selections under that ID need to be selected once more.

New features

• Connect to OpenAI, Anthropic, Google, xAI, OpenCode, OpenRouter, and other hosted providers, or use Ollama and LM Studio locally.
• Sign in with supported subscriptions or API keys. The model picker refreshes live catalogs and lets you save a preferred model scope.
• Resume with e -c or /resume, and branch from an earlier prompt with /tree. Older linear sessions load without conversion.
• Steer an active turn by sending another message. Review queued prompts above the composer.
• Attach PNG, JPEG, GIF, and WebP images. Supported models receive them, and sessions retain them for resume.
• Open full tool output and edit diffs with Ctrl+O. Run shell commands directly with !.
• Use /compact or automatic compaction to continue long conversations. The previous session remains available in /resume.
• Add tools, slash commands, hooks, and typed launch flags through executable JSONL extensions. Examples include MCP tools, delegated agents, and worktree launches.
• Customize themes, composer keybindings, models, prompt templates, and skills under ~/.e/. Trusted projects can supply their own instructions, skills, and prompts.
• Use e ask for a headless turn or e rpc for JSONL automation. e doctor and e providers provide redacted diagnostics.
• Install checksum-verified binaries for macOS and Linux on ARM64 or x86-64. e update downloads updates, and /reload switches to an installed update while resuming your session.
• Read the bundled configuration and extension guides with e docs.

Improvements

• Tool calls render as connected trees with live command output, outcome counts, and edit statistics. Full write content stays in Ctrl+O rather than filling the conversation.
• Markdown supports highlighted code, tables, task lists, nested blockquotes, and terminal hyperlinks. The composer wraps drafts, supports selection, and collapses large pastes.
• Provider retries show their cause and backoff and can be cancelled immediately. Billing and quota failures stop without retrying; retry_max_attempts controls the retry budget.
• Sessions record timestamps and real token usage. Footer counts use provider reports rather than estimated output, and Show thinking is off by default.
• Reasoning effort reaches supported models across all four provider dialects. Anthropic prompt caching covers the growing conversation.
• Write and edit results send a short confirmation to the model instead of repeating the new content. Long shell results retain their output tail.
• Model, skill, and session pickers have filter tabs. /help opens the searchable command picker.
• Terminal rendering preserves native scrollback and redraws changed rows. Long syntax-highlighted lines no longer stall input.

Bug fixes

• Resume follows the selected branch rather than replaying abandoned messages. Torn final records can be recovered; interior corruption reports an error.
• Interrupted replies retain text already shown, and unfinished tool calls get explicit results so history can replay.
• Provider streams preserve split UTF-8, reasoning blocks, and tool-call identity. Truncated or filtered replies report warnings instead of appearing successful.
• Shell commands enforce their timeout, stop their process group, and drain output without deadlocking. Non-regular files fail instead of hanging tools.
• Edits reject stale file observations, and parallel edits to the same file no longer silently overwrite each other.
• Extension crashes and blocked pipes cannot leave calls waiting indefinitely. Failed startup cleans up child processes.
• Startup preserves typed input, and exit paths restore terminal modes. Wide characters and long drafts keep the cursor in the right column.
• Session and configuration write failures report warnings. Unknown configuration keys survive updates, and corrupt configuration files are preserved for recovery.

Security

• Directory trust controls project instructions, skills, and prompts, not tool execution. Trusted ancestors cover their children unless a workspace has its own recorded choice.
• Model output and extension notices cannot inject terminal controls. Hyperlinks reject control bytes and oversized URLs.
• Pasted API keys stay out of composer recall history. Diagnostics omit credential values.
• Releases include checksums, a CycloneDX SBOM, and signed build provenance. CI checks allowed network hosts, configuration write paths, and pinned workflow actions.

Detailed development history.