ocodey releases — version history

What changed in every version of ocodey, the multi-model CLI coding agent. Update any time with /update — login, keys, and settings always carry over.

v0.1.1 — Recovery commands & clearer messages (current)

ocodey now fixes itself from outside the TUI: update, reinstall, doctor, and uninstall all work as terminal commands — even when providers.json is broken. The --force bug is fixed, and every error surface now gives you the exact next step.

  • NEW: Out-of-the-TUI recovery commands. ocodey update, reinstall, uninstall, doctor, and version now run as terminal commands — before any config loads. A broken providers.json can never block recovery. The Go TUI dispatches them natively (ocodey help even works without a backend binary).
  • FIXED: --force now actually re-downloads. ocodey update --force and ocodey reinstall were supposed to re-download the current release, but the force flag was never passed through to performSelfUpdate. Fixed — both commands now always re-download.
  • IMPROVED: Startup crash messages with recovery steps. When the backend exits before the TUI finishes starting, the crash message now lists the exact recovery commands to try (ocodey doctor, ocodey update, ocodey reinstall) instead of a vague hint.
  • IMPROVED: ocodey doctor — structured diagnostics. The --doctor command now prints aligned, scannable diagnostics: platform, install path, each provider key status (✓/✗), active provider, and friendly explanations for config errors with fix suggestions.
  • IMPROVED: Clearer CLI messages everywhere. --version shows platform and tarball name. --help lists every subcommand. --self-update, --uninstall, and in-TUI /update messages now mention the terminal alternative and give actionable next steps.
  • IMPROVED: Friendlier TUI chrome. Status bar hints are more compact (^y copy · /help · ^c quit). The welcome banner shows copy shortcuts upfront. In-TUI update messages suggest ocodey update as a fallback when /update fails.

v0.1.0 — Control while it runs

A release about staying in control of a working agent: steer a task mid-flight, stop it instantly, keep permissions you set, and glance at its progress — plus import your Claude Code history and new connector, plugin, and session commands.

  • NEW: Import Claude Code sessions. Already mid-conversation in Claude Code? /import reads its transcripts from ~/.claude/projects, matches the ones for your current project, converts the full message log — text, tool calls, results, images — and resumes it live in ocodey. It lands as a normal session, so it shows up in /sessions like any other.
  • FIXED: /new never deletes your sessions. /new (and /new-session) now saves the current conversation first, then starts a clean session under a new id — so your previous chat stays fully resumable in /sessions instead of being overwritten. /clear still resets the current session in place; only /sessions delete ever removes one.
  • NEW: Steer a running task with /note. While the agent is working, just type and press Enter to leave a note. It reaches the agent at its next step and it adjusts course — no cancelling, no losing the work in progress. The input tells you when notes are live.
  • FIXED: Permissions that stick (/permission). "Always allow" used to reset every turn, so the agent kept re-asking. Now it persists across turns and restarts, and /permission opens an editable list of every tool the agent may run without asking — revoke any of them anytime.
  • FIXED: Instant cancel. Esc now stops a turn right away, even mid-command. A shell that backgrounds a process (like starting a dev server) no longer leaves you stuck watching "cancelling…" — the command is killed and the turn ends in milliseconds.
  • NEW: Task board on Ctrl+T. Pull up the agent's live todo list — done, in-progress, and pending, with a progress count — while it works, the same way Ctrl+L shows loop status. (The mouse-scroll toggle moved to Ctrl+O.)
  • NEW: Connectors and plugins. /mcp lists your connected MCP servers — now loadable globally from ~/.ocodey/mcp.json so they're available in every project — and /plugin lists loaded plugins with the tools and commands each one contributes.
  • IMPROVED: Cleaner reasoning display. Models that stream a chain-of-thought (GLM-5.2, DeepSeek and the like) no longer leak a stray </think> tag into the answer. Reasoning renders as its own dim block, visibly separate from the reply, and is kept out of copies.

v0.0.9 — A clearer live view

The terminal now shows its work more honestly: precise status words instead of one blanket "thinking", reasoning kept visually apart from the answer, a real task checklist, and pasting big blocks of text without the wait.

  • IMPROVED: Status words that mean something. "thinking…" now only shows while the model is actually reasoning about your prompt. Everything else gets its own word — "reading…", "editing…", "executing…", "browsing…", "planning…" while a tool runs, "writing…" while the reply streams in, "updating…" during /update — so the status bar reflects what's actually happening.
  • NEW: Reasoning, visibly separate from the answer. <think>…</think> content now renders as its own dim, gutter-marked "✻ thinking" block instead of blending into the reply — and it's excluded from ^y/^b copies, so you copy the answer, not the model's scratchpad.
  • IMPROVED: A real task checklist. The agent's todo list now renders as an actual checklist — ✓ done (struck through), ◐ in progress, ○ pending — instead of getting flattened into an unreadable one-line tool result.
  • FIXED: Instant paste, no streaming delay. Pasting a large or multi-line block used to animate in line by line, with a visible lag for big content. It now lands instantly as a compact "[Pasted text #N +M lines]" placeholder, with the full text sent when you hit Enter. Small pastes still insert normally.

v0.0.8 — Eight new providers, one switch

The provider catalog more than doubles: Kimi, GLM, Qwen, Grok, DeepSeek, and MiniMax join via their official APIs — and /switch-provider makes hopping between them a two-keystroke affair.

  • NEW: Eight providers on their original routes. Moonshot Kimi (global api.moonshot.ai and China api.moonshot.cn), Zhipu BigModel (open.bigmodel.cn), Z.ai GLM global (api.z.ai), Alibaba Qwen via DashScope International, xAI Grok (api.x.ai), DeepSeek (api.deepseek.com), and MiniMax (api.minimax.io) — each talking to the company's own endpoint, no middleman.
  • NEW: Current flagships on the strength dial. Each provider ships with today's models mapped to fast / balanced / powerful: Kimi K2.7 Code and K2.6, GLM-5.2, Qwen3.7-Max, Grok 4.3, DeepSeek V4 flash & pro, MiniMax M3.
  • IMPROVED: /switch-provider. The new name for changing providers, built around the interactive picker: every provider shows its kind, the model it will run, and whether its key is already saved — and selecting one without a key opens the paste-key prompt on the spot. /provider, /switch, and /p still work.
  • IMPROVED: The provider list keeps itself current. Providers added in a release now appear automatically, even if you have a customized providers.json — your own entries and your active choice are never overridden.

v0.0.7 — Faster setup, reliable copy, one-step skills

Three rough edges, smoothed: adding a provider key is now interactive, copying actually lands on your clipboard, and installing a skill is a single step.

  • IMPROVED: Interactive provider setup. Switching to a provider with no key set — like /provider anthropic — now opens a paste-your-key prompt right there. Type or paste the key and it’s saved and activated in one step. No more copying a /key ENV value command by hand.
  • FIXED: Copy that actually works. ^y (copy last reply) and ^b (copy code block) now write to your OS clipboard directly — clip on Windows, pbcopy on macOS, xclip/wl-copy on Linux — on top of OSC 52. So copying lands even in terminals that don’t support OSC 52.
  • NEW: One-step skill injection. /skills add now accepts a local folder, a SKILL.md file, or a raw SKILL.md URL — not just a git repo — and the skill name is optional (read from the pack’s frontmatter). Dropping in a new capability is a single command.

v0.0.6 — Connection resilience

Long turns on a flaky connection used to die outright. Now they recover on their own — a dropped socket mid-reply is treated as a hiccup, not the end of the turn.

  • FIXED: Dropped connections auto-recover. A socket closing mid-reply — the “The socket connection was closed unexpectedly” error, common on long reasoning turns through a relay — no longer ends the turn. ocodey keeps what it received and automatically resumes where it left off.
  • IMPROVED: Retries with backoff. Transient failures — socket resets, timeouts, and 429/5xx responses from a provider or relay — are retried automatically with exponential backoff, bounded so a genuinely broken request still surfaces clearly instead of looping.
  • IMPROVED: Permanent errors still surface. Real problems — a bad request or an auth failure like an expired key — are shown immediately with their hint, never silently retried, so you always know when it’s something you need to fix.

v0.0.5 — Built-in skills & terminal copy

Two things you asked for: ocodey now ships with skills already installed, and you can finally copy generated content straight out of the terminal.

  • NEW: Five skills, built in. Every install now comes with default skill packs already enabled — code-review-rigor, doc-reader, pdf-figure-analysis, research-paper, and a new general frontend-design pack. /skills has real value the moment you install, with nothing to add first.
  • NEW: Copy from the terminal. Press ^y to copy the last reply, ^b to copy just its code block, or ^t to drop into selection mode and drag-select with your terminal. It uses OSC 52, so copying works over SSH too.
  • NEW: Frontend-design skill. A framework-agnostic guide to building distinctive, production-grade UIs — type scales, spacing systems, restrained color, accessible states — so generated interfaces look intentional instead of templated.

v0.0.4 — Clean uninstall & fresh-start flush

Leaving should be as easy as arriving — and coming back even easier. This release adds a proper uninstaller with two speeds: keep your data for a future reinstall, or flush everything for a truly fresh start.

  • NEW: Uninstall from inside the app. /uninstall opens a picker: remove the app but keep your login, keys, and sessions in ~/.ocodey — or choose the full flush. On Windows, cleanup finishes itself the moment ocodey exits.
  • NEW: Full flush, double-confirmed. The fresh-install reset also signs you out server-side and deletes ~/.ocodey — login, API keys, saved sessions, trusted folders, settings. Because it removes everything, ocodey asks twice before doing it.
  • NEW: One-line uninstallers. Same one-liner style as installing: uninstall.sh and uninstall.ps1 are hosted at /dl/ and work even when the app itself is broken — add --purge (or OCODEY_PURGE=1) for the full flush.

v0.0.3 — Cleanup controls & clearer errors

A quality-of-life release: remove what you no longer need, and when a provider rejects your key, ocodey now tells you exactly why and what to do about it.

  • NEW: Delete stored keys. A dead or expired API key no longer sticks around: /key delete <ENV> removes it from the key store immediately, so you can paste a fresh one cleanly.
  • NEW: Delete saved sessions. Prune your history: /sessions delete <id> removes any saved session you no longer want to keep.
  • IMPROVED: Auth errors that explain themselves. Cryptic provider replies like “401 token status is unavailable” now come with a plain-English hint: whether your key is missing, mistyped, disabled, or out of credit — and the exact command to fix it.

v0.0.2 — Strength, skills & a self-updating CLI

This release is about control: dial how hard the agent works, teach it new capabilities on demand, point it at the newest frontier models — and never reinstall by hand again.

  • NEW: Capability strength. One dial for how hard ocodey works: /strength low-fast, normal, high, max, or ultra. Higher levels think longer per step, take on more authority, orchestrate sub-agents, and keep going on long tasks until the board is clear.
  • NEW: Skill packs. Teach ocodey new tricks without waiting for a release: /skills installs progressive-disclosure skill packs from any git URL or from curated groups, and you can toggle each pack per project.
  • NEW: Self-updating CLI. ocodey now checks for a new version when it starts and updates itself with /update — download, checksum-verify, swap, done. Your login, API keys, and settings live in ~/.ocodey and are never touched by an update.
  • IMPROVED: Newest frontier models. The model catalog now covers the latest generation — Claude Fable 5, Opus 4.8, Sonnet 5, GPT-5.5 Pro, Gemini 3.5, GLM 5.2, Kimi K2.7, MiniMax M3, Doubao Seed 2.1 — pick any with /model and it auto-routes to the right provider.
  • IMPROVED: Smarter agent internals. The internal tools got a tune-up: long answers survive output cut-offs, autopilot checkpoints keep multi-step tasks moving between turns, and output budgets now scale with your strength level.
  • IMPROVED: Update channel. Follow a release channel (/update channel) and opt into auto-update on startup (/update auto on) — or stay manual and let ocodey simply tell you when something new ships.

v0.0.1 — First public release

The foundation: a coding agent that lives in your terminal, reads your whole repo, writes reviewed multi-file diffs, and runs your tests — with any model, on your keys.

  • NEW: The agent loop. Describe the change in plain English; ocodey maps the repo, edits across files, runs your tests, reads the failures, and iterates — you only step in to approve the diff.
  • NEW: Any model, your key. Multi-model from day one: switch providers and models with one command, and let relays route anything else. Your keys, your budget, your choice of engine.
  • NEW: Review before apply. Every change arrives as a clean diff to approve, edit, or reject. Nothing touches disk silently.
  • NEW: Sessions, themes & tools. Persistent sessions you can resume, six built-in themes, and a full toolbelt — file edits, search, shell, web, memory, todos, MCP servers, and media generation.
  • NEW: Multi-agent & loops. /mult-agent splits big work across a controller-led fleet of models working in parallel; /loop repeats or iterates a prompt toward a goal until it is met.
  • NEW: One-line install & sign-in. Standalone builds for macOS, Linux, and Windows — no Node, Python, or Go required — plus passwordless one-time-code login with OpenBNet.