Releases

TL;DR — MiniMax M3.1 Flash is the new MiniMax default ($0.15/$0.60, 1M context), and DeepSeek's deepseek-flash is now V4.1-Flash-1017 with no change to price or id.

TL;DR — A terminal run can now tell your phone right away, or after 15 s, 30 s, 1 or 5 min (Telegram run notices after, in /settings). 1 minute stays the default.

TL;DR — A launcher can hand Codeep its prompt in a file with --prompt-file, so it stays out of the process list. The Omarchy bar widget's Ask field does that from plugin 1.1.0.

TL;DR — Gemini 4 Argon is the new Google default ($2/$10 until 1 December), tried against Google's API. Its thinking tokens are counted, streamed replies once, and 3.7 and 3.5 Flash move to 3.8 and…

TL;DR — On Omarchy, Codeep now sends a desktop notification when a run ends or waits for you (opt-in on other Linux), codeep usage --json prints your usage, and bar plugins can see its status.

TL;DR — Claude Haiku 5.5 arrives at $0.10/$0.50 (five times that past 100K prompt tokens), far fewer runs end "without a summary", and Always Allow now lasts the whole chat.

TL;DR — A new machine gets its API keys with codeep account sync --keys, account push skips keys from environment variables, the welcome block follows the session, and agent runs no longer end on…

TL;DR — On Omarchy, Codeep now appears in the Agents panel with its plan, and its models count toward the panel's "Mostly …" line. A new local usage ledger in ~/.codeep/usage keeps no prompt text or…

TL;DR — codeep --help and the README now say what --yolo leaves unchecked: a shell command runs without asking and can write the files the file tools still ask about.

TL;DR — Fixes found on Omarchy: a one-word answer is no longer pushed to "continue", text and Enter typed together send, y/n answer safe questions at once, the header shows granted access, setup…

TL;DR — codeep --yolo -- "<prompt>" starts Codeep for a launcher with nothing left to answer, Omarchy's skills in ~/.agents/skills load, and Codeep's colours follow the Omarchy theme.

TL;DR — While a thinking model works between steps, the agent screen now says it is waiting for the model and for how long, not "Listing" a tool that already finished. Auto-verify's checks show as…

TL;DR — GPT-6.1 Sol arrives as the new OpenAI default: near-Astra quality at $2/$10, with cache reads at half GPT-6 Sol's price. A setting on GPT-6 Sol keeps it.

TL;DR — Claude Sonnet 5.5 arrives (Opus 5.5 stays the default), Claude replies that start with thinking no longer come back empty, and a request Claude declines now says so instead of going quiet.

TL;DR — A project can no longer replace one of your own MCP servers, a failed MCP server now says why, and /mcp browse drops the servers that are no longer maintained.

TL;DR — OpenAI agent turns now use the Responses API: GPT-6 Astra is back with tool calls, GPT-6 Sol and Luna reason while they work, and GPT-6 Sol is the new OpenAI default.

TL;DR — Claude Opus 5.5 is the new Anthropic default, Grok 4.7 and GPT-6 Sol and Luna arrive, DeepSeek V4 Pro is back, and saved settings on retired models now move to their replacements.

TL;DR — Security release: a repository's own .git/config could make git run a program of its choosing, and opening the folder was enough. Codeep now neutralises those settings on every git call it…

TL;DR — Security and reliability release: a repo can no longer grant itself access or start MCP servers, secrets stay out of prompts, MCP tools ask first, and /commit, /scan, /undo and sessions stop…

TL;DR — Security release. A web page could redirect fetch_url into your own machine or network (localhost, cloud metadata, Tailscale), and the agent runs that tool without asking. Update now.

TL;DR — DeepSeek cached tokens were billed at five times their price, and "saved via caching" was wrong for every provider not priced like Anthropic. Also a correction: 3.3.0 said Codeep could not…

TL;DR — /account answered "Unknown command", /stats has printed the /cost report since June, fresh installs started on an older model, and 40 commands were missing from /help. All fixed, and each is…

TL;DR — DeepSeek V4.1 Flash replaces both V4 models before DeepSeek routes V4 Pro to it on 14 September. Qwen 3.8 Max and Flash, GLM-5.3 in China, and four cost figures corrected — DeepSeek's by…

TL;DR — linking a server no longer means signing in to GitHub on it (it never did, nothing said so), and a device that failed to register once is no longer invisible and impossible to revoke.

TL;DR — usage reports now include one boolean: whether the run was started from your phone. It exists to answer whether a Codeep iOS app is worth building with behaviour rather than a poll.

TL;DR — GPT-6 Astra and Gemini 3.8 Flash are in the model picker. Astra is not the default: it costs twice what GPT-5.6 Sol does and is still rolling out by organization.

TL;DR — a long answer from a phone-started run now arrives whole, across up to three messages, instead of being cut at 3000 characters.

Send Codeep an instruction from Telegram and it runs, with the answer coming back to the same chat — plus Claude Fable 5.1, a Sonnet 5 price that was 50% too high, and cache reads on Kimi and Qwen…

Approval on your phone, on every platform — and the version numbers across the CLI, the Mac app and the VS Code extension now line up.

GLM-5.3-Flash: the same million-token window and tool calling as the flagship, at a twentieth of the price.

A rejected API key made an agent run look like it had simply run out of steps: every request was retried, the budget drained, and the report blamed the iteration limit.

Security rules had never once looked at a .mjs file, forEach + await only counted when it wasn't an arrow function, and the CI fix agent ran out of steps before it could finish.

Every run now records what it touched, including what the boundary refused — and four bugs that only a real session could surface.

Agents you delete on the dashboard now disappear from your machines, sync says when it failed instead of going quiet, and a dependency advisory is closed.

GLM-5.3 reached Z.AI's pay-per-use API and now carries the rate they publish, instead of no price at all.

Custom bots you can actually trust: pin a model, grant only the capabilities you choose, and scope a bot to the projects it belongs to — enforced at runtime across CLI, Mac, VS Code and the dashboard.

The last place /cost still priced plan usage: the prompt-caching savings line quoted dollars for subscription providers that bill a flat fee.

Subscription plans stop inventing dollar amounts — GLM Coding Plan, MiniMax, Kimi and Qwen now read "included in plan". Adds Grok 4.6, Gemini 3.7 Flash and GLM-5.3, and halves a Gemini rate stored…

A redesigned agent view: a live PLAN → READ → EDIT → VERIFY timeline with changed files and checks in a side rail, a persistent header, and energy/water estimates beside cost. Plus a self-migrating…

@mentions inline file context — type @src/file.ts anywhere in your message and the file's contents are attached to that message. No more /add + /drop dance for one-off file references.

Cross-tool rules + MCP config parity: Codeep now reads AGENTS.md (Claude Code / Cursor / Kilo Code standard) as a third project-rules source, and .mcp.json at the workspace root as a fourth MCP…

Claude Sonnet 5 replaces Sonnet 4.6 in the Anthropic picker (1M context, the full low→max reasoning-effort range), plus correctness fixes from a full-project audit: the ACP edit-approval diff,…

Cloud stats: the CLI now reports Anthropic prompt-caching breakdown (cache-creation and cache-read token counts) alongside the existing input/output/cost totals, so the dashboard can show "saved $X…

Repo hygiene: dropped dead React-hooks dir, locked the package manager to npm, made npm run build rebuild dist/ from scratch, added a command registry as the single source of truth for autocomplete…

Three new providers — Kimi (Moonshot), Grok (xAI), and Qwen (Alibaba) — covering the major coding models. Kimi and Qwen include their flat-fee coding-plan subscriptions alongside pay-per-use; Grok…

New /thinking (alias /effort) reasoning-effort control — auto · low · medium · high · max, shown beside the model in the status bar and clamped per provider+model so it never sends a value the API…

Trimmed the model pickers (Claude Fable 5 is de-listed — unavailable under the US export ban — and a few older variants drop off), and editor clients (VS Code, Zed) now see API retry/backoff instead…

Hotfix: the Z.AI default was glm-5.2[1m], but the API rejects that id ("Unknown Model", code 1211) — so a fresh Z.AI session failed on its first request. The default is now plain glm-5.2 (which…

New default model GLM-5.2 (1M-context glm-5.2[1m]) across every Z.AI provider, plus TUI polish: ↑ recalls history, diffs render green/red, full / autocomplete, and /settings values now stick.

/tasks add now matches the dashboard: tag a task as a bug or feature and give it a description inline (--bug / --feature / --desc), and the list tags each row with its project when global.