Changelog

Tudo o que entregámos, com data.

Cada versão, com as alterações incompatíveis primeiro porque são as que lhe custam tempo.

v0.0.868 de setembro de 2026PATCH

AskCodi comes to Windows — and the app is free to use

The biggest release yet: AskCodi Desktop launches on Windows, the app is now free to use with no sign-in gate, and the whole workspace has been rebuilt around a manager shell — the app opens straight to your work, conversations read like a team channel with threads, and you can talk to an engineer directly from its task.

NOVO7
  • Windows launch: AskCodi Desktop now ships for Windows x64 — installer and portable builds, with auto-updates. (Builds are unsigned for now, so Windows SmartScreen asks once on first install: More info → Run anyway.)
  • Free to use: the app opens with no sign-in and no plan. Signing in with AskCodi exists for one reason — connecting the AskCodi gateway as a model source with its token balance.
  • The manager shell: the app opens straight to the product. One sidebar answers everything — your private 1:1 with Codi, Needs you (the only interrupt), project channels, Running (the whole fleet, including your own terminal sessions), Work (the cross-project ledger), and History.
  • Channels read like a team chat: your asks are the channel; every reply lands in that message's thread with participants, reply count, and last-reply time on the ask itself. Day dividers, member grammar, and working signals included.
  • Talk to the engineer from its task: open any task from Work and type — your message goes straight to that engineer in its own worktree, no Codi turn spent on a one-line nudge. Runs and reviews still flow through Codi.
  • Pick your harness once: onboarding is three steps — the CLI that types (Claude Code or Codex), the accounts that pay (logins, API keys, or AskCodi credits), and Codi's own model — with a live model catalog that knows what every connected source can serve.
  • Sign in with just a gateway key: paste an AskCodi API key and it is validated for real, persisted, and counted as a connected source — no browser round-trip required.
MELHORADO5
  • A naming counts as a tag: with one project, an untagged ask in the 1:1 resolves to it automatically — the 'Tag a project first' refusal only remains where scope is genuinely ambiguous.
  • One answer, one 'Read more': an engineer's continued runs used to render as a stack of clamped fragments; a task's reply stream now folds as a single block with a single Read more.
  • Uninstalling on Windows now offers to delete your local data (chats, projects, settings) — and defaults to keeping it, so updates and reinstalls never wipe anything.
  • Switching Codi's harness carries its model along instead of dead-ending, and every model picker orders newest first from the live catalog.
  • Open-source groundwork: the app is now Apache-2.0 licensed under the hood, with hardened OAuth, stricter renderer permissions, and a 40,000-line dead-code cleanup.
CORRIGIDO4
  • Orientation with only an AskCodi account connected could show 'thinking' for a few seconds and then go silent forever; the engine path now fails loudly and retries cleanly, and the whole class of malformed streaming tool-calls is gone.
  • The 'New Project' dialog (and every popup) now renders in the app's own theme — no more light-on-dark or black-on-black portals, on Windows or Mac.
  • Windows: the bundled Claude and Codex CLIs are found correctly (.exe, no executable-bit assumption), window controls are present, and the sign-in artwork renders.
  • A turn that ends mid-tool-call is now recorded as the failure it is — never as a silent success that leaves the session looking idle.
Notas completas
v0.0.8527 de agosto de 2026PATCH

Codi runs the team — and nothing merges without you

Every finished task now goes through Codi's review before it counts as done, every outcome lands visibly in your chat, and no agent can merge into your working branch — approving the merge is your click, always. Also in 0.0.85: shared workspaces with members and presence, session-based activity, one Codi turn per batch of answers, and a spent model no longer takes its whole engine offline.

NOVO6
  • The review lane: a worker finishing a task hands it to Codi for review — with the worker's findings in front of it — and Codi approves, requests changes, or escalates to you. 'Done' is the review's verdict, never the worker's own claim.
  • Merge consent: agents approve the WORK, you approve the MERGE. When reviewed work carries a branch, you get one message in chat and one Inbox card — nothing touches your branch until you approve the task from the Board or task header.
  • Shared workspaces with members and presence: see who's in the workspace and who's online.
  • Activity is session-based: every ask is a session with a truthful status, and each expands into its full trail — runs, tool calls, tasks, costs, errors.
  • Answer a batch of Codi's questions and you get ONE follow-up turn, not one per answer — Codi waits for you to stop clicking before it starts.
  • Inbox cards carry Codi's summary and its specific question, with the raw findings one click away on the task's Findings tab — you decide from a briefing, not a wall of comments.
MELHORADO6
  • A model hitting its weekly limit no longer takes the whole engine offline — 'Fable is spent' now leaves Opus and Sonnet working instead of showing 'rate limited, try again in 32 hours'.
  • Work you trigger shows on YOUR ask: answering orientation questions no longer leaves the thread reading 'done' while Codi is still working.
  • Codi owns task budgets: a task running out escalates to Codi, which raises the budget itself within bounds — you only hear about it when it actually needs you.
  • Blocked handoffs are loud in every direction: the engineer is told at handoff time (with the exact command to fix it), your channel names the branch where the work is safe, and Codi is woken to deal with it.
  • Orientation got faster and Codi speaks in the thread while it works.
  • Deliveries land on the branch the work was branched from — no more finished work quietly merging to main while you work on dev.
CORRIGIDO5
  • Codi finished one task and never moved to the next: a completion event arriving while Codi's previous run was winding down could be silently swallowed. Event wakes now queue behind a live run instead of vanishing, so dependent tasks always start.
  • Tasks could jump straight to done without review, stranding their dependent tasks in backlog. A completed worker run now always hands off to review.
  • A repeated handoff of the same task was a silent no-op that read as success — the commit sat on its branch forever. Every handoff now answers with what actually happened.
  • Sub-agent questions appeared only in the Inbox; they are now announced in the chat as well, exactly once.
  • Stale engine verdicts ('Claude is rate limited') could persist after the limit cleared; availability is now evidence-based, self-healing, and the banner names the evidence and its age.
Notas completas
v0.0.8430 de julho de 2026PATCH

Codi survives dead engines (and the cost meter tells the truth)

Your Claude limit resetting in 3 hours no longer stops Codi — it fails over to your next connected engine mid-conversation and tells you it did. Also in 0.0.84: a cost-metering bug that inflated Codex/gateway spend by ~100x is fixed, sessions are deletable, and your Claude CLI login is never imported without asking.

NOVO6
  • Engine failover: when an engine hits a usage limit or auth/credit failure, Codi retries the same turn on your next connected engine and notes the switch in chat. Sticky per conversation, so one dead engine doesn't tax every turn.
  • When no connected engine can serve, a persistent banner in the chat explains what failed and offers a retry — no more silent blank replies.
  • Delete a session: hover your message → ⋯ → Delete session. Removes the message, Codi's replies, tool calls and threads, locally and from the synced workspace.
  • Sharper 1:1 grammar: @ lists your projects only (that's how an ask reaches a project), # opens files, skills and MCP. Project tags render by name.
  • Codebase Map tab: your repo as a navigable graph, built locally during orientation.
  • Using your Claude CLI login is now an explicit onboarding choice — the app no longer imports it from the keychain as a side effect of opening a chat.
MELHORADO4
  • Slack-style scrolling in Codi conversations: new messages land at the bottom and stay in view.
  • Codi delegates instead of doing: repo-wide code sweeps are an engineer's job, and project asks are grounded in your Charter, feature map and codebase map first.
  • Quick actions (Standup and friends) now align with the composer below them.
  • Models with rotating ids are priced by family, so no spend silently disappears from the meter.
CORRIGIDO5
  • Cost metering for Codex and gateway tasks was inflated ~100x and double-billed cached tokens (a $0.17 run could show as $314). Budgets, the spend meter and the savings widget now reflect reality.
  • The standup conversation could go blank mid-run — your 1:1 was being silently converted into a project channel. Fixed, and the 1:1 can't be rebound anymore.
  • Task cards in threads disappeared after an app restart. They survive now.
  • Tagging a project from the @ menu was rejected by the 'tag a project first' gate.
  • Orientation could stall after the scout finished; Whiteboard and Charter now get written reliably, and orientation-created issues always land in backlog instead of auto-running.
Notas completas
v0.0.8328 de junho de 2026PATCH

Opt-in routing telemetry (so we can prove the savings story)

No new features. 0.0.83 adds an anonymous, opt-in feed of CTO routing decisions so we can show real savings numbers on the public landing instead of a generic pitch. Off by default in the EU/UK until you answer the consent dialog; one toggle in Settings → Privacy turns it off everywhere.

MELHORADO2
  • Settings → Privacy 'Detailed analytics' toggle now also controls the new routing telemetry. One switch, two pipelines.
  • Privacy tab copy spells out exactly what gets sent (model name, token counts, who chose the model) and what doesn't (no prompts, no code, no file paths, no project or branch names).
Notas completas
v0.0.8226 de junho de 2026PATCH

Honest model display + receipts on the dashboard

The CTO chat header used to say Opus while the engineer ran on Sonnet. 0.0.82 makes the model selector honest, turns on cost tracking for every chat turn, and adds a few quality-of-life wins for working across multiple workspaces.

NOVO6
  • Model selector in CTO and engineer chats now shows the model that's actually running, not your last-picked default.
  • Cost dashboard records every chat turn, not just background agent runs — the 'Saved vs Opus' card finally moves with your real usage.
  • Sidebar 'Saved this month' widget is always visible — green when you have savings, muted otherwise.
  • Create multiple workspaces on the same folder. Useful for parallel spikes against one repo.
  • Delete a workspace from the sidebar trash icon. Cleans up only AskCodi-created worktrees; your source folder is never touched.
  • Nested sub-chats in CTO sessions and issue chats — jump to any sub-thread from the sidebar.
MELHORADO5
  • Cleaner boot for users without Codex installed — no more 'Codex warmup failed' on launch.
  • Engineer-on-issue is now a single continuous conversation. No more orphan sub-chats fragmenting context.
  • Workspace slug derives from the workspace name, so two workspaces on the same folder get distinct worktree directories.
  • Auth modal now has a backup toast with a 'Sign in' button when your Claude token expires — no more silent failures.
  • Console output stripped of ~820 noisy log lines. Real errors now stand out.
CORRIGIDO4
  • Months-old milliseconds-vs-seconds bug in the cost queries — the 'Saved vs Opus' card was permanently stuck at $0 for everyone. Fixed.
  • CTO '+ new session' button now works on non-git projects (used to fail silently when worktree creation hit a snag).
  • Chat no longer fails with a misleading 'Claude binary not found' when the worktree was cleaned up — falls back to the project path automatically.
  • Engineer model assigned by the CTO at hire time now persists across ChatView remounts.
Notas completas
v1.0.1530 de maio de 2026MAJOR

AskCodi Desktop is here

An AI engineering team that runs on your machine. A CTO agent plans the work, specialists build in isolated git worktrees, and nothing leaves your laptop unless you ship it.

NOVO7
  • A CTO agent reads the task, plans the work, and briefs specialist sub-agents before any file changes happen.
  • Each sub-agent runs in its own git worktree so they can't step on each other's edits — review the diffs, then merge.
  • Plan mode lets you scope a task in chat first and approve the plan before code is written.
  • Local-first by default. Conversations, plans, and run history live in a SQLite database on your machine.
  • Bring your own provider keys: Anthropic, OpenAI, Fireworks, Google, xAI, OpenRouter. The desktop app talks to them directly.
  • MCP server config per project — GitHub, Slack, web search, anything else you've got wired up.
  • Built on the Claude Code SDK, so the agent loop matches what our gateway runs in the cloud.
MELHORADO3
  • Signed arm64 macOS build with auto-update via GitHub Releases.
  • tRPC over IPC for fast, type-safe round-trips between the main process and the renderer.
  • Drizzle ORM on better-sqlite3 for the local store.
Notas completas
v1.0.1415 de maio de 2026MAJOR

A web chat that keeps up with the work

Per-chat models and keys, a prompt library on the slash key, attachable skills, MCP servers, side-by-side model compare, fork from any message.

NOVO7
  • Type "/" in the input to drop in a prompt from your library — variables become editable fields right in the chip row.
  • Attach skills to a chat so the model gets focused domain context before it answers.
  • Bring your own MCP servers per chat — web search, GitHub, Slack, anything you wire up. The gateway turns them into tools for that session.
  • Pick the model and API key per chat, not globally. Keep one Claude tab and one GPT-5 tab open side by side.
  • Compare two models on the same prompt with a split view. Useful when you can't decide between Claude Sonnet 4.6 and Kimi K2.
  • Fork any message into a new chat to try a different direction without losing the original thread.
  • Drag, drop, or paste images and files straight into the input. Vision models read them inline.
MELHORADO3
  • Markdown rendering with syntax-highlighted code blocks and one-click copy on every snippet.
  • Conversation history restores the original model, API key, and prompt stack — switching chats no longer leaks state.
  • Export any conversation as Markdown.
Notas completas
v1.0.1321 de março de 2026MAJOR

AskCodi Agent Builder and Marketplace

Implementation of a visual agent builder and a comprehensive marketplace for AI assets.

NOVO2
  • Released AskCodi Agent Builder with visual canvas for sub-agent wiring and shell execution routing.
  • Launched AskCodi Marketplace for installing prompts, skills, Model Context Protocol tools, and Agent Collections.
MELHORADO2
  • Expanded API access with an OpenAI-compatible endpoint for custom agents.
  • Stabilized asset collection installations and GitHub repository imports.
Notas completas
v1.0.1220 de março de 2026MAJOR

AskCodi High-Efficiency Reasoning Models Update

Integration of new specialized models for fast execution, background tasks, and multi-file reasoning.

NOVO5
  • Added support for GPT 5.4 Mini and Nano models with 400,000 token context.
  • Added support for Grok 4.20 featuring a 4-agent internal squad and 2,000,000 token memory.
  • Added support for Nvidia Nemotron 3 Super with 1,000,000 token memory and Multi-Token Prediction.
  • Added support for Zhipu GLM-5-Turbo with 202,000 token context and 131,100 output streaming.
  • Added support for MiniMax M2.7 tuned for multi-file refactoring with 204,800 token window.
MELHORADO2
  • GPT 5.4 Mini scored 54.38% on SWE-Bench Pro.
  • Grok 4.20 internal agent debate reduces hallucination rates below 4%.
Notas completas
v1.0.107 de janeiro de 2026MINOR

New in AskCodi: MiniMax M2.1 Model Integration

AskCodi now supports MiniMax's advanced M2.1 model, bringing exceptional programming capabilities and improved reasoning to your development workflow. This lightweight yet powerful model provides significant upgrades in multi-language coding and complex task execution.

NOVO3
  • MiniMax M2.1 Model Support: A state-of-the-art 10B activated parameter model optimized for coding, agentic workflows, and modern application development
  • 197K token context window for handling large codebases and complex projects
  • 68.85 tokens per second processing speed for responsive development experience
Notas completas
v1.0.97 de janeiro de 2026MINOR

New in AskCodi: GLM-4.7 Model

AskCodi now supports Z.AI's newest model, GLM-4.7, bringing enhanced programming capabilities and improved reasoning to your development workflow. This model provides significant upgrades in multi-language coding and complex task execution.

NOVO1
  • Z.AI's latest model with enhanced programming capabilities and stable multi-step reasoning
MELHORADO2
  • More reliable performance on complex agent tasks and terminal operations
  • Better Reasoning: 42.8% on the HLE benchmark - a 41% increase over GLM-4.6
Notas completas
v1.0.826 de novembro de 2025MINOR

Claude Opus 4.5 - Anthropic's Frontier Reasoning Model Added

AskCodi now supports Claude Opus 4.5, Anthropic's frontier reasoning model optimized for complex software engineering, agentic workflows, and long-horizon computer use. This advanced model brings substantial improvements in reasoning capabilities and efficiency to your development workflow.

NOVO1
  • Claude Opus 4.5 Model Access: Anthropic's frontier reasoning model optimized for complex software engineering and agentic workflows
MELHORADO1
  • Opus 4.5 comes with 1/3rd Price Reduction: 5$ per million input, 25$ per million output - 4x input and 20x output efficiency compared to predecessor
Notas completas
v1.0.617 de novembro de 2025MAJOR

:Free Models Available on AskCodi

We've added new :free models to AskCodi that require no credits to use, providing more accessibility for users to experiment with AI capabilities. These models offer a cost-effective option for tasks that don't require premium performance.

NOVO2
  • Added :free models which require no credits to use but come with slower/uncertain response times and data can be used to train by Assistiv.ai and model providers
  • Introduced new range of free model options to expand accessibility and experimentation
Notas completas
v1.0.529 de outubro de 2025MINOR

Frontier Open Source Models Added - Kimi K2 Thinking and MiniMax M2

AskCodi has added two cutting-edge frontier open source models: Kimi K2 Thinking and MiniMax M2. These advanced models bring powerful reasoning capabilities and efficient coding assistance to enhance your development workflow.

NOVO2
  • Kimi K2 Thinking (Moonshot AI): Trillion-parameter MoE model. Uses 32 billion parameters per pass with 256K-token context. Built for complex reasoning across hundreds of steps. Handles 200-300 tool calls without losing focus. Works for autonomous research, coding, and writing.
  • MiniMax M2 (MiniMax): 10 billion active parameters (230 billion total). Built for coding and agentic tasks. Strong in code generation, multi-file editing, compile-run-fix cycles, and test-validated repair.
MELHORADO4
  • Persistent workflows: Both models stay focused across hundreds of sequential actions
  • Fast performance: MiniMax M2 runs with low latency despite strong reasoning
  • Complex reasoning: Kimi K2 handles long research and coding tasks
  • Top benchmarks: Both models rank high on HLE, BrowseComp, SWE-Bench, and related tests
Notas completas
v1.0.416 de outubro de 2025MINOR

Claude Haiku 4.5 Added with Model Updates

AskCodi has added Claude Haiku 4.5, Anthropic's fastest and most intelligent Haiku model with near-frontier intelligence at blazing speeds. The update also includes model retirements and recommendations for migration to newer models.

ALTERAÇÕES INCOMPATÍVEIS3
  • Sonnet 4 model retirement: Users should migrate to Haiku 4.5 or Sonnet 4.5
  • glm-4.5 model retirement: Users should migrate to glm-4.6
  • grok 4, grok-4-fast-reasoning and grok-4-fast-non-reasoning retirement: All combining into one model grok-4-fast
NOVO1
  • Claude Haiku 4.5: Anthropic's fastest and most intelligent Haiku model with near-frontier intelligence at blazing speeds, extended thinking, and exceptional cost-efficiency
MELHORADO1
  • Same performance as Sonnet 4 at 1/3rd the price with Claude Haiku 4.5
Notas completas
v1.0.38 de outubro de 2025MAJOR

Now /v1/responses OpenAI Endpoint Compatible

AskCodi is now /v1/responses OpenAI endpoint compatible, enabling users to access all AskCodi models through the OpenAI Codex CLI. This update enhances integration capabilities for developers using the OpenAI Codex CLI with AskCodi's model offerings.

NOVO1
  • /v1/responses OpenAI Endpoint Compatibility: AskCodi now supports the OpenAI endpoint format for enhanced compatibility
MELHORADO1
  • OpenAI Codex CLI Integration: Users can now access all AskCodi models through the OpenAI Codex CLI
Notas completas
v1.0.22 de outubro de 2025MINOR

11 New Models Added Including GPT-5 Codex

AskCodi has added 11 new models including OpenAI's GPT-5 Codex to enhance building, debugging, and refactoring capabilities. The update includes models from OpenAI, Moonshot AI, Z.ai, and x.ai with specialized features for various development tasks.

NOVO4
  • GPT-5 Codex (OpenAI): Purpose-built for code review superpowers, from quick pair-programming to 7-hour autonomous refactors
  • Kimi Models (Moonshot AI): kimi-k2-0905-preview with extended context for massive codebases
  • GLM Models (Z.ai): glm-4.6 for advanced reasoning, glm-4.5 as reliable all-purpose coding assistant, glm-4.5v with vision-enabled capabilities for turning screenshots/mockups into code, and glm-4.5-flash completely free for instant completions and syntax fixes
  • Grok Models (x.ai): grok-code-fast-1 for rapid prototyping, grok-4 for step-by-step reasoning, grok-4-fast as cheaper and faster option, and grok-4-fast-non-reasoning for direct-to-code without reasoning overhead
Notas completas
v1.0.130 de setembro de 2025MINOR

Claude Sonnet 4.5 Integration Now Available

Anthropic's Claude Sonnet 4.5 is now available in AskCodi, bringing smarter code reviews, better support for long-running tasks, and stronger reasoning capabilities to your development workflow. The model is accessible through all AskCodi integrations including Cline, Continue.dev, Kilo Code, Zed, and Roo Code.

NOVO1
  • Claude Sonnet 4.5 Model Access: Anthropic's latest advanced reasoning model now available through AskCodi
Notas completas
v1.0.08 de setembro de 2025MAJOR

AskCodi Platform: Major Update - New Platform Launch

The AskCodi platform has undergone a complete transformation with a new API-first architecture. This update brings comprehensive tool integrations, direct access to leading AI models, custom model configuration options, and a transparent pricing model with zero markup. Existing users must migrate to the new platform to continue using AskCodi.

ALTERAÇÕES INCOMPATÍVEIS3
  • The old platform is no longer available - all users must migrate to continue using AskCodi
  • Plan transitions: Lifetime & AppSumo users receive equivalent one-time tokens equivalent to their original purchase
  • Subscribing users migrated to an equivalent "Pay as you go" plan
NOVO5
  • Full API Access: Integrate AskCodi directly into favorite dev tools (Cline, Continue.dev, Aider, n8n, and more)
  • Provider Models: Access to cutting-edge models including GPT-5, Claude, and other leading AI providers
  • Custom Model Configuration: Build and configure custom AI models tailored to specific needs
  • Transparent Pricing: Pay same cost as original providers with zero markup
  • Token-Based System: More transparent and flexible than previous credit system
Notas completas
Acompanhe

Boa parte disto entregou-se a si própria.

A AskCodi mantém a AskCodi, por isso estas notas descrevem trabalho que também tivemos de rever.