What’s new in OxCode

Checking its own work

• NewA code review now ends with one report instead of scattered notes. Every finding from every step is collected into a single report ordered by severity, most serious first, each naming its file and line and what it costs you. It opens by saying what kind of system it is looking at and where the problems concentrate, and closes by naming the one thing to fix first.
• NewOxCode disproves its own findings. In review work it re-reads what it found against what it actually observed while probing, and drops anything its own tests contradict. In testing it formed a theory about a validation bug, ran a probe that disagreed, checked four more ways, and retracted the finding rather than reporting it.
• NewIt says when two problems hide each other. A fallback that quietly supplies a value when the broken thing fails, a catch that swallows the error it would have raised, a missing log: these are reported as pairs, and the hidden one is rated with the hiding counted, because a defect nobody can detect is worse than the same defect in the open.
• FixedA completed review is no longer reported as "nothing changed". A read-only review that found real problems could end with a summary saying no files were created or edited and nothing else, throwing away everything it had found one step before you saw it.
• FixedLine numbers in a review are the ones the code is actually on. A finding could be reported correctly by the step that found it and then cited twelve lines off in the summary. Every line the summary quotes is now checked against what the steps reported, and a summary that moves one is sent back to be corrected.
• FixedA long review no longer loses its last findings. On a review with five or more steps the last steps were dropped from the summary for length, and the summary then described work it could no longer see. The budget now gives every step a share.
• FixedTest scripts are no longer written into your project. When OxCode writes a throwaway script to probe a running service it goes to a scratch folder instead of your source tree, so a review that was told to change nothing changes nothing.
• FixedA new session starts where you pressed the button. Pressing "+ New" with the panel docked in the secondary sidebar opened a chat in an editor tab instead. OxCode was guessing which sidebar it was in by measuring the panel width, which cannot tell them apart.
• ImprovedSeverity ratings are more consistent. The rating scale now reaches the part of OxCode that rates a finding, which it never did before. Ratings are judged on who is affected and when they find out, so a problem that fails silently outranks one that crashes in front of you.
• ImprovedReading a file returns numbered lines, in one consistent format across every tool that reads one, so a finding that cites a line cites the right one. The numbers are display only and are never treated as part of the file.
• ImprovedA review examines each file once. Plans that read a file in one step and analysed it in another produced the same finding two or three times, at different line numbers and different severities. Each file now gets one step that owns it.
• ImprovedFindings say what they cost, not just what they are. A review traces a problem through to its consequences and stops at the last one, quotes the code at the line it cites, and marks whether it observed the behaviour or reasoned it from the code.

Knowing when it is finished

• FixedA finished task now stops. A one-step job could run for eleven minutes, reading the whole project and writing the same file four times over, because nothing told the run it was done: once every step was complete the plan still said "work the next step" and there was no next step. The run now ends after its last step, with one turn left to write its summary.
• FixedOxCode remembers what it already did. On a long task, older turns get dropped to make room, and it used to lose the fact that it had already written a file or already read one, so it would do both again from scratch. The list of files it has written and read now travels with it and cannot be dropped.
• FixedAsking it to go ahead works. Replying "can you resolve all those" or "do it" after a list of findings would come back with a plan containing no steps and a question asking what to continue. Short instructions that point back at the previous message are now understood as instructions.
• FixedCorrect work is no longer marked failed. A run could finish the job and then fail its own check, because the check named a file the task never mentioned and nothing had produced. A check can no longer demand a file that appears nowhere in the request or the plan.
• FixedA finished step is no longer reported as unfinished. Work that was done but never formally marked complete was graded as incomplete. OxCode now asks once before ending, instead of failing the run over a bookkeeping gap.
• FixedSteering during a fix is followed. Asking mid-run for something specific, such as a PNG instead of an SVG, was refused as "not in the task", because only your original message counted. What you say during a run now counts too.
• FixedOxCode stops describing its own settings to you. Sitting in Agent mode, it would answer "I cannot run commands in read-only chat mode", naming an internal state that contradicted the mode picker on your screen. It now says what it cannot do without narrating how it is configured.
• FixedA document opened with no project folder is handled honestly. That case runs read-only, and OxCode was neither told so nor given a working way out, so it could describe edits it had not made. It now says plainly that it cannot change anything and tells you to open a folder, which is the thing that actually helps.
• FixedASCII diagrams and their copy button survive. A diagram in a reply could break out of its code block and lose the copy button, and prose after a diagram could get trapped inside one.
• ImprovedMachine learning tasks are planned as pipelines. A request to train or evaluate a model is now planned as data, split, features, train, evaluate, report, with four things it must get right: fit only on the training split, set a seed, plan a baseline to compare against, and report the number the run actually printed. Ordinary work is untouched.
• ImprovedA plan can no longer assume a tool is missing. The planner is now told which tools the run will actually have, after it turned down an image request on the grounds that image generation was unavailable when it was not.
• NewThe panel has been rebuilt to match your editor theme. Colour, contrast and spacing now come from the theme you are using rather than from fixed values, including light, dark and high contrast. Reduced-motion settings are respected throughout.
• NewThe run surface shows what is happening. A plan card with real progress, a context meter that shows how full the window is and when it was compacted, sub-agent steps on the same rail as everything else, and a changes card listing every file touched, each of which can be put back on its own.
• NewDrag a file onto the panel, or paste an image into it. Both are read straight into the conversation.

Saying which one it is

• FixedFull access now works in Chat, Debug and Security modes. Those three modes ignored the access level you picked and fell back to asking about anything outside a short built-in list of safe commands, so choosing Full access changed nothing in them. They now use the level you selected. Agent mode is still under investigation and is not covered by this fix: if you are on Full access in Agent mode and still see a prompt, the card now tells us why, and we would like to see it.
• FixedThe access picker no longer claims Full access when it is not. Opening the panel without touching the control showed "Full access" while OxCode was actually running at a lower level, so you were told you had something you did not, and would never think to change a setting that already said what you wanted. The label now reflects what is actually in effect.
• NewAn approval card says which access level asked for it. A card now carries a line such as "Partial access asked before this." If you are on Full access and still get a card, those two facts together say so on screen, instead of leaving you to guess which one is wrong.
• FixedA document you open without a project folder can be read. Opening a PDF on its own and asking about it was answered with "I cannot read files", even though OxCode had already extracted the text and was holding it. Attaching a document now works with no folder open. Asking for work on a project you have not opened is still turned down, because there is nothing there to work on.
• FixedA waiting run explains itself in Chat mode. When your account's request limit was reached, Agent mode said it was waiting and Chat mode showed nothing at all, so an ordinary pause looked like the product had stopped responding. Both now say the same thing. Several other status messages had the same gap and were only ever shown in Agent mode.
• FixedAn approval card you never answer no longer holds a slot forever. Walking away from a card left a run occupying one of the three concurrent slots indefinitely, so a later run in another session could be refused for no visible reason. An unanswered card now parks its run and releases the slot, and parking is not treated as a refusal.
• ImprovedHitting your request limit is now waited out once, not once per run. Three runs at the same time each waited out the limit separately against a limit that counts your whole account, so the pause was longer than it needed to be. They now share one wait.

Finishing what it starts

• SecurityCredentials could appear in error messages shown in the chat. When a request failed, the error text was scrubbed by looking for keys that begin with a known prefix, and four of the six credential types OxCode handles do not have one: your sign-in token, your session refresh token, and two provider keys. Any of them quoted in an error was displayed in full. Error text is now scrubbed by recognising what a credential looks like rather than what it starts with, so a key we have never seen before is still removed. The same gap was fixed in the OxCode service, where error bodies are logged. If you saw a token in an error message, rotate it: sign out and back in to replace your session tokens.
• FixedErrors now say what actually went wrong. A rate limit, a dropped connection, a refused request and an expired sign-in used to read alike, and an internal bug could be reported as a provider problem, sending you to check the wrong thing. Each now says which it is and what to do next.
• FixedA question no longer edits your code, and it can no longer talk itself into it. Asking OxCode something could result in it deciding to carry on and change files, install packages and rewrite sources. Typing "hi" was enough to start a build. Questions now get read-only tools, and when OxCode asks to start making changes anyway, that request is turned down for anything you did not actually ask to have changed. It answers your question instead.
• FixedOxCode no longer picks up an old request you never repeated. A request from an earlier conversation could be read as something still outstanding, so a short message could start work you had not asked for. Earlier conversations are now a record of what happened, used to answer questions about past work and never acted on.
• FixedStop now stops. Pressing Stop while OxCode was thinking could leave the work running: it finished what it was doing, wrote files after you had moved on, and anything you typed next was absorbed into the task you had just cancelled instead of being read as a new message. Stop ends the work immediately, and your next message is your own.
• FixedPlan mode builds what it plans. It used to produce a plan you could read but not act on, and then tell you to switch modes to continue, which lost the plan. Plan mode now produces a real plan with steps and checks that you can approve, and carries out that plan when you do. Nothing is written until you approve, which makes it the safest way to start something large.
• FixedThe wrap-up no longer claims work it did not do. It could invent a command-line option that was never written, report a problem that had already been fixed during the run, or miss files a build step created. It now describes what actually changed, and says plainly when something did not finish rather than reporting success.
• NewA way back to the latest message. Scroll up to read while OxCode is working and a button now appears to take you back to the newest message. It tells you when new messages arrived while you were away, so you can read at your own pace without losing your place in a long run.
• FixedCompaction says what it cost. When a long run runs out of room, OxCode now tells you whether it simply rendered fewer turns, summarised the earlier part, or dropped it, instead of showing the same sentence for all three.
• ImprovedLong writes no longer look like a hang. Creating or appending to a file shows the file and the lines written as it happens, and a step that is waiting on the model says so rather than showing an empty thinking block.
• ImprovedA busy moment no longer ends a run. If the model OxCode reached for is at capacity, it now moves to the next one and keeps going, and if every option is busy it pauses and offers to pick up where it left off instead of failing. Under heavy load you may see a note saying it switched to an alternate model: the run continues normally.
• ImprovedReaching your account's request limit pauses a run instead of ending it. OxCode used to treat that as though every model it could reach had run out of capacity, so it worked through all of them, exhausted the list and stopped, reporting a problem that was never the cause. It now recognises the limit, waits the short interval, and carries on with the same model. The limit itself was also raised: it had been set for someone typing in a chat box, and one ordinary task makes far more requests than that.

Less re-reading, quieter runs, a wrap-up worth reading

• FixedSteps in a planned build no longer re-read the same files. What one step reads is now carried to the next, so a five-step build stops paying five times for the same file.
• FixedLong tasks are no longer cut off part way. OxCode used to take its own tools away after a stretch of reading, which could strand it right before the edit it was working up to.
• FixedAllowing a command now takes effect immediately. "Allow this session" and "Always allow" used to apply only to the next task, so the same command kept asking for the rest of the run.
• FixedTyping an instruction into the approval box actually reaches OxCode. It previously went nowhere, so it looked like you were being ignored. There is also a Send button now, and your message appears in the chat.
• FixedInternal reminders meant for the agent no longer appear in your chat as "Noted: ..." lines.
• FixedReopening a session shows the context meter where you left it, instead of dropping to near zero.
• ImprovedSmall requests stay small. Asking for a short script no longer triggers a full plan and approval step; planning is reserved for several features at once, one feature spanning frontend and backend, or a change to how the project is laid out.
• ImprovedOxCode reads more like an engineer: it searches for the part it needs and reads that, rather than opening whole files to find one function.
• ImprovedPlans are grounded in your actual codebase before they are written, so steps name real files and say where new code connects.
• ImprovedThe end-of-run summary now tells you what the change does, the decision worth knowing about, and any rough edge it noticed, instead of repeating the list of files you can already see.
• ImprovedApproval prompts and the changes-applied card are clearer: keyboard shortcuts are readable on every button, and the card clears itself once you keep or undo.

Read PDFs and Office documents

• NewAttach a document and OxCode reads it. Drop or paste a PDF, Word, PowerPoint, or Excel file into the chat and OxCode pulls the text out and works from it, so you can ask what a report says or what a slide deck covers. It all runs inside the extension, with nothing extra to install.
• FixedEach session keeps its own memory. Working in two sessions at once no longer lets one session's files or conversation slip into the other.
• FixedAsking about a document gets a straight answer. "What is in this PDF?" is now answered directly instead of being treated as a build task with a plan to approve.
• FixedAttachments and file paths render cleanly in the chat instead of leaving a stray placeholder.
• ImprovedAn attached document shows as a tidy file chip with its name and length, while OxCode still reads the full contents.

Image input and semantic code search

• NewShow OxCode a screenshot. Paste or drop an image straight into the chat and it can see it, so you can point at a bug, a design, or an error on screen and ask about it. OxCode automatically switches to a vision-capable model when an image is present, in both chat and agent mode. Images are downscaled before sending and are never kept after the turn.
• NewOxCode now finds code by meaning, not just keywords. Ask where something happens in plain language and it locates the right files even when your words never appear in them. Everything runs locally: a one-time model download, then fully on your machine, nothing leaves it.
• ImprovedThe codebase map understands relationships across your project, so it recalls the decisions and notes most relevant to what you are doing right now.
• ImprovedSemantic search is easy to turn on and stays out of your way: a gentle prompt the first time, plus a status bar button that is always one click away if you miss it.
• ImprovedOxCode remembers the whole conversation. Follow-up questions and your next task now build on the real previous turns, so it stops re-reading and re-explaining work it just did.
• FixedApproving a command once now sticks for the session. Running the same command from a different folder, or against a different file or test, no longer pops the approval prompt a second time.
• FixedChat in a side panel no longer goes silent. When OxCode is docked in the secondary sidebar, messages you send now respond right there instead of quietly routing to a closed editor tab.
• FixedStay signed in. Reopening your editor no longer greets you with a login screen, and a brief session hiccup mid-task refreshes quietly in the background instead of signing you out.
• FixedAsk "what did you change?" right after an edit and OxCode answers from memory, quoting the actual change, instead of re-reading the whole file it just edited.
• FixedVerification is honest when there is nothing to verify: a project with no tests now reports that clearly instead of claiming test validation passed.
• FixedOxCode will not let a file, terminal output, or web page it reads talk it into destroying your work. Deleting a file, overwriting one wholesale, and destructive shell commands such as removing files now ask first, even in full-access mode, and everything OxCode reads is treated as information to analyze, never as commands to follow.
• New"Restore File From Before This Task" brings a file back to its state before the current task, from an automatic local snapshot OxCode takes before it first changes the file.
• FixedOxCode stays inside the folder you opened: a symlink that points outside the workspace can no longer be used to read or write files beyond it.
• FixedIt stops you committing a secret by accident: a git commit that would include a .env, private key, or credentials file is blocked with a clear note on how to proceed, while .env.example and other scaffolding are left alone.
• ImprovedThe end-of-run summary shows exactly what changed in each file, as a colored, expandable diff per file, so you can review a whole run at a glance.
• FixedOxCode knows what system it is on. On Windows it uses the right commands and real workspace paths instead of assuming a Unix shell and a made-up /workspace folder.
• FixedEach session is its own. Opening a new or different session while a task is running no longer shows that task's work in the wrong place or leaves the original stuck, and a running task's replies stay in the panel you started it from even if you open OxCode in an editor tab midway.
• FixedA blocked command reads as blocked. When you decline a command, OxCode says plainly that it did not run and nothing changed, instead of ever reporting a blocked action, like deleting a folder, as if it had succeeded.
• ImprovedOxCode treats saved memory as notes to check, not gospel. Something you once told it, such as "the project already passed all tests," is confirmed against the actual code before it is relied on or repeated as fact.
• ImprovedA clearer status bar shows the current mode, how much of the context window is in use, and when the conversation is compacted, so you always know what OxCode is doing.
• FixedOxCode re-reads a file that changed on disk before it edits it, so it never writes over a version it has not actually seen.

OxCode learns your codebase

• NewCodebase index: OxCode builds a local, private map of your project (files, symbols, import relationships) that persists across sessions and finds the right files for cross-file changes without you attaching anything. It stays fresh through saves, branch switches, pulls, and rebases, and never leaves your machine.
• NewDecisions you state in chat, like which database or framework to use, are captured automatically and future plans respect them without being told again.
• FixedConversation memory: answers you give mid-task are remembered, threads no longer recall only their first and last message, and follow-ups reuse what was already learned this session instead of re-reading the whole codebase.
• FixedVerification is honest about its environment: checks that cannot run where you are (missing tools, no browser) show as unverified instead of failing the run, and fix rounds stop with a clear note instead of repeating the same failure.
• NewWeb search works out of the box with no key setup, and it is available in chat mode too.
• ImprovedCommand approvals nag less: an approved command covers harmless output variations, common test runners are pre-approved, and force pushes are always blocked.
• ImprovedCleaner runs: thinking and tool details collapse so the answer is what you see, writing a file shows as writing, refusals render clearly, and duplicate icons are gone.
• FixedFiles no longer land in a doubled folder when a plan repeats the workspace folder name in a path.

• ImprovedOxCode describes itself with more confidence about its strengths and where it is heading, and keeps the focus on OxCode rather than the infrastructure behind it.

Identity, privacy, and a calmer agent

• NewOxCode now introduces itself consistently as your AI Software Engineer and stays model-agnostic; it no longer reveals which underlying model is answering.
• ChangedModel names and gateway details are hidden from the Settings view.
• FixedLarge tasks no longer loop re-surfacing the same issue. Each new problem is shown to the agent once instead of on every step.
• FixedThe Marketplace listing image now loads correctly.

• FixedThe Google sign-in button now works in the editor chat on a fresh install, not only from the Command Palette.
• ChangedNew circular mascot icon on the Marketplace.

• ImprovedRaised the agent step budget to 200 so longer tasks can finish.
• NewAdded Marketplace cover art.

• FixedThe command-approval selector (ask / auto / full) is visible again.
• ImprovedClearer thinking and output signals, hardened chat rendering, and a proper sign-in view when a session expires.

• ChangedChat moved to a full editor panel; the sidebar now focuses on Sessions and Memory.

• ImprovedHardened the agent's prompts so it stops confabulating tools, memory, or file paths and uses the real ones.

• FixedA reliable Stop button, plus runaway and tool-calling fixes.
• ChangedThe orchestrated workflow is now on by default.

• ChangedNew Marketplace icon and an updated sign-in callback.

First release

• NewOxCode arrives: an agentic AI Software Engineer for VS Code with Dev Mode, editor chat, sessions, memory, and command approval.