Skip to content

AI and workspace

Direct the AI, then inspect the outcome

Files, terminal and AI, in one workspace

Work in your browser. The AI reads files, makes edits and runs commands under mission rules. You review the changes, test them and decide what to submit.

WorkspaceSample data. Commands do not execute.
Retry investigationChat 1
Read the evidence first. Find where a retry can charge the same order twice.
Codex

I’ll inspect the retry worker and the existing tests before making a change.

Read src/retry-worker.ts

The retry path calls the provider again after a timeout. I’ll check how the operation key is passed through.

Search: idempotencyKey
Ask the agent to inspect the changed retry path
Work

src/retry-worker.ts

return provider.charge({ amount: order.amount,+ idempotencyKey: key, });

Change for review

From a question to work you can inspect

Sample workflow only—not a guarantee that one prompt always finds the right fix.

  1. Ask for evidence first

    For example, ask where a retry can duplicate a request. Tie the question to mission files and supplied evidence.

  2. Inspect the change

    Open available files or diffs. Compare the AI’s explanation with saved content.

  3. Check the outcome

    Read actual command output and available tests. A chat reply guides inspection; it does not certify a mission pass.

Edit files yourself and see conflicting changes

  1. A local draft is separate from a saved file

    You can read and edit mission files directly. Saves check the file's revision and warn you if it changed underneath you instead of silently overwriting. A page-local draft isn't a submitted file—check save confirmation before leaving.

    Before treating an edit as savedDiagram
    Draft
    Page-local content
    Save
    Check expected revision
    Conflict
    Reconcile with the current file

The terminal belongs to the attempt, not your computer

What you seeMeaningWhat to inspect
Requested commandAn intended operation in the mission environmentAttempt permissions, resources and deadline
Command outputThe actual tool responseSuccess, failure and resulting changes
Your test runChecks available in the workspaceNot a verdict from the mission evaluator

Commands remain within mission authority. They do not grant access to your computer or a real company system.

Choose from this attempt’s model menu

  1. Current models, within the mission’s allowed menu

    Several AI models are supported, though each mission's menu may be narrower than the full catalog. Pick from the models shown for that attempt—personal subscriptions don't add models or expand limits. Even the same model can give different answers.

    What defines the menuDiagram
    Mission version and preset
    This attempt’s policy
    Allowed models
    Choose within the menu
    AI output
    Still needs inspection

AI requests reserve budget before usage settles

  1. Reserve before dispatch

    A new request needs enough capacity for admission. If it does not fit, it is refused rather than starting without a limit.

  2. Record actual usage afterwards

    Accepted requests settle actual usage once finished, so it's not a guarantee of an instant mid-reply cutoff the moment usage hits the cap.

  3. Manual tools have their own gates

    AI exhaustion alone does not submit. File editing, terminal and Submit remain subject to their own time, permission and runtime gates.

Credits, AI budget and time are separate limits

LimitControlsWhere to read it
CreditsFunding the missionStart and Credits; the Pricing page explains funding
AI budgetAI-request admission for the attemptThis attempt’s budget and usage
TimeServer-kept mission timeAttempt clock, which continues when the page closes

Pricing and credits

Read the action when approval is requested

  1. Approval is followed by an actual outcome

    The AI asks for approval when the attempt's policy requires it. Not every command prompts, and mode switches aren't always available. Check what it wants to read, edit, or run. Approving allows the action—it doesn't mean it succeeded, so inspect the result afterwards.

    Permission and outcome are separateDiagram
    One action
    1. Read request
    2. Approve or deny
    3. Inspect outcome

Request Stop, then check the confirmed state

Interrupt to change direction

Use the attempt’s interrupt path during an active turn. A stop request is not yet a confirmed stop.

Check confirmation

Read the returned lifecycle state rather than assuming immediate cessation from a click.

Inspect partial changes

Earlier edits or commands may remain. Usage and elapsed time do not rewind. Inspect, then edit or use supported Undo.

Example: what Undo reverses and what stays

Example scenario: the AI edits a file and then runs a command.

What happenedSupported UndoStill needs attention
File patchCan reverse a supported patch if it does not conflictCheck the resulting file and status
Command side effectDoes not reverse terminal side effectsInspect current state and repair within permissions
AI usage and elapsed timeDoes not rewind budget or timeUse the actual remaining limits
Accepted submissionDoes not undo submissionFollow submission in the mission guide

How it works

Saved revisions do not capture every draft

  1. Capture after work changes

    Completed AI work and supported workspace mutations request asynchronous revisions. A capture request is separate from completion.

  2. Return to the same attempt

    Reloading during a reply may not restore every unfinished line, so inspect saved files after reconnecting.

  3. Verify saved files

    A page-local draft is not the workspace file. Saved revisions do not replace deliberate submission.

How it works

Submission has a review step

  1. Open review

    Open the review screen from the workspace. If it fails to open, check the displayed status and retry option.

  2. Inspect or return

    Inspect the intended work. Return to the workspace before confirming if you are not ready.

  3. Check acceptance

    After confirming, check accepted submission rather than the click alone. Follow the mission’s next stage.

How it works

Prepare a comfortable device and language setup

  1. Work through the browser

    Mission tools do not need installation on your computer. Terminal work runs in the mission environment.

  2. The workspace is in English

    Surrounding pages such as Start and results support Thai and English. Read the mission’s published content language.

  3. Use room for files and terminal

    Files, code and terminal benefit from screen space. Do not assume full editing parity on a phone.

No camera, with a record of the work

  1. Know what enters the record

    The workspace denies camera, microphone and screen capture. Supported AI messages, replies, commands and file changes can appear in bounded work records. This is not a video or every keystroke. Account and security processing remain separate under the Privacy notice. Viewing follows result-specific permissions.

    Capture boundariesDiagram
    Camera, mic, screen capture
    Denied
    Workspace work
    Bounded supported capture
    Readers
    Result-specific access

Who sees what

A few more questions

Does using more AI lower the result?

AI volume isn't the pass rule. Your work is graded against the mission's requirements—see How we measure for details.

Is “done” in chat enough to submit?

Inspect files and output first. AI replies can be wrong, and submission still needs review and acceptance.

What if the workspace becomes unavailable?

Read the state and use the attempt’s recovery or reporting path. Do not start a new attempt to resolve an unknown outcome.

Direct, inspect, then submit

Read the brief and attempt permissions before directing the AI.