CORVID

Install an extension

Installing an extension means your own Corvid builds it into your copy. A normal agent follows the extension's text on its own branch. You review the result and click to merge it.

Before you start, you need Corvid's own source checkout and its build tools. The page's Before installing list says if something is missing.

Import the file#

Importing checks the text and stores it. It never runs or builds anything. Use any of these:

  • Drop one or more .corvid files anywhere on the Corvid window.
  • Click Import in the Extensions tab and pick the files.
  • Copy the text of a .corvid file, then click Paste in the Extensions tab.

Corvid checks each file against the text rules on the safety scan page. A file that breaks a rule is not added, and Corvid lists the problems in plain words. A single new extension opens on its page with Review and install?. Several files, or any problem, show the results in a dialog.

If an extension with the same id is already in the list:

  • An identical file is refused as a duplicate.
  • A newer version can replace the old one while the old one is not installed. You can also choose Keep both.
  • A different file with the same or an older version can be kept beside the old one.

Scan it#

Every extension needs a finished scan before it can be installed. See The safety scan.

  1. On the extension's page, find Safety scan and click Scan…. Corvid shows the rough cost of the scan.
  2. Click Scan now. The page shows "Scanning…", then the report with its verdict: clear, warn or block.
  3. If the verdict is warn, read the warnings. To go on, type I have read the warnings and click Confirm.

A blocked extension cannot be installed.

Check the estimate and confirm#

Building an extension spends your Claude usage. Corvid makes its own estimate first, and nothing is built until you confirm it.

Caution

a build uses your Claude usage, and the estimate can be wrong. It excludes retries and failed builds. A build can stall when your plan's limit is reached.

  1. Click Install…. The dialog "Build name into your Corvid?" opens.
  2. Read the estimate. It shows a range of output tokens, the share of your 5-hour window and your week, the time, the model and how confident the estimate is.
  3. Read the soft cap line. It says when Corvid stops to ask you during the build.
  4. Click Confirm and build. If your usage is above the limit you set in Settings, a second warning shows and the button reads Build anyway. If that limit is set to block new agents, the button is off.

The author's own effort numbers show beside the estimate as a hint. Corvid does not use them for the estimate, because they can be wrong. The estimate gets better as builds finish on your PC: once there are 3, the Extensions tab says how far off the estimates have been.

The soft cap#

The soft cap is a multiple of the likely estimate. The default is 2×. Set it with Soft budget cap on the Extensions tab, from 1.5× to 5×.

When a build passes the cap, Corvid asks you in its own window whether to continue or stop. The build keeps working while the dialog is open.

  • Continue raises the cap by one more estimate. Corvid asks again when the build passes it. Closing the dialog counts as Continue.
  • Stop ends the agent. Its branch and the work so far are kept.

The build#

After you confirm, Corvid creates a task "Build extension: name version" in the Corvid project. A normal agent builds the feature on its own branch, in its own folder. The usual agent limit, usage gates and time limit apply.

  • The agent's file tools are denied Corvid's protected paths, its agent and git settings, and your main checkout. A confirmed tier 3 extension unlocks only the protected areas you ticked.
  • The agent has no merge tool. It can only send a build report.
  • The card and the page show the stage, the time, the tokens against the estimate and cap, and whether the agent needs you. Stop build ends the agent and keeps the branch.

Note

the deny rules cover the agent's file editing tools. A shell command or a script the agent writes can still change a file. The diff scan after the build looks for that.

Review and merge#

When the agent reports, Corvid scans the real code change (see the diff scan). The extension moves to Needs review. Its page shows the build report, the scan's flags with file and line, the changed files by area, and Show the full diff.

  1. Read the flags and the diff.
  2. If anything was flagged, tick I've read the flags.
  3. Click Approve and merge. Corvid scans the branch again. If the branch changed since you reviewed it, nothing is merged and the page shows the new scan.

A blocked diff scan can never be merged. Instead of merging you can:

  • click Send back with a note, and the agent continues on the same branch
  • click Discard the build, and the extension goes back to "Not installed yet" with the branch kept
  • click Open the branch to open its folder

Rebuild and restart#

A merge changes Corvid's source. The running app does not change until you rebuild it.

Caution

Rebuild and restart Corvid ends every running agent. If the new build fails to start, the deploy rolls back to the previous one.

Click Rebuild and restart Corvid when you are ready. Nothing deploys by itself.

When a build fails#

A failed build keeps its branch and never touches main. The page says why: the agent ended without a report, a test or check failed, the branch has no commits, you chose Stop at the cap, or Corvid restarted during the build. From there you can install again, send it back with a note, discard it, or open the branch.