CLI reference
Every codev command, grouped by what it’s for. If you’re looking for a narrated
walkthrough instead of a reference table, start with
the tutorials
or the onboarding guide.
Install, update, remove the bundle
Section titled “Install, update, remove the bundle”| Command | Purpose |
|---|---|
codev init --target <path> [--agent-platform <platform>] [--programming-language <lang>] |
Install the bundle into a repository |
codev diff --target <path> |
Preview what update would change, without writing anything |
codev update --target <path> |
Apply a previewed update |
codev remove --target <path> [--dry-run] |
Remove the installed bundle and integrations |
All four preflight the entire operation first. A locally modified managed file becomes a visible conflict; CoDev never silently overwrites it.
--agent-platform accepts opencode, junie, antigravity, claude, or all (the
default). OpenCode and Claude Code get the full workflow; Junie and
Antigravity get a single narrower assistant agent for bounded, surgical edits — see
ADR-0031
for why. Pass it more than once, or a comma-separated list, to select several platforms at
once instead of all.
--programming-language selects which language-specific style-audit skill gets installed:
python, typescript, all (both), or none (the language-agnostic audit only, and the
default). An update with no --programming-language preserves whatever was already
selected, recorded in .codev/lock.json.
Health and status
Section titled “Health and status”| Command | Purpose |
|---|---|
codev status [--verbose] [--json] [--since <date>] |
Bundle health, installed adapters, open tasks, WIP-per-owner, changed-file overlap, per-task size vs. review.max_lines/review.max_files, stacked-task depth, and (--verbose/--json) a gate_decisions count of Claude Code guardrail-hook asks by hook and decision |
codev adapter list |
Show which platform adapters are installed |
codev adapter add <platform> |
Add one adapter to an existing installation |
codev adapter remove <platform> |
Remove one adapter (still works for a platform an older CoDev version installed, even after the current version drops it — see ADR-0031’s migration note) |
codev adapter verify <platform> |
Check one adapter’s structural conformance: lifecycle wiring present, no unrestricted shell access, no retired review-scale patterns |
codev check and codev doctor still work as deprecated aliases for status and
status --verbose — each prints a warning and will be removed in a future major version.
Configuration
Section titled “Configuration”| Command | Purpose |
|---|---|
codev config get <key> |
Read one configuration value |
codev config set <key> <value> [--global] |
Write one configuration value |
codev config list [--global] |
Show all configuration values |
Configuration is layered: command-line flags override environment variables, which
override project config, which overrides global config, which overrides the built-in
default. codev config set git.pr_base <branch> is the one most people need early — it
sets the pull-request base branch once, repository-wide, instead of repeating it on every
task. git.workflow defaults to trunk; set it to feature-branch to let plan-wave and
build-change split a task at an engineering-dependency boundary instead of only a
usefulness boundary, provided the task states its own containment.
The task lifecycle
Section titled “The task lifecycle”| Command | Purpose |
|---|---|
codev task start --id <id> --base <sha> [--entry takeover|direct-review] |
Open a new task |
codev task record ... |
Record one builder or reviewer round (normally automated by your agent, not typed by hand) |
codev task check --id <id> --head <sha> |
Ask whether the task may proceed to a pull request |
codev task status [--target <path>] |
List tasks in progress |
codev task log --id <id> |
Show one task’s full round history |
codev task close --id <id> --outcome approved|abandoned|escalated |
Close a task |
codev task triage | escalate | escalations | waive | reopen | relink |
Outer-loop and recovery operations — read ADR-0001 before scripting against any of these |
This tracks one task’s round state as local JSON under .codev/task/; it never writes
product source itself.
Git and GitHub
Section titled “Git and GitHub”| Command | Purpose |
|---|---|
codev git issue-create ... |
Create a GitHub issue (no task precondition) |
codev git issue-view --number <n> |
Print an issue’s body and all comments as JSON (read-only, no task precondition) |
codev git branch --id <id> [--base <sha> | --stack-on <task-id>] [--allow-dirty] |
Create the task’s own branch; --base defaults to git.pr_base, then the repository’s default branch; --stack-on targets another task’s own branch instead (ADR-0034, trunk workflow only) |
codev git commit --id <id> --message <msg> |
Commit on that task’s branch |
codev git push --id <id> |
Push that task’s branch |
codev git open-pr --id <id> --title <title> |
Open the pull request as a draft |
codev git mark-ready --id <id> |
Mark the PR ready for human review once the outer loop says so |
codev git restack --id <id> |
Rebase a stacked task’s branch onto its recorded parent’s current head and force-push with --force-with-lease (ADR-0034) |
This is the only path for an agent to mutate the repository or GitHub — raw git commit/git push are denied to every role for exactly this reason
(ADR-0002).
Skill evaluation
Section titled “Skill evaluation”| Command | Purpose |
|---|---|
codev eval doctor [--target <path>] |
Zero-cost readiness check before a real trial run |
codev eval task create <name> --target <path> --include <path> |
Scaffold a new evaluation task |
codev eval task run <name> --target <path> --output <dir> |
Run one task once |
codev eval benchmark run <skill> --target <path> --output <dir> |
Run every task tagged with a skill, with and without it, and report the pass-rate delta |
codev eval report <output-dir> |
Render a trial’s or benchmark’s output directory as plain text |
codev eval show <skill> [--target <path>] |
Render a skill’s packaged eval trace |
codev eval nvidia <verb> |
Second, independent engine wrapping the externally installed NVIDIA SkillEvaluator CLI against a skill directory itself |
Bring your own skill or agent, in your own repository, and test it with OpenCode — no
CoDev-hosted execution, no credentials read or stored. --sandbox docker opts a task with
its own declared environment block into container isolation
(ADR-0027);
worktree isolation on the host stays the default.
| Command | Purpose |
|---|---|
codev codeowners init |
Scaffold a starter .github/CODEOWNERS — human-run directly, never agent-invoked |
codev self version |
Show the installed CoDev version |
codev self update |
Show how to upgrade the installed CoDev tool |
Reviewing an existing GitHub Pull Request
Section titled “Reviewing an existing GitHub Pull Request”The installed pr-review skill reviews an existing GitHub Pull Request and can prepare
validated inline comments for the exact PR head, using the GitHub CLI credential store by
default. See PR review setup
for installing and authenticating the GitHub CLI (Windows-specific steps included) and the
full fetch/publish command reference.