- name
- gh-cli-agentic
- description
- gh CLI commands with JSON output for agents. Use when querying PRs, issues, runs, or repo metadata; inspecting PR checks; fetching failed CI logs; or mapping a github.com URL to an API call.
- user-invocable
- false
- allowed-tools
- Bash(gh pr *), Bash(gh run *), Bash(gh issue *), Bash(gh repo *), Bash(gh workflow *), Bash(gh api *), Read
- created
- 2025-01-16T00:00:00.000Z
- modified
- 2026-11-04T00:00:00.000Z
- reviewed
- 2026-04-25T00:00:00.000Z
# GitHub CLI Agentic Patterns
## When to Use This Skill
| Use this skill when... | Use the alternative when... |
|---|---|
| Querying PRs, issues, runs, or repo metadata with `gh` and need JSON output | Use `git-cli-agentic` for porcelain-mode local git queries (status, diff, log) |
| Inspecting PR checks, mergeable status, or fetching failed CI logs | Use `gh-workflow-monitoring` to actively watch a run until it completes |
| Resolving a github.com URL into an `gh api` call or working with sub-issues | Use `git-issue-hierarchy` for sub-issue add/remove/dependency-graph operations |
| Looking up label, repo, or workflow metadata via `gh` JSON commands | Use `github-labels` to actually apply or create labels on issues/PRs |
Optimized `gh` commands for AI agent consumption using JSON output and structured field selection.
## Core Principle
Always use `--json <fields>` for machine-readable output. The `--jq` filter is built-in (no jq installation required).
### `gh api` sends every field as a string unless you use `-F`
`-f key=value` types the value as a **string**, so an endpoint expecting a
number rejects it:
```
gh api -X POST repos/O/R/issues/2380/sub_issues -f sub_issue_id=5471095685
Invalid property /sub_issue_id: "5471095685" is not of type `integer`. (HTTP 422)
```
`-F key=value` reads the value as a typed literal (number, boolean, null, or
`@file`) and the identical call succeeds. Observed 2026-09-16: six sub-issue
links failed this way before the flag changed. The error quotes the value and
names the property, so it reads as a bad id — the id was right and the flag was
wrong. Use `-F` for every numeric id (`sub_issue_id`, `after_id`) and `-f` for
free text.
## Pull Request Operations
### Check Status
```bash
# Get all check statuses
gh pr checks $PR_NUMBER --json name,state,conclusion,detailsUrl
# Filter to failed only
gh pr checks $PR_NUMBER --json name,state,conclusion --jq '.[] | select(.conclusion == "FAILURE")'
```
**Fields**: `name`, `state`, `conclusion`, `detailsUrl`, `startedAt`, `completedAt`
### PR Details
```bash
# Essential PR info
gh pr view $PR_NUMBER --json number,title,state,mergeable,statusCheckRollup
# Full context
gh pr view $PR_NUMBER --json number,title,body,state,author,labels,assignees,reviewDecision,mergeable,statusCheckRollup
```
**Key Fields**:
| Field | Description |
|-------|-------------|
| `mergeable` | `MERGEABLE`, `CONFLICTING`, `UNKNOWN` |
| `reviewDecision` | `APPROVED`, `CHANGES_REQUESTED`, `REVIEW_REQUIRED` |
| `statusCheckRollup` | Array of check statuses |
### List PRs
```bash
# Open PRs
gh pr list --json number,title,author,labels
# PRs by author
gh pr list --author @me --json number,title,state
# PRs needing review
gh pr list --search "review-requested:@me" --json number,title
```
## Workflow Run Operations
### Run Details
```bash
# Get run status with jobs
gh run view $RUN_ID --json conclusion,status,jobs,createdAt,updatedAt
# List recent runs
gh run list --json databaseId,status,conclusion,name,createdAt -L 10
```
**Status Values**: `queued`, `in_progress`, `completed`
**Conclusion Values**: `success`, `failure`, `cancelled`, `skipped`, `neutral`
### Watch Run Until Completion
```bash
# Watch and wait for run to complete (blocking, no timeout needed)
gh run watch $RUN_ID --compact --exit-status
# Find and watch latest run
RUN_ID=$(gh run list -L 1 --json databaseId --jq '.[0].databaseId')
gh run watch $RUN_ID --compact --exit-status
```
See **gh-workflow-monitoring** skill for comprehensive workflow watching patterns.
### Failed Logs
```bash
# Get only failed step logs (most useful for debugging)
gh run view $RUN_ID --log-failed
# Full logs (verbose)
gh run view $RUN_ID --log
```
### Workflow Triggers
```bash
# Trigger workflow manually
gh workflow run $WORKFLOW_NAME
# Trigger with inputs
gh workflow run $WORKFLOW_NAME -f param1=value1 -f param2=value2
# List workflows
gh workflow list --json name,state,path
```
## Issue Operations
### Issue Details
```bash
# Full issue context
gh issue view $ISSUE_NUMBER --json number,title,body,state,labels,assignees,comments
# Minimal
gh issue view $ISSUE_NUMBER --json number,title,state,labels
# With sub-issue progress
gh issue view $ISSUE_NUMBER --json number,title,state,subIssuesSummary
```
**Read a thread with `--json title,body,comments`, not `--comments`.** In a non-TTY context (every agent Bash call), `gh issue view N --comments` prints only the comments, with no title or body. When the issue has zero comments it prints nothing and exits 0, so an empty thread looks like an empty issue (observed on gh 2.102.0, #3125). `gh pr view --comments` likely behaves the same way, but that is unverified.
### List Issues
```bash
# Open issues
gh issue list --json number,title,labels,assignees
# By label
gh issue list --label "bug" --json number,title
# Assigned to me
gh issue list --assignee @me --json number,title,state
```
### Issue Types
```bash
# Create issue with type (requires repo/org with issue types configured)
gh issue create --title "..." --body "..." --type "Bug"
gh issue create --title "..." --body "..." --type "Feature"
gh issue create --title "..." --body "..." --type "Task"
```
### Sub-Issues
```bash
# List sub-issues of a parent
gh api repos/{owner}/{repo}/issues/{parent}/sub_issues --jq '.[].number'
# Add existing issue as sub-issue
gh api repos/{owner}/{repo}/issues/{parent}/sub_issues -F sub_issue_id={child_id}
# Remove sub-issue
gh api repos/{owner}/{repo}/issues/{parent}/sub_issues/{sub_issue_id} -X DELETE
# Reprioritize sub-issue (move after another sub-issue)
gh api repos/{owner}/{repo}/issues/{parent}/sub_issues -X PATCH \
-F sub_issue_id={id} -F after_id={after_id}
# Get sub-issue summary via issue view
gh issue view {N} --json title,subIssuesSummary
# Returns: {"total": 5, "completed": 3, "percentCompleted": 60}
```
## Repository Operations
```bash
# Get repo info
gh repo view --json nameWithOwner,defaultBranchRef,description
# Just owner/name
gh repo view --json nameWithOwner --jq '.nameWithOwner'
```
## API Direct Access
For operations not covered by subcommands:
```bash
# Get specific data
gh api repos/{owner}/{repo}/actions/runs --jq '.workflow_runs[:5]'
# With pagination
gh api repos/{owner}/{repo}/issues --paginate --jq '.[].number'
```
## Agentic Optimizations
| Context | Command |
|---------|---------|
| CI diagnosis | `gh pr checks $N --json name,state,conclusion,detailsUrl` |
| Get failure logs | `gh run view $ID --log-failed` |
| PR merge status | `gh pr view $N --json mergeable,reviewDecision,statusCheckRollup` |
| Quick issue list | `gh issue list --json number,title,labels -L 10` |
| Full issue thread | `gh issue view $N --json title,body,comments` (non-TTY `--comments` omits the body) |
| Sub-issue progress | `gh issue view $N --json title,subIssuesSummary` |
| List sub-issues | `gh api repos/{o}/{r}/issues/{N}/sub_issues --jq '.[].number'` |
| Add sub-issue | `gh api repos/{o}/{r}/issues/{N}/sub_issues -F sub_issue_id=M` |
| Transfer issue | `gh issue transfer N target-repo` |
| Create dev branch | `gh issue develop N --checkout` |
| Workflow trigger | `gh workflow run $NAME` |
## Failed Calls: Exit Code First, Output Second
Three scripts misread a failed `gh` call as an answer (2026-10): a merge chain
read a 504 as "PR not OPEN", `worktree-audit` parsed a 422 error body as data
and classified branches with unpushed work as merged, and a retried
`gh issue create` that had errored after creating the issue filed a duplicate.
**`gh api` error bodies land on stdout, and `--jq` is skipped.** Observed on gh
2.102.0:
```
$ gh api repos/cli/cli/pulls/999999999 --jq '.state // "none"'
{"message":"Not Found","documentation_url":"...","status":"404"} # stdout, filter not applied
gh: Not Found (HTTP 404) # stderr; exit 1
```
`2>/dev/null` hides only the second line, so a script that skips the exit-code
check parses the error JSON as the result. Test the exit status before using
the output.
**A 5xx, timeout, or empty output is UNKNOWN state.** Re-query with backoff and
fail if the state never resolves; never map the failure onto a state value:
```bash
state=
for delay in 2 4 8 16; do
state=$(gh pr view "$N" --json state --jq .state) && break
state=; sleep "$delay"
done
[ -n "$state" ] || { echo "PR #$N: state UNKNOWN after retries" >&2; exit 1; }
```
A 500 from `gh pr merge` has its own control: see `git-merge-hazards`
([merge-endpoint-failures.md](../git-merge-hazards/references/merge-endpoint-failures.md)).
**After a create errors, look for the object before retrying.** The write can
succeed server-side while the response is lost. List rather than search, since
the search index lags new objects:
```bash
gh issue list --state all -L 5 --json number,title,createdAt # newest first
gh pr list --head "$BRANCH" --state all --json number,url
```
## Error Handling in Context
Use `2>/dev/null` to suppress errors in context expressions (do NOT use `||` fallbacks - blocked by Claude Code 2.1.7+):
```markdown
- PR checks: !`gh pr checks $PR --json name,state,conclusion`
- Run status: !`gh run view $ID --json status,conclusion`
```
For custom issue fields, issue management (transfer/pin/lock/develop), GitHub URL resolution (file contents, diffs, patches), and complete JSON field lists, see [REFERENCE.md](REFERENCE.md).
View on GitHub