GitHub Agentic Workflows: one useful triage

A public Xebia demo for GitHub Universe To You, 11 November 2026: Turning Repository Work into Repeatable Agent Tasks.

The contract is deliberately small: a maintainer selects one reviewed open issue; the agent reads repository context, requests up to two justified labels and one useful comment; a separate permission-scoped job applies allowed outputs. It never fixes code, closes/assigns issues, sweeps the backlog, or creates/merges PRs. The fictional Workshop Queue product supplies realistic issue context, not a claim of a deployed application.

See observed rehearsal evidence for actual readiness and run links. Never treat the evaluation rubric below as a recorded output.

What to show

1. Open a sample issue and explain what a maintainer needs to decide.
2. Open the source: trigger, read permissions, tools, bounded outputs, then the natural-language task.
3. Open the generated workflow. Markdown is authored; GitHub Actions executes the committed .lock.yml.
4. Dispatch the workflow with that issue number. Inspect the agent, detection, and safe-output jobs, then the actual labels/comment on the issue.
5. Ask whether the output reduced maintainer work. Successful execution alone is not evidence of correct triage.

Pinned tooling and setup

Use Git, GitHub CLI, and gh-aw v1.0.1 (stable release checked 5 November 2026). This source was checked against v1.0.1 docs/source, not an unpinned gallery example. Do not compile it with an older extension.

Set-Location C:\Repos\Xebia\github-agentic-workflows-demo
gh extension install github/gh-aw --pin v1.0.1
gh aw --version
gh aw compile triage-one-issue --strict --validate --no-check-update
git diff -- .github/workflows

If already installed at another version, choose an explicit pinned install in your own environment, or download the v1.0.1 platform binary and verify its SHA-256 against that release's checksums.txt. Do not casually upgrade someone else's global tools. Commit both source and generated lock after any change. Compiler validation checks syntax/contracts, not model correctness or runtime entitlement. Docker image validation needs Docker; absence is not proof that images were independently scanned.

Run the dependency-free deterministic gate and compiled-contract tests with Node.js 22 or newer:

node --test tests\demo-contract.test.mjs

These cover invalid inputs, approval/open-issue/PR checks, propagated API failures, generated read permissions, tool allowlists, action pins, and output targets/caps. They do not substitute for a real inference/output rehearsal.

Actions must be enabled and the organization must already support Copilot inference through copilot-requests: write with centralized billing. No PAT is required for this configured path. Do not create credentials or change billing or organization policy to make the demo run. If inference is denied, record the real failure and ask the owner to choose an authorized setup separately. An existing COPILOT_GITHUB_TOKEN does not override this inference path.

Labels and inputs

Create the demo taxonomy deterministically (not through the model), once:

gh label create demo-approved --repo xebia/github-agentic-workflows-demo --color 5319E7 --description "Human-reviewed synthetic demo input"
gh label create needs-info --repo xebia/github-agentic-workflows-demo --color D876E3 --description "Specific information needed for useful triage"

GitHub's default bug, enhancement, and question labels are used unchanged; if absent, create them according to the policy. Review samples before marking them approved; never put real confidential records or secrets in this public repository.

gh issue create --repo xebia/github-agentic-workflows-demo --title "CSV export splits a speaker name containing a comma" --body-file samples\well-formed.md --label demo-approved
gh issue create --repo xebia/github-agentic-workflows-demo --title "Export looks wrong" --body-file samples\vague.md --label demo-approved
gh issue create --repo xebia/github-agentic-workflows-demo --title "Allow pending sessions in draft CSV export" --body-file samples\related-not-duplicate.md --label demo-approved

These intentionally demonstrate a clear reported bug, appropriate uncertainty, and related-but-not-duplicate behavior. Use the issue numbers returned by GitHub, not numbers assumed from this README.

Run and rehearse

From Actions > Triage one approved demo issue > Run workflow, select main and enter one approved open issue number. Or use the CLI (replace 1 with the actual number):

gh workflow run triage-one-issue.lock.yml --repo xebia/github-agentic-workflows-demo --ref main -f issue_number=1
gh run list --repo xebia/github-agentic-workflows-demo --workflow triage-one-issue.lock.yml --limit 5
gh run view RUN_ID --repo xebia/github-agentic-workflows-demo
gh issue view 1 --repo xebia/github-agentic-workflows-demo --comments
gh aw audit RUN_ID

Capture the real run URL, commit SHA, job results, requested/applied outputs, elapsed wall-clock time, and any inference failure. Read artifacts before sharing them; never publish raw credentials/logs wholesale.

Case	Human evaluation rubric (not an exact scripted answer)
Well-formed	bug; no unnecessary information request; distinguish the pending-export feature if linked; no invented root cause.
Vague	needs-info; focused questions about input, steps, and expected/actual output; abstain from a guessed primary label.
Related feature	enhancement; explain that pending selection differs from malformed CSV; no duplicate classification.
Repeat unchanged issue	No extra comment; unchanged useful conclusions do not justify noise.
Invalid number or unapproved issue	Fail the deterministic gate before inference, without issue writes.

The first rehearsals took about five minutes per complete run (including detection and output jobs). Open a verified earlier run for the main stage walkthrough; treat a fresh live dispatch as optional, not an instant response. See the evidence for exact outcomes and timings.

Do not repeatedly dispatch a failing engine. Keep a verified earlier run as the stage fallback; if no run succeeded, say explicitly that runtime readiness is blocked, rather than presenting a fabricated outcome.

Safeguards and limits

Only manual dispatch; no schedules, issue-open triggers, or unrestricted public comment commands. Role allowlist is explicitly admin/maintainer/write. Input validation, open-issue check, and human approval label are deterministic and run before inference. Output targets are bound to the selected input, not *, and handlers recheck the approval label. Repository reads are restricted to this public demo; only three read tools are exposed. The prompt budgets two context reads, six issue/comment reads, and two searches; these are instructions, not hard per-tool counters (v1.0.1's object-form tool counters emit a compiler frontmatter parsing warning, so this demo uses its compatible string allowlist). No shell, editing, browser, custom MCP, or cross-repository access is enabled for the agent. No automatic status comments or reactions inflate the output count. Automatic failure-issue reporting is disabled; failures remain visible in Actions. The compiler may grant framework/output jobs broader issue/PR scopes for its generic handlers; the agent remains read-scoped and the selected open-issue gate plus fixed output targets constrain this demo's write path.

The agent has an eight-minute job timeout, twelve turns, a 20-AI-credit best-effort limit, and no harness retries. Concurrency serializes this workflow; do not queue many runs. Detection and output processing are additional jobs, so eight minutes is not a whole-run latency or invoice guarantee. AI-credit estimates are not a paid-account spending limit; Actions and inference billing are separate.

Safe outputs restrict actions, not truth. Prompt injection filtering/detection are defense in depth, not a correctness guarantee. Humans still inspect applied labels, rationales, citations, and whether the comment is useful.

Primary references

• Pinned release
• v1.0.1 source and docs
• Workflow structure
• Safe outputs
• Copilot authentication

Live website references may describe newer releases. The compiled lock is the reviewable executable artifact for this demo.