selfheal-demo

name

selfheal-demo

description

Run or reset the self-heal walkthrough demo in this repo. USE FOR: showing the @-mention-PR-author handoff pattern end-to-end (inject a CI failure, watch the demo workflow open a tracking issue and PR comment, then assign Copilot manually). Trigger phrases: "run the self-heal demo", "demo the self-healing pattern", "show the selfheal-experiments demo", "reset the self-heal demo", "clean up the self-heal demo". Wraps run-demo.ps1 and reset-demo.ps1 (next to this SKILL.md). DO NOT USE FOR: experiment trials (use scripts/inject.ps1 + experiments/trials.jsonl); fixture changes; workflow changes.

selfheal-demo

Drives the customer-facing walkthrough packaged with this skill. The skill lives at .github/skills/selfheal-demo/ and VS Code auto-discovers it from that location — no manual install step.

Operations

run — start a demo run

Run when the user asks to start, run, kick off, or trigger the demo.

1. Pre-flight — verify all of:
   • Current working directory is the selfheal-experiments repo root (check for .github/skills/selfheal-demo/run-demo.ps1).
   • gh auth status reports an authenticated user (not just an App). If not, stop and tell the user — App tokens cannot assign Copilot, which is the entire point of the demo.
   • No open [demo] PR already exists. (gh pr list --state open --search '"[demo]" in:title' --json number --jq 'length' returns 0.) If one exists, ask whether to reset first; do not silently -Force.
2. Invoke — run pwsh .github/skills/selfheal-demo/run-demo.ps1 from the repo root. (On non-Windows: bash .github/skills/selfheal-demo/run-demo.sh.) The user's working tree is never modified — the script uses a sibling worktree.
3. Report — surface the script's printed output to the user, in particular:
   • The demo PR URL.
   • The expected wait (~30 s for CI to fail; ~8 min median for the agent's fix PR after manual assignment).
   • The exact gh issue edit <NUMBER> --add-assignee copilot-swe-agent snippet, with <NUMBER> left as a placeholder until the tracking issue actually appears.
4. Do NOT attempt to assign Copilot programmatically. The manual click is the demo's payload.

reset — clean up after a run

Run when the user asks to reset, clean up, tear down, or repeat the demo.

1. Invoke pwsh .github/skills/selfheal-demo/reset-demo.ps1 -Force (skill always passes -Force — the user already asked for the reset by invoking the skill).
2. Surface the printed output. The script is idempotent; "nothing to clean" is a success path.

Pre-flight checklist (both operations)

• git and gh on PATH.
• Repo has the App installed and Demo Self-heal workflow present at .github/workflows/demo-self-heal.yml.
• Authenticated gh user has access to the repo.

Files this skill touches

• Reads / executes (next to this SKILL.md): run-demo.ps1, run-demo.sh, reset-demo.ps1, reset-demo.sh.
• Reads only (for context if the user asks why something happened): README.md, ../../workflows/demo-self-heal.yml, ../../DEMO_CI_FAILURE_TEMPLATE.md.

Anti-patterns

• Do not edit the workflow or the inject scripts to make the demo "more reliable." The current fixture has 5/5 success in Batch 1; the variability sits with the agent, not the harness.
• Do not attempt to bypass the manual assignment step by introducing a PAT or OAuth user-to-server token. That defeats the point of the walkthrough — see README.md.
• Do not run reset while a Copilot fix-PR session is in progress; the agent will lose context. Wait for the fix PR to be opened (or abandoned) first.