Ship issues, PRs, and review replies
Everything human-visible follows one iron rule (the gate): the agent drafts, you see the exact payload and target, nothing is sent until you approve. These flows produce the text; the k-github skill performs the click, gated.
Prerequisites: a session; gh authenticated for the actual side effects.
Draft an issue
draft an issue: todo.py crashes with a TypeError traceback when the db file is the old string format
The draft comes back with Problem / Expected / Actual / numbered Reproduction — with your local paths, hostnames, and session junk scrubbed into portable steps (that sanitization is the skill's job, not yours). If the repro or environment detail is missing, it says so explicitly instead of inventing one.
Then: file it → the agent shows the final title/body/repo/labels once more and creates it only after your yes.
Draft a PR body
draft the PR text for this branch
The Test Plan section is the strict part: it contains commands actually run with their observed results — if manual verification didn't happen, the draft says so and gives reviewers the steps instead of pretending. Closes #N only appears when the issue is genuinely closed by the change.
Reply to review threads
address the review comments on PR #4321
Fixable comments get fixed in code; the rest get drafted replies, shown to you per-thread. Human-authored threads are never auto-replied — you approve each send. Only verified bot threads inside a flow you explicitly invoked can auto-resolve.
Present a PR before review
/k-present-pr
Builds a self-contained HTML walkthrough of the PR (what changed, in what order to read it, which files are load-bearing vs mechanical) that you can hand to reviewers. It verifies its own output renders (browser-checked, zero console errors) before showing you. Nothing is posted anywhere.
Pivots from here
- Issue you're drafting turns out to already exist → the necessity check in spec catches duplicates before writing; ask
has anyone filed this?any time. - PR body needs the change verified first →
verify this works end-to-endbefore drafting, so the Test Plan writes itself from real runs.