Appraise GitHub Issues

One read-only agent turn per issue answers two questions and stops: is it still real, and how big is it. The answer lands on the issue, and that is the whole contract. Nothing reads this process; the burndown, a person, or another skill reads the tracker.

This feature is about an issue; its children are appraisals: one verdict about one issue, judged against the fetched base ref at one moment.

The second opinion on a close

A close is the one appraisal verdict that is a write with consequences: a wrongly kept issue costs a re-read, a wrongly closed one disappears. So before an issue closes, a confirmer on another engine, with none of the appraiser's context, re-checks the receipt against the base ref: the commit is an ancestor and does what was asked, or the file as the base holds it no longer has the premise. Agreement closes the issue with both receipts on the thread. Disagreement labels it needs-human with both opinions, and a person breaks the tie. A confirmer that crashes or writes nothing is not evidence either way; the close is not made and the issue stays unsized.

Sizing and hand-offs stay single-opinion. They are labels a person can change.

What happens after a size is not the appraiser's to say

A producer puts files named on-size-<N>, on-size-over-<M>, or on-size into a callbacks directory, and the appraiser looks one up for the points it just applied and runs it: an executable with the appraisal as JSON on stdin, a .md prompt as one agent turn on the callback seat (a seat is a role an agent fills; this one runs callback prompts), or both. The burndown writes the callbacks it ships there on every start, and that is how carve-github-issue, which breaks an oversized issue into smaller ones, is reached without the appraiser knowing it exists. The ladder, the two forms, and the payload are in references/callbacks.md.

Keep the appraiser and the confirmer on different engines; the driver refuses to start when they match. --no-confirm (or confirmCloses: false) closes on the appraiser alone, for a tracker whose authors you trust to that degree, and then the confirmer's engine does not matter.

Run it

From inside the target repository:

bun run <this-skill-dir>/appraise.ts --dry-run             # select and print; no lock, no agent, no GitHub write (only runs/appraise.log)
bun run <this-skill-dir>/appraise.ts --limit 12            # up to 12 unsized issues in the window
bun run <this-skill-dir>/appraise.ts --age-days 90         # widen the window to 90 days for this run
bun run <this-skill-dir>/appraise.ts --issue <n>           # one issue, whatever its age or size
bun run <this-skill-dir>/appraise.ts --all                 # the whole open backlog, not only the window
bun run <this-skill-dir>/appraise.ts --all --include-sized # re-judge everything, sized or not
bun run <this-skill-dir>/appraise.ts --every <minutes>     # heartbeat: appraise whatever is new, on a cadence you choose
bun run <this-skill-dir>/appraise.ts --appraiser codex:gpt-5.6-sol --confirmer claude:claude-opus-5
bun run <this-skill-dir>/appraise.ts --callbacks <dir>     # where a producer's size callbacks live; --callback-seat engine[:model]

<this-skill-dir> is the filesystem path of this directory wherever the collection is checked out; it is a path, not a skill name. The command reads appraise-github-issues.config.ts at the repository root, or the burndown's burn-down-github-issues.config.ts when that is what the repository has (references/adopting.md). Every run is teed to <worktreeRoot>/runs/appraise.log.

Read next

Standards this skill enforces

• Read-only. The appraiser and the confirmer edit no file, create no branch, install nothing, and run no suite. Their working directory is an empty scratch directory, not a checkout.
• The base ref is the evidence. "Already in the base" is judged against the fetched base branch, never the main checkout's working state.
• A decision beats a small diff. An issue that turns on a product or domain ruling nobody has made is needs-decision, however small the change.
• The written convention beats the issue's prescription. When an issue prescribes a remedy the repository's convention docs rule out, the appraiser sizes the convention-respecting remedy and says so; it does not stop for a human.
• Closes carry receipts, and receipts get re-checked. Narrative alone never closes an issue.

Hard dependencies

• The sibling fix-github-issue skill, imported for the agent runner, config loader, prompt renderer, and label helpers. It ships in the same collection.
• gh authenticated with issue-edit rights on the target repository.
• The CLIs the seats name (by default codex and claude) on PATH.

The ceiling, named

The appraiser executes the tracker's claims against the code, not your intent: an issue that is internally coherent and wrong is sized competently. Issue bodies and comments are an untrusted instruction channel read by agents with their approval gates bypassed; the read-only contract is a prompt, not a sandbox, so run this only on trackers whose authors you trust as far as the credentials the agents hold. The confirmer's independence is model-level, not identity-level; it runs as the same GitHub account. And a size is a judgement on the configured scale, not a measurement; disagreements between runs are recorded on the issue, not resolved by it.