triage

Triage one batch of untriaged issues: issues without the triaged marker, which apply stamps on every processed issue (title patterns and kind labels prove nothing — reporters set both). The issue templates add needs-triage as a visible awaiting signal; apply removes it. The state lives in the labels, so there is no bookkeeping file.

The classification needs codex logged in with quota (codex login status). Without it, stop and say so: issue bodies are untrusted input, so the classification never moves to an unsandboxed agent.

1. Snapshot

python3 .claude/skills/triage/scripts/snapshot.py

Fetches every open untriaged issue (body + comments) into /tmp/clice-triage/issues/chunk-N/ (25 per chunk), plus digest.json (new issues in the last 7 days, needs-info/needs-repro threads with no activity for over 14 days), existing-labels.json, titles.json and states.json (what apply checks for drift). --state all sweeps closed issues too, paired with apply's --include-closed, for a backfill. With zero untriaged issues, skip straight to the digest section of the report.

2. Classify

One codex call per chunk; run chunks as parallel background jobs:

codex_root="$(cd "$(dirname "$(command -v codex)")/.." && pwd)"
systemd-run --user --pipe --wait --collect --same-dir \
  --setenv=PATH="$codex_root/bin:/usr/bin:/bin" \
  -p ProtectHome=tmpfs \
  -p BindReadOnlyPaths="$PWD" \
  -p BindReadOnlyPaths="$codex_root" \
  -p BindPaths="$HOME/.codex" \
  codex exec -m gpt-6-astra -c model_reasoning_effort=xhigh \
  --sandbox read-only \
  -o /tmp/clice-triage/verdicts-N.md \
  "Read .claude/skills/triage/rules.md, .github/labels.yml, and every
issue file in /tmp/clice-triage/issues/chunk-N/. Work ONLY from these
local files — no gh, no network. Classify every issue per the rules and
reply with ONLY the JSON array defined by the rules' output schema." < /dev/null

Both sandbox layers are mandatory, never the usual full bypass. Codex's --sandbox read-only blocks writes and command network access but still lets model-run commands read the whole filesystem, so an injected issue could exfiltrate credentials through the verdict text. The systemd-run wrapper masks $HOME and rebinds only the repo (read-only), the codex install prefix, and ~/.codex (codex's own state — the one residual exposure).

3. Validate

python3 .claude/skills/triage/scripts/validate.py /tmp/clice-triage/verdicts-*.md

A deterministic gate over the model output: every label exists in .github/labels.yml, exactly one kind:, no good first issue or help wanted, no os:wsl beside a native os label, every snapshot issue covered. It writes /tmp/clice-triage/validated.json with add (proposed minus existing labels) and suggest_remove (existing labels the model omitted — reported, never removed). A verdict that fails validation is reported, never applied, and never hand-edited back in.

4. Report

Return to the main conversation:

• Proposed changes, one line per issue: #N [conf] +labels — rationale, with title → / ask → sub-lines where the model proposed them.
• Validation failures and taxonomy gaps.
• Digest: new issues this week, stale waiting threads (with day counts), open/untriaged totals.

Do NOT apply anything in the forked run — the maintainer reviews the proposals first.

5. Apply (main conversation, after approval)

python3 .claude/skills/triage/scripts/apply.py /tmp/clice-triage/validated.json \
  [--only N,N | --skip N,N] [--retitle N,N | --retitle all] [--include-closed]

apply refetches each issue's live labels and title. It skips an issue whose state changed since the snapshot, that gained triaged meanwhile, a closed one without --include-closed, and one that needs manual resolution (several maintainer kinds, os:wsl beside a native os label). While needs-triage is present the model's kind replaces the template's; once it is gone the existing kind is the maintainer's, and when the model proposes another kind, the maintainer's wins and the model's kind, status: labels and title are dropped. A closed issue loses every status: label. Beyond that apply only adds labels — removals stay manual via suggest_remove — and stamps triaged in a separate final call, so a failed edit leaves the issue visible to the next snapshot. Titles change only for the issues --retitle names whose title did not change since the snapshot. ask_reporter suggestions are never posted — the maintainer sends them personally if worthwhile.