Issue-work — complete the next most important issue, witnessed

The pick is an admission decision; the completion is a claim; only a witness closes it. Importance alone does not pick an issue — importance × feasibility does (an issue whose fix surface the SELF_MODIFY guard protects is not yours to take, however urgent). And "done" is never your narration: it is the issue's done-condition made TRUE, pinned by a test, landed in a commit whose subject dos commit-audit witnesses, with Fixes #N left for git ancestry — not you — to close. Worked example: #12 → commit 8e7f2f5 (the reference run of every step below).

Layering. Dev tooling that operates ON the repo (the /release tier) — it names a vendor (gh/GitHub), so it lives in .claude/skills/, never src/dos/skills/. The closing half is issue-verify; this skill is the completing half.

Public-repo note. Issue comments and labels are public documents — no dev-machine paths, hostnames, or private-process prose. Leak-scan every drafted body (python scripts/leak_scan.py --text-file <draft>) before posting; if the scanner is absent, the hand rule is the floor.

Step 1: Triage from the deterministic floor (skip if given an issue number)

python scripts/backlog_triage.py --top 12    # the typed, ordered queue (docs/315)

The script types every open issue into a closed disposition set and orders the offerable rows deterministically: priority tier → ready-label bias → freshness (the kernel's cooldown + pick-priority folds over the lane-journal OP_ATTEMPT history, unit id issue-N) → FIFO. The top row is the default pick. The floor is deterministic-first, advisory-only (the JUDGE-rung split): you may deviate, but state the one-line reason — the operator reads it to audit the triage, not to re-do it.

What each disposition hands you:

• READY (code) — implement it; the steps below.
• READY (execute-plan) — a design issue whose docs/NN plan exists: ship the plan's next unshipped phase (dos verify it first — never trust the plan's own status prose).
• NEEDS_PLAN (write-plan) — the unit of work IS the docs/NN plan: write it, link it from the issue, and stop there (implementation is a later pick).
• COOLING / T1_GATED / OPERATOR_GATED — not yours; surface it, don't pick it.

Two residues the floor cannot read, still on you: (a) the T1 detection UNDER-matches (it fires only when the issue text literally names a guarded runtime file) — before building, re-check your pick's real fix surface against _DISPATCH_RUNTIME_FILES in src/dos/self_modify.py; if the fix must edit a file in that set, the PreToolUse hook will deny you and the issue is operator-gated (say so in your report; a NEW file under src/dos/ is not in the set and is fine). (b) external blockage (a not-yet-published contract, a waiting upstream PR) lives in prose the floor doesn't parse — skip those with a stated reason.

Step 2: Verify the issue's claims before building

An issue body is a CLAIM about current behavior — probe it, don't trust it. Reproduce the defect and test the done-condition's examples against today's code first; pin what is actually true (#12's done-condition asserted a command "still DENIES today" that in fact never denied — the test that shipped pins reality, not the issue text).

Probe trap: the live hook adjudicates YOUR tool calls with the OLD code. A Bash command (or heredoc) that merely names a kernel runtime path may be denied — write probes and drafts to a file OUTSIDE the repo and run them by path, and keep kernel-path literals out of command strings until your fix is live.

Step 3: Implement inside the lane

• Stay in the issue's lane; out-of-scope findings become NEW issues (gh issue create, with a done-condition), never commit riders.
• Find the twins before declaring done. A change to hook/sensor logic has a Go fast-path twin (go/internal/hook/) plus a differential parity corpus: port the change, regenerate LF-safe (python gen_corpus.py | tr -d '\r' > corpus.jsonl), and run go test ./... from go/. Grep for the Python symbol name in go/ — the ports cite their Python originals.
• Closed-set changes prefer under-matching: an entry you are unsure of stays OUT (the conservative direction must be preserved by construction, and the PR should be able to say so).

Step 4: Test with the hot-tree discipline

Run the touched test file first, then the FULL suite foreground (python -m pytest -q, ~4 min — wait for the verdict). On failures:

1. Re-run the failing files in isolation. The shared working tree carries sibling loops' in-flight edits; README/workspace-config failures that pass isolated are the documented false positives.
2. A failure that touches YOUR surface is yours — fix it. A failure that doesn't is reported as "pre-existing, not mine", with the isolation evidence.

Step 5: Commit — scoped, atomic, pathspec'd

The index here is SHARED STATE and the sweep cuts both ways: a plain git commit after a narrow git add absorbs whatever a sibling left staged. One Bash call, pathspec on the commit itself:

git add <your files> && git commit -F - -- <your files> <<'EOF'
<type>(<scope>): <subject in the repo grammar>

<body: what was wrong, what the fix does, what is preserved>

Fixes #N
EOF

• Fixes #N in the BODY; the subject keeps its grammar. NO co-author trailer.
• Immediately verify: git show --stat HEAD contains exactly your files. Swept a sibling's staged work in anyway? git reset --soft HEAD~1, then re-commit with the pathspec form — their staged entries survive.
• Do not push, tag, or release — those are outward-facing and stay with the operator (the commit closes the issue when someone else's push lands it).

Step 6: Witness, then leave the evidence trail

dos commit-audit --workspace . <SHA>   # by SHA — HEAD may have moved already

A witnessed row is the honesty floor; an unwitnessed row means your subject claims what your diff doesn't show — fix the subject before anything else. Then one comment on the issue (drafted OUTSIDE the repo, leak-scanned, posted with --body-file): the fixing SHA, where the done-condition is pinned, and any operational caveat (e.g. "pre-fix binaries keep the old behavior until rebuilt"). Never close the issue yourself — closure is Fixes #N reaching master, or the issue-verify skill's evidenced path.

Step 7: Record the attempt (the cross-run memory)

python scripts/backlog_triage.py --record-attempt N --outcome shipped
#                                              outcome ∈ shipped|drained|blocked|error

Whatever happened, record it — shipped (landed), drained (nothing left to do in it), blocked (a gate/claim/external wall), error (partial progress, retry-worthy). This appends the lane-journal OP_ATTEMPT for unit issue-N, so the next sweep's cooldown fold holds a just-tried issue (the anti re-pick-storm memory) and the freshness fold prefers untouched work. A recorded blocked on a kernel-surface issue is what keeps the NEXT loop from walking into the same SELF_MODIFY storm.

Step 8: Priority-tag every issue you touched

The tier (create once if absent): priority:high (blocking or actively costing the fleet/operator) / priority:medium (real value, nothing bleeding) / priority:low (safe to defer).

gh issue edit <N> --add-label "priority:<tier>"

"Touched" = the issue you completed AND every issue you read deeply enough to rank during triage — the sweep's judgment is worth keeping, so write it down as labels, not just prose in a session log.

Step 9: Report

• The pick and why; the runner-ups and why not (one line each).
• What shipped: SHA, files, the done-condition's before/after.
• Witness results: suite verdict (with the isolated-rerun evidence for any pre-existing failures), parity/twin verdict, commit-audit row.
• Labels applied; any new issues filed for out-of-scope findings; any operator-gated work surfaced (T1 surfaces, pushes, external parties).