feat(work-items): decision-session skill (/work-items:decide) — batched operator decision loop over needs-human/needs-decision

[kyle-sexton]

Problem

The operator decision queue (needs-human + status: needs-decision) has no owning skill. Processing it today is improvised: a pasted operating prompt drives a one-off pass (classification, decision briefs, re-intake transitions), and the interactive part — the operator actually answering — has no repeatable mechanism at all. Decisions are the pipeline's scarcest resource; the loop that consumes them should be as codified as triage and work.

Proposal

New skill in the work-items family: /work-items:decide — a batched operator decision session.

Flow per invocation:

1. Load the queue via the work-item-tracker seam: open items labeled needs-human or status: needs-decision, ranked priority-first then oldest.
2. Ensure briefs. Any item without a decision brief (RECOMMENDED option + verified today-state + reversibility per option) gets one dispatched to a fresh-context brief agent before it enters a batch. Brief supply contract is work-items(triage): pre-planning enrichment path — attach explore/research/plan/design artifacts to issues before the human decision point #515's enrichment path — this skill consumes it, and enforces brief-before-gate: nothing sits in the decision queue without an answerable brief.
3. Present ranked batches (5-8 items) interactively: per item — one-line problem, one-line RECOMMENDED, reversibility flag. Operator answers per item: accept / modify: <words> / discuss / defer.
4. Record decisions durably. Each accept/modify becomes a decision comment on the issue in the operator's words (quote-attributed relay), then the re-intake transition applies: remove needs-human, add priority: needs-triage so the triage lane routes the unblocked work.
5. Discuss tier items get one-axis-at-a-time interactive resolution in-session; deep design items route to /planning:design or /planning:interview instead of being forced through the batch.
6. Summary telemetry: counts per disposition posted to the lane-telemetry issue.

Quality bar (from the working prototype)

• Every re-intake transition defensible from quoted evidence alone — explicit decisions only; inferred = stays gated.
• Every brief answerable in under 2 minutes without opening files.
• Operator control is absolute: the skill never selects an option; it relays the operator's words.

Prototype evidence

The 2026-07-19 needs-human queue pass is the working prototype: 5 re-intake transitions on quoted operator decisions, 11 briefs posted, 10 duplicate-brief skips, summary at #502 (comment).

Cross-links

• work-items(triage): pre-planning enrichment path — attach explore/research/plan/design artifacts to issues before the human decision point #515 — brief/enrichment supply (this skill is its consumer; brief-before-gate is the shared contract)
• ops: morning-ritual dashboard — one command for queues, merge-ready, decision list #507 — morning-ritual dashboard (surfaces the decision list this skill processes)
• work-items(triage): absorb v4 loop-prompt routing rules into the skill #478 / work-items(work): absorb v4 loop-prompt execution rules into the skill #479 / source-control(babysit-prs): absorb v4 loop-prompt operating rules into the skill #477 — loop-prompt rule absorption (same rule-lifecycle path: prompt → proven → skill)
• .work/leave-week-priorities/operating-procedures.md decision protocol — the codified source this skill absorbs

[kyle-sexton]

This was generated by AI during triage.

T3 — routing needs-human + wayfind: design. Genuinely open design space: a new skill surface (/work-items:decide) with multiple integration points (#515 brief-supply contract, #507 dashboard, #478/#479/#477 rule-absorption precedent) and product-shape questions (batch size, discuss-tier routing to /planning:design vs /planning:interview, decision-comment attribution format) that aren't reducible to a single maintainer-vetoable fork. Prototype evidence is strong (cited #502 pass) but this needs a design pass, not a default.

[kyle-sexton]

This was generated by AI (decision brief for the wayfind: design design pass on this issue).

Decision brief — /work-items:decide skill shape

Problem

The operator decision queue (needs-human + status: needs-decision) is the pipeline's scarcest resource, yet it has no owning skill — the interactive part (the operator actually answering) is improvised from a pasted prompt each time. This issue asks us to codify that loop into a /work-items:decide skill; the open design space is three orthogonal product-shape questions (batch size, discuss-tier routing, decision-comment attribution) plus one load-bearing cross-plugin constraint (#440 outward-write posture), none reducible to a single maintainer-vetoable fork.

Today (verified)

• Skill family shape is established. plugins/work-items/skills/ holds triage, work, track, decompose, scan-todos, setup — decide would be a seventh sibling. Skills live under plugins/*/skills/<name>/SKILL.md (verified via the repo tree).
• The tracker seam this skill would load the queue through exists. plugins/work-items/tools/work-item-tracker/ with CONTRACT.md, work-item-tracker.sh, a GitHub adapter under adapters/github/, and plugins/work-items/reference/tracker-seam.md. This is the same seam planning: wayfind hardcodes GitHub Issues/gh as the only tracker backend, bypassing the work-item-tracker seam #416 flags planning:wayfind for bypassing (it hardcodes gh).
• AI-attribution and operator-relay conventions are real and consistent. Every AI-authored comment carries the italic disclaimer *This was generated by AI …* (contextualized per pass), verified in plugins/work-items/skills/triage/SKILL.md and across the prototype pass. Operator/maintainer rulings are relayed with a bold header + > blockquote of the operator's words — e.g. on source-control: no read/consume-only posture — "safe" babysit tier and pull-request monitor write to GitHub autonomously #440: **Maintainer decision (2026-07-19, batch-approved; relayed by planning session).** followed by a quoted enum, and the re-intake carrier *This was generated by AI during needs-human queue processing.* **Re-intake: the human gate here is already answered.** with a link to the deciding comment. There is no literal "Operator decision" string anywhere in the repo today — the attribution axis below is about whether to coin one, not reuse an existing one.
• Correction to the triage note's premise: ops: morning-ritual dashboard — one command for queues, merge-ready, decision list #507's morning-brief dashboard is not yet shipped. PR feat(claude-ops): add morning-brief operator dashboard skill #569 is OPEN (mergedAt: null), and plugins/claude-ops/skills/morning-brief/ does not exist on the default branch (404). ops: morning-ritual dashboard — one command for queues, merge-ready, decision list #507 itself is still open and status: needs-decision. So "ops: morning-ritual dashboard — one command for queues, merge-ready, decision list #507 surfaces the queue this skill processes" is an in-flight integration point, not a settled one — the two are being designed in the same window.
• Prototype evidence is strong and real. The 2026-07-19 needs-human pass (ops: loop-lane telemetry — running per-lane status (babysit / triage / work) #502 summary): 5 re-intake transitions on quoted operator decisions, 11 briefs, 10 dedup skips. A later live decide-lane cycle already ran under /loop and posted lane telemetry to ops: loop-lane telemetry — running per-lane status (babysit / triage / work) #502 — so a v1 of the briefing half is running in prototype; the interactive operator-answer half is what remains uncodified.

The load-bearing constraint (must be decided, not deferred)

This skill's whole job is to write: decision comments and the needs-human → priority: needs-triage label transition. #440 proposes a canonical per-plugin outward_write_mode: report-only | write with default report-only, and the prototype summary explicitly flagged that "report-only collides with any mechanism that posts to GitHub." So /work-items:decide cannot ship without a stance:

• RECOMMENDED: the decision-comment write is exempt from report-only — it is an interactive relay the operator authorizes in real time, in-session, one decision at a time (not an autonomous side effect), which is precisely the category source-control: no read/consume-only posture — "safe" babysit tier and pull-request monitor write to GitHub autonomously #440's fail-safe is designed to catch and this is not. But the label mutation half (needs-human off / needs-triage on) MUST honor outward_write_mode: under report-only the skill computes and prints the transition for the operator to apply, and only auto-applies under write. Rationale: the comment is the operator's own words going out under their authorization; the label transition is an autonomous state change and belongs behind the posture gate. Reversibility: high — both halves are behind a single config key; posture can flip per-repo without touching skill logic. This resolution should be coordinated with source-control: no read/consume-only posture — "safe" babysit tier and pull-request monitor write to GitHub autonomously #440 before decide flips to ready.

Product-shape decisions (orthogonal — decide each independently)

A. Batch size for the decision loop

• RECOMMENDED: default 5, operator-overridable, hard cap 8. Reversibility: high (a single default constant). Rationale: the proposal says 5-8; the prototype ran briefing caps of 10 and 15, but those were non-interactive brief-drafting passes, not interactive answer sessions. Interactive attention degrades faster than batch drafting, so anchor at the low end of the proposal's own range and let the operator raise it. Start conservative because raising a cap is trivial and lowering it after operators habituate to 15 is not.
• Alt: fixed 8 (matches upper bound, fewer round-trips) — reversibility high, but front-loads more items per decision context.

B. Discuss-tier routing — /planning:design vs /planning:interview

The proposal has three tiers: accept/modify resolved inline, discuss resolved one-axis-at-a-time in-session, and deep items routed out. Two sub-questions:

• Inline discuss vs route-out threshold — RECOMMENDED: route out only when in-session one-axis resolution stalls or the item needs artifact production (explore/research/design drafts), i.e. work-items(triage): pre-planning enrichment path — attach explore/research/plan/design artifacts to issues before the human decision point #515's enrichment territory; otherwise keep it inline. Reversibility: high (a per-session judgment, no persisted state).
• When routing out, design vs interview — RECOMMENDED: follow the item's existing wayfind: signal rather than inventing a new classifier. Reversibility: high. Rationale (root-cause aligned): the repo already carries this taxonomy — wayfind: interview = "contract/requirements call the user must make" (e.g. planning: wayfind hardcodes GitHub Issues/gh as the only tracker backend, bypassing the work-item-tracker seam #416), wayfind: design = "design-space or domain-model decision" (e.g. source-control(babysit-prs): absorb v4 loop-prompt operating rules into the skill #477, and this issue feat(work-items): decision-session skill (/work-items:decide) — batched operator decision loop over needs-human/needs-decision #556). That maps verbatim onto the two skills' own charters (planning:interview locks a contract / reaches shared understanding; planning:design explores design space). So requirements/contract-shaped → /planning:interview; design-space/domain-model-shaped → /planning:design. Items lacking a wayfind label default to interview (cheaper to discover the shape than to over-invest in design). This reuses a proven signal instead of adding a second, drift-prone classifier.

C. Decision-comment attribution format

• RECOMMENDED: coin a distinct **Operator decision (<date>; relayed by /work-items:decide)** header, keeping the verified *This was generated by AI …* disclaimer + > blockquote of the operator's exact words. Reversibility: medium (the string convention gets referenced by the morning-brief/dashboard extractor and any future audit tooling, so a later rename has a small blast radius — cheap now, mildly sticky later). Rationale: the relay mechanics (disclaimer + blockquote) are already proven precedent and are not up for debate; the only open choice is the header label. Coining Operator decision (vs reusing the existing Maintainer decision … relayed by …) distinguishes the routine decide-loop operator role from a repo-maintainer batch ruling, which keeps the two readable apart in history and greppable for the dashboard's RECOMMENDED/decision extraction (ops: morning-ritual dashboard — one command for queues, merge-ready, decision list #507/feat(claude-ops): add morning-brief operator dashboard skill #569). Be explicit in the skill that this string is new, not inherited.
• Alt: reuse **Maintainer decision … relayed by …** verbatim (zero new convention, but conflates operator-loop decisions with maintainer batch rulings) — reversibility high.

Cross-links

• work-items(triage): pre-planning enrichment path — attach explore/research/plan/design artifacts to issues before the human decision point #515 — brief-supply / enrichment path. Core consumer relationship: decide enforces brief-before-gate (nothing sits in the queue without a <2-min answerable brief) and consumes work-items(triage): pre-planning enrichment path — attach explore/research/plan/design artifacts to issues before the human decision point #515's artifacts; the "route-out for artifact production" threshold in (B) is exactly work-items(triage): pre-planning enrichment path — attach explore/research/plan/design artifacts to issues before the human decision point #515's territory.
• source-control: no read/consume-only posture — "safe" babysit tier and pull-request monitor write to GitHub autonomously #440 — outward_write_mode. The load-bearing constraint above; must be reconciled before ready.
• planning: wayfind hardcodes GitHub Issues/gh as the only tracker backend, bypassing the work-item-tracker seam #416 — tracker-seam bypass. Same seam-vs-raw-gh tension: decide must load the queue through work-item-tracker, not hardcode gh, or it reintroduces the exact coupling planning: wayfind hardcodes GitHub Issues/gh as the only tracker backend, bypassing the work-item-tracker seam #416 is filed against.
• ops: morning-ritual dashboard — one command for queues, merge-ready, decision list #507 / feat(claude-ops): add morning-brief operator dashboard skill #569 — morning-brief dashboard. Surfaces the decision list decide processes; in-flight (PR feat(claude-ops): add morning-brief operator dashboard skill #569 open), co-designed in this window. The attribution string in (C) is what its RECOMMENDED/decision extractor greps.
• work-items(triage): absorb v4 loop-prompt routing rules into the skill #478 / work-items(work): absorb v4 loop-prompt execution rules into the skill #479 / source-control(babysit-prs): absorb v4 loop-prompt operating rules into the skill #477 — loop-prompt rule absorption. Lifecycle precedent (prompt → proven → skill), not an integration point: decide is the same prompt→skill graduation these apply to triage/work/babysit, so follow their absorption shape.
• issue-metadata conventions: operationalize native types, sub-issues hierarchy, dependencies, priority re-baseline, close-reason + documented conventions (umbrella) #552 — issue-metadata conventions (light). The re-intake transition mutates needs-human/priority: needs-triage; if issue-metadata conventions: operationalize native types, sub-issues hierarchy, dependencies, priority re-baseline, close-reason + documented conventions (umbrella) #552 re-bases priority or moves to native fields, decide's transition step inherits that change.

(Deliberately omitted #476 — it was an item the prototype re-intook, an example decision, not a design integration point for this skill.)

[kyle-sexton]

This was generated by AI (decide lane — brief freshness supplement).

Freshness supplement — not a replacement

This supplements the decision brief above; it does not replace it. The brief's Today (verified) section carried a self-flagged correction ("#507 dashboard is not yet shipped, PR #569 is OPEN, morning-brief skill does not exist") that was accurate when written and is now stale. Three load-bearing facts flipped in the same design window the brief anticipated.

What changed (load-bearing)

• PR #569 is MERGED (2026-07-20 04:40 UTC), not open.
• #507 is CLOSED (2026-07-20 04:40 UTC), not open + needs-decision.
• The morning-brief skill now EXISTS on the default branch: plugins/claude-ops/skills/morning-brief/SKILL.md.

The brief treated #507 / #569 as an in-flight, co-designed integration point. It is now a settled, shipped surface — so the integration edge can be re-derived against what actually shipped rather than a moving target.

Re-derivation against current reality

RECOMMENDED: the /work-items:decide proposal STANDS at full scope — shipping morning-brief does not shrink or reshape it, and confirms the read/write division of labor the brief assumed. Two shipped-surface facts drive this:

1. Complementary read/write split over a shared queue — not competition. Both surfaces enumerate the same needs-human/needs-decision queue, but on opposite sides. The shipped morning-brief skill is read-only and never mutates (its own SKILL.md: "never mutates labels, comments, issues, or PRs"); its "Parked decisions" section surfaces open status: needs-decision issues with their RECOMMENDED lines — a gh-direct glance, the read side ops: morning-ritual dashboard — one command for queues, merge-ready, decision list #507 owned. decide is the write/answer side: the interactive operator-answer loop, decision-comment relay, and needs-human → needs-triage re-intake, loaded through the tracker seam (per planning: wayfind hardcodes GitHub Issues/gh as the only tracker backend, bypassing the work-item-tracker seam #416) for mutation-safe processing rather than raw gh. morning-brief surfaces the queue; decide processes it. So nothing feat(work-items): decision-session skill (/work-items:decide) — batched operator decision loop over needs-human/needs-decision #556 proposed is now covered by a shipped surface — and the natural "why not have decide just consume morning-brief?" is answered by the seam requirement: decide must load through work-item-tracker, which morning-brief's read-only gh glance deliberately does not.

2. Correction to the original brief's decision (C) rationale. The brief recommended coining an **Operator decision (…; relayed by /work-items:decide)** header partly because it would be "greppable for the dashboard's RECOMMENDED/decision extraction (ops: morning-ritual dashboard — one command for queues, merge-ready, decision list #507/feat(claude-ops): add morning-brief operator dashboard skill #569)." Now that the skill shipped, that justification does not hold: morning-brief's extractor greps the uppercase RECOMMENDED marker (SKILL.md line 44: "the uppercase marker wins over an incidental lowercase mention"), which lives in the issue/brief body. The **Operator decision** header lives in a separate decision-record comment and carries no RECOMMENDED marker, so morning-brief's extractor cannot and does not surface it — regardless of label state. The dashboard never greps the decision header. Coining the distinct header is still RECOMMENDED, but now on history-readability / audit-greppability grounds alone (distinguishing routine operator-loop decisions from maintainer batch rulings in issue history) — drop the "dashboard extracts it" rationale as falsified.

One shape addition, from this live decide-lane loop as prototype. This very supplement is a decide-lane brief-freshness sweep — the loop detected that #556's brief had gone stale (its OPEN/unshipped assertions overtaken by the #569 merge and #507 close) and re-derived against current state. That is direct first-hand evidence for a step the original brief did not name: decide must re-verify brief freshness immediately before presenting an item, not trust a brief drafted earlier in the queue's life. Briefs go stale between drafting and answering — demonstrated here, where #569 merged and #507 closed between the brief being written and this pass. RECOMMENDED: add a staleness guard to the flow — before an item enters an interactive batch, re-verify the brief's load-bearing "Today (verified)" claims (or timestamp + max-age them); stale ones get re-briefed by a fresh-context agent before the operator sees them. This is the same brief-writer-subagent / staleness-sweep / telemetry machinery the 2026-07-19 prototype (#502) and this running loop already exercise — evidence the skill's proposed shape holds and gains one guard.

Unchanged: the #440 outward_write_mode load-bearing constraint (comment-write exempt as interactive relay; label mutation honors the posture gate) still stands and still must be reconciled before decide flips to ready — shipping morning-brief does not touch it. Batch-size (A) and discuss-tier routing (B) recommendations are unaffected.

Net

Proposal stands at full scope. Adjust the brief on two points: (C)'s dashboard-greppability rationale is falsified — keep the coined header on audit grounds only; and add a brief-staleness guard to the flow, evidenced by this very sweep. Cross-links: #569 (merged), #507 (closed), #440, #515.