work-items(attend-queue/triage): interactive direction gate lacks a brief-before-ask requirement — operator asked to decide without item context

[kyle-sexton]

This was generated by AI during triage.

Context

Observed during an attended-queue pass in melodic-software/provisioning. Both attend-queue ("Working the queue") and triage ("Direction gate") mandate recommend-then-wait for operator direction, but neither requires restating the item's identity, story, and each option's consequences at the decision point. Mid-flow, a fork question about one item's recheck infrastructure was posed with option labels only; the operator had to halt the pass and demand full context before they could answer ("what issue are you looking at... I need full context before I can answer questions"). The gap compounds under terse output styles: compression drops exactly the load-bearing context the gate never mandated.

Proposed work

Add a brief-before-ask rule to the interactive gates: every operator-facing decision question restates (1) which item, (2) the decision being asked, (3) the consequence of each option — before asking. Applies to attend-queue row-working and triage's interactive direction gate.

Acceptance criteria

• Skill text requires full context restatement before any operator decision question in the interactive gates

References

• work-items(tracker-seam): recurring-schedule mechanism presumed present in consuming repo — no entry invariant or remediation when the automation is absent #1201 (same pass, sibling observation)

[kyle-sexton]

This was generated by AI during an attended-queue pass.

Second occurrence, melodic-software/ci-runner attended pass (2026-07-23): the #64 route question was posed twice (option labels detached from the earlier context text), and the operator halted both times — "you give me no brief or show what the ticket is about... how can I answer that?" Identical failure shape to this issue's provisioning observation, again compounded by a terse output style.

Operator direction from that pass sharpens the acceptance bar beyond "restate context":

• Every operator-facing decision prompt must carry a link to the item plus a one-paragraph purpose summary — enough context to select without leaving the prompt.
• The recommendation is marked RECOMMENDED and listed first.
• The prompt is self-contained at the surface where the operator answers — never split across an earlier message and a detached question UI.
• Presentation design is explicit: what the operator is shown, in what order, and how many questions per round.

Routing the question rounds themselves through /planning:interview (plus a size gate and pre-ask discovery) is filed separately with a cross-reference, since it is a process-composition gap rather than this presentation gap.

[kyle-sexton]

This was generated by AI during triage.

Triage notes

Verified: attend-queue's and triage's interactive direction gates both instruct recommend-then-wait-for-direction before mutating, but neither mandates restating the item's identity, a link to it, or each option's consequence at the point a decision question is posed. The reported gap (operator forced to halt and demand context mid-flow) is a real skill-text omission, confirmed by two independent occurrences cited in this item's comments.

Agent Brief

Type: Task
Summary: Add a brief-before-ask requirement to the interactive direction gates (attend-queue row-working, triage's interactive gate) so every operator-facing decision question is self-contained at the point it's asked.

Current behavior:
The interactive direction gates say to recommend an option and wait for operator direction, but say nothing about what the question itself must contain. A decision question can be posed with option labels only, detached from the item's identity and the earlier context that explained it — forcing the operator to halt and ask "what item is this" or "what's the context" before they can answer.

Desired behavior:
Every operator-facing decision question in an interactive direction gate is self-contained at the surface where the operator answers — never split across an earlier message and a detached question. Each such question includes: which item is under discussion (with a link to it), a one-paragraph summary of what it's about, the decision being asked, and the consequence of each option. The recommended option is marked RECOMMENDED and listed first. Presentation is explicit about what's shown, in what order, and how many questions run per round.

Key interfaces:

• attend-queue's row-working procedure — the point where it poses a decision question about the row under evaluation
• triage's interactive direction gate (step 2, "Recommend category + state") — the point where it presents a recommendation and waits

Acceptance criteria:

• attend-queue's row-working text requires the item link + one-paragraph summary + decision + per-option consequences before any operator decision question
• triage's interactive direction gate text carries the same requirement
• Both texts require the RECOMMENDED option marked and listed first, consistent with the existing global communication convention

Out of scope:

• Routing question rounds through /planning:interview (filed separately as a distinct process-composition gap)
• Any change to the autonomous-lane direction-gate behavior (which does not pose operator questions)

[kyle-sexton]

This was generated by AI on operator direction.

Ratified — admitted for autonomous dispatch.

The operator delegated this ratification decision explicitly. Ratified on the basis that the item is briefed to a behavioral contract, its scope is bounded, and its classification was verified against the issue body rather than its title: it is genuinely C3 scoped/bug-fix-shaped, with no open design decision and no path/topic hard-gate surface (no dependency pins, no checksum recomputation, no declared-inviolable ground rules).

Restored to the autonomous frontier: agent-ready applied, needs-human removed.

first_drain_complete is deliberately left false. This is the lane's first cycle, with zero completions and therefore no evidence base; the autonomy contract's promotion apparatus expects an evidence predicate over telemetry, not a blanket flip. Ratifying per item admits the work that is actually verified while keeping the gate armed for future unratified intake — and demotion stays fail-closed if these items produce contrary evidence.