work-items(triage): absorb v4 loop-prompt routing rules into the skill

[kyle-sexton]

Category: behavioral · Source: 2026-07-19 deep-dive audit (external audit: 13/13 routes correct; improvements below are queue-reduction + consistency, not accuracy fixes). Session-scoped prompt rules that are skill gaps.

Absorb into the skill

• Ready-with-noted-decision route: single-fork items where the T2 plan states a well-grounded RECOMMENDED answer and the alternative is maintainer-vetoable → status: ready + "Decision defaulted: X — veto before merge", instead of needs-human. Reserve needs-human for open design space / product intent / cross-repo policy.
• Cluster-aware routing: several issues sharing ONE underlying decision → route one representative as the decision carrier (members listed in its body); members get "blocked by #carrier decision" without individual needs-human labels. One human touch per decision.
• Multi-surface T1 stub: T1 fixes spanning 3+ surfaces get a one-line "sites + fix pattern" comment.
• Severity sub-sort: when a directive/category rule sets priority above a finding's self-labeled severity, record the severity in the triage comment for implementer sub-sorting.
• Autonomous-mode contract: reconcile step-2 "wait for user direction" with standing-lane operation (= work-items(triage): step-2 "wait for user direction" gate contradicts the skill's own autonomous/agent triage mode #459) — codify that standing lane rules constitute direction.
• Cross-lane label semantics: define whether/how triage applies wayfind:* types (self-audit applied them inconsistently without mandate; sanctioned-uniform in v4 prompt pending maintainer veto).

Related

#459 gate contradiction · #463 in-progress marker · github-iac#176 complexity labels

Stays prompt-level

Week priority directive (north-star plugin list + weighting); repo scope.

[kyle-sexton]

This was generated by AI during triage.

Absorbing the v4 loop-prompt routing rules into the triage skill. Five of the six rules are real gaps in the triage skill's routing workflow. Bullets 1–4 are mechanical absorption with at most maintainer-vetoable forks — each maps onto an existing section, label, or native edge with no new taxonomy. Bullet 5 is out of scope here. Bullet 6 is genuinely open: it is a cross-repo label-policy question this skill cannot answer unilaterally, which gates the whole item to a human.

Per-bullet checklist

• 1. Ready-with-noted-decision route — absorb (mechanical). Not currently documented. The briefed state has exactly two exits today: the delegable outcome (autonomous-eligible role, default agent-ready) and the needs-human-judgment outcome (human-gated role, default needs-human). A single-fork item whose brief carries a well-grounded RECOMMENDED answer with only a maintainer-vetoable alternative currently has to fall to human-gated. Change: add a third briefed outcome to the "Apply outcome" table and a matching target-state option in "Recommend category + state" — briefed, decision-defaulted: route to the autonomous-eligible role with status:ready, and record a triage comment line "Decision defaulted: X — veto before merge". Tighten the human-gated outcome's definition in the "Triage states" side-exits to reserve it for genuinely open design space, product intent, or cross-repo policy. Add the routing test to step 2: is the alternative reversible/maintainer-vetoable (→ defaulted) or open (→ human-gated)?
• 2. Cluster-aware routing — absorb (mechanical; uses existing primitives). Not documented; triage evaluates one item at a time and has no decision-carrier concept. Change: in "Gather context" add a cross-reference step — when the item shares one underlying decision with other open intake, designate one representative as the decision carrier (human-gated, member numbers listed in its body) and link each member to it via the native blocked-by edge with a triage comment "blocked by # decision", instead of applying the human-gated role to each member. One human touch per decision. No new labels; reuses human-gated + the native dependency edge already named in "Triage states".
• 3. Multi-surface T1 stub — absorb (mechanical; one small fork). Not documented; the delegable outcome requires a full agent brief. Change: add a lightweight briefing variant to the "Apply outcome" table for trivial (T1) fixes spanning 3+ surfaces — in place of a full brief, post a one-line "sites + fix pattern" comment and still apply the autonomous-eligible role. Maintainer-vetoable fork: keep the agent-brief durability rule (name sites by interface/symbol/domain concept, not file paths or line numbers). Recommended default: symbol-level naming, consistent with the existing brief durability principle.
• 4. Severity sub-sort — absorb (fully mechanical, additive). Not documented. Change: in "Recommend category + state" (priority-label step) and the outcome comment, when a directive or category rule sets the priority: label above the finding's self-labeled severity, record the original severity in the triage comment (e.g. "priority set to pX by ; reporter severity: ") so implementers can sub-sort within a priority band. No new labels.
• 5. Autonomous-mode contract — defers to work-items(triage): step-2 "wait for user direction" gate contradicts the skill's own autonomous/agent triage mode #459. The step-2 "wait for the user's direction before mutating anything" vs standing-lane operation reconciliation is that atomic issue's fix; not re-planned here.
• 6. Cross-lane wayfind:* label semantics — open design; gates this item. Triage documents no mandate for applying wayfind:* types, and the label taxonomy carries no such axis. Two open questions block absorption: (a) whether wayfind:* is a legitimate triage-applied axis at all, and (b) if so, its members and write policy. Both are label-as-code decisions the taxonomy reference routes to the declared owner — triage stays read-only over labels. This is cross-repo label policy (github-iac; cf. github-iac#176 complexity labels), not a call this skill can default.

Acceptance criteria

• "Triage states" documents three briefed exits (delegable, decision-defaulted, human-gated) and the human-gated definition is narrowed to open design / product intent / cross-repo policy.
• "Recommend category + state" and "Apply outcome" include the decision-defaulted route with the verbatim "Decision defaulted: X — veto before merge" comment convention.
• "Gather context" documents cluster detection; the outcome section documents the decision-carrier pattern using the native blocked-by edge and a "blocked by # decision" member comment, with no per-member human-gated label.
• "Apply outcome" documents the T1 multi-surface stub variant and states its durability-naming rule.
• The priority-label step documents recording self-labeled severity when a rule-assigned priority diverges upward.
• Bullet 5 is captured only as a pointer to work-items(triage): step-2 "wait for user direction" gate contradicts the skill's own autonomous/agent triage mode #459; no fix text lands here.
• wayfind:* semantics are NOT changed in this skill until the cross-repo label-policy decision returns.

Needs human because: bullet 6 (wayfind:* type semantics) is genuinely open — defining whether triage applies wayfind:* and what its members/write policy are is cross-repo label policy owned by melodic-software/github-iac (label-as-code; related github-iac#176), which the taxonomy reference forbids this skill from setting unilaterally. Recommended human decision: either (a) split bullet 6 into its own github-iac-scoped issue and land bullets 1–4 now, or (b) hold #478 until the policy returns. Bullets 1–4 carry only maintainer-vetoable forks and are ready to absorb once the human resolves bullet 6's routing.