source-control:babysit-prs: SKILL.md sits at 499/500 lines, so the next contract change cannot reach the always-loaded surface

[kyle-sexton]

This was generated by AI on operator direction.

Lane: work (issue-execution loop) · Category: operational (maintainability constraint)

Observation

Surfaced while implementing #704. plugins/source-control/skills/babysit-prs/SKILL.md is 499 lines on main, against plugins/skill-quality/scripts/check-skill.sh's LINE_HARD_CAP=500 (the gate errors at >= 500, so 499 is the last passing value). CI runs this gate via scripts/check-changed-skills.sh "origin/$BASE_REF".

The practical effect: this skill's always-loaded surface has zero budget. Any future change that needs a Gotchas bullet, a mode-table row, or a new rule in SKILL.md fails CI on line count alone and must either be diverted into a reference/ file or bundled with an unrelated compression pass.

How it showed up

#704 mirrored #702's fix pattern, which landed in two surfaces of implement-dispatch: the brief-composition contract and a ## Gotchas reminder. In babysit-prs only the contract half fit. The Gotchas bullet was written, tripped the cap at 507 lines, and was reverted — the mitigation now lives only in reference/orchestration.md's Worker Contract and Worker Prompt Template.

That is defensible for #704 specifically (the Worker Contract is mandatory reading for any orchestrator composing a brief, so the fix still reaches every worker without a third copy to keep in sync). It is not a general answer: the next change may genuinely need the always-loaded surface.

Candidate directions (not yet evaluated)

• Compress SKILL.md — e.g. /docs-hygiene:compress, or relocate a section wholesale into an existing reference/ file behind a pointer.
• Re-derive whether 500 is the right cap for a skill of this complexity, or whether the cap should count differently for skills that already use progressive disclosure via reference/.
• Accept the constraint and document it, so the next author diverts to reference/ deliberately instead of discovering it through red CI.

No direction is recommended here — this is a filed observation for triage, not a briefed fix.

Related

• source-control:babysit-prs: Worker Contract has the same worktree-cwd-persistence gap #566 fixed in implement-dispatch #704 — where this surfaced
• docs(implementation): require worktree-cwd clause in dispatch briefs #702, loops/work: Bash tool cwd does not persist across calls despite tool docs — caused a worker's edits to land in the canonical checkout #566 — the sibling implement-dispatch fix whose two-surface pattern could not be fully mirrored

[kyle-sexton]

This was generated by AI during triage.

Verified

Both reported facts hold exactly, confirmed by direct read on main (7cad5cae):

• plugins/source-control/skills/babysit-prs/SKILL.md is 499 lines.
• plugins/skill-quality/scripts/check-skill.sh sets LINE_HARD_CAP=500, and Check 4 errors on
  ((LINE_COUNT >= LINE_HARD_CAP)) — so 499 is the last passing value and the always-loaded
  surface has zero remaining budget. The soft-cap warning (LINE_SOFT_CAP=200) is already firing on
  this file.

The observation is real and correctly stated. The consequence the item draws — that the next change
needing the always-loaded surface fails CI on line count alone — follows directly from the >=
comparison.

Classification

• Work class: none applied. The three candidate directions are not variants of one fix; they
  have materially different blast radii, and the class is a consequence of which direction is
  chosen, not a property of the item as filed:

  • Compress or relocate a section within babysit-prs behind a pointer — bounded, single-skill.
  • Re-derive whether 500 is the right cap, or whether it should count differently for skills that
    already use reference/ progressive disclosure — repo-wide policy, changing the gate every
    skill in the repo is measured against.
  • Accept and document the constraint so the next author diverts deliberately — docs-only.

  Rather than fabricate a class the evidence does not support, none is stamped (fail-closed).

• Priority: medium. priority: high is reserved here for work that blocks other work or
  carries an imminent external deadline. This blocks no in-flight item — it constrains the next
  change to one skill.

• Disposition: human-gated. The item explicitly declines to recommend a direction, and one of
  the three is a cross-cutting policy call on the shared skill-quality gate. This is genuine open
  design space, not a decision-defaulted single fork: the alternatives are not merely
  maintainer-vetoable, and choosing between them changes what the work is.

Cluster check

Searched the open population for an existing owner of the SKILL.md line-cap / body-budget policy
before gating this standalone. The nearest neighbor is #1271, which concerns the skill
listing/description budget — a different budget on a different surface — so this is not a cluster
member and no decision carrier applies.

Note for whoever decides

The item's own history is evidence for the third direction being insufficient on its own: #704's fix
pattern landed in two surfaces of implement-dispatch but only one of babysit-prs, and the
Gotchas half was reverted after tripping the cap at 507 lines. That mitigation now lives only in
reference/orchestration.md. Documenting the constraint records the tradeoff but does not restore
the budget.

Role-label resolution

.work-item-tracker.json carries no config.role_labels, so the canonical roles resolved to their
documented defaults (agent-ready / needs-human / recurring). Surfacing that explicitly rather
than defaulting silently — a remapped repo would otherwise be mislabeled. needs-human was verified
present in this repo's live label set before application.

[kyle-sexton]

Closing as fixed on main.

Evidence: babysit-prs/SKILL.md is now 429 lines (was 499/500 at filing). Headroom restored by extracting the runbook spoke (#2424).

wc -l plugins/source-control/skills/babysit-prs/SKILL.md → 429

Related duplicate #1626 is already closed.