source-control/claude-ops: absorb control-tower merge-sitting + checkpoint flow-gauge rituals into skills (evals from tonight's real catches)

[kyle-sexton]

Category: skill extraction · Source: control-tower operations 2026-07-19/20 (proven over ~15 digest-verified merges)

Problem

The control-tower merge sitting is proven, repeatable ritual knowledge living only in .work docs and session context: eligibility query (green CI + zero unresolved threads + not draft + no do-not-merge) → per-PR digest agents (delivered-vs-issue, contract conformance, version-vs-main, scope/regression/leak check, merge-order constraints) → verdict (MERGE-CLEAR/HOLD with evidence) → thread-resolution under the explicit-defer pattern → squash-merge → closure verification. Same for the checkpoint flow gauge (created/closed, opened/merged, queue depths, projection, drain-or-grow verdict, two-checkpoint throttle rule).

Un-absorbed, this dies with the session and cannot be audited, eval'd, or improved by the pipeline that improves everything else.

Direction

Absorb into skill(s) — e.g. source-control:merge-sitting (or a babysit tier) + a claude-ops checkpoint/flow-gauge skill:

• Deterministic parts scripted (eligibility query, flow-gauge math, closure verification) per the script-the-deterministic-work discipline (loops: script the deterministic lane mechanics — machine-behavior block, #502 telemetry upsert, wire script-the-deterministic-work into prompt ritual #538).
• Judgment parts (digest, HOLD reasoning) stay model, with the digest brief as the skill's documented contract.
• Evals from tonight's real cases: the feat(ci): add skill-quality static gate lane for changed skills (#530) #593 umbrella auto-close catch, the fix(source-control): require per-thread pins for autonomous bulk babysit thread resolves #561 false-guarantee HOLD, the fix(work-items): reconcile triage step-2 wait-gate with autonomous mode (#459) #564-before-feat(work-items): absorb v4 loop-prompt execution rules into work skill (#479) #563 ordering call.
• Relation to feat(source-control): babysit autopilot merge tier — bot-account review + criteria-gated merge #476: autopilot merge tier subsumes the sitting for criteria-clean PRs; the sitting skill remains the manual/exception path and the digest contract feeds feat(source-control): babysit autopilot merge tier — bot-account review + criteria-gated merge #476's criteria.

Relations

#476, #534, #538, #477-#479 (absorption pattern), operating-procedures merge-protocol + decision-authority sections.

[kyle-sexton]

This was generated by AI during triage.

T3 — routing needs-human + wayfind: design. Genuinely open design space: skill boundaries (one skill or two — merge-sitting vs checkpoint flow-gauge), the deterministic-vs-judgment split for each ritual, and how the digest contract feeds #476's autopilot criteria are all product-shape questions, not a single maintainer-vetoable fork. Real evidence base (3 named catches) but needs a design pass to scope correctly.

[kyle-sexton]

Decision brief (design pass). ~2-min read. Not a decision — options for the human.

Problem

Two proven control-tower rituals — merge-sitting (per-PR digest → MERGE-CLEAR/HOLD verdict → resolve → squash → verify) and the checkpoint flow-gauge (created/closed, opened/merged, queue depths, projection, drain-or-grow, two-checkpoint throttle) — live only in .work docs and session context, so they die with the session and can't be eval'd or improved by the pipeline. The open design questions are the skill boundary (one skill or two), the deterministic-vs-judgment split per ritual, and how the digest contract feeds #476's autopilot criteria.

Today

• Merge-sitting is run by the operator by hand each cycle. Its deterministic front half already exists as code in babysit-prs's gate machinery: plugins/source-control/scripts/babysit-readiness-gate.sh (finding-classification completeness — the "unresolved threads" slice) plus the babysit_checks.py / babysit_merge.py scripts (CI-green, draft, do-not-merge, merge step), and plugins/source-control/skills/babysit-prs/SKILL.md already ships tiered merge autonomy (safe / worker / autopilot) behind that gate. What is not absorbed is the judgment half — the per-PR digest and HOLD reasoning.
• Checkpoint flow-gauge is also hand-run. Its natural neighbor is claude-ops telemetry: plugins/claude-ops/skills/observability/SKILL.md (read-only trend reporting) and the ops: morning-ritual dashboard — one command for queues, merge-ready, decision list #507 morning-brief dashboard (PR feat(claude-ops): add morning-brief operator dashboard skill #569), which already wants queue counts and the merge-ready list in one view.
• The 3 named catches read as judgment catches, not mechanical ones (characterizing from the issue's own labels): feat(ci): add skill-quality static gate lane for changed skills (#530) #593 umbrella auto-close (a merge would wrongly close an umbrella issue via linkage), fix(source-control): require per-thread pins for autonomous bulk babysit thread resolves #561 false-guarantee HOLD (a guarantee the PR claimed that the diff didn't deliver — a semantic delivered-vs-issue check), fix(work-items): reconcile triage step-2 wait-gate with autonomous mode (#459) #564-before-feat(work-items): absorb v4 loop-prompt execution rules into work skill (#479) #563 ordering (sequencing two green PRs by dependency). This is the core evidence for where the deterministic/judgment line falls — all three sit on the judgment side.

Internal precedent (the load-bearing facts)

• source-control(babysit-prs): absorb v4 loop-prompt operating rules into the skill #477-work-items(work): absorb v4 loop-prompt execution rules into the skill #479 set the pattern: absorb operating rules into the existing babysit-prs skill, not a parallel skill.
• source-control:babysit-prs: one shared authorship/finding/approval classifier + golden fixtures; gate script owns the merge-ready boolean (6-issue class) #534 rules that the gate script owns the merge-ready boolean — a second skill would fork or re-import that machinery.
• docs/PLUGIN-PHILOSOPHY.md: a plugin is one cohesive capability; plugins stay horizontally decoupled and cooperate only through a documented public seam (artifact contract). It also fixes skill names as imperative verbs — merge-sitting / flow-gauge are nouns and could not be minted as standalone skills without a verb.
• loops: script the deterministic lane mechanics — machine-behavior block, #502 telemetry upsert, wire script-the-deterministic-work into prompt ritual #538 / script-the-deterministic-work: script the mechanical parts, keep judgment in the model.

Together these point away from "two new noun-named skills" and toward absorb-where-a-home-exists.

---

Options

✅ RECOMMENDED — Option A: Absorb each ritual into the capability that already owns its deterministic half; cooperate via a documented digest contract

• Merge-sitting → into source-control:babysit-prs as the digest + verdict layer riding the existing readiness/merge gate (a capability, not a new source-control:merge-sitting skill). Deterministic (scripted, extends the existing gate): eligibility query, head-SHA-unchanged, version-vs-main diff, post-merge closure verification. Judgment (model, documented as the skill's digest contract): delivered-vs-issue conformance (catches fix(source-control): require per-thread pins for autonomous bulk babysit thread resolves #561), scope/regression/leak read, merge-order call (fix(work-items): reconcile triage step-2 wait-gate with autonomous mode (#459) #564-before-feat(work-items): absorb v4 loop-prompt execution rules into work skill (#479) #563), linkage-consequence check (feat(ci): add skill-quality static gate lane for changed skills (#530) #593 umbrella). All 3 catches land on the judgment side — exactly what the issue proposes and what the evidence supports.
• Flow-gauge → a claude-ops reporting surface (fold into observability / the ops: morning-ritual dashboard — one command for queues, merge-ready, decision list #507 morning-brief, or a new verb-named skill). Deterministic: the counts and projection math. Judgment: drain-or-grow verdict + the two-checkpoint throttle call.
• Digest → feat(source-control): babysit autopilot merge tier — bot-account review + criteria-gated merge #476 seam: because the merge tier (feat(source-control): babysit autopilot merge tier — bot-account review + criteria-gated merge #476) and the sitting path live in the same skill, the digest's per-criterion signals (delivered-vs-issue, contract-conforms, no-scope-leak, merge-order-clear) extend feat(source-control): babysit autopilot merge tier — bot-account review + criteria-gated merge #476's mostly-mechanical criteria set (checks-green, issue-linked, no blocking thread, no do-not-merge, SHA-unchanged) with the semantic judgments that set lacks — producer and consumer colocated, cooperation with claude-ops via artifact seam only (no cross-plugin import, per PLUGIN-PHILOSOPHY).
• Tradeoff / reversibility: most reversible — both moves are additive to skills that already exist; if merge-sitting later outgrows babysit-prs it can be extracted. Honors source-control(babysit-prs): absorb v4 loop-prompt operating rules into the skill #477, source-control:babysit-prs: one shared authorship/finding/approval classifier + golden fixtures; gate script owns the merge-ready boolean (6-issue class) #534, loops: script the deterministic lane mechanics — machine-behavior block, #502 telemetry upsert, wire script-the-deterministic-work into prompt ritual #538, and the naming grammar. Cost: babysit-prs SKILL grows; mitigate with progressive-disclosure context/ files (repo norm).

Option B: Two new dedicated skills — source-control:merge-sitting + claude-ops:<verb>

Cleaner conceptual separation, but forks/re-imports the readiness-gate + merge machinery babysit-prs already owns (fights #534), splits the digest contract away from the #476 merge tier that consumes it (a cross-skill seam where Option A has none), and fights the #477 absorb precedent and the verb-only naming rule. Reversible but higher standing coupling and more surface to maintain.

Option C: One combined skill for both rituals

Rejected on PLUGIN-PHILOSOPHY cohesion: merge-sitting is a PR-advancement capability (source-control), flow-gauge is pipeline telemetry (claude-ops); one skill cannot cohesively live in both plugins and would couple two independently useful capabilities. Least reversible.

---

Cross-links: #476 (autopilot merge tier — consumes the digest criteria), #507 / PR #569 (morning-brief dashboard — natural home/consumer of the flow-gauge signal), #477 / #534 / #538 (absorption + gate-ownership + script-deterministic precedents this leans on).

One-line recommendation: take Option A — absorb merge-sitting into babysit-prs (digest = documented contract feeding #476), put the flow-gauge on a claude-ops reporting surface, and let the two plugins cooperate only through the digest/signal artifact, never a cross-import.

[kyle-sexton]

Frontier disposition (drain session)

Ruling: advisory absorption proposal, not adopted. The merge-sitting and checkpoint flow-gauge rituals were tower practices from a specific operating period; absorbing them into skills presumes they remain load-bearing, and the babysit/loop surfaces have since evolved past the session that motivated this.

Reopen trigger: the same class of catch (merge sitting unnoticed, flow gauge missed) recurs in a current session, demonstrating the ritual is still needed and unowned.