loops/skills: self-paced /loop idles between non-empty cycles — cadence should be work-state-driven (no delay when queue non-empty), not a guessed clock

[kyle-sexton]

Lane: triage (applies to every self-paced /loop lane) · Category: operational (throughput) · Source: live triage loop, 2026-07-19, user course-correction.

Triggering example

Triage loop, dynamic (self-paced) mode. Two consecutive cycles ended with the queue non-empty (14 items, then 9 items) and each scheduled the next cycle with a 600s delay ("short delay when queue has items" + "let concurrent filings settle"). The operator flagged it: with real work already queued, the delay is pure idle latency — waiting never improves triage quality, and "let filings settle" is a rationalization (intake arrives continuously; there is no settled state to wait for).

Observed vs expected

• Observed: the standing loop-prompt rule ("self-paced: short delay when queue has items, ~30 min when empty") and the /loop skill's dynamic-mode guidance ("lean 1200–1800s fallback") push the lane to insert a non-zero wall-clock delay between cycles even when the queue is non-empty. Cycle boundaries exist to refresh context (avoid degradation) — not to burn wall-clock. A delay-when-work-exists adds latency with zero quality benefit.
• Expected: loop cadence keys to work state, not a guessed clock. Queue non-empty → start the next cycle immediately (delay ≈ 0; the cap-per-cycle + fresh-context-per-cycle already bound each cycle's size — see loops/skills: orchestrator context economy — subagent return-payload contracts + GitHub-externalized state #496). Delay applies only when the queue is empty (idle poll, ~30 min) or when genuinely gated on an external event (CI, a Monitor). "Short delay when queue has items" should read "no delay when queue non-empty."

Category

operational (throughput). Every idle delay between non-empty cycles is dead time on a lane whose whole value is turning intake around.

Related

• source-control:babysit-prs + loop: /loop 15m fires overlap in-flight fan-out cycles when batch-cap-10 dispatch runs long #472 — the sibling pacing bug from the opposite direction: a fixed-interval loop (/loop 15m) whose interval is shorter than the cycle, causing overlap. Together these frame one principle: loop cadence must derive from queue/work state and cycle duration, not a fixed or guessed delay — too-short fixed intervals overlap (source-control:babysit-prs + loop: /loop 15m fires overlap in-flight fan-out cycles when batch-cap-10 dispatch runs long #472); too-long dynamic delays idle (this issue).
• feat(claude-ops or session-flow): loop-prompt authoring skill — generate and maintain standing loop-lane prompts #480 (loop-prompt authoring skill) — the canonical prompt template should encode "no delay when queue non-empty" as an invariant so hand-authored lanes get it for free.
• loops/skills: orchestrator context economy — subagent return-payload contracts + GitHub-externalized state #496 (context economy) — the reason a cycle boundary exists at all is fresh context, which makes wall-clock delay between non-empty cycles unnecessary.

[kyle-sexton]

This was generated by AI during triage. · Category: operational (pipeline throughput) · Priority: priority: high (week 2nd-priority: the pipeline improving itself). Routed status: ready.

Plan

Encode the corrected pacing so self-paced /loop lanes stop idling between non-empty cycles:

1. Loop-prompt template (feat(claude-ops or session-flow): loop-prompt authoring skill — generate and maintain standing loop-lane prompts #480's domain): the canonical standing-prompt template's batch/pacing rule states the invariant — queue NON-EMPTY → schedule the next cycle immediately (delaySeconds at the runtime's 60s floor); the cycle boundary exists only for fresh context (loops/skills: orchestrator context economy — subagent return-payload contracts + GitHub-externalized state #496), never idle wall-clock. Delay (~30 min) ONLY when the queue is empty, or when genuinely gated on an external event (CI, a Monitor). Hand-authored lanes then inherit it instead of re-inventing "short delay when queue has items" (the wording that produced the idle).
2. /loop skill dynamic-mode guidance: clarify that the "lean 1200–1800s" figure is the Monitor fallback-heartbeat, NOT the cadence for a work-queue loop with no Monitor. For a queue-draining loop, cadence is work-state-driven: immediate while the queue is non-empty, idle-poll only when empty.

Acceptance

• The loop-prompt template and /loop dynamic-mode guidance both state "no wall-clock delay when the queue is non-empty"; the only delays are empty-queue idle-poll and external-event gating.
• A queue-draining loop following the template schedules back-to-back cycles (fresh context per cycle) with no inter-cycle idle while work remains.

Related

• source-control:babysit-prs + loop: /loop 15m fires overlap in-flight fan-out cycles when batch-cap-10 dispatch runs long #472 — the opposite-direction pacing bug (fixed 15m interval shorter than a long cycle → overlap). Together: loop cadence must derive from queue/work-state + cycle duration, never a fixed or guessed clock.
• feat(claude-ops or session-flow): loop-prompt authoring skill — generate and maintain standing loop-lane prompts #480 (loop-prompt authoring skill — the template home) · loops/skills: orchestrator context economy — subagent return-payload contracts + GitHub-externalized state #496 (context economy — why the boundary exists at all).

Decision defaulted: encode "no delay when queue non-empty; delay only when empty or externally gated" into the loop-prompt template + clarify the /loop skill's Monitor-fallback vs work-queue-cadence distinction — veto before merge

[kyle-sexton]

Triggering example (work lane, this session, dogfood): cycle 1 closed and the loop scheduled a fixed 30-min wakeup off a point-in-time 'serialization wall' snapshot (every plugin appeared occupied by an open PR). Within minutes, 5 of the cycle's 7 PRs merged, freeing those plugins — 15+ ready issues became immediately dispatchable, yet the loop would have idled ~30 min on a guessed clock before noticing. Confirms the ask: cadence must be work-state-driven — recompute the dispatchable set (ready minus plugins-with-open-PRs) on merge events / immediately after a cycle, and apply a delay ONLY when that set is genuinely empty. Caught here by a human nudging 'is the loop inefficient?' — the loop should self-detect it. Related: #504 (engine recommended_cadence), #496 (context economy).

[kyle-sexton]

Parking — lane work-2, cycle 2026-07-19.

Scoped this before dispatch (rule 7: no guessing on contract-shaped calls). The plan comment routes the fix to two surfaces, neither actionable inside this repo as currently scoped:

1. feat(claude-ops or session-flow): loop-prompt authoring skill — generate and maintain standing loop-lane prompts #480's domain (loop-prompt authoring skill) — feat(claude-ops or session-flow): loop-prompt authoring skill — generate and maintain standing loop-lane prompts #480 is itself an unbuilt proposal (priority: needs-triage, wayfind: design, explicit "Placement (claude-ops vs session-flow vs autonomy) = design decision for triage/human"). There's no existing skill file to edit; authoring one now to hold this fix would be scope creep past loops/skills: self-paced /loop idles between non-empty cycles — cadence should be work-state-driven (no delay when queue non-empty), not a guessed clock #500 and would preempt feat(claude-ops or session-flow): loop-prompt authoring skill — generate and maintain standing loop-lane prompts #480's own placement decision.
2. The /loop skill's dynamic-mode guidance — that's the Claude Code CLI built-in /loop skill, not distributed by claude-code-plugins (repo-wide grep for loop-prompt/fallback.heartbeat finds nothing outside plugins/source-control/skills/babysit-prs/). Out of this repo's write surface entirely.

RECOMMENDED: Hold #500 until #480 lands and has a placement decision — #500 becomes one of the first invariants encoded into that template. Re-check status: ready on #480 to unblock the chain.

Options:

1. (RECOMMENDED) Block loops/skills: self-paced /loop idles between non-empty cycles — cadence should be work-state-driven (no delay when queue non-empty), not a guessed clock #500 on feat(claude-ops or session-flow): loop-prompt authoring skill — generate and maintain standing loop-lane prompts #480 landing; leave status: ready off until then.
2. Split loops/skills: self-paced /loop idles between non-empty cycles — cadence should be work-state-driven (no delay when queue non-empty), not a guessed clock #500: file a narrow in-repo issue for any lane-specific doc (e.g. plugins/source-control/skills/babysit-prs/reference/loop.md, mirroring source-control:babysit-prs + loop: dynamic /loop wakeups ignore the engine's own recommended_cadence field #504's fix) that duplicates the "no delay when queue non-empty" invariant locally now, accepting the drift risk once feat(claude-ops or session-flow): loop-prompt authoring skill — generate and maintain standing loop-lane prompts #480 centralizes it.
3. Fast-track feat(claude-ops or session-flow): loop-prompt authoring skill — generate and maintain standing loop-lane prompts #480 placement as a same-cycle decision so loops/skills: self-paced /loop idles between non-empty cycles — cadence should be work-state-driven (no delay when queue non-empty), not a guessed clock #500 has a concrete target this week.

cc: sibling issue #504 (babysit-specific recommended_cadence consumption — scoped and dispatched separately, not blocked the same way).