Cadence

Description

The temporal rhythm of a process — the regular interval or event-driven beat at which a system, review, or feedback loop executes. Cadence is the form that governs when things happen rather than what happens. It is a primitive in its own right rather than a specialization of feedback-loop because its structural properties are distinct: cadence is about periodicity, phase alignment, and the cost of being out of sync with upstream or downstream processes.

Cadence has two flavors: interval-driven (daily, weekly, every N deployments) and event-driven (on commit, on user action, when condition fires). Mixing them without intentional reconciliation produces phase-mismatch bugs — a weekly cadence receiving event-driven data, or an event-driven subscriber on an interval-driven publisher, accumulates systematic lag or stale state. The cadence-alignment move is to make the mismatch visible before it accumulates.

Cadence doubled in frequency in the late period of the full corpus, tracking the increasing complexity of multi-agent coordination work where sync rhythms became a recurring design concern.

Encounters

• Daily sync agents — dotfiles-sync and claude-transcripts-sync share script shape and cadence. Both run on a daily interval; aligning their cadence enabled joint reasoning about their failure modes.
• Insights capture cadence — the Stop-hook captures insights at session end (event-driven). The engine’s MAC retrieval may run on a different cadence (batch vs. real-time). The mismatch needs a reconciliation design.
• Standup rhythm — /standup command produces a 24h cross-worktree summary. The cadence is daily; the data is continuous. Phase alignment: the window is fixed (last 24h), not session-based.
• PR review cadence — some teams review PRs in daily batches; others review on submission. The cadence choice affects how long invalid state persists in the review queue.
• Evaluation cadence for the engine — “how often does the engine evaluate its projection quality?” is a cadence question. Weekly held-out eval vs. per-session vs. per-projection. Each has different feedback-loop properties.
• Audit cadence in active-gate-vs-passive-audit — the audit pole requires specifying cadence: when does the review happen? Cadence-free audits pile up unreviewed signal.
• Agent orchestration — the orchestrator process’s heartbeat, the subagent dispatch cadence, and the corpus-index refresh cadence need explicit alignment.

When it applies / triggers on

User-initiated: User is discussing when something runs, how often, what triggers it, or whether two processes are in sync. Common formulations: “this runs every day,” “we do this weekly,” “it fires on every commit,” “we batch these.”

Agent-initiated: Engine detects a process with an implicit or unspecified rhythm — a rule that “should happen periodically” without stating the period. Candidate inference: “this process has no explicit cadence — what’s the right interval or trigger, and is it in phase with its data sources and consumers?”

Vocabulary cues: “cadence,” “schedule,” “interval,” “frequency,” “periodic,” “daily,” “weekly,” “on event,” “batch,” “real-time,” “how often,” “when does this run,” “trigger,” “rhythm,” “tempo.”

Situation-shape signals: Any process that runs repeatedly and consumes or produces data for other processes. The form is most useful when two or more processes interact and their timing relationship is implicit. Also: when an audit lacks a cadence, or a feedback loop has no stated frequency.

Composes with

• feedback-loop (specialization relationship) — feedback loops require cadence to work: how often does the signal arrive and how often does the corrector respond? Cadence specifies the timing of a feedback loop. A feedback loop with mismatched producer/consumer cadences degrades or breaks.
• active-gate-vs-passive-audit (composition relationship) — the audit pole of this form is cadence-dependent: the audit must have a cadence to be actionable. “Audit daily” vs. “audit weekly” vs. “audit on alert” are distinct postures with different signal-to-lag tradeoffs.
• trigger-rule-pair (composition relationship) — cadence-driven triggers are a specialization of trigger-rule-pair where the trigger is temporal rather than event-based. “At session end, run X” is a cadence-shaped trigger.
• grain (composition relationship) — cadence choices are implicitly grain choices (event-grain vs. day-grain vs. week-grain). Cadence and grain co-determine the unit at which a process operates.
• uniformity-dividend (composition relationship) — processes with aligned cadences earn a uniformity dividend: shared debugging, shared monitoring, shared reasoning. Misaligned cadences are a uniformity tax.
• loop-completion (composition relationship) — loop-completion diagnostics often surface cadence mismatches: gaps visible because the loop’s cadence doesn’t match the user’s mental model of the journey’s rhythm.

When it doesn’t apply

• One-shot processes — a process that runs exactly once has no cadence to design. The form earns its keep when recurrence is part of the design.
• Fully real-time systems — systems where every event is processed at arrival and no batching occurs have cadence determined by event rate, not by design choice. The form is moot unless the question is “what’s the event arrival rate?”
• When timing doesn’t matter — pure batch transforms on static data have no cadence-alignment concerns. The form earns its keep when timing relationships between producers and consumers are load-bearing.

Sources

• Musical metaphor: cadence as rhythm, tempo, beat — the temporal structure that organizes a sequence.
• Database literature: batch vs. streaming processing, event-driven vs. scheduled ETL.
• Agent-framework design: orchestrator heartbeat, subagent dispatch intervals, corpus-index refresh.
• 57 hits in full corpus, doubled in late period tracking multi-agent coordination complexity.

Canonical exemplars from corpus (T2 2026-05-17)

Mined via scripts/mine_new_form_exemplars.py from data/backfilled-insights.jsonl. 190 backfill-only matches at score ≥ 2 — strong support, higher than the 57-hit catalog count because the new heuristic catches “daily/weekly” cadence phrasings the original count under-sampled. Top co-occurrences: doctrine (31), grain (10), active-gate-vs-passive-audit (7).

• Weekly run cadence as economic load-bearing (cwd: campconnect, 2026-05-08): “The content-addressed cache is the cluster’s silent quality multiplier. Without it, every weekly run would pay $15-25\times 3markets=$50-75 in LLM calls. With it, $3\times 3=$9.” — weekly cadence × per-call cost makes the cache load-bearing for economic viability.
• Cadence label as scannable shape on dual-display cards (cwd: campconnect, 2026-05-10, task-notification): “Cards now have two distinct semantic layers: Subtitle (cadence): scannable shape — ‘what kind of thing is this?’; Date row (schedule): concrete planning info — ‘when does it happen?‘” — cadence as the recurring-shape layer distinct from per-event schedule.
• Fallback issue when cadence and date row collide (cwd: campconnect, 2026-05-10, task-notification): “Cards WITHOUT date+time data fall back to the cadence label in the date row position, creating the ‘twice’ appearance: subtitle (‘Short one-day activity’) + date row (‘Short one-day…’).” — cadence × grain mismatch produces a visible UI bug.
• Cycle-complete diagnostic (brainstorm → design → plan → impl → verification → PR) (cwd: campconnect, 2026-05-09, task-notification): “Jackrabbit adapter cycle complete — brainstorm → design → plan → impl → verification → PR all ran end-to-end.” — cadence as the rhythm of a multi-stage development cycle, used as a completeness check.

Trigger pattern (T2): Cadence surfaces when the user discusses when/how-often something runs, OR when the agent notices a process with implicit timing that needs explicit cadence design (audit cadence, sync interval, batch vs. real-time). The 31 co-occurrences with doctrine indicate many doctrines specify their cadence-shaped trigger explicitly.