weekly-coaching-patterns

Weekly Coaching Patterns

Scheduled-check skill: reads the last 4 compiled/activity-*.md artifacts for steady sessions and detects whether cardiac drift is trending upward over time. Returns a fixed findings block for reflect --scheduled-checks to classify and route.

Contract: idempotent, read-only, no self-scheduling, short-running. Returns findings or silence — never creates proposals itself.

Steps

1. Run the trend check. Issue a single Bash call:

   bun ${CLAUDE_PLUGIN_ROOT}/scripts/fitness-lab.ts weekly-patterns
   

   The script does the whole deterministic pass — filesystem only, no Strava token: globs .claude-code-hermit/compiled/activity-*.md, keeps entries whose frontmatter has type: activity-note AND session_kind: steady, takes the 4 most recent by created, reads each note's signed drift from the cardiac_drift_bpm frontmatter field, falling back to the body's Cardiac drift: line for notes written before that field existed (neither present → excluded; a negative value compares as smaller than any positive — the sign is never stripped), and evaluates a strict oldest-to-newest monotonic rise across all 3 adjacent pairs. It emits:

   {"steady_sessions_found": <int>, "series": [{"file": "…", "date": "…", "drift": <signed int>}], "trend": "upward" | "none" | "insufficient-data"}
   

   series is ordered oldest-to-newest. trend: "upward" means the 4-value bpm series is strictly increasing (the deliberately strict v1 bar — avoids false positives from noisy week-to-week variation). "none" = 4 values but not monotonic; "insufficient-data" = fewer than 4 drift values (expected on interval/strength-heavy weeks — not an error).

   Anti-duplication guard: emit a finding ONLY on trend: "upward" (the quantitative numeric rise). Do NOT emit on label recurrence alone.

2. Output the findings block. Always output a plain-text findings block to stdout, regardless of outcome. reflect --scheduled-checks classifies the result from this block.