activity-deep-dive - SKILL.md Agent Skill

name: activity-deep-dive description: Per-activity coaching analysis from Strava. Detects interval vs steady-state sessions and road vs trail terrain, branching the metrics accordingly. Computes zone breakdown and recovery estimate; saves a compiled artifact and returns a compact summary. Run after a workout or to retro-analyze a specific activity. allowed-tools: - Read - Write - Edit - mcpstravacheck-strava-connection - mcpstravaget-athlete-zones - mcpstravaget-recent-activities - mcpstravaget-activity-details - mcpstravaget-activity-laps - mcpstravaget-activity-streams

Activity Deep-Dive

Produces a standardised per-activity coaching note. Detects interval vs steady-state sessions and branches accordingly: interval sessions get work-interval HR progression and between-bout recovery quality; steady sessions get pace/HR efficiency and cardiac drift. It also classifies terrain (road vs trail): trail sessions swap pace/HR efficiency for VAM and a grade-adjusted-pace estimate, reframe cardiac drift against the altitude profile, and extend the recovery window for descent load. Both get zone breakdown, recovery estimate, and a coaching note. Saves a compiled artifact and returns a compact summary.

Usage

/claude-code-fitness-hermit:activity-deep-dive <activity-id>
/claude-code-fitness-hermit:activity-deep-dive latest

Steps

Call mcp__strava__check-strava-connection — abort if disconnected.
Fetch athlete zones via mcp__strava__get-athlete-zones (needed for zone calculations).
Resolve activity:
- If "latest": call mcp__strava__get-recent-activities with limit 1, extract the activity ID.
- Otherwise: use the provided activity ID directly.

3b. Read .claude-code-hermit/state/activity-notes.json. If the file exists and contains an entry for the resolved activity ID, hold rpe and notes in context for steps 6 and 7.

Issue the following three calls in a single turn so they execute concurrently:
- mcp__strava__get-activity-details — name, type, sport_type, date, distance, duration, moving time, avg/max HR, avg pace, total elevation gain (m)
- mcp__strava__get-activity-laps — lap splits
- mcp__strava__get-activity-streams with keys heartrate,velocity_smooth,altitude,cadence (add watts if sport type is Ride — use altitude,cadence as the baseline keys and add watts only when sport type is known to be Ride)

4b. Classify session kind — using the lap data and HR stream already fetched:

Primary signal (HR alternation): does the HR stream show ≥ 3 repeated high→low→high cycles? Divide the stream into segments aligned with laps (or into 8–12 equal windows if lap data is sparse). If alternating segments show a ≥ 15 bpm differential between "work" and "recovery" phases, the session is a candidate for interval.
Corroborating (lap clustering): do laps cluster into two groups — high-HR laps and low-HR laps — with the groups alternating?
Default to steady when the alternation pattern is absent or ambiguous (e.g. a tempo run with gradually rising HR, a progression run, or a steady run with 1 km auto-laps that have flat HR across all laps).
Hold the result as session_kind (interval or steady) and the interval structure — N work bouts (the high-HR lap group identified by the clustering signal) and their avg duration in minutes — for steps 5 and 6.

4c. Classify terrain — compute elev_gain_per_km = total_elevation_gain_m / distance_km. Set terrain:

trail if sport_type == TrailRun and elev_gain_per_km >= 10, OR as a fallback if elev_gain_per_km >= 20 for a running activity (≥ 100 m gain per 5 km — significant gradient that decouples pace from effort).
road otherwise. This includes a flat TrailRun (elev_gain_per_km < 10, e.g. a canal or gravel path), where pace stays a valid effort proxy and VAM/GAP add no signal — road metrics are the right lens. Hold terrain and elev_gain_per_km for steps 5–7.

Compute metrics:

Zone breakdown — from HR stream vs athlete zone boundaries. Calculate % of stream datapoints in each zone (Z1–Z5). If HR stream absent: note "HR data unavailable".

Cadence (running sport types only — Run, TrailRun, VirtualRun) — from the cadence stream. Convert to steps-per-minute first: if the median stream value is < 130 it is single-leg RPM — multiply every value by 2; otherwise use as-is. Compute avg, standard deviation σ, and coefficient of variation CV = σ/μ × 100. Flag if avg < 170 spm (over-striding risk) and/or CV > 8% (high within-run variability — neuromuscular fatigue signal). On trail terrain, report avg and CV for reference but suppress both ⚠ flags — the < 170 spm and CV > 8% thresholds are road-calibrated, and naturally lower, more variable cadence is expected on technical terrain (climbs, descents, footing) rather than a fault. If the cadence stream is absent or empty, or the sport type is not a running type, skip the cadence line entirely (mirror the "HR data unavailable" guard — do not emit a placeholder).

Pace/HR efficiency (steady sessions only, road terrain only) — average pace (min/km) divided by average HR. Lower = more efficient. Call mcp__strava__get-recent-activities with perPage: 5 now (only needed for this metric): filter for activities of the same sport type, exclude the current activity ID, take up to the 4 most recent. Compute the same ratio for each. Report delta vs prior mean. If terrain == trail, skip this metric entirely and emit no placeholder — gradient decouples pace from aerobic effort on technical terrain.

Cardiac drift (steady sessions only) — on road (terrain == road): compare average HR in first 20% of HR stream vs last 20%. Flag if difference > 10 bpm at similar pace (± 15 sec/km). Report: drift: +N bpm. On trail (terrain == trail): elevation changes confound the first/last HR split, so instead report a qualitative HR/altitude note: describe whether HR broadly tracked the climb/descent profile (expected) or showed a decoupled trend worth noting.

VAM (trail terrain only) — vertical ascent meters per hour: total_elevation_gain_m / moving_time_h. Report VAM: N m/h. Contextual benchmarks (heuristic): < 500 m/h is hiking pace, 500–800 is easy trail, 800–1200 is solid trail running, > 1200 is race/VO2 effort.

GAP estimate (trail terrain only) — session-level grade-adjusted pace: converts the run to an equivalent flat effort using total elevation gain only (not per-point grade). Heuristic: equiv_flat_km = distance_km + 0.008 × total_elevation_gain_m (100 m of ascent ≈ 0.8 km of flat effort). GAP = moving_time / equiv_flat_km — use moving time, not elapsed, so the estimate is comparable to Strava's moving-time-based actual pace. Report GAP ~M:SS/km (est, vs actual P:SS/km). Label as heuristic in output.

Work-interval HR progression (interval sessions only) — per-lap avg HR across work laps in sequence (e.g. I1 157 → I4 168 bpm — progressive ✓). Include the activity max HR (from get-activity-details) to note the peak bout (e.g. "peaked at 174 bpm on I3"). Use activity max HR as HRmax — do NOT use the Z5 floor from get-athlete-zones as HRmax.

Recovery between intervals (interval sessions only) — using per-lap avg HR for recovery laps: did HR return toward a recovery zone (Z1–Z2) between bouts? Report as a qualitative observation (e.g. "recovered to ~130 bpm between bouts — adequate" or "incomplete recovery: HR only dropped to 155 bpm before next bout").

Recovery estimate (scale 1–5):
- 1 = easy (< 5% Z3+, avg HR < Z3 floor, < 60 min)
- 2 = moderate (5–20% Z3, < 60 min)
- 3 = quality (20–50% Z3, or any Z4, < 90 min)
- 4 = hard (> 50% Z3 or > 10% Z4, or > 90 min hard)
- 5 = race-level (> 20% Z4+, or peak HR > 95% max) Include recommended recovery window: 1→24h, 2→36h, 3→48h, 4→72h, 5→5–7 days. For trail activities, extend the recommended window based on elev_gain_per_km to account for the muscular load of descents and technical footwork not captured by HR zones (heuristic):
- elev_gain_per_km < 15 → no extension
- elev_gain_per_km 15 to < 30 → add 1 day to the window (e.g. 48h → 72h, 72h → 4 days)
- elev_gain_per_km >= 30 → add 1–2 days (e.g. 72h → 4–5 days, 5–7 days stays at 5–7 days) Append (+trail vert) to the recovery line when an extension is applied.
Coaching note — 2–3 sentences grounded in the numbers. Highlight what was executed well and one concrete thing to monitor or adjust next time. Reference specific metrics (e.g. "cardiac drift of +14 bpm suggests pacing started too hot").

Format output (8–10 lines):

Steady session (road):

Activity: <name> | <date> | <distance>km in <duration>
Session type: Steady
Zones: Z1 N% / Z2 N% / Z3 N% / Z4 N% / Z5 N%
Cadence: N spm avg (CV: X%) [⚠ over-striding | ⚠ high variability]   ← omit line when cadence absent or non-running; show ⚠ only on a tripped flag
Pace/HR efficiency: X.XX min·km⁻¹·bpm⁻¹ (vs prior 4: ±X%)
Cardiac drift: +N bpm (flag if > 10 bpm)
Recovery: N/5 — recommended rest: Xh
Subjective: RPE N/10 — <notes>          ← include only when RPE data exists from step 3b
Coaching: <2–3 sentences>

Steady session (trail):

Activity: <name> | <date> | <distance>km in <duration> | <elevation>m gain
Session type: Steady · Trail
Zones: Z1 N% / Z2 N% / Z3 N% / Z4 N% / Z5 N%
Cadence: N spm avg (CV: X%)   ← reference only on trail (no ⚠ flags); omit line when cadence absent or non-running
Trail: VAM N m/h | GAP ~M:SS/km (est, vs actual P:SS/km)
HR/altitude: <qualitative note on whether HR tracked the climb/descent profile>
Recovery: N/5 — recommended rest: Xh[(+trail vert)]
Subjective: RPE N/10 — <notes>          ← include only when RPE data exists from step 3b
Coaching: <2–3 sentences>

Interval session (road):

Activity: <name> | <date> | <distance>km in <duration>
Session type: Interval (N×~Xmin)
Zones: Z1 N% / Z2 N% / Z3 N% / Z4 N% / Z5 N%
Cadence: N spm avg (CV: X%) [⚠ over-striding | ⚠ high variability]   ← omit line when cadence absent or non-running; show ⚠ only on a tripped flag
Intervals: I1 NNNbpm → IN NNNbpm — <progressive ✓ / regressive / flat>; peaked at NNNbpm on IN
Between-bout recovery: HR to ~NNNbpm — <adequate / incomplete>
Recovery: N/5 — recommended rest: Xh
Subjective: RPE N/10 — <notes>          ← include only when RPE data exists from step 3b
Coaching: <2–3 sentences>

Interval session (trail):

Activity: <name> | <date> | <distance>km in <duration> | <elevation>m gain
Session type: Interval (N×~Xmin) · Trail
Zones: Z1 N% / Z2 N% / Z3 N% / Z4 N% / Z5 N%
Cadence: N spm avg (CV: X%)   ← reference only on trail (no ⚠ flags); omit line when cadence absent or non-running
Trail: VAM N m/h | GAP ~M:SS/km (est, vs actual P:SS/km)
Intervals: I1 NNNbpm → IN NNNbpm — <progressive ✓ / regressive / flat>; peaked at NNNbpm on IN
Between-bout recovery: HR to ~NNNbpm — <adequate / incomplete>
Recovery: N/5 — recommended rest: Xh[(+trail vert)]
Subjective: RPE N/10 — <notes>          ← include only when RPE data exists from step 3b
Coaching: <2–3 sentences>

Save compiled artifact to .claude-code-hermit/compiled/activity-<id>-<YYYY-MM-DD>.md:

---
title: "Activity Note — <name> <date>"
type: activity-note
created: <ISO 8601>
session: <current session ID from SHELL.md>
source: manual
tags: [activity-analysis]
activity_id: <id>
sport_type: <Run|TrailRun|Ride|WeightTraining|…>
terrain: <road|trail>
session_kind: <interval|steady>
rpe: <int>                               # include only when RPE data exists from step 3b
subjective_notes: "<string>"             # include only when notes exist from step 3b
---

Body: the full output above.

7b. Write signal-only coaching observations to .claude-code-hermit/sessions/SHELL.md Findings.

From the computed metrics and coaching note, derive 0–N observations that carry a coaching signal worth tracking across sessions: a flagged cardiac drift, a zone-distribution anomaly, a recovery estimate that conflicts with subjective RPE, an efficiency regression vs the prior mean, a flagged cadence (low average or high within-run variability), a notable VAM value, or a trail recovery extension. Do NOT write routine confirmations ("session completed normally") unless they represent a pattern break. If nothing clears the signal bar, skip this step.

First Read .claude-code-hermit/sessions/SHELL.md (Edit requires the file in context, and you need its current ## Findings content to dedup). For each qualifying observation, anchor on the HTML comment and append one line:

old_string: "<!-- Anything unexpected found during work. Proposal-worthy items get their own file. -->"
new_string:  "<!-- Anything unexpected found during work. Proposal-worthy items get their own file. -->\nCoaching observation [<label>] (activity <id>): <one-line description grounded in a specific metric>"

Skip the append if a line with the same [<label>] (activity <id>) already exists in ## Findings; re-running the deep-dive must not duplicate observations. If the anchor comment is absent (operator edited SHELL.md), append directly under the ## Findings heading instead.

Labels are kebab-case and reused across sessions for consistency. Prefer an existing label over inventing a synonym. Seed vocabulary: cooldown-hr-elevated, vo2max-stimulus-confirmed, cardiac-drift-high, interval-pacing-inconsistent, recovery-insufficient, efficiency-regression, cadence-low, cadence-variability-high, trail-recovery-extended, vam-notable. Add a new kebab-case label only when none fit.

These lines feed reflect's current-session evidence path; the label convention lets reflect recognize recurrence across sessions, and recurring observations graduate to proposals through the normal reflection-judge / proposal-triage gates.

Return the formatted output to the caller.