description: Sport-aware workout analytics from FIT files: pace and splits, interval detection with evidence, training load, and machine-readable insights.¶

Analyze workouts¶

The analytics layer turns honest streams into meaning — in each sport's own language. It is optional (chiptime.metrics), deterministic, and never estimates what it wasn't given.

The one-call version¶

chiptime analyze ride.fit --ftp 250 --max-hr 185 --resting-hr 48

session 1: cycling/virtual_activity
  55:11 · 29.61 km · 32.2 km/h (moving) · avg 164 W · weighted 175 W · avg HR 136
  structure [laps:manual]: 3 x 10:00 @ 194 W rest 3:24
  load 45 [power+ftp]
  PACING_NEGATIVE_SPLIT: Second half 7.6% faster than the first.

Add --json for the machine-readable report.

Sports speak differently¶

Sport	Pacing	Primary signal
Running	min/km	pace (power if a running power stream exists)
Cycling	km/h	watts when a power meter is present
Pool swim	min/100m + SWOLF	pace from lengths × pool size — never GPS
Open water	min/100m	GPS pace (with its known optimism)
Rowing	/500m split	watts ↔ split via the Concept2 relation

A sport-profile registry routes every session; unknown sports get an honest generic treatment rather than a wrong specific one.

from chiptime import metrics

profile = metrics.profile_for(session)     # SportProfile(key="pool_swim", pace_style="per_100m", ...)
metrics.format_pace(105.0, "per_100m", suffix=True)    # "1:45/100m"
metrics.watts_to_split_500m(200)                       # 120.7 → a 2:00.7 split

Pace is treated as presentation: all math happens in m/s and seconds, because averaging paces directly is a classic error (pace is an inverse).

Intervals: an evidence ladder¶

detect_structure reads workout structure the way athletes think in it — and every result names its evidence:

st = metrics.detect_structure(session, result.messages)
st.basis            # "steps:workout" | "laps:manual" | "lengths:sets"
                    # | "detected:power-steps" | "detected:speed-steps" | "none"
st.repeats[0].label # "6 x 0:30 @ 300 W rest 0:30"
st.note             # when basis == "none": WHY (too few reps, too varied, ...)

The ladder: structured-workout steps → manual lap presses → swim sets grouped from lengths → deterministic detection on smoothed power/speed. Detection refuses to hallucinate: fewer than 3 similar work efforts, or wildly varied durations, and the answer is none with the reason — a valid result, not a failure.

Thresholds are yours, never guessed¶

settings = metrics.AthleteSettings(
    ftp_w=250, max_hr=185, resting_hr=48,
    hr_zone_bounds=(115, 135, 155, 172, 188),
)
report = metrics.analyze(result, settings)

Zones and load come from your settings or from zone messages inside the file — never estimated from the workout itself. Without inputs, the report says exactly what was omitted and why:

report.sessions[0].omissions
# ["load: hr covers only 17% of the session (< 50%); trimp would understate load"]

Load, named¶

Every load number carries its basis: power+ftp (duration × intensity² × 100, from 4th-power-weighted watts), or hr-trimp (Banister TRIMP, sex-labeled coefficient) — with an HR-coverage guard so a swim that only logged HR at the wall can't produce a confidently wrong number. For trends, fitness_fatigue_form() runs the classic 42-day / 7-day impulse-response over your daily loads.

Insight codes¶

Notable observations come with stable codes and numeric evidence — built for both humans and agents:

Code	Fires when
`PACING_NEGATIVE_SPLIT` / `PACING_POSITIVE_SPLIT`	halves differ by > 2%
`HR_DRIFT_HIGH`	output per heartbeat fell > 5% (aerobic decoupling)
`COASTING_HIGH`	> 25% of ride samples at 0 W
`WORKOUT_STRUCTURE`	repeated intervals found (label in evidence)