Agent: training-data-coach¶

Persistent Claude Code subagent at ~/.claude/agents/training-data-coach.md. Owns the interpretation side of Casey's training.

What it owns¶

Training advice, protocol recommendations, race-build interpretation
Finding synthesis (findings/*.md) — when to add, when to update, when to mark superseded
Reconciling podcast/research claims against Casey's actual data
Reading findings/_active.md to know what's currently in flight
"What does my data show about X" questions where X is qualitative

When to invoke¶

User mentions any of: - Training, mileage, MPW, threshold, CTL/ATL/eFTP - Race builds, OTQ, taper, peak - Fueling protocols, ferritin/iron - Podcast insights, findings/

Or proactively: when the user starts working on training data or opens a finding file.

When NOT to invoke¶

Pure data queries ("how many threshold sessions in March") → librarian
Pipeline issues ("why didn't sync run") → conductor
Code unrelated to training → main agent

Source files¶

~/.claude/agents/training-data-coach.md — the agent definition
~/garmin-warehouse/findings/_active.md — workstream state (read on every session start)
~/garmin-warehouse/findings/README.md — index of curated findings
~/garmin-warehouse/findings/_todo.md — Casey's checklist
~/garmin-warehouse/ARCHITECTURE.md — 8-layer model

Memory¶

Coach has access to the user-scoped memory at ~/.claude/projects/-Users-caseymanos/memory/. Relevant entries:

garmin_warehouse_project.md — pipeline shape
intervals_icu_setup.md — athlete config
training_data_coach_agent.md — agent ergonomics
feedback_active_archive_convention.md — _active.md trim convention
warehouse_hardening_2026_05.md — kb migrations, prompt versioning, rebuild path