xunji

name	xunji
description	Read, analyze, create, and update 训记 (Xunji) training records through the public Open API v2 endpoints `/api_trains_for_llm_v2` and `/api_upsert_trains_for_llm_v2` with local caching, action-name alignment, dry-run validation, and safe upsert handling.

Xunji Training Skill

Use the bundled helpers instead of calling the API manually. This single skill covers both read and write workflows for 训记 training data and follows the official train_open_api_v2 schema.

It handles:

• Bearer-token authentication from XUNJI_API_KEY (header Authorization: Bearer ...; also accepts x-api-key)
• POST /api_trains_for_llm_v2 for reading records (lightweight by default; pass --full for include_full_data: true)
• POST /api_upsert_trains_for_llm_v2 for writing records with client_request_id and optional server-side dry_run
• gzip JSON decoding and res.trains array parsing
• persistent per-date local caching
• action library maintenance for safer write-back
• same-day validation, normalization, and dry-run checks before writes

Choose The Workflow

Use the read workflow when the user wants to inspect, summarize, analyze, compare, or reuse existing Xunji workout data.

Use the write workflow only when the user clearly wants to save, create, update, or write finalized training content back into Xunji.

Do not use the write helper for brainstorming, formatting-only, or read-only analysis that stops short of actual persistence.

Read Command

python3 "$CLAUDE_SKILL_DIR/scripts/fetch_xunji_trains.py" --date 2026-04-02

Add --full only when the user needs unchecked sets, per-set RPE, per-set remarks, or movement remarks (sends include_full_data: true):

python3 "$CLAUDE_SKILL_DIR/scripts/fetch_xunji_trains.py" --date 2026-04-02 --full

Use line output when the user wants raw training text directly:

python3 "$CLAUDE_SKILL_DIR/scripts/fetch_xunji_trains.py" --date 2026-04-02 --format lines

Refresh only when the user explicitly asks for fresh data:

python3 "$CLAUDE_SKILL_DIR/scripts/fetch_xunji_trains.py" --date 2026-04-02 --refresh

Read Rules

1. Always pass read dates as YYYY-MM-DD. Compact YYMMDD is no longer accepted.
2. Send the API key in Authorization: Bearer ...; the helper reads it from XUNJI_API_KEY.
3. Every request body includes schema_version: "train_open_api_v2".
4. Default lightweight mode returns nearly v1-equivalent fields; use --full for unchecked sets, per-set RPE, per-set remarks, and movement remarks.
5. Preserve returned training data exactly when it may later be written back.
6. Keep localid (id:... in line format) and start/end (train_time:...) tokens unchanged.
7. Treat localid as the training record's identifier for later upsert writes.
8. Cardio, timed, and Tabata movements expose per-set metrics (distance/kcal/bpm/steps).
9. Prefer cached data for repeated analysis of the same date; the API rate-limits one read per training day to roughly once per 90 seconds.
10. If the API returns an auth, membership, or frequency error such as apikey missing, apikey invalid, 仅VIP可用, or too frequent, retry after 90s, surface it clearly and respect the retry hint.
11. A read response with empty res.trains does not prove there are no unfinished or unchecked training items that day.
12. Use read data as comparison context for modifying existing records.

Action Library

The read helper maintains the action library at:

• ~/.cache/opencode/xunji-reader/action-library.json

This path is kept for compatibility with earlier versions of the split reader/writer skills.

Use this file as the first lookup source when checking whether a user-entered action name already exists. If it is missing or incomplete, run the read helper so it can refresh history and rebuild the library before deeper action-name analysis.

Write Commands

For clearly new records with explicit lines:

python3 "$CLAUDE_SKILL_DIR/scripts/upsert_xunji_trains.py" \
  --date 2026-04-02 \
  --allow-new-records \
  --line '2026-04-02,胸部训练,状态不错,1.卧推,1组,60kg,10次,2组,60kg,8次'

With a JSON file containing either a raw array or an object with res:

python3 "$CLAUDE_SKILL_DIR/scripts/upsert_xunji_trains.py" \
  --date 2026-04-02 \
  --res-file /tmp/xunji-res.json

Before a real write, prefer validating the payload without sending it:

python3 "$CLAUDE_SKILL_DIR/scripts/upsert_xunji_trains.py" \
  --date 2026-04-02 \
  --res-file /tmp/xunji-res.json \
  --dry-run

Write Rules

Stop and ask the user before writing whenever the outgoing structure is not uniquely determined.

Clarify before write when:

• the user has not clearly said to save, create, update, or write back into Xunji
• the target date is missing or unclear
• it is unclear whether this modifies an existing record or creates a new one
• an input token is ambiguous compact shorthand such as 10x3
• a numeric token could mean weight, reps, or sets
• action names cannot be aligned safely against the canonical Chinese names from https://github.com/Foveluy/movements (the Xunji-movements reference) or the local action library

Mandatory write rules:

1. Every train in the outgoing res (or res.trains) must belong to the same target day.
2. A single call may carry at most 4 trains; each train may carry at most 15 movements; each movement may carry at most 20 sets. The server rejects anything larger.
3. Always send schema_version: "train_open_api_v2" and a fresh client_request_id.
4. When updating existing records, preserve localid, start, and end exactly. (In line format these surface as id:... and train_time:start-end.)
5. If cached data for the same localid contains train_time:..., preserve that exact value unless you intentionally pass --allow-train-time-drop.
6. If any non-rest-day train omits localid, require explicit create intent with --allow-new-records; the server will then assign a new id.
7. The upsert endpoint writes by localid; old same-day records omitted from the request are not automatically deleted.
8. For a new plan, combine same-plan actions into one top-level training instead of splitting them across multiple new records.
9. If one weight/reps payload is repeated across multiple sets, expand it into explicit 1组 ... 2组 ... segments instead of compressed shorthand like 60kg,3组.
10. For ordinary strength sets, every set must include at least one of weight (60kg), reps (10次), time (time:60s), or 自重 (selfWeight).
11. Send only Chinese movement name values. Never send the internal key. When unsure, look the name up against the canonical Xunji-movements table before writing.
12. Unfinished sets must be sent with done: false. Do not silently drop unchecked sets that were read in --full mode.
13. Cardio, timed, and Tabata sets keep their indicators inside sets[].metrics (distance/kcal/bpm/steps).
14. Prefer the local --dry-run flag for validation: it never contacts the server. The --server-dry-run flag still sends dry_run: true in the body, but live testing on 2026-05-19 showed the Xunji v2 server currently persists those calls anyway, so treat --server-dry-run as equivalent to a real write until upstream confirms is honored.

Training Line Format

Examples:

2026-04-02,id:123456,胸部训练,train_time:1744010000000-1744013600000,状态不错,1.卧推,1组,60kg,10次,2组,60kg,8次
2026-04-02,有氧,1.跑步,5km,300kcal,time:1800s,140bpm
2026-04-02,休息日

Use YYYY-MM-DD everywhere — compact YYMMDD is no longer accepted in either the --date argument or the leading segment of a line. The helpers convert these lines into the v2 structured payload (datestr, localid, title, start, end, movements[].sets[] with metrics for cardio) before sending.

Normalization Rules

Prefer exact structured values such as:

• 1组
• 60kg
• 10次
• time:60s
• cardio metrics such as 5km, 300kcal, 140bpm

Key behavior:

• 12-15次 normalizes to 15次.
• 20-25kg normalizes to 25kg.
• 60kg,3组,10次 expands to explicit per-set segments.
• 热身 or 热身组 is converted into ordinary writable set structure when safe.
• 10x3 is ambiguous and must be clarified before writing.
• Discarded content is not written into remarks unless the user explicitly asks for that and --preserve-invalid-in-remarks is used.

Output

Default helper output is JSON with:

• datestr
• cached
• fetched_at
• count
• res — CSV-style training lines reconstructed from the v2 train objects
• trains — original v2 train objects returned by the API, for clients that want structured data

Use --format lines when the user wants line output. Local cache stores both res and trains so callers may consume either shape.

Action Name Reference

When a Chinese movement name is not in the local action library, fall back to the canonical Xunji movement reference at https://github.com/Foveluy/Xunji-movements and pick the closest Chinese name from that table before retrying the write. Never send a key field — only the Chinese name is allowed.

Configuration

The helpers read credentials from:

1. XUNJI_API_KEY environment variable

The API endpoints are hardcoded. Do not put API keys in this repository or normal responses unless the user explicitly asks to inspect configuration.

Always reply in the same language the user is currently using.