/training-check

name: training-check
description: Periodically check WandB metrics during training to catch problems early (NaN, loss divergence, idle GPUs). Avoids wasting GPU hours on broken runs. Use when training is running and you want automated health checks.
argument-hint: "[wandb-run-path]"
allowed-tools: Bash(*), Read, Grep, Glob, Write, Edit, mcp__codex__codex, mcp__codex__codex-reply, mcp__antigravity__antigravity, mcp__antigravity__antigravity-reply, mcp__grok__grok, mcp__grok__grok-reply, mcp__claude-review__review, mcp__claude-review__review_reply, mcp__claude-review__review_start, mcp__claude-review__review_reply_start, mcp__claude-review__review_status

Training Check

> **Before each reviewer slot**, follow [Selectable Reviewers](../shared-references/reviewer-backends.md): > resolve invocation → `.aris/reviewer.json` → Claude Code default (Codex MCP) > with `reviewer_route.py` through the canonical helper resolver, preserving this > slot's tier, fresh contexts, source/image inputs and deterministic gates. > Codex examples/pins below apply only to `backend=codex`; other routes use their > own first-call/reply contract. Save the route and actual returned identity; > unavailable selections are `REVIEW_UNAVAILABLE`. Classify acceptance from actual > families/completion/scope. Forward invocation reviewer/model/effort overrides > to reviewer-bearing children. Generation workers are unchanged.

Periodically read WandB metrics during training to catch problems early. Do not wait until training finishes to discover it was a waste of GPU time.

> ⏱ This skill is **correctly** cron-wired (see below): it polls > machine-checkable training health (NaN / divergence / idle GPU) — the additive > external-wait shape in > [`shared-references/external-cadence.md`](../shared-references/external-cadence.md). > The occasional Codex call for an ambiguous metric is a **one-shot** check per > tick, not a multi-round verdict loop, so it stays additive — it never grows > into a wrapped verdict skill.

Context: $ARGUMENTS

Constants

• WANDB_ENTITY and WANDB_PROJECT: read from CLAUDE.md or passed as argument (format: `entity/project/run_id`)
• CHECK_INTERVAL: starts at 10 minutes, then gradually increases if consistently healthy: 10 min → 20 min → 30 min → 60 min (cap)
• REVIEWER_MODEL = `gpt-6-astra` — used via Codex MCP for ambiguous cases only

When to Use

• After training is confirmed running (session alive, loss decreasing for first few steps)
• Set up via CronCreate to fire periodically during training
• **This skill checks training QUALITY, not process HEALTH.** Process health (session alive, GPU utilization) is [watchdog.py](../../tools/watchdog.py)'s job.

Workflow

Step 1: Read WandB Metrics

import wandb
api = wandb.Api()
run = api.run("<entity>/<project>/<run_id>")
history = run.history()

If WandB is unreachable (API error, network issue), fall back to reading the log file directly via SSH:

ssh server "tail -100 /path/to/training.log"

Check these signals:

• **Loss trend**: Is training loss decreasing over the last N steps?
• **Eval metrics**: Are evaluation metrics improving (or at least not degrading)?
• **NaN / Inf**: Any NaN or Inf values in loss or gradients?
• **Spikes**: Sudden large jumps in loss (>10x normal variance)?
• **Learning rate**: Is the schedule behaving as expected?
• **Gradient norm**: Exploding or vanishing?

Step 2: Judgment

| Signal | Judgment | Action | |--------|----------|--------| | NaN/Inf in loss | **Clearly bad** | Stop training, investigate | | Loss diverging (increasing for >N steps) | **Clearly bad** | Stop training, investigate | | Eval metrics significantly worse than baseline | **Clearly bad** | Stop training, investigate | | Loss decreasing, metrics improving | **Clearly fine** | Continue, increase check interval | | Loss flat but not diverging | **Unsure** | → Step 3 (Codex judgment) | | Metrics noisy, can't tell trend | **Unsure** | → Step 3 (Codex judgment) | | Slightly worse than baseline but still early | **Unsure** | → Step 3 (Codex judgment) |

Step 3: Codex Judgment (only when unsure)

Only escalate to Codex when the signal is ambiguous. For clearly good or clearly bad signals, act directly.

> Resolve this reviewer slot first; the following is the `backend=codex` example.

mcp__codex__codex:
  model: gpt-6-astra
  config: {"model_reasoning_effort": "xhigh"}
  prompt: |
    TRAINING HEALTH CHECK — need your judgment on ambiguous metrics.

    Run: <entity>/<project>/<run_id>
    Current epoch/step: X / Y total
    Training loss (last 10 checkpoints): [values]
    Eval metrics (last 3 evals): [values]
    Baseline reference: [numbers from paper/reproduction]

    What I'm unsure about: [specific concern]

    Please respond with exactly one of:
    - STOP: clearly problematic, should kill training
    - CONTINUE: looks fine, check again next interval
    - WAIT: not enough data to judge, check again sooner

Step 4: Act

| Decision | Action | |----------|--------| | **Stop** | Kill the training session. Save the WandB run URL, key metrics, and reason for stopping. Log to project notes for debugging. | | **Continue** | Do nothing. Will be invoked again at next interval (increase interval if consistently healthy). | | **Wait** | Do nothing but keep the current short interval (don't increase). |

Integration with Watchdog

Training-check and [watchdog.py](../../tools/watchdog.py) operate at different levels:

| Layer | Tool | What it checks | Frequency | |-------|------|----------------|-----------| | Process health | watchdog.py | Session alive? GPU active? | Every 60s (continuous) | | Training quality | training-check | Loss trend? Metrics improving? | Every 10-60 min (periodic) |

Use both together:

• Watchdog catches crashes and idle GPUs immediately
• Training-check catches subtle quality issues (loss plateau, metric degradation)

Rules

• Do not stop training on first sign of