chore: add flash agent skill co-located with source code

[TimPietruskyRunPod]

Summary

• Adds flash/SKILL.md (588 lines) rewritten around the unified Endpoint class API
• Replaces the old skill in runpod/skills which documented the deprecated 8-class resource hierarchy (LiveServerless, CpuLiveServerless, etc.)
• Co-locating the skill with source code ensures it stays in sync with the codebase
• Discoverable via npx skills add runpod/flash

Skill contents

1. Quick reference and getting started (flash login, flash init, flash run, flash deploy)
2. Four Endpoint modes: QB decorator, LB decorator, external image client, existing endpoint client
3. Full constructor parameter table (verified against endpoint.py)
4. EndpointJob API
5. GPU groups & types, CPU instance types (verified against enums)
6. Cloudpickle scoping rules
7. All CLI commands (verified against cli/main.py)
8. Common patterns matching skeleton templates
9. Architecture overview and common gotchas

Test plan

• Verify npx skills add runpod/flash discovers the skill
• Verify code examples match skeleton templates (gpu_worker.py, cpu_worker.py, lb_worker.py)
• Verify all constructor parameters match Endpoint.__init__ in endpoint.py
• Verify CLI commands match registrations in cli/main.py
• Verify GPU/CPU enums match current definitions

[runpod-Henrik]

PR #254 Review: chore: add flash agent skill co-located with source code

1. Bug: Fabricated CLI subcommands (high confidence — verified against cli/main.py and flash deploy --help)

The CLI section documents subcommands that don't exist:

flash deploy new staging      # ❌ not a real command
flash deploy send staging     # ❌ not a real command
flash deploy list staging     # ❌ not a real command
flash deploy info staging     # ❌ not a real command
flash deploy delete staging   # ❌ not a real command

The actual flash deploy is a single command with --env:

flash deploy --env staging           # build + deploy to staging
flash deploy --exclude torch,pkg2    # exclude packages

Environment management is via flash env:

flash env list
flash env create
flash env get
flash env delete

2. Issue: execution_timeout_ms missing from the constructor table

Endpoint.__init__ has execution_timeout_ms: int = 0 but it's absent from the skill's parameter table. This is user-facing (needed for long-running jobs) and was a known fix (executionTimeoutMs → execution_timeout_ms snake_case rename in 1.7.0).

3. Question: "Auto GPU switching requires workers >= 5"

Gotcha #8 states this as a rule. Is this a documented platform policy? If not verified, it could mislead users into setting unnecessarily high worker counts.

4. Nit: ADA_80_PRO VRAM labeled "80GB" but includes H100 NVL

The source docstring says: "NVIDIA H100 PCIe, NVIDIA H100 80GB HBM3, NVIDIA H100 NVL". H100 NVL is 94GB, not 80GB. Worth noting the label is approximate.

5. Nit: Missing less-common constructor params

accelerate_downloads, datacenter, scaler_type, scaler_value are in __init__ but absent from the table. Fine to omit as advanced, but accelerate_downloads=False is a useful workaround for slow dep installs.

---

Verdict: NEEDS WORK — The CLI section needs to be corrected before merge. An agent using this skill will generate flash deploy new/send/list commands that fail immediately.

🤖 Reviewed by Henrik's AI-Powered Bug Finder

[TimPietruskyRunPod]

we continue in runpod/skills#7