name: scillm
description: >
Managed scillm Paved Path execution.
Provides strict contract-compliant tools for Text Batch, VLM, and Lean4 Proving.
Uses parallel_acompletions_iter for reliable large-scale processing.
allowed-tools: Bash, Read
triggers:
- batch LLM calls
- parallel completions
- describe image
- prove mathematically
- extract JSON
metadata:
short-description: Scillm Paved Path (Text, VLM, Proofs)

Scillm Paved Path Skill

This skill provides contract-compliant wrappers around scillm for robust agent operations.
It enforces the patterns defined in SCILLM_PAVED_PATH_CONTRACT.md.

Features

• Strict uv run Execution: No global environment dependencies (outside of uv).
• Parallel Iterators: Uses parallel_acompletions_iter for fault-tolerant batch processing.
• Multimodal Standards: Correctly formats VLM payloads.
• VLM Inputs: Accepts file paths, HTTPS URLs, or data: URIs; --inline-remote-images (or SCILLM_INLINE_REMOTE_IMAGES=1) downloads remote assets before dispatch, and --dry-run previews payloads without live calls.
• Preflight Helpers: run.sh preflight ... shells into scillm.paved.sanity_preflight and list_models_openai_like for Step 07 readiness checks.
• JSON Strict by Default: --json automatically enables SCILLM_JSON_STRICT, with optional --schema, --retry-invalid-json, and repair flags.

Usage Guide

1. Text Batch Processing (batch.py)

Use for large-scale text extraction or summarization.

Command:

.pi/skills/scillm/run.sh batch --input prompts.jsonl --output results.jsonl --json

Input Format (JSONL):

{"prompt": "Summarize this article..."}
{"prompt": "Extract names from...", "id": "123"}

Code Pattern (Python):
The skill implements this Paved Path pattern:

from scillm.batch import parallel_acompletions_iter

reqs = [
    {"model": "model-id", "messages": [{"role": "user", "content": "prompt"}]}
]

async for res in parallel_acompletions_iter(reqs, concurrency=6):
    if res["ok"]:
        print(res["content"])

2. VLM / Multimodal (vlm.py)

Use for describing images, diagrams, or Tables.

Command:

.pi/skills/scillm/run.sh vlm describe image.png --prompt "Extract table data" --json

• Supports --inline-remote-images (with optional --inline-remote-timeout) to download HTTPS assets when the gateway cannot reach them, and --dry-run to print the payload without making an API call (used by sanity scripts).

Batch Command:

.pi/skills/scillm/run.sh vlm batch --input images.jsonl

Code Pattern (Python):
The skill enforces the correct VLM message structure:

messages = [{
    "role": "user",
    "content": [
        {"type": "text", "text": "Describe this..."},
        {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}}
    ]
}]
await acompletion(..., messages=messages)

3. Lean4 Proving (prove.py)

Use for formal verification steps.

Command:

.pi/skills/scillm/run.sh prove "Prove that n + 0 = n"

4. Preflight + Model Discovery (preflight.py)

Use to run paved-step sanity_preflight and list models without bespoke scripts.

Commands:

# Model availability + auth style
.pi/skills/scillm/run.sh preflight preflight --model "$CHUTES_MODEL_ID" --json

# List models (returns JSON array)
.pi/skills/scillm/run.sh preflight models --json

These commands exit non-zero when the model is unavailable, making them CI-friendly.

Infrastructure

• Entry Point: run.sh (dispatches via uv run)
• Dependencies: Defined in pyproject.toml (scillm, typer)
• Verification: sanity.sh verifies CLI help and structural integrity.