gencast-source-check

gencast source-check

Preflight a gencast notebook (or candidate source files) before spending money. Reports source-token counts, predicted cost, fits-in-budget verdict, and a subjective topical-coherence read. Identifies opportunities to split the notebook, summarise first, or pick a cheaper model.

Available episode profiles

gencast list-profiles --type episodes --json 2>/dev/null | jq -r '.[] | "- **\(.name)**: \(.description // "(no description)")"' || echo "(gencast not installed)"

Version check (need gencast>=1.2.0): !gencast --version 2>/dev/null | python3 -c "import sys,re; v=sys.stdin.read().strip(); m=re.search(r'(\d+)\.(\d+)', v); sys.exit(1) if not v else print(v if m and (int(m[1]),int(m[2]))>=(1,2) else f'TOO OLD: need gencast>=1.2.0, found {v} — run: pipx upgrade gencast')" 2>/dev/null || echo "MISSING — install with: pipx install gencast"

Prerequisites

• gencast>=1.2.0 on PATH
• Either an existing notebook.yaml, OR a list of candidate source files the user wants to evaluate.

Workflow

1. Identify the input. If the user gave a notebook.yaml, use it directly. If the user only named source files, scaffold a minimal notebook in a temp file (using bundled defaults) so gencast estimate has something to operate on. Do not prompt the user for profiles in this case — defaults are fine for a preflight.

2. Run gencast estimate.

   gencast estimate <NB.yaml> --json
   

   Parse the JSON output: total_usd, per-stage breakdown, source_tokens, suggestions.

3. Read the source content. For each source file, read the actual text (~first 2,000 chars is enough for a coherence read). Do not load entire large PDFs.

4. Report findings in this order:

   • Cost verdict. "$X estimated (±25%). Fits within typical $0.10–$0.50 range." OR "expensive — $Y; consider cheaper outline model."
   • Token verdict. "Source is N tokens; well within budget." OR "Source exceeds budget — gencast will run map-reduce summarisation first (adds cost)."
   • Coherence. Subjective read — "all three sources are about Topic X" or "source 2 is off-topic, consider splitting".
   • Suggestions. Surface any suggestions from the estimate JSON (cheaper-model swaps), and add the user's own based on coherence (split into multiple notebooks, summarise first, etc.).

5. Output format. Plain-language summary in 5–8 lines. No tables unless the user asks.

What to NOT do

• Do not run gencast generate. This skill is preflight only.
• Do not edit the user's notebook. Suggestions are advisory.
• Do not load entire PDF binaries — use the first ~2,000 chars per source for the coherence read.
• Do not re-run gencast estimate multiple times. One call per notebook.