| name | variance-analysis |
| description | Compute month-end variance — driver decomposition (volume/price/mix/one-off heuristic), anomaly flags, and a matplotlib bar chart of variance % by cost centre saved to /tmp for download. |
| enabled | true |
Instructions
Use this skill whenever the controller asks for a variance review, a chart, or any "is this number reasonable?" sanity check. This is the computation surface — it uses code_interpreter against the numbers returned by sap-s4. Do NOT do mental math on financial figures; always send the numbers through this skill.
Workflow
- Pull the raw rows via
sap_get_variance_analysis for the requested period.
- Pass the resulting JSON into
code_interpreter with the analysis script below — DO NOT retype the numbers.
- The script emits:
- A compact table per flag bucket (
investigate, watch, normal)
- A matplotlib bar chart saved to
/tmp/variance-<period>.png
- A CSV of the underlying data at
/tmp/variance-<period>.csv for the close pack
- Reference both file paths in your response so the file-sharing convention picks them up.
Reference script (variance_analysis.py)
The full script is in scripts/variance_analysis.py in this skill's directory. Read it once with the file_read tool, then invoke it via code_interpreter — DO NOT inline the script body in your response.
The script signature:
python /app/use-cases/finance-close/skills/variance-analysis/scripts/variance_analysis.py \
--period 2026-05 \
--rows-json '<the JSON returned by sap_get_variance_analysis>' \
--out-dir /tmp
Outputs to stdout: the per-bucket tables. Outputs to disk: /tmp/variance-2026-05.png and /tmp/variance-2026-05.csv.
Driver decomposition heuristic
The script also classifies each investigate row with one of:
- price —
gl_account.type == "OpEx" and prior-year run-rate close, but unit-rate moved (Sentinel-style licence uplift)
- volume — variance scales with a known cost-centre headcount / production-order change
- one-off — single JE >50% of the variance (manual / accrual / reclass), look up via
sap_list_journal_entries
- mix — none of the above; flag for owner explanation
This is a suggestion to seed owner commentary, not a determination. Always present it as Suggested driver: … and let the human override.
When NOT to use
- For a single number / single GL → just cite
sap_get_variance_analysis inline; don't build the whole chart.
- For non-finance computation (e.g. parse an email body) → use
code_interpreter directly without this skill.
