| name | mcp-geo-long-horizon-summary |
| description | Generate a Long Horizon-style Codex metrics summary for mcp-geo from local Codex session logs. |
MCP Geo Long Horizon Summary
Use this skill when a user asks for an OpenAI "Long horizon tasks" style summary
for MCP Geo work (runtime, tokens, tool calls, patch volume, compactions, etc.).
Inputs
- Codex session logs under
$CODEX_HOME (defaults to ~/.codex).
- Repo filter (default:
mcp-geo).
Output
- Summary-card SVG image in
docs/reports/.
- Markdown report in
docs/reports/ that opens with the summary image.
- Optional JSON metrics payload for downstream dashboards.
Runbook
- Generate report:
python3 scripts/codex_long_horizon_summary.py \
--repo-filter mcp-geo \
--summary-svg-output docs/reports/mcp_geo_codex_long_horizon_summary_$(date -u +%F).svg \
--summary-title "Codex MCP-Geo Summary" \
--output docs/reports/mcp_geo_codex_long_horizon_summary_$(date -u +%F).md \
--json-output docs/reports/mcp_geo_codex_long_horizon_summary_$(date -u +%F).json
- If needed, point to a non-default Codex home:
python3 scripts/codex_long_horizon_summary.py \
--codex-home /path/to/.codex \
--repo-filter mcp-geo
- Validate that the report includes these headline metrics:
- Active runtime
- Wall-clock span
- Token usage (input/output)
- Cached input reused
- Tool calls (shell + patch calls)
- Patch volume (added lines from
apply_patch)
- Peak single-step tokens
- Auto context compactions
Notes
- Sessions are included when
session_meta.payload.cwd contains the repo filter.
- Patch volume is estimated from
apply_patch payloads and will not include
edits made through other methods.
- Token totals use the latest
token_count.total_token_usage snapshot per session.
- Graphic output uses deterministic template file
skills/mcp-geo-long-horizon-summary/templates/summary_card.svg.tmpl with
slot-based number substitution.
- Template is tuned for slide import (
1920x1080, 16:9) and uses
PowerPoint-friendly inline text styling (no CSS font shorthand) to keep
text sizing consistent.