Backlog Session Scorecard¶
Generated from a repository-tracked measurement ledger so issue #1361 has operational data instead of an advisory-only target.
- Generated: 2026-10-10T12:53:23.888Z
- Source data:
.github/backlog-session-metrics.json - Output:
docs/reference/BACKLOG_SESSION_SCORECARD.md - Refresh command:
python scripts/generate-backlog-efficiency-scorecard.py - Tracking window: latest 5 recorded entries for practice metrics and ledger; latest 5 baseline-comparable sessions for target and savings
- Target band: 35,000,000-45,000,000 tokens
- Expensive baseline: 68,000,000-84,000,000 tokens
- Best measured reference: 9,100,000 tokens, 101 events, 207x ratio
Scorecard Summary¶
Tracking in progress: 0/5 baseline-comparable session(s) recorded; 0/0 in the target band so far.
| Metric | Value |
|---|---|
| Recent ledger entries | 1 |
| Baseline-comparable sessions | 0 |
| Comparable sessions remaining to fill window | 5 |
| Baseline-comparable sessions in target band | 0/0 |
| Average tokens (baseline-comparable) | - |
| Average events | 950.0 |
| Average ratio | 161.6x |
| Average savings vs 76M midpoint (baseline-comparable) | - |
| Average phase compactions | 0.0 |
| Average max captured user-message size | 1.1 KB |
Baseline target-band and savings calculations use only sessions marked
baselineComparable: true. Other entries remain visible in the ledger but are
excluded from single-wave comparisons. Prompt-size values are captured user-message
measurements; check each entry's notes for coverage limitations.
Policy Adoption¶
| Practice | Threshold | Adoption |
|---|---|---|
| Phase-boundary compaction | >= 2 compactions | 0/1 |
| Sprint template reuse | usedSprintTemplate = true |
0/1 |
| File-reference-only prompts | usedFileReferencesOnly = true |
0/1 |
| Delegated scan work | delegatedScanWork = true |
0/1 |
| Max captured user-message size | <= 10 KB | 1/1 |
Session Ledger¶
| Date | Session | Tokens | Events | Ratio | Baseline Comparable | Compactions | Max Captured User Message | Template | File Refs | Delegated Scan | Notes |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 2026-10-10 | Oldest-first backlog wave (umbrella-session snapshot) | 117,030,409 | 950 | 161.6x | no | 0 | 1.1 KB | no | no | no | Refreshed local assistant_usage_events snapshot at usage_cut 2026-10-10T12:53:23.888Z UTC. Replaces the same umbrella's 2026-10-09T22:04:15.215Z cut (75,187,703 tokens, 580 recorded calls), not a new independent session. Current arithmetic is 116,310,470 input + 719,939 output = 117,030,409 total tokens; input already includes cache usage, so cache counts are not added again. Ratio round(input/output, 1) = 161.6; 950 events are recorded API calls, not GitHub workflow events. This spans earlier backlog waves, repository-only cleanup and current read-only observation/conflict-fixture/consumer-label delivery, not one wave. Prior maxPromptKb 1.1 remains the historical partial captured-message observation, not a fresh measurement or complete payload maximum. False optimization fields remain conservative: automatic summaries are not /compact evidence, and prior separate-context artifact actors do not prove umbrella compaction or savings. Raw event history and exact per-wave boundaries remain unavailable. This cut is not the final session total and excludes subsequent work. baselineComparable remains false; never compare this umbrella directly to a single-wave baseline or infer savings. |
Logging Contract¶
Record one object per backlog session with these required fields:
{
"date": "2026-06-13",
"label": "Backlog sprint execution",
"tokens": 41200000,
"events": 188,
"ratio": 233,
"baselineComparable": true,
"phaseCompactions": 2,
"maxPromptKb": 6.4,
"usedSprintTemplate": true,
"usedFileReferencesOnly": true,
"delegatedScanWork": true,
"notes": "Compact at triage->implementation and implementation->merge-waiting."
}
Entries that omit baselineComparable are treated as comparable for
backward compatibility; mark umbrella or otherwise non-like-for-like runs
false so they remain visible without affecting baseline calculations.
Use this scorecard together with docs/templates/sprint-structure.md and
.github/instructions/cost-optimization.instructions.md to keep the next
5 backlog sessions inside the 35-45M token target band.