[spending-forecast] Daily spending forecast

Overview

Forecast date: 2026-08-23 · History window: last 30 days · Period basis: monthly (with weekly derived views)
Source: gh aw forecast structured output (forecast.json), 49 workflows evaluated, 23 active (≥1 sampled run in the last 30 days), 26 inactive/dormant with no runs in the window.

Executive summary

Total observed AIC (sum of actually-run samples across all active workflows, last 30 days): 37,118.3
Total sampled runs: 385 across 23 active workflows
Weekly forecast total — P10 (10th percentile — optimistic: 9/10 weeks cost at least this much): 2,775.7 · P50 (50th percentile — median): 7,219.5 · P90 (90th percentile — conservative: only 1/10 weeks expected to exceed): 13,473.5
Monthly forecast total — P10 (optimistic): 19,812.9 · P50 (median): 32,043.0 · P90 (conservative): 47,424.3

The three highest-spend workflows — Go Logger Enhancement, Agentic Workflow Audit Agent, and CLI Version Checker — together account for roughly 58% of total observed AIC and monthly P50 projection, making them the primary budget drivers.

Charts

Key metrics — workflow table

All 23 workflows with ≥1 sampled run in the 30-day window, sorted by monthly P50 (median) projected AIC descending. "Reliable" (✅) marks Monte Carlo projections the forecaster considers stable (is_reliable=true); ⚠️ marks projections built from very small samples (typically ≤4 runs) where the confidence interval is wide relative to the estimate.

Workflow	Sampled runs	Observed AIC (30d)	P50 AIC/run	P95 AIC/run	Weekly P50 AIC	Monthly P50 AIC	Success rate	Monthly P10–P90 range
Go Logger Enhancement	28	9924.5	359.8	628.9	1677.2	7461.0	75%	4812–10551 ✅
Agentic Workflow Audit Agent	25	7138.7	282.1	408.0	1415.2	6279.2	88%	4053–8913 ✅
CLI Version Checker	27	4064.4	134.8	300.6	888.6	3917.3	96%	2575–5488 ✅
Copilot Agent PR Analysis	23	1798.3	78.3	94.5	389.5	1726.2	96%	1097–2460 ✅
Tidy	29	2016.7	67.1	150.6	373.1	1670.8	83%	1079–2357 ✅
Lockfile Statistics Analysis Agent	27	1637.6	59.3	89.2	357.7	1574.7	96%	1052–2190 ✅
Dev	27	1426.5	26.1	101.8	319.4	1423.9	100%	914–2032 ✅
Duplicate Code Detector	26	1446.2	51.9	141.4	308.1	1383.0	96%	855–2017 ✅
Smoke Copilot	34	2065.9	56.4	93.8	255.7	1144.2	56%	743–1615 ✅
Terminal Stylist	31	925.0	27.5	48.5	211.4	926.8	100%	640–1258 ✅
Daily News	21	833.5	33.3	83.5	185.2	836.7	100%	516–1200 ✅
Smoke Claude	11	825.8	75.9	105.6	181.4	830.5	100%	428–1361 ✅
GitHub MCP Remote Server Tools Report Generator	3	731.4	244.3	248.7	238.4	731.4	100%	238–1706 ⚠️
Weekly Workflow Analysis	3	439.6	137.7	214.7	87.1	439.6	100%	87–1084 ⚠️
Documentation Unbloat	30	427.2	13.6	20.3	94.6	415.0	97%	284–565 ✅
Daily Documentation Updater	15	372.6	23.5	33.8	81.9	368.9	100%	213–569 ✅
Smoke Codex	9	323.1	41.7	56.0	58.2	286.2	89%	126–506 ⚠️
Weekly Issue Summary	4	288.9	67.2	113.5	39.7	213.9	75%	40–493 ⚠️
Artifacts Usage Report	4	159.8	39.5	46.5	39.5	159.8	100%	44–341 ⚠️
Scout	2	150.6	59.2	91.3	0.0	150.6	100%	0–420 ⚠️
Repository Tree Map Generator	4	79.6	17.4	26.1	17.4	60.8	75%	17–132 ⚠️
Smoke OpenCode	1	22.9	22.9	22.9	0.0	22.9	100%	0–92 ⚠️
Plan Command	1	19.6	19.6	19.6	0.0	19.6	100%	0–78 ⚠️

Data quality and accuracy

26 workflows had zero sampled runs in the 30-day window (avg_aic=0, sampled_runs=0) — including CI, CodeQL, Dependabot Updates, several "Test *" and "Smoke *" workflows, and content workflows like "Poem Bot" and "Notion Issue Summary". These are excluded from all totals and the table above rather than shown as zero-cost, since a true zero cost cannot be distinguished from "did not run" or "no telemetry collected" in this dataset. Forecast impact: none on the totals reported (they are correctly excluded), but budget owners should confirm these workflows are genuinely dormant rather than silently failing to emit AIC telemetry.
9 workflows have is_reliable: false Monte Carlo projections (flagged ⚠️ above): GitHub MCP Remote Server Tools Report Generator, Weekly Workflow Analysis, Smoke Codex, Weekly Issue Summary, Artifacts Usage Report, Scout, Repository Tree Map Generator, Smoke OpenCode, Plan Command. All of these have ≤9 sampled runs (several have only 1–4), so their P10–P90 ranges are proportionally very wide (e.g., Scout: 0–420, a >∞% relative spread since P10=0). Forecast impact: their contribution to the aggregate totals (roughly 2,225 of the 32,043 monthly P50, ~7%) carries much higher uncertainty than the headline P50 implies; treat any single-digit-sample workflow's monthly figure as directional only, not a committed budget line.
History window is consistently 30 days across all 49 workflows (history_days=30 uniformly) — no inconsistent date windows detected.
No implausible run frequencies were found: sampled run counts range from 1–34 over 30 days, all consistent with normal trigger cadences (daily/weekly cron plus PR-driven triggers).
No negative or outlier AIC values were found in the observed run_samples; per-run AIC values are all in a plausible $0.02–$630 range for this repository's workflow mix.
No follow-up gh aw forecast --eval backtest or MCP-server evidence gathering was required — the prepared forecast.json was internally consistent (totals reconcile, no missing workflow entries, exit code recorded in forecast-metadata.txt was clean) and did not exhibit the "sampled_runs=0 for all workflows" failure mode that would have triggered escalation.

Full data reconciliation notes

forecast.json top-level: period=month, as_of=2026-08-23T09:48:18Z, 49 workflow entries — all parsed successfully.
Verified sampled_runs matches the length of each workflow's run_samples array for a sample of entries (Go Logger Enhancement: 28 samples listed and reported; CLI Version Checker: 27 vs 27).
Weekly and monthly aggregate totals above are sums of each active workflow's own weekly_monte_carlo / monthly_monte_carlo P10/P50/P90 fields (not re-derived from a separate combined simulation), so cross-workflow correlation is not modeled — the true portfolio-level P90 could be somewhat lower than a naive sum of individual P90s if workflow costs are not perfectly correlated.
26 dormant workflows (full list): Mergefest, .github/workflows/test-proxy, Doc Build - Deploy, Notion Issue Summary, Commit Changes Analyzer, Poem Bot - A Creative Agentic Workflow, Q, Rebuild the documentation after making changes, Go Pattern Detector, Resource Summarizer Agent, Copilot Setup Steps, Sentry Issue Analyzer, CodeQL, MCP Inspector Agent, CI Failure Doctor, Dependabot Updates, CI, Test, Test Claude, Test Copilot CLI Engine, Test Copilot GitHub Integration, Basic Research Agent, Video Analysis Agent, Format, Lint, Build and Commit, Dev Hawk, copilot only.

Assumptions

AIC figures are taken as-is from gh aw forecast's per-run cost estimates; no currency conversion or cost-model adjustment was applied.
"Observed AIC" = sum of run_samples[].aic for each workflow, i.e., actual spend recorded for sampled runs in the last 30 days (not a projection).
"Projected" P10/P50/P90 figures come directly from each workflow's weekly_monte_carlo / monthly_monte_carlo blocks (Monte Carlo simulation, 10,000 iterations per workflow).
Workflows with zero sampled runs are treated as having no forecastable near-term cost and are excluded from aggregate totals, not assumed to cost $0 going forward.
Report generated from run §32631819113.

Next actions

Monitor Go Logger Enhancement (75% success rate, highest observed AIC) — its retry/failure rate may be inflating spend; investigate failure causes to reduce wasted runs.
Treat the 9 low-sample (⚠️) workflows' monthly projections as low-confidence; re-run this forecast in 1–2 weeks once more samples accumulate before using them for budget commitments.
Periodically confirm the 26 zero-run workflows are intentionally dormant (e.g., disabled, manual-trigger-only) rather than silently failing to execute or report telemetry.
No follow-up MCP evidence gathering or forecast re-run was needed this cycle; data was internally consistent.

Generated by 📈 Daily Spending Forecast · auto · 37.1 AIC · ⌖ 4.72 AIC · ⊞ 11.3K · ◷

expires on Aug 30, 2026, 1:54 AM UTC-08:00