[spending-forecast] Daily spending forecast - 2026-08-25

The forecast engine itself flags is_reliable: false for both weekly and monthly
Monte Carlo projections on: GitHub MCP Remote Server Tools Report Generator,
Scout, Weekly Workflow Analysis, Weekly Issue Summary, Artifacts Usage
Report, Q, Repository Tree Map Generator, Smoke OpenCode, and Plan
Command. Root cause: each has only 1–7 sampled runs in the 30-day window, too few
for a statistically meaningful weekly/monthly distribution — several show a P10 of
0.0, meaning the lower bound collapses to "may not run at all this week," which is
plausible for low-frequency or manually-triggered workflows but should not be treated
as a hard floor for budgeting. Forecast impact: their contribution to the P10 total
is likely understated; their contribution to P90 carries wide uncertainty. No
follow-up evidence was pulled since sample counts are inherently small (these are
low-frequency workflows, not a collection failure) and the totals above already
reflect the raw Monte Carlo output as-is.

The following workflows returned sampled_runs: 0 in the 30-day window: Copilot
Setup Steps, .github/workflows/test-proxy, Notion Issue Summary, MCP Inspector
Agent, Poem Bot - A Creative Agentic Workflow, Copilot cloud agent, Rebuild the
documentation after making changes, Go Pattern Detector, Resource Summarizer Agent,
Doc Build - Deploy, Commit Changes Analyzer, Sentry Issue Analyzer, CodeQL,
Mergefest, 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.

Many of these are conventional CI/CD or setup workflows (CodeQL, CI, Test, Format
Lint Build and Commit, Dependabot Updates) that are expected to have zero or
near-zero AIC since they are not LLM-driven agentic runs, so their exclusion from the
forecast is expected and does not indicate a collection failure. A smaller subset
(Notion Issue Summary, MCP Inspector Agent, Poem Bot, Copilot cloud agent, Go
Pattern Detector, Resource Summarizer Agent, Sentry Issue Analyzer, Basic Research
Agent, Video Analysis Agent) are agentic-style workflows with zero recent runs,
suggesting they are currently disabled, unscheduled, or simply idle in this window —
this is consistent with intermittent/manual triggers rather than a data-collection
gap, since 25 other workflows across the same window did produce non-zero samples
from the same gh aw forecast collection pass. No rerun was performed because
sampled_runs was not zero for all workflows (only a subset), which is the
documented threshold for escalating to direct run/artifact inspection.

Smoke Copilot shows the lowest success rate among active workflows (62%,
23/37 runs), consistent with it being a smoke-test workflow that is expected to
surface engine failures. No other active workflow falls below 75% success.
Success rate does not directly bias the AIC forecast (both successful and failed
runs incur cost), but a low success rate for a production (non-smoke) workflow
would warrant investigation; here it is limited to explicitly-named smoke/test
workflows.

All 25 active workflows report history_days: 30, consistent with the forecast
period. sampled_runs matches observed_runs_per_period for every active
workflow, and p50_aic_per_run values fall within the observed min/max of each
workflow's run_samples[].aic, confirming internal consistency of the reported
percentiles. No AIC values were zero, negative, or missing for any run sample
among the 25 active workflows.