Index of every page on this site: llms.txt
All skills
Run it yourself

Audits a repository's GitHub Actions workflows for CI optimization opportunities — missing caches, redundant setup, sleep-based readiness, long test jobs without sharding, full-history checkout, dead env vars, build-cache misconfig, and ~60 more patterns across caching, redundancy, parallelization, conditional execution, trigger scope, and hidden failures. Use when: (1) analyzing a repo's CI for optimization opportunities, (2) producing a prioritized report with measured wall-clock and runner-minute savings, (3) re-auditing after upstream CI changes. Do not trigger for: general CI setup help, writing new workflows from scratch, non-GitHub-Actions CI systems, or security/posture audits — use `ci-secure` for those.

Use this Skill: https://skilld.dev/gh/starslingdev/skills/ci-speedup

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

referencesreport-defaults.md

≈12k tokens on demand. Your agent reads this file only when SKILL.md points to it.

Report defaults review

Status: awaiting owner approval

This document records the choices behind the report's sampling, ordering, and finding thresholds. A value marked keep has a reason we can defend to a reader. A value marked change is a proposal until the owner approves this document and the named public issue lands. Approval should cover every proposal, including the values rather than only the direction of change.

Contents

Gate sample

pr-gate-sample-size

Current value: a minimum target of 100 completed PR commits (pr_sample_contract.PR_SAMPLE_TARGET). The sample can exceed it while spreading across days, and the report then says "N PRs sampled (minimum target 100)"; it stops short when the page or job-list read cap binds first. The unrelated --max-runs option bounds per-workflow timing history and does not set this population.

Report statements: PR-gate P50/P90 durations, gate frequency, pole order, and every claim that describes a check as typical or rare.

Reason: One hundred keeps the API cost bounded while reducing sensitivity to one short traffic burst.

Proposed decision: Keep 100 completed PR commits. #315, relanded by #333, centralizes the value as pr_sample_contract.PR_SAMPLE_TARGET; it remains awaiting owner approval.

pr-gate-eligibility

Current value: only completed PR commits are eligible; excluded in-flight commits are disclosed rather than treated as zero or clean.

Report statements: sample completeness, "finishes last," gate frequency, pole order, and the Bottom line.

Reason: A last-finisher cannot be known until every sampled run for the commit finishes. Partial commits remain useful coverage evidence.

Proposed decision: Keep completed PR commits only and retain the in-flight exclusion disclosure.

Approval status: awaiting owner approval.

pr-gate-span

Current value: at least 5 distinct days when history permits (pr_sample_contract.PR_SAMPLE_MIN_DISTINCT_DAYS). On the busiest repositories the page and job-list read caps bind first: the Home Assistant and Next.js golden replays stop at job_list_cap with 100 commits over 2 days, and the report states that stop reason.

Report statements: every frequency, interval, "typical PR," and pole-order claim derived from the gate population.

Reason: Multiple days reduce sensitivity to one traffic burst, weekday, or temporary path mix. Five days covers most of a working week.

Proposed decision: Keep at least 5 distinct days and continue printing the first and last sampled dates. #315, relanded by #333, centralizes the value in pr_sample_contract.py; it remains awaiting owner approval.

Approval status: awaiting owner approval.

pr-candidate-pages

Current value: 12 pages per workflow (pr_sample_contract.PR_SAMPLE_CANDIDATE_PAGE_CAP), with an early stop once 100 eligible commits over 5 days are available, and at most 1,200 job-list reads for the sample per audit (pr_sample_contract.PR_SAMPLE_JOB_LIST_READ_CAP), reported with the stop reason job_list_cap when it binds.

Report statements: sample stop reason, candidate exhaustion, sample count, time span, and every downstream gate-frequency claim.

Reason: A fixed cap prevents an unbounded walk, while twelve pages leave room for incomplete or ineligible commits before reaching the target.

Proposed decision: Keep the cap at 12 pages and 1,200 job-list reads. #315, relanded by #333, centralizes the cap in pr_sample_contract.py. Approve the API cost ceiling together with the sample target; the value remains awaiting owner approval.

config-era-boundary

Current value: a post-change era needs 6 observations to stand alone (_RARE_PRESENCE_MIN_PR); candidate paging is disabled when any config-era fact exists, including a boundary outside the workflow being measured.

Report statements: previous-configuration, current-configuration, and thin era warnings, plus every post-era P50.

Reason: Mixing workflow definitions makes a timing population incoherent. One workflow's change does not invalidate another workflow's history.

Proposed decision: Change paging under #306 so an era limits only its own PR-triggered workflow. Keep 6 as the minimum sufficient post-change population pending a separate evidence-led review.

Approval status: awaiting owner approval.

Lookbacks and cohorts

rolling-volume-window

Current value: 30 days (collect_runs.py::_window_30d).

Report statements: monthly run volume, runs per month, runner-minutes per month, default-branch activity, and modeled monthly savings.

Reason: Thirty days is recent enough to reflect current CI and long enough to include weekly or release-adjacent jobs. It is a rolling window, not a calendar-month claim.

Proposed decision: Keep 30 days and print the exact endpoints when available.

Approval status: awaiting owner approval.

run-list-depth

Current value: 100 all-status runs per workflow (_COST_RUNLIST_MAX).

Report statements: run-volume sampling, cost-spine inputs, push/PR lane classification, retry evidence, and duplicate-run evidence.

Reason: One full GitHub API page is a bounded, cheap first view and avoids silently accepting the API's smaller default page.

Proposed decision: Keep 100. Deeper walks need their own explicit cap and coverage disclosure.

Approval status: awaiting owner approval.

default-branch-push-cohort

Current value: 10 commits (_DEFAULT_BRANCH_PUSH_SAMPLE_MAX).

Report statements: after-merge lane P50/P90, slowest jobs, and projected push-lane savings.

Reason: Ten provides a bounded recent cohort without mixing the push lane into the merge gate.

Proposed decision: Keep 10. Continue disclosing fetch failures and the sample cap.

Approval status: awaiting owner approval.

Pole selection and order

rendered-pole-count

Current value: 5 distinct workflow/job matrices (_TOP_WORKFLOWS).

Report statements: Long pole list, drill-down count, captured logs, and hidden-pole coverage.

Reason: Five keeps the report readable while retaining several competing paths; the collector captures the same number of drill logs.

Proposed decision: Keep 5 and retain the collector/renderer equality invariant.

Approval status: awaiting owner approval.

holds-up-threshold

Current value: before #318, a check was recurrent after finishing last on 2 sampled PRs (_POLE_RECUR_FLOOR) without an explicit two-group presentation. #318 provisionally keeps 2 as gate_frequency_contract.POLE_HOLDS_UP_MIN_COUNT, while #317 provisionally requires 5 as gate_frequency_contract.GATE_MIN_COUNT_FOR_RANKING.

Report statements: recurring-gate framing and, after #305, membership in "Checks that hold up most PRs."

Reason: Two rejects a single outlier, but it is too weak for a larger population and can elevate roughly one-percent events.

Proposed decision: Change to 5 last-finisher observations under #305 and #311. The interval and span must also support ordering. The fixed 2 in #318 and minimum 5 in #317 govern one ordering decision with two values; both remain awaiting owner approval, and the implementations must converge after approval.

rare-presence-rule

Current value: at most 50% presence is rare, but presence demotion is inactive below 6 sampled PRs (_RARE_PRESENCE_FRAC, _RARE_PRESENCE_MIN_PR). Required checks are exempt.

Report statements: minority/conditional labels, rare-pole caveats, and the typical-PR floor.

Reason: The median PR does not run a check present on half or fewer of the sample, while six observations guards against a one- or two-run guess.

Proposed decision: Keep 50% and 6 for presence framing. Gate-group membership uses last-finisher count and uncertainty as a separate rule.

Approval status: awaiting owner approval.

within-group-order

Current value: legacy selection prioritizes typical/recurrent checks and then duration; it does not promise two groups sorted by displayed P50.

Report statements: Long pole numbering, "longest first," the biggest-lever link, and the Bottom-line check.

Reason: Within one reader-facing group, descending P50 is direct and visible from the row itself.

Proposed decision: Change to descending displayed P50 within each labelled group under #305. Equal P50 uses stable workflow/check identity only as a deterministic tiebreaker.

Approval status: awaiting owner approval.

needs-chain-handling

Current value: serial needs: members add to a measured chain; the chain headline is bounded by observed makespan, with a 25% divergence threshold (_CHAIN_MAKESPAN_DIVERGENCE_PCT).

Report statements: chain P50, stage numbering, addressable chain headroom, and the Bottom-line chain explanation.

Reason: Serial stages form one path rather than competing parallel poles; the makespan guard keeps the sum consistent with observed developer wait.

Proposed decision: Keep the chain arithmetic and 25% divergence guard. After #305 the Bottom line also identifies Long pole 1, or names a chain that contains it.

Approval status: awaiting owner approval.

bookkeeping-cutoff

Current value: several reader-facing floors already shape related output: fast-workflow triage skips job lists for a workflow whose sampled wall stays under _TRIAGE_WALLCLOCK_FLOOR_S = 90 seconds, then recovers any triaged workflow whose run-list wall reaches _TRIAGE_RECOVER_GATE_FRAC = 0.5 of the measured gate, so a seconds-scale gate's competing checks are still ranked; aggregate-gate handling uses _AGG_GATE_TRIVIAL_S = 30; wall-clock saving presentation uses _WALL_CLOCK_LONG_POLE_FLOOR_S = 30; and no-op classification uses _NOOP_FLOOR_S = 30. They do not yet form one general bookkeeping rule.

Report statements: which short jobs appear as developer-wait poles and which are disclosed separately.

Reason: A short last-finisher can be bookkeeping, but duration alone cannot demote a confirmed gate or a stage on its required chain.

Proposed decision: Add a conditional 60-second cutoff under #311: list a short job separately only when it finished last and is outside the confirmed gate chain; always retain a qualifying gate.

Approval status: awaiting owner approval.

ranking-minimum-count

Current value: no ranking floor beyond the 2-occurrence recurrence rule. #317 provisionally adds 5 through gate_frequency_contract.GATE_MIN_COUNT_FOR_RANKING; #318 provisionally uses 2 through POLE_HOLDS_UP_MIN_COUNT for its first group.

Report statements: pole 1, ranked close guidance, and Bottom-line gate identity.

Reason: A check last only a handful of times has a wide frequency interval; a precise order overstates the sample.

Proposed decision: Require at least 5 last-finisher observations under #311, in addition to the multi-day and interval rules. This is the same owner decision as holds-up-threshold; both provisional values remain awaiting owner approval.

gate-frequency-interval

Current value: none is displayed.

Report statements: "finishes last N/M," whether two checks are ranked, and whether the chat close may name one above another.

Reason: A rough binomial interval shows when a small count difference is not evidence of stable order without pretending to forecast future CI.

Proposed decision: Add a two-sided 95% Wilson score interval (z = 1.96) under #311. Overlap makes the order unsupported while observations remain shown.

Approval status: awaiting owner approval.

Other output-shaping defaults

workflow-timing-depth

Current value: job lists are fetched for the first 10 runs of each workflow (_SHALLOW_RUNS), then the owners of at most 12 leading checks are deepened to the full timing depth (_DEEPEN_TOP_CHECKS). Run lists and other measurements remain at full breadth.

Report statements: workflow and check P50/P90 values, long-pole evidence, and whether a row has enough observations for a measured saving.

Reason: A shallow pass bounds API cost while targeted deepening spends reads on checks most likely to shape the report.

Proposed decision: Keep 10 shallow runs and 12 deepened checks. Both values remain awaiting owner approval.

required-suite-eligibility

Current value: when the required suite is observable, a representative PR commit is eligible only when all observable required checks ran; the selection is implemented by collect_runs.py::_select_repr_shas.

Report statements: sample size and stop reason, required-suite scope, gate share, pole order, and the Bottom line.

Reason: Mixing docs-only or path-filtered PRs into a full-suite population changes the denominator and can make required checks look rare.

Proposed decision: Keep full-required-suite eligibility and disclose unobservable required checks. This semantic default remains awaiting owner approval.

Approval status: awaiting owner approval.

typical-floor-thresholds

Current value: when per-pole co-occurrence data is available, a floor needs strictly more than 50% co-occurrence (_FLOOR_COOCCUR_MAJORITY_FRAC). When it is unavailable, the fallback needs 80% presence relative to the most-present sibling (_FLOOR_NEAR_UNIVERSAL_FRAC).

Report statements: typical-PR floor, recurrent chain framing, and which check is treated as part of the common wait.

Reason: The floor should describe a common path rather than a minority combination of checks.

Proposed decision: Keep the more-than-50% co-occurrence path and its 80% presence fallback. Both values remain awaiting owner approval.

Approval status: awaiting owner approval.

severity-cutoffs

Current value: 300 seconds is HIGH and 120 seconds is MEDIUM (_SEV_HIGH_S, _SEV_MED_S).

Report statements: the colored severity dot on Long pole headings and their Contents rows.

Reason: These cutoffs make the measured long-pole wait visually scannable without changing generic finding severity, section order, or measured duration.

Proposed decision: Keep 300 and 120 seconds. Both values remain awaiting owner approval.

Approval status: awaiting owner approval.

success-only-caveat

Current value: the success-only caveat needs 10 runs and a 20% failure rate (_SUCCESS_ONLY_CAVEAT_MIN_RUNS, _SUCCESS_ONLY_CAVEAT_FAILURE_RATE).

Report statements: whether a duration is labelled success-only and whether failed-run behavior is called out beside the estimate.

Reason: Sparse failures should not dominate the report, while a sustained failure share must be visible before a success-only timing is used.

Proposed decision: Keep 10 runs and 20%. Both values remain awaiting owner approval.

Approval status: awaiting owner approval.

minimum-timed-runs

Current value: a per-run compute mean needs 3 timed runs (_MIN_TIMED_RUNS).

Report statements: compute means, monthly runner-minute projections, and whether a saving is measured or withheld.

Reason: Fewer than three observations cannot show even basic run-to-run variation.

Proposed decision: Keep 3. The value remains awaiting owner approval.

account-weekly-paired-minimum

Current value: at least 5 explicit paired run observations per side (before_after.py::_ACCOUNT_WEEKLY_MIN_PAIRED_RUNS), with equal before and after counts, before an account weekly export states a runner-minute effect.

Report statements: the paired before/after median runner-minute effect in the optional account-level weekly section, or the reason that comparison is withheld.

Reason: A handful of runs per side cannot show ordinary run-to-run variation. Five per side is enough for each side's median to be meaningful and matches OPT81's minimum of 5 runs per runner size.

Proposed decision: Require 5 paired observations per side.

Approval status: owner-approved 2026-11-20 (5).

rare-push-lane

Current value: fewer than 8 default-branch push runs in 30 days is treated as rare (_PUSH_LANE_RARE_PER_30D).

Report statements: push-lane confidence, rarity caveats, and whether the after-merge lane is promoted as a recurring opportunity.

Reason: A very small monthly push cohort should not carry the same weight as a routinely exercised lane.

Proposed decision: Keep 8 per 30 days. The value remains awaiting owner approval.

Approval status: awaiting owner approval.

lifecycle-minimum

Current value: PR lifecycle analysis needs 10 measured PRs (pr_lifecycle.MIN_MEASURED).

Report statements: lifecycle queue and merge-wait claims and their insufficient-sample caveats.

Reason: Lifecycle distributions need a modest population before their percentiles are presented as representative.

Proposed decision: Keep 10. The value remains awaiting owner approval.

merge-queue-proof-thresholds

Current value: OPT98 requires 30 merges per day, 15 successful runs per event, a queue/PR median ratio from 0.8 through 1.5, a 30-PR churn sample, and a 30-day window (_OPT98_*).

Report statements: whether the merge queue is said to repeat the full suite and whether that result receives measured savings.

Reason: The finding should require both high merge pressure and enough comparable PR and queue runs, while rejecting workflows that do materially different work.

Proposed decision: Keep 30 / 15 / 0.8 / 1.5 / 30 / 30 days. Every value remains awaiting owner approval.

timeout-recommendation-thresholds

Current value: OPT57 treats 95% of GitHub's 360-minute default as near the ceiling, needs 3 successful samples, recommends max(15 minutes, P99 + 10 minutes, 1.5 × P99) of the observed successful runs, and accepts 15 minutes of kill grace (_OPT57_*).

Report statements: timeout-risk findings, the recommended timeout, and whether a failed job is described as reaching the platform ceiling.

Reason: Recommendations need successful duration evidence and headroom, while remaining below the platform's hard stop.

Proposed decision: Keep 95% / 3 samples / 15 minutes / 10 minutes / 1.5x / 15 minutes. Every value remains awaiting owner approval.

dependency-cache-floor

Current value: dependency-cache evidence needs a median miss rate of at least 50% before it can support the corresponding cache claim (_DEPENDENCY_CACHE_FLOOR_PCT).

Report statements: dependency-cache confidence, saving credit, and withheld-evidence wording.

Reason: A cache that misses on fewer than half of the relevant runs is mostly warm and should not receive cold-cache saving credit.

Proposed decision: Keep 50%. The value remains awaiting owner approval.

cold-cache-miss-threshold

Current value: 99.5% is the cold-miss threshold (_CACHE_COLD_MISS_PCT).

Report statements: cold-cache labels, cache saving credit, and the warning that a cache is effectively never warm.

Reason: The label should be reserved for cohorts with virtually no useful hits while allowing small measurement noise.

Proposed decision: Keep 99.5%. The value remains awaiting owner approval.

Approval status: awaiting owner approval.

cache-partial-miss-floor

Current value: 40% is the general partial-miss floor (blocking_path.py::_PARTIAL_MISS_FLOOR_PCT).

Report statements: whether a cache is described as mostly warm rather than churning, and whether its remediation can become the report's BIGGEST LEVER.

Reason: A cache whose miss share stays below two fifths is predominantly warm; treating that cohort as churn would overstate both the problem and its available saving.

Proposed decision: Keep 40%. The value remains awaiting owner approval.

Approval status: awaiting owner approval.

ceiling-minimum-gates

Current value: the verifier needs 5 gating PRs to re-derive a stable saving ceiling (verify_report.py::_CEILING_MIN_GATING_PR).

Report statements: whether a wall-clock saving ceiling is verified or reported as thin-sample evidence.

Reason: A ceiling from only a few gating occurrences is too sensitive to one commit.

Proposed decision: Keep 5. The value remains awaiting owner approval.

partial-majority-rule

Current value: strictly more than one half through pr_sample_contract.PR_SAMPLE_PARTIAL_MAJORITY_NUMERATOR and PR_SAMPLE_PARTIAL_MAJORITY_DENOMINATOR.

Report statements: whether the sample is labelled mostly partial and whether gate frequency may be headlined.

Reason: A population with more incomplete than complete commits cannot support a confident last-finisher ranking.

Proposed decision: Keep strictly more than 1/2. The #315 value, relanded by #333, remains awaiting owner approval.

Finding thresholds and caps

default-branch-batching-window

Current value: one hour (collect_runs.py:_OPT107_BATCH_WINDOW_S), with at least two batchable runs (_OPT107_MIN_BATCHABLE_RUNS) comprising at least 25% of the readable default-branch push sample (_OPT107_MIN_BATCHABLE_SHARE).

Report statements: whether dense post-merge validation is a measured runner-minute opportunity, which observed runs contribute to its signed sample-window saving, and its monthly estimate (the sample saving scaled by the workflow's 30-day default-branch push-run count over the sampled runs, capped at its measured 30-day validation compute) or, when that count is unreadable, that the monthly saving is unsized and why.

Reason: the proposed replacement is an hourly current-tip validation, so a successor inside that interval is evidence that a schedule can coalesce work. The count/share gates keep isolated busy periods quiet; the detector also requires three completed timed push runs, reserves conservative hourly replacement slots, and credits only removable runs with their own readable completed-job compute.

Proposed decision: Keep one hour, two runs, and 25%. This remains awaiting owner approval.

magnitude-noise-band

Current value: a magnitude range wider than 25% of its median triggers a wider cross-run probe (_MAG_WIDE_REL).

Report statements: whether a drilled log's magnitude is stable enough for a short probe and whether more runs are sampled.

Reason: A quarter-of-median spread is large enough that three observations cannot distinguish ordinary variation from warm/cold or other modes.

Proposed decision: Keep 25%. It widens evidence collection and does not claim statistical significance.

Approval status: awaiting owner approval.

bimodal-noise-band

Current value: at least 8 observations; each mode at least 30%; slow/fast median ratio at least 1.6; slow median at least 60 seconds (_bimodal_split defaults).

Report statements: fast/slow disclosure, minority-mode pole selection, and worst-case reconciliation.

Reason: These gates reject lone outliers and sub-minute fluctuations while retaining a substantial, visibly separated slow mode.

Proposed decision: Keep 8 / 30% / 1.6x / 60 seconds. A later change needs corpus evidence because it changes which checks enter the spine.

Approval status: awaiting owner approval.

ten-percent-sizing-rules

Current value: 10% modeled occurrence for OPT29, OPT35, OPT38, and OPT44 in collect_runs.py::_SIZING.

Report statements: modeled runner-minute savings for merge-queue step skips, failed-shard fallback, PR-edit triggers, and restrictive concurrency.

Reason: These are conservative placeholders for qualitative static findings, not measured rates; the report labels them modeled and caps them to affected compute.

Proposed decision: Keep the four 10% rules for now, but never describe them as measured frequency. Replace each with measured or uncredited evidence when a bounded observable cohort exists.

Approval status: awaiting owner approval.

per-pattern-minimums

Current value: representative report-bearing floors are OPT64: 3 retried runs; OPT79: 3 hits, 3 misses, 5 seconds and 20% waste, a 25% tail share, and 1 second of long-pole headroom; OPT80: 6 occurrences, a tail threshold equal to max(3 × P50, P50 + 30 seconds), 2 tail runs, 2 proven tail runs, a 20-second gap, and at most 4 log probes; OPT81: 8 samples per label, a 30-second and 25% gap, 50% covered time, and a 30-second cache-lever floor; its same-class downsize half needs 5 samples per size and allows at most a 5% absolute median-duration difference, and is never priced (its runner-cost saving is unsized); OPT82: 60 seconds; OPT83: 15% and 20 seconds; OPT84: 0% minimum share and 20 seconds, and 15% of the pole for its install row (npm, pnpm or Yarn install to bun install); OPT106: OPT83's 15% and 20 seconds for a runtime-loader step on the default (unsized) path, and 30 seconds and 20% of the job wall for a profiled loader; OPT108: a successful comment-triggered run under 30 seconds counts as a rejected comment (the drill's self-skip floor, _NOOP_FLOOR_S), and at least 3 such runs must be priced before its runner-minute saving is credited.

Report statements: whether each named pattern is promoted, withheld, or uncredited, plus its evidence and saving.

Reason: Each floor prevents a sparse observation from becoming a finding or ranked saving while keeping the withheld reason visible.

Proposed decision: Keep these current per-pattern minimums. Each future pattern must add its own named minimum and contract entry rather than borrowing an unrelated global threshold.

Approval status: awaiting owner approval.

go-package-dominance

Current value: OPT24 classifies a long go test step for list-then-run sharding only when one package accounts for at least 60% of the package time printed by the sampled command (_OPT24_GO_PACKAGE_DOMINANT_SHARE).

Report statements: that one Go package dominates the long test step and is therefore a concrete candidate for internal test-name partitioning.

Reason: A majority substantially above one half distinguishes one package that can move the pole from an already distributed package set. Missing timing or a balanced package set stays unclassified.

Proposed decision: Keep 60%.

Approval status: owner-approved 2026-11-20 (60%).

go-fixed-wait-thresholds

Current value: the measured OPT19 Go poll route needs at least two consistent wait intervals; the poll total, or the constant test-body sleep total, must account for at least 50% of the measured test (or package) time and 20% of the drilled long-pole job. The race-helper route needs at least three synchronous call sites that re-exec the test binary (_OPT19_GO_MIN_INTERVALS, _OPT19_GO_MIN_TEST_SHARE, _OPT19_GO_MIN_JOB_SHARE, and _OPT19_GO_RACE_HELPER_MIN_SPAWNS).

Report statements: that a Go test is dominated by a fixed production wait, or that a race job repeatedly pays the race runtime's default helper-exit sleep.

Reason: Repeated equal gaps plus both dominance tests distinguish a fixed wait from ordinary slow work. Three helper call sites avoid recommending a race option change for a one-off subprocess.

Proposed decision: Keep two intervals, 50%, 20%, and three helpers.

Approval status: awaiting owner approval.

opt25-universal-housekeeping

Current value: 60 seconds per matrix leg (_OPT25_UNIVERSAL_HOUSEKEEPING_MAX_S). A shared step named as a setup action (Run <owner>/setup-*@), an artifact upload or download, an upload, or a cleanup counts as housekeeping at any duration.

Report statements: whether work shared by every active matrix leg is treated as short housekeeping or retained as real work when diagnosing OPT25 imbalance.

Reason: universally present setup and upload steps should not keep otherwise idle legs in the matrix, but a substantial shared test phase is useful work and must keep those legs in the comparison.

Proposed decision: Keep 60 seconds. This matches the reader-facing bookkeeping cutoff while making longer shared work visible to the imbalance detector.

Approval status: owner-approved 2026-11-21 (60 seconds).

saving-caps

Current value: modeled cache savings use at most 90% of the available cache budget (_CACHE_BUDGET_FULL_FRACTION); cache-tail framing starts at 25% of qualifying runs (_CACHE_TAIL_MIN_FRAC). Every saving is also bounded by measured affected compute and wall-clock headroom.

Report statements: wall-clock and runner-minute totals, "up to" ceilings, cache-tail wording, biggest-lever ranking, and the Bottom line.

Reason: A model must leave room for irreducible work and cannot exceed the compute or critical-path wait observed in the same population.

Proposed decision: Keep 90%, 25%, and the physical compute/headroom caps. Measured detectors may use stamped exact evidence but cannot cross physical bounds.

Approval status: awaiting owner approval.

rendered-row-caps

Current value: 12 rows each for Also noticed, promoted runner-minute findings, and the cost spine (_ALSO_NOTICED_CAP, _TIER2_CAP, _RUNNER_MINUTE_SPINE_CAP), plus 5 displayed withheld job names (_WITHHELD_JOBS_SHOWN).

Report statements: visible row counts and every "+N more" disclosure.

Reason: Twelve keeps each secondary section scannable while the hidden-row count preserves coverage honesty.

Proposed decision: Keep 12 / 12 / 12 / 5. A cap may hide rows, never erase them from findings or omit the hidden count.

Approval status: awaiting owner approval.

Machine-readable contract

The suite compares this block with named engine constants. Proposed values that do not exist in current behavior remain in the decision sections until their implementation lands.

{
  "account-weekly-paired-minimum": {
    "decision": "account-weekly-paired-minimum",
    "sources": {
      "before_after.py:_ACCOUNT_WEEKLY_MIN_PAIRED_RUNS": 5
    }
  },
  "bimodal-noise-band": {
    "decision": "bimodal-noise-band",
    "symbols": [
      "collect_runs.py:_bimodal_split"
    ]
  },
  "bookkeeping-cutoffs": {
    "decision": "bookkeeping-cutoff",
    "planned_sources": {
      "gate_frequency_contract.py:GATE_BOOKKEEPING_CUTOFF_S": 60.0
    },
    "sources": {
      "blocking_path.py:_AGG_GATE_TRIVIAL_S": 30.0,
      "blocking_path.py:_WALL_CLOCK_LONG_POLE_FLOOR_S": 30.0,
      "collect_runs.py:_AGG_GATE_TRIVIAL_S": 30.0,
      "collect_runs.py:_NOOP_FLOOR_S": 30.0,
      "collect_runs.py:_TRIAGE_RECOVER_GATE_FRAC": 0.5,
      "collect_runs.py:_TRIAGE_WALLCLOCK_FLOOR_S": 90.0,
      "verify_report.py:_VR_AGG_GATE_TRIVIAL_S": 30.0,
      "verify_report.py:_WALL_CLOCK_LONG_POLE_FLOOR_S": 30.0
    }
  },
  "ceiling-minimum-gates": {
    "decision": "ceiling-minimum-gates",
    "sources": {
      "verify_report.py:_CEILING_MIN_GATING_PR": 5
    }
  },
  "chain-divergence": {
    "decision": "needs-chain-handling",
    "sources": {
      "blocking_path.py:_CHAIN_MAKESPAN_DIVERGENCE_PCT": 25.0,
      "verify_report.py:_VR_CHAIN_DIVERGENCE_PCT": 25.0
    }
  },
  "cold-cache-miss-threshold": {
    "decision": "cold-cache-miss-threshold",
    "sources": {
      "collect_runs.py:_CACHE_COLD_MISS_PCT": 99.5,
      "verify_report.py:_VR_CACHE_COLD_MISS_PCT": 99.5
    }
  },
  "cache-partial-miss-floor": {
    "decision": "cache-partial-miss-floor",
    "sources": {
      "blocking_path.py:_PARTIAL_MISS_FLOOR_PCT": 40.0,
      "verify_report.py:_VR_PARTIAL_MISS_FLOOR_PCT": 40.0
    }
  },
  "config-era-minimum": {
    "decision": "config-era-boundary",
    "sources": {
      "blocking_path.py:_RARE_PRESENCE_MIN_PR": 6,
      "collect_runs.py:_RARE_PRESENCE_MIN_PR": 6,
      "verify_report.py:_VR_RARE_PRESENCE_MIN_PR": 6
    }
  },
  "dependency-cache-floor": {
    "decision": "dependency-cache-floor",
    "sources": {
      "collect_runs.py:_DEPENDENCY_CACHE_FLOOR_PCT": 50.0,
      "verify_report.py:_VR_CACHE_FLOOR_BY_FIX_KEY": {
        "dependency-cache": 50.0
      }
    }
  },
  "default-branch-batching-window": {
    "decision": "default-branch-batching-window",
    "sources": {
      "collect_runs.py:_OPT107_BATCH_WINDOW_S": 3600,
      "collect_runs.py:_OPT107_MIN_BATCHABLE_RUNS": 2,
      "collect_runs.py:_OPT107_MIN_BATCHABLE_SHARE": 0.25
    }
  },
  "gate-frequency-interval": {
    "decision": "gate-frequency-interval",
    "planned_sources": {
      "gate_frequency_contract.py:GATE_INTERVAL_Z": 1.96
    }
  },
  "go-package-dominance": {
    "decision": "go-package-dominance",
    "sources": {
      "collect_runs.py:_OPT24_GO_PACKAGE_DOMINANT_SHARE": 0.6
    }
  },
  "go-fixed-wait-thresholds": {
    "decision": "go-fixed-wait-thresholds",
    "sources": {
      "collect_runs.py:_OPT19_GO_MIN_INTERVALS": 2,
      "collect_runs.py:_OPT19_GO_MIN_JOB_SHARE": 0.2,
      "collect_runs.py:_OPT19_GO_MIN_TEST_SHARE": 0.5,
      "collect_runs.py:_OPT19_GO_RACE_HELPER_MIN_SPAWNS": 3
    }
  },
  "holds-up-threshold": {
    "decision": "holds-up-threshold",
    "planned_sources": {
      "gate_frequency_contract.py:GATE_MIN_COUNT_FOR_RANKING": 5,
      "gate_frequency_contract.py:POLE_HOLDS_UP_MIN_COUNT": 2
    },
    "sources": {
      "blocking_path.py:_POLE_RECUR_FLOOR": 2,
      "collect_runs.py:_POLE_RECUR_FLOOR": 2,
      "verify_report.py:_VR_POLE_HOLDS_UP_MIN_COUNT": 2,
      "verify_report.py:_VR_POLE_RECUR_FLOOR": 2
    }
  },
  "lifecycle-minimum": {
    "decision": "lifecycle-minimum",
    "sources": {
      "pr_lifecycle.py:MIN_MEASURED": 10
    }
  },
  "magnitude-wide-band": {
    "decision": "magnitude-noise-band",
    "sources": {
      "blocking_path.py:_MAG_WIDE_REL": 0.25,
      "collect_runs.py:_MAG_WIDE_REL": 0.25
    }
  },
  "merge-queue-proof-thresholds": {
    "decision": "merge-queue-proof-thresholds",
    "sources": {
      "collect_runs.py:_OPT98_CHURN_SAMPLE": 30,
      "collect_runs.py:_OPT98_DIFFERENT_RATIO": 1.5,
      "collect_runs.py:_OPT98_MIN_MERGES_PER_DAY": 30,
      "collect_runs.py:_OPT98_MIN_RUNS": 15,
      "collect_runs.py:_OPT98_REPROVE_RATIO": 0.8,
      "collect_runs.py:_OPT98_WINDOW_DAYS": 30
    }
  },
  "minimum-timed-runs": {
    "decision": "minimum-timed-runs",
    "sources": {
      "collect_runs.py:_MIN_TIMED_RUNS": 3
    }
  },
  "opt25-universal-housekeeping": {
    "decision": "opt25-universal-housekeeping",
    "sources": {
      "collect_runs.py:_OPT25_UNIVERSAL_HOUSEKEEPING_MAX_S": 60.0
    }
  },
  "opt64-retried-runs": {
    "decision": "per-pattern-minimums",
    "sources": {
      "blocking_path.py:_OPT64_MIN_RETRIED_RUNS": 3,
      "collect_runs.py:_OPT64_MIN_RETRIED_RUNS": 3,
      "verify_report.py:_VR_OPT64_MIN_RETRIED_RUNS": 3
    }
  },
  "opt108-comment-runs": {
    "decision": "per-pattern-minimums",
    "sources": {
      "collect_runs.py:_OPT108_SHORT_RUN_S": 30.0,
      "collect_runs.py:_OPT108_MIN_PRICED_RUNS": 3,
      "verify_report.py:_VR_OPT108_SHORT_RUN_S": 30.0,
      "verify_report.py:_VR_OPT108_MIN_PRICED_RUNS": 3
    }
  },
  "opt79-minimums": {
    "decision": "per-pattern-minimums",
    "sources": {
      "collect_runs.py:_OPT79_MIN_HITS": 3,
      "collect_runs.py:_OPT79_MIN_MISSES": 3,
      "collect_runs.py:_OPT79_MIN_WASTE_S": 5.0,
      "collect_runs.py:_OPT79_MIN_WASTE_FRAC": 0.2,
      "collect_runs.py:_CACHE_TAIL_MIN_FRAC": 0.25,
      "collect_runs.py:_OPT79_POLE_MIN_HEADROOM_S": 1.0,
      "verify_report.py:_VR_OPT79_MIN_HITS": 3,
      "verify_report.py:_VR_OPT79_MIN_MISSES": 3,
      "verify_report.py:_VR_OPT79_MIN_WASTE_S": 5.0,
      "verify_report.py:_VR_OPT79_MIN_WASTE_FRAC": 0.2,
      "verify_report.py:_VR_OPT79_TAIL_MIN_FRAC": 0.25,
      "verify_report.py:_VR_OPT79_POLE_MIN_HEADROOM_S": 1.0
    }
  },
  "opt80-occurrences": {
    "decision": "per-pattern-minimums",
    "sources": {
      "collect_runs.py:_OPT80_MIN_SAMPLED_OCCURRENCES": 6,
      "collect_runs.py:_OPT80_MIN_TAIL_RUNS": 2,
      "collect_runs.py:_OPT80_MIN_PROVEN_TAIL_RUNS": 2,
      "collect_runs.py:_OPT80_MIN_GAP_S": 20.0,
      "collect_runs.py:_OPT80_LOG_PROBE_MAX": 4,
      "verify_report.py:_VR_OPT80_MIN_SAMPLED_OCCURRENCES": 6,
      "verify_report.py:_VR_OPT80_MIN_TAIL_RUNS": 2,
      "verify_report.py:_VR_OPT80_MIN_PROVEN_TAIL_RUNS": 2,
      "verify_report.py:_VR_OPT80_MIN_GAP_S": 20.0,
      "verify_report.py:_VR_OPT80_LOG_PROBE_MAX": 4
    }
  },
  "opt80-tail-multiple": {
    "decision": "per-pattern-minimums",
    "sources": {
      "collect_runs.py:_OPT80_TAIL_P95_MULTIPLE": 3.0,
      "collect_runs.py:_OPT80_TAIL_P95_ABS_S": 30.0,
      "verify_report.py:_VR_OPT80_TAIL_P95_MULTIPLE": 3.0,
      "verify_report.py:_VR_OPT80_TAIL_P95_ABS_S": 30.0
    }
  },
  "opt81-samples-and-gap": {
    "decision": "per-pattern-minimums",
    "sources": {
      "collect_runs.py:_OPT81_MIN_GAP_FRAC": 0.25,
      "collect_runs.py:_OPT81_MIN_GAP_S": 30.0,
      "collect_runs.py:_OPT81_MIN_SAMPLES_PER_LABEL": 8,
      "collect_runs.py:_OPT81_DOWNSIZE_MIN_SAMPLES_PER_LABEL": 5,
      "collect_runs.py:_OPT81_DOWNSIZE_MAX_ABS_DELTA_FRAC": 0.05,
      "collect_runs.py:_OPT81_COVERED_FRAC": 0.5,
      "collect_runs.py:_OPT81_CACHE_LEVER_MIN_S": 30.0,
      "verify_report.py:_VR_OPT81_MIN_GAP_FRAC": 0.25,
      "verify_report.py:_VR_OPT81_MIN_GAP_S": 30.0,
      "verify_report.py:_VR_OPT81_MIN_SAMPLES_PER_LABEL": 8,
      "verify_report.py:_VR_OPT81_DOWNSIZE_MIN_SAMPLES_PER_LABEL": 5,
      "verify_report.py:_VR_OPT81_DOWNSIZE_MAX_ABS_DELTA_FRAC": 0.05,
      "verify_report.py:_VR_OPT81_COVERED_FRAC": 0.5,
      "verify_report.py:_VR_OPT81_CACHE_LEVER_MIN_S": 30.0
    }
  },
  "opt82-minimum-seconds": {
    "decision": "per-pattern-minimums",
    "sources": {
      "collect_runs.py:_OPT82_MIN_LINT_P50_S": 60.0
    }
  },
  "opt83-minimums": {
    "decision": "per-pattern-minimums",
    "sources": {
      "collect_runs.py:_OPT83_MIN_SHARE": 0.15,
      "collect_runs.py:_OPT83_MIN_STEP_S": 20.0,
      "verify_report.py:_VR_OPT83_MIN_SHARE": 0.15,
      "verify_report.py:_VR_OPT83_MIN_STEP_S": 20.0,
      "verify_report.py:_VR_OPT106_MIN_STEP_SHARE": 0.15,
      "verify_report.py:_VR_OPT106_MIN_STEP_S": 20.0
    }
  },
  "opt84-minimum-seconds": {
    "decision": "per-pattern-minimums",
    "sources": {
      "collect_runs.py:_OPT84_MIN_SHARE": 0.0,
      "collect_runs.py:_OPT84_MIN_STEP_S": 20.0,
      "verify_report.py:_VR_OPT84_MIN_SHARE": 0.0,
      "verify_report.py:_VR_OPT84_MIN_STEP_S": 20.0,
      "collect_runs.py:_OPT84_INSTALL_MIN_SHARE": 0.15,
      "verify_report.py:_VR_OPT84_INSTALL_MIN_SHARE": 0.15
    }
  },
  "opt105-hook-duplication-minimums": {
    "decision": "per-pattern-minimums",
    "sources": {
      "collect_runs.py:_OPT105_MIN_DUPLICATE_STEP_S": 20.0,
      "collect_runs.py:_OPT105_MIN_JOB_SHARE": 0.15
    }
  },
  "opt106-profile-minimums": {
    "decision": "per-pattern-minimums",
    "sources": {
      "collect_runs.py:_OPT106_MIN_PROFILE_S": 30.0,
      "collect_runs.py:_OPT106_MIN_JOB_SHARE": 0.2,
      "blocking_path.py:_OPT106_MIN_PROFILE_S": 30.0,
      "blocking_path.py:_OPT106_MIN_JOB_SHARE": 0.2,
      "verify_report.py:_VR_OPT106_MIN_PROFILE_S": 30.0,
      "verify_report.py:_VR_OPT106_MIN_JOB_SHARE": 0.2
    }
  },
  "partial-majority-rule": {
    "decision": "partial-majority-rule",
    "sources": {
      "pr_sample_contract.py:PR_SAMPLE_PARTIAL_MAJORITY_DENOMINATOR": 2,
      "pr_sample_contract.py:PR_SAMPLE_PARTIAL_MAJORITY_NUMERATOR": 1
    }
  },
  "pr-candidate-pages": {
    "decision": "pr-candidate-pages",
    "sources": {
      "pr_sample_contract.py:PR_SAMPLE_CANDIDATE_PAGE_CAP": 12,
      "pr_sample_contract.py:PR_SAMPLE_JOB_LIST_READ_CAP": 1200
    }
  },
  "pr-sample-eligibility": {
    "decision": "pr-gate-eligibility",
    "symbols": [
      "collect_runs.py:_select_repr_shas"
    ]
  },
  "pr-sample-span": {
    "decision": "pr-gate-span",
    "sources": {
      "pr_sample_contract.py:PR_SAMPLE_MIN_DISTINCT_DAYS": 5
    }
  },
  "pr-sample-target": {
    "decision": "pr-gate-sample-size",
    "sources": {
      "pr_sample_contract.py:PR_SAMPLE_TARGET": 100
    }
  },
  "push-cohort": {
    "decision": "default-branch-push-cohort",
    "sources": {
      "collect_runs.py:_DEFAULT_BRANCH_PUSH_SAMPLE_MAX": 10,
      "verify_report.py:_VR_DEFAULT_BRANCH_PUSH_SAMPLE_MAX": 10
    }
  },
  "ranking-minimum-count": {
    "decision": "ranking-minimum-count",
    "planned_sources": {
      "gate_frequency_contract.py:GATE_MIN_COUNT_FOR_RANKING": 5,
      "gate_frequency_contract.py:POLE_HOLDS_UP_MIN_COUNT": 2
    },
    "sources": {
      "collect_runs.py:_POLE_RECUR_FLOOR": 2,
      "verify_report.py:_VR_POLE_RECUR_FLOOR": 2
    }
  },
  "rare-presence-rule": {
    "decision": "rare-presence-rule",
    "sources": {
      "blocking_path.py:_RARE_PRESENCE_FRAC": 0.5,
      "blocking_path.py:_RARE_PRESENCE_MIN_PR": 6,
      "collect_runs.py:_RARE_PRESENCE_FRAC": 0.5,
      "collect_runs.py:_RARE_PRESENCE_MIN_PR": 6,
      "verify_report.py:_VR_RARE_PRESENCE_FRAC": 0.5,
      "verify_report.py:_VR_RARE_PRESENCE_MIN_PR": 6
    }
  },
  "rare-push-lane": {
    "decision": "rare-push-lane",
    "sources": {
      "collect_runs.py:_PUSH_LANE_RARE_PER_30D": 8
    }
  },
  "rendered-poles": {
    "decision": "rendered-pole-count",
    "sources": {
      "blocking_path.py:_TOP_WORKFLOWS": 5
    }
  },
  "rendered-row-caps": {
    "decision": "rendered-row-caps",
    "sources": {
      "blocking_path.py:_ALSO_NOTICED_CAP": 12,
      "blocking_path.py:_RUNNER_MINUTE_SPINE_CAP": 12,
      "blocking_path.py:_TIER2_CAP": 12,
      "blocking_path.py:_WITHHELD_JOBS_SHOWN": 5,
      "verify_report.py:_VR_RUNNER_MINUTE_SPINE_CAP": 12,
      "verify_report.py:_VR_TIER2_CAP": 12,
      "verify_report.py:_VR_WITHHELD_JOBS_SHOWN": 5
    }
  },
  "required-suite-eligibility": {
    "decision": "required-suite-eligibility",
    "symbols": [
      "collect_runs.py:_select_repr_shas"
    ]
  },
  "rolling-volume-window": {
    "decision": "rolling-volume-window",
    "symbols": [
      "collect_runs.py:_window_30d"
    ]
  },
  "run-list-depth": {
    "decision": "run-list-depth",
    "sources": {
      "collect_runs.py:_COST_RUNLIST_MAX": 100
    }
  },
  "saving-caps": {
    "decision": "saving-caps",
    "sources": {
      "collect_runs.py:_CACHE_BUDGET_FULL_FRACTION": 0.9,
      "collect_runs.py:_CACHE_TAIL_MIN_FRAC": 0.25,
      "verify_report.py:_VR_CACHE_TAIL_MIN_FRAC": 0.25
    }
  },
  "severity-cutoffs": {
    "decision": "severity-cutoffs",
    "sources": {
      "blocking_path.py:_SEV_HIGH_S": 300.0,
      "blocking_path.py:_SEV_MED_S": 120.0
    }
  },
  "success-only-caveat": {
    "decision": "success-only-caveat",
    "sources": {
      "collect_runs.py:_SUCCESS_ONLY_CAVEAT_FAILURE_RATE": 0.2,
      "collect_runs.py:_SUCCESS_ONLY_CAVEAT_MIN_RUNS": 10
    }
  },
  "ten-percent-sizing-rules": {
    "decision": "ten-percent-sizing-rules",
    "sources": {}
  },
  "timeout-recommendation-thresholds": {
    "decision": "timeout-recommendation-thresholds",
    "sources": {
      "collect_runs.py:_GHA_DEFAULT_JOB_TIMEOUT_S": 21600.0,
      "collect_runs.py:_OPT57_DEFAULT_KILL_CEILING_S": 22500.0,
      "collect_runs.py:_OPT57_KILL_GRACE_S": 900.0,
      "collect_runs.py:_OPT57_MIN_TIMEOUT_S": 900.0,
      "collect_runs.py:_OPT57_NEAR_DEFAULT_TIMEOUT_S": 20520.0,
      "collect_runs.py:_OPT57_TIMEOUT_BUFFER_S": 600.0,
      "collect_runs.py:_OPT57_TIMEOUT_MULTIPLIER": 1.5,
      "verify_report.py:_VR_OPT57_DEFAULT_KILL_CEILING_S": 22500.0,
      "verify_report.py:_VR_OPT57_DEFAULT_TIMEOUT_MIN": 360.0,
      "verify_report.py:_VR_OPT57_MIN_SUCCESS_SAMPLES": 3,
      "verify_report.py:_VR_OPT57_MIN_TIMEOUT_S": 900.0,
      "verify_report.py:_VR_OPT57_NEAR_DEFAULT_TIMEOUT_S": 20520.0,
      "verify_report.py:_VR_OPT57_TIMEOUT_BUFFER_S": 600.0,
      "verify_report.py:_VR_OPT57_TIMEOUT_MULTIPLIER": 1.5
    },
    "symbols": [
      "collect_runs.py:_opt57_recommended_timeout_s"
    ]
  },
  "typical-floor-thresholds": {
    "decision": "typical-floor-thresholds",
    "sources": {
      "blocking_path.py:_FLOOR_COOCCUR_MAJORITY_FRAC": 0.5,
      "blocking_path.py:_FLOOR_NEAR_UNIVERSAL_FRAC": 0.8
    }
  },
  "within-group-order": {
    "decision": "within-group-order",
    "symbols": [
      "blocking_path.py:_render_core"
    ]
  },
  "workflow-timing-depth": {
    "decision": "workflow-timing-depth",
    "sources": {
      "collect_runs.py:_DEEPEN_TOP_CHECKS": 12,
      "collect_runs.py:_SHALLOW_RUNS": 10
    }
  }
}

Approval checklist

  • Approve or edit every keep/change decision above.
  • Approve 100 completed PR commits over at least 5 days.
  • Approve the 12-page cap with early stopping and the 1,200 job-list read cap.
  • Approve the conditional 60-second bookkeeping rule.
  • Approve the 60% go-package-dominance share (owner-approved 2026-11-20).
  • Approve 5 paired runs per side for account-weekly-paired-minimum (owner-approved 2026-11-20).
  • Approve the 60-second opt25-universal-housekeeping cutoff (owner-approved 2026-11-21).
  • Approve 5 last-finisher observations for supported ranking/grouping.
  • Approve two-sided 95% Wilson intervals and overlap as unsupported order.
  • Confirm #305, #306, and #311 may merge after their implementations match the approved values.

Source: SKILL.md on GitHub

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

1 alert2d3 checks · Risk SAFE
  • Gen Agent Trust Hub2d

    The skill is a robust CI optimization tool that audits GitHub Actions workflows. It identifies performance bottlenecks using deterministic logic and provides analysis of job logs. While it processes untrusted log data, it includes strong defensive instructions to mitigate indirect prompt injection. A minor best-practice issue regarding unsafe YAML loading was identified in a utility script.

  • Socket2d

    No alerts

  • Snyk2d

    Risk: CRITICAL · 2 issues

Signed by skilld at 9a84115. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub 16 hours ago.

Activeupdated yesterday

README badge

README badge for starslingdev/skills/ci-speedup