Context
With the Orb v2 fleet collector live (#1256/#1257) and computeFleetAnalytics surfaced read-only three ways — the MCP tool gittensory_get_fleet_analytics (#1268), the internal route /v1/internal/fleet/analytics (#1256), and the operator-dashboard fleetMetrics panel (#1268) — the natural next step is to let fleet calibration inform gate thresholds. This issue captures the design and, deliberately, defers building it.
Adversarial verdict: DESIGN-ONLY for now
A scoping pass found three reasons not to build a parallel "fleet suggestions" surface yet:
- The read-out already exists (MCP tool + internal route + operator dashboard). An operator can already read fleet
mergePrecision / fpRate / outliers today.
- A data-driven tuning loop already exists.
GITTENSORY_REVIEW_SELFTUNE + src/review/selftune-wire.ts computes tuning recommendations, shadow-soaks any tightening change, and auto-promotes after a soak — audited, tighten-only. (src/review/auto-tune.ts is the accuracy circuit-breaker — it only ever disables auto-merge — not a tunable-knob engine; don't conflate them.) A second fleet-median suggestions endpoint would duplicate the read-only half of both, with a different and less-trustworthy data source.
- The data isn't there yet.
computeFleetAnalytics needs per-instance decided >= MIN_DECIDED (5) to count toward the fleet median; instanceCount < 3 / windowDays < 30 is explicitly nascent. Post-launch the fleet median is almost certainly null or single-instance, and the gaming / aggregate-vs-per-repo / precision-paradox risks all bite hardest at tiny n.
Data-readiness trigger (build gate)
Do not implement until prod computeFleetAnalytics returns a non-null fleet.mergePrecision with instanceCount >= 3 over a >= 30-day window.
When the bar is met — preferred design
- Prefer feeding the existing
GITTENSORY_REVIEW_SELFTUNE loop a fleet-derived signal over a new parallel suggestions endpoint — one tuning path, one audit trail, the tighten-only safety property preserved.
- If a standalone surface is still wanted: a pure
buildFleetThresholdSuggestions(analytics) fn behind a default-OFF GITTENSORY_REVIEW_FLEET_AUTOTUNE flag, zero-mutation (read-only, never applies to the live gate), returning per-instance context + windowDays/instanceCount in every suggestion, and null below MIN_DECIDED.
- Always label suggestions ADVISORY; never auto-apply; log outlier instances (
OUTLIER_BAND) for human review.
Tracking under the Orb fleet arc (#1255).
Context
With the Orb v2 fleet collector live (#1256/#1257) and
computeFleetAnalyticssurfaced read-only three ways — the MCP toolgittensory_get_fleet_analytics(#1268), the internal route/v1/internal/fleet/analytics(#1256), and the operator-dashboardfleetMetricspanel (#1268) — the natural next step is to let fleet calibration inform gate thresholds. This issue captures the design and, deliberately, defers building it.Adversarial verdict: DESIGN-ONLY for now
A scoping pass found three reasons not to build a parallel "fleet suggestions" surface yet:
mergePrecision/fpRate/ outliers today.GITTENSORY_REVIEW_SELFTUNE+src/review/selftune-wire.tscomputes tuning recommendations, shadow-soaks any tightening change, and auto-promotes after a soak — audited, tighten-only. (src/review/auto-tune.tsis the accuracy circuit-breaker — it only ever disables auto-merge — not a tunable-knob engine; don't conflate them.) A second fleet-median suggestions endpoint would duplicate the read-only half of both, with a different and less-trustworthy data source.computeFleetAnalyticsneeds per-instancedecided >= MIN_DECIDED (5)to count toward the fleet median;instanceCount < 3/windowDays < 30is explicitly nascent. Post-launch the fleet median is almost certainly null or single-instance, and the gaming / aggregate-vs-per-repo / precision-paradox risks all bite hardest at tiny n.Data-readiness trigger (build gate)
Do not implement until prod
computeFleetAnalyticsreturns a non-nullfleet.mergePrecisionwithinstanceCount >= 3over a >= 30-day window.When the bar is met — preferred design
GITTENSORY_REVIEW_SELFTUNEloop a fleet-derived signal over a new parallel suggestions endpoint — one tuning path, one audit trail, the tighten-only safety property preserved.buildFleetThresholdSuggestions(analytics)fn behind a default-OFFGITTENSORY_REVIEW_FLEET_AUTOTUNEflag, zero-mutation (read-only, never applies to the live gate), returning per-instance context +windowDays/instanceCountin every suggestion, andnullbelowMIN_DECIDED.OUTLIER_BAND) for human review.Tracking under the Orb fleet arc (#1255).