Trust Fleet Topology¶
The Trust Fleet is the layer of caretaker loops that hardens HydraFlow's delivery pipeline against silent regression. Per ADR-0045, it consists of 8 autonomous loops plus 2 non-loop subsystems, all running concurrently and feeding each other through events, issue labels, and shared state.
This page is hand-curated — it captures the intent of the trust fleet's
internal relationships. The corresponding loop list is auto-generated from
source at Loop Registry; per-area assignments live in
functional_areas.yml.
Internal data flow¶
flowchart LR
subgraph rc_boundary["RC Boundary Gates"]
SP["StagingPromotionLoop"]
SB["StagingBisectLoop"]
CR["ContractRefreshLoop"]
SP -- "RC red event" --> SB
SB -- "auto-revert PR + retry issue" --> SP
end
subgraph corpus["Adversarial Corpus + Skill Eval"]
CL["CorpusLearningLoop"]
SPE["SkillPromptEvalLoop (Quality Gates)"]
CL -- "grows corpus" --> SPE
SPE -- "drift signals" --> CL
end
subgraph fakes["Fake-Adapter Contracts"]
FCA["FakeCoverageAuditorLoop"]
CR -- "weekly cassette refresh" --> FCA
FCA -- "uncovered method issues" --> CR
end
subgraph audit["Foundational Audit"]
PA["PrinciplesAuditLoop"]
FT["FlakeTrackerLoop"]
RB["RCBudgetLoop"]
end
subgraph meta["Meta-Observability"]
TFS["TrustFleetSanityLoop"]
TFS -. "watches all 7" .-> SP
end
PA -- "gates onboarding" --> rc_boundary
PA -- "gates onboarding" --> fakes
FT -- "filters flaky from RC red" --> SB
RB -- "RC wall-clock spikes" --> SP
Loops in detail¶
| Loop | Role | Triggers / Reads | Emits / Writes |
|---|---|---|---|
| StagingPromotionLoop | Cuts RC snapshots from staging; auto-promotes on green CI. | merges to staging branch |
RC_RED event on failure; promotion PR on success |
| StagingBisectLoop | Bisects culprit PR on RC red; auto-revert + retry issue. | RC_RED event |
revert PR; staging-bisect-stuck escalation |
| ContractRefreshLoop | Weekly refresh of fake-adapter cassettes against live gh/git/docker/claude. |
weekly tick | refresh PR; fake-drift issues |
| CorpusLearningLoop | Grows the adversarial skill corpus from production escapes. | escape signals from skill chain | new test cases in tests/trust/adversarial/cases/ |
| FakeCoverageAuditorLoop | Flags un-cassetted fake methods, un-exercised helpers. | scan tick | fake-coverage-gap issues |
| PrinciplesAuditLoop | ADR-0044 principle conformance audit. | weekly | drift PRs / issues; gates other trust loops on green audit |
| FlakeTrackerLoop | Detects persistently flaky tests across last N RC runs. | RC results | flake-tracker issues |
| RCBudgetLoop | Detects RC wall-clock bloat (rolling-median + spike-vs-max signals). | RC timing | rc-budget hydraflow-find issues |
| TrustFleetSanityLoop | Meta-observability — watches the other 8. | other loops' status events | one-attempt escalations on issue-storm / repair-failure / staleness / cost spikes |
Non-loop subsystems¶
- Adversarial skill corpus at
tests/trust/adversarial/— read by bothCorpusLearningLoop(writer) and the post-implementation skill chain (reader). - VCR-style contract cassettes at
tests/trust/contracts/cassettes/— committed YAML; refreshed byContractRefreshLoop.
See also¶
- ADR-0042 — the staging→RC→main promotion model that the RC-boundary gates protect.
- ADR-0045 — the architecture this page renders.
- ADR-0048 — auto-revert mechanism used by
StagingBisectLoop. - Functional Area Map → Trust Fleet — auto-joined membership table.
- Loop Registry — live truth for each loop's source path, kill-switch, ADR refs.