Repository navigation
Build smaller Agentplane CI contracts and narrow deployment inputs - #16102
Draft
agentydragon wants to merge 30 commits into
Draft
agentydragon wants to merge 30 commits into
agentydragon wants to merge 30 commits into
Conversation
7 of 16 tasks
Contributor
agentydragon
added a commit
that referenced
this pull request
Nov 13, 2026
Publish a concrete Agentplane test architecture plan: smaller production-service contracts, explicit deployment/image slices, retained app/Pod/VM/Firecracker acceptance, and six scoped implementation steps with coverage and feedback admission criteria. The native pilot composes real Gateway/Dispatcher, Runtime and History without K3s or the app. Three fresh positive runs passed; final rebased-source physical execution was 42.475s. Discarding its Runtime journal was rejected at the retained-Session assertion. This proves feasibility, not complete-CI savings or replacement coverage. The current merge gate stays unchanged. - Source pilot and plan: draft #16102, branch `agent/test-architecture-pivot`. - Implementation and future shadow measurements: #16099; earlier exposure work remains in #16034. - Real RBE visual validation: [8/8 cases](https://app.buildbuddy.io/invocation/999b8b3a-b5e5-4e49-b851-0b954463fbb7), desktop/mobile light/dark, real pinned CDN loading, live theme changes, all PR selectors and six expanded steps. Reviewed representative screenshots; proof binds 12 frozen assets to the rendered bytes. - Root navigation and README updated. Existing latency/reliability snapshots retain their original windows. After merge, GitHub Pages serves the plan at https://agentydragon.github.io/ducktape/agentplane-test-architecture/. This PR does not deploy or adopt the proposed test-selection policy.
agentydragon
changed the title
agentydragon
added a commit
that referenced
this pull request
Nov 13, 2026
…#16144) Continue the Agentplane fast-CI plan with the implementation and measurements on draft source PR #16102, head `e81dd61bc9`. The page retains the intended native/deployment/app boundaries and marks P2–P4 partial and P5–P6 pending. The same four recovery targets passed a source-pinned current-base pair: 342.108s control versus 271.736s candidate, 70.372s (20.57%) lower Bazel wall in this one warm observation. The fixture omits three unused OAuth workloads and only the budget profile configures 45s instead of the production 120s default. The page discloses the changed duration exposure, rejected five-second trial, unchanged assertions, definitive-session-loss branch and withdrawn Secret-projection experiment. Native input CAS proves Codex removal with identical Claude retained; the native ABBA sample remains too noisy/sparse for a causal speed claim. All required source checks passed, but this one whole-PR head still took 10m39s and the 21-case Pod suite finished last. The page names its actual largest cases and the next ownership split. It does not claim complete-CI or reliability improvement. Source, GitHub synthetic-merge and control provenance remain separate; tracker #16099 retains 30-pair/100-execution shadows and coverage/recurrence follow-ups. Validation: eight real RBE desktop/mobile and light/dark browser cases passed [085f5574](https://app.buildbuddy.io/invocation/085f5574-c7f8-4504-b651-a46e117ba7ca), including live theme changes, all selectors, populated comparison tables and expanded implementation steps. Frozen asset hashes and all 36 screenshots are recorded in visual-validation.json after visual review. Earlier harness-only lint/readiness failures are retained as superseded attempts. Root navigation/README preserve the independently merged organization update. Updates https://agentydragon.github.io/ducktape/agentplane-test-architecture/ after merge. This report changes no source test policy or deployed infrastructure.
…st-architecture-pivot
agentydragon
added a commit
that referenced
this pull request
Nov 13, 2026
Refresh the architecture plan with implemented Pod service closure, independent harness concurrency, reusable native fixtures and a same-Pod Secret cadence screen. Preserve the earlier pilot/recovery experiments, explicit coverage limits, failed controls and adoption criteria. All three full-suite variants retain the same 21 cases. Six cadence trials verify actual kubelet configuration and complete cleanup; their sample is too small for a reliability claim. Ordinary CI and its different synthetic merge tree remain separate from literal-source experiments. MCP services is the next measured long pole. Validation: eight RBE browser cases, desktop/mobile light/dark, real CDN loading and live theme changes; 44 screenshots captured and representative captures reviewed; all new tables, retained selectors and implementation disclosures exercised; final rendered asset hashes recorded in visual-validation.json. Updates [the published plan](https://agentydragon.github.io/ducktape/agentplane-test-architecture/) when merged. Experimental source remains in draft PR #16102; measurement follow-ups stay in #16099.
agentydragon
changed the title
agentydragon
changed the title
agentydragon
added a commit
that referenced
this pull request
Nov 13, 2026
…16200) Publish the reviewed Agentplane architecture iteration 4 measurements and concrete next probe. Four manual native recovery contracts pass 25 fresh actions (100 correlated case exposures); the selected-graph ABBA screen retains the VM infrastructure re-execution and teardown failure. The report adds bounded PVC-read and nested-outcome controls, full VM validation, exact required-CI provenance and independent devel-backport tracking. No qualified whole-CI saving or gate adoption is claimed. Only the architecture namespace changes. Historical iterations remain labeled; root navigation and sibling reports remain accurate. The source architecture stays on draft #16102. CaseBatch backport #16198 is separate and unmerged. Validation: pinned real CDN browser execution passed 4 desktop/mobile light/dark cases with 52 screenshots, live theme change and all selector/expanded-plan states. Representative final screenshots were visually reviewed. All five frozen report/shared asset hashes match the browser-tested bytes; report/evidence manifest hashes and Prettier checks pass. Visual proof: https://app.buildbuddy.io/invocation/cdaaeb05-6c5f-43c8-8047-340c992a0afd. Page: https://agentydragon.github.io/ducktape/agentplane-test-architecture/ (redeploys after merge).
agentydragon
changed the title
agentydragon
added a commit
that referenced
this pull request
Nov 13, 2026
Pod scenarios, VM and Firecracker fixtures render and import MCP-edge even though their scenarios do not use external Connections. This gives MCP-edge source changes an unnecessary dependency path into three costly deployment suites. Give those fixtures a `NO_MCP_EDGE` profile and exclude its OCI image from their declared inputs. Keep Actions, all retained rendered resources, scenario bodies and test frequency. Default, Dispatcher, deletion, real-model and MCP acceptance retain the full service. Renderer tests require every retained object to match the full profile and verify that default/MCP fixtures still include the edge. Normal hooks pass. Required PR CI validates this independent backport. On architecture source `91f72591a1`, the unchanged Pod/full-VM/both Firecracker cases passed required CI; a [source-pinned synthetic selector comparison](https://app.buildbuddy.io/invocation/c51d318c-90ea-4fe0-a892-6110260728ef) selected 15 rather than 18 tests, removing exactly these three suites while retaining MCP acceptance. That is graph evidence for one footprint, not a measured CI-time saving or a representative workload frequency. This leaves the architecture branch's native contracts, case concurrency, Secret cadence, reconnect-budget and candidate split experiments in #16102; none is needed for this input cut. Revert this profile change to restore full fixture/image ownership if a retained scenario acquires an MCP dependency.
This was referenced Nov 13, 2026
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Agentplane PRs repeat cluster/service setup and inherit service images that some scenarios never use. This branch adds smaller real-service contracts and explicit deployment compositions so CI can follow the guarantees each target owns.
Current source is
135dba720768d9dfdb60a122145d9d23f9699dd0, integrated with develf2a98e950d5ed8b7b1d28b85f0b87e719b5ee7de. Required CI and fresh literal-head focused validation are running. Normal commit hooks pass.The configured guest working-tree pair passes healthy readiness and rejects wrong client trust, each in one fresh physical action. Existing native recovery and admission-reply-loss targets also pass after fixture extraction. This guest creates no Session or executor and uses fixture-issued credentials; History delivery, real controller/issuer, KubeVirt, PVC, executor and App acceptance remain in their retained tests.
No complete-feedback saving is established. The first compatible historical Runtime replay selects the same 51 tests (117 versus 124 affected build/test labels); manual contracts do not yet shorten routine Runtime selection. A renderer batching spike reduced aggregate CPU but did not shorten its observed parallel window, so its patch was withdrawn. Candidate contracts remain manual and required scenario frequency remains unchanged until assertion ownership, at least 30 paired replays, fresh contract exposures, busy windows and a four-PR burst qualify a specific boundary.
Independent devel repairs have merged as #16198 (CaseBatch failure propagation), #16201 (bounded read-only PVC LIST retry), #16218 (peer-TLS startup-test import) and #16220 (unused MCP inputs). Their population-level reliability/speed effects remain measurement follow-ups. Managed executor kernel retries are tracked separately in #16222.
Assertion ownership · Investigation and reproduction · Architecture report · Tracker #16099.