harmony/ROADMAP/pr-archive-2026-07.md

PR knowledge archive, July 2026

This document preserves useful knowledge from the PR cleanup. It is not an architecture decision and does not keep the old branches alive. Any new work starts from current master, checks the current code first, and extracts only the part needed by the active roadmap.

Application deployment notes

OpenWebUI with external Ollama (#277)

The branch was reported as a working deployment used by real users. Preserve the recipe, not the stale framework change.

• Chart: open-webui/open-webui from https://open-webui.github.io/helm-charts/.
• Deployment shape: one replica, external Ollama, internal service on port 80 to container port 8080, 5 GiB RWO persistence, and an OKD edge-terminated ingress.
• Disabled bundled services: Ollama, pipelines, Tika, and websocket support.
• Security intent: non-root pod, RuntimeDefault seccomp, no privilege escalation, and all capabilities dropped.
• Known manual step: the generated ingress did not work and was repaired manually.
• Before restoring the example: pin the chart/image, source WEBUI_SECRET_KEY from secret management, supply the real Ollama URL, fix and test ingress, add resources and a smoke check, and decide the required auth and backup behavior.

The old branch's Helm interpreter edit is obsolete. Current Helm deployment already supports upgrade/install behavior.

Qwen3.5-27B-FP8 with vLLM on RTX 5090 (#254)

This branch was also reported as a working deployment. Its 523-line raw Kubernetes example does not fit ADR-023, but the hardware and runtime settings are useful.

• Image used: vllm/vllm-openai:latest; pin a tested release or digest before reuse.
• Model: Qwen/Qwen3.5-27B-FP8 on one NVIDIA RTX 5090 with 32 GiB VRAM.
• Starting context: 16,384 tokens. The full 262k context did not fit this hardware.
• Runtime flags: --reasoning-parser qwen3, automatic tool choice, --tool-call-parser qwen3_coder, and --language-model-only.
• Resources: one GPU, CPU request/limit 2/10, memory request/limit 10/30 GiB, 50 GiB RWO Hugging Face cache, and 4 GiB memory-backed /dev/shm.
• Placement: nvidia.com/gpu.product=NVIDIA-GeForce-RTX-5090.
• Rollout and health: one replica, Recreate, /health readiness after 120s, and liveness after 300s.
• Exposure: service and OKD edge-terminated Route on port 8000.
• Known manual step: the official image required the anyuid SCC because it ran as root and wrote under /root.
• Before restoring the recipe: use a non-root image, remove the SCC exception, pin vLLM, revalidate GPU utilization and parser flags, add API authentication, source the Hugging Face token from secret management, and add metrics and a smoke check.

The replacement should be a focused deployment Score or use an existing Score. Do not restore the raw manifest builder.

Fleet knowledge to retain

PR	Useful knowledge	Missing or unsafe part	Current use
#328	Operator restart recovery, replay signal, orphan cleanup scenarios	Failed deletes can be forgotten; watcher failures are hidden; E2E is not in normal CI	Input to post-release Fleet recovery work
#329	Two-second HTMX device log tail with conditional auto-scroll	No load bound, fan-out, or browser test	Input to a small operational logs feature
#330	ADR-022 upgrade state machine, version heartbeat, staged binary verification and operator handoff	Draining, signed artifacts, attempt-scoped stop, rollback guarantees, and device E2E are incomplete	Deferred until the production release is stable
#331	Container-ID tracking and graceful workload replacement	Restart adoption and partial multi-service failures can preserve or report the wrong state	Keep as upgrade/convergence test cases, not code to port
#293	Route authorization middleware and typed claim parsing	Uses the retired role-map vocabulary rather than ADR-025/027 groups	Input to dashboard authorization design
#294	Twenty operator restart/recovery scenarios	Only one weak scenario was implemented	Preserve the scenario list for production E2E planning
#299	A smaller smoke companion than #292	Still abstract, unenforced, and has no production consumer	Input to the dedicated readiness/idempotency design
#319/#332	Evolution from JWT deployment claims to entity policies	Both models were superseded at Fleet scale	Historical rationale is preserved in ADR-025
#313	Staging OpenBao/config bootstrap experience	Old config paths and secret assumptions conflict with current Fleet code	Recheck only for missing operational steps
#300	Tag-driven image and chart publication	Superseded implementation	Current release path owns this behavior

Framework and infrastructure knowledge to retain

PR	Useful knowledge	Missing or unsafe part	Current use
#261	Multiple named instances of one config type are needed for firewall pairs and similar deployments	Old API, unvalidated names, path/env collisions, no current consumer	Backlog under ROADMAP/11-named-config-instances.md
#255	Brocade VLAN CRUD, interface speed, port-channel IDs, and LACP configuration	Broad API break, reachable todo!(), weak idempotency and tests	Input to the OPNsense/bare-metal track
#285	OPNsense bootstrap sequence, firmware/package setup, NIC pinning, and LAN migration	Too broad, stale config path, failing CI	Split only when the bare-metal track resumes
#297	Rook operator/cluster/storage/object-store resource requirements	Runtime todo!(), handwritten schema drift, raw example manifests, no E2E	Storage backlog; no current sprint work
#238	Monitoring installation-plan idea across several backends	127-file redesign before selecting the capability boundary	Design input for ROADMAP/14-tenant-application-monitoring.md
#212	Missing certificate capability documentation	Old text is stale and too long	Small documentation backlog item

Closed work with no remaining implementation to recover

PR	What happened
#340	Compose build arguments and Zitadel readiness landed through #341
#316	Login return-path behavior reached master through later work
#311	SSO example cleanup became stale after config and identity changes
#292	The first smoke framework was replaced by the smaller #299 design
#295	Log querying was incomplete and attached to the wrong Score abstraction
#281	Redpanda benchmark knowledge was tied to raw kubectl apply; recreate through Scores only if prioritized
#260	Arbitrary shell desired state was replaced by typed Fleet reconciliation
#253	SQLite configuration support landed and evolved on master
#247	KVM support landed, gained tests, and now supports Fleet E2E
#236	Earlier version of #238 with no separate knowledge to preserve
#215	Docker autoinstall was incomplete, unsafe, and based on obsolete tooling

Prioritization rule

The current return order is:

1. production Fleet deployment UX/DX;
2. next customer release;
3. Fleet deployment E2E on the test cluster;
4. focused readiness and idempotency design;
5. small framework extractions proven by the Fleet work;
6. application recipes and infrastructure backlog when they serve an active deployment.

Branch history remains available in Git if a detail is needed. Do not copy an old implementation wholesale; reproduce the proven behavior with the current framework.