Orchestrators rename + dashboard, KG recall stabilization, and plan resume#385
Merged
Conversation
… predictable
Addresses three reported failures: the KG "not working between sessions",
recalling knowledge unpredictably, and not using plans.
- bootstrap: self-heal a dual-active knowledge-graph state (both inmemory and
neon installed) by removing the non-selected sibling, keeping the backend
that actually holds a DSN (vault or env) rather than choosing by env alone.
Previously left both providers active so reads/writes routed
non-deterministically between backends.
- bootstrap: reconcile the embeddings plugin's ollama_base_url/ollama_model/
max_concurrent from env on every boot, not just first install. Enabling the
Ollama overlay on an existing deployment was a silent no-op, leaving semantic
recall, the durable tier and process-reuse inactive (FTS-only).
- recall judge: make the relevance judge fail-deterministic with an internal
verdict cache, so an identical query replays the same verdict instead of
oscillating between filtered and unfiltered. Errors/abstains keep-all and are
not cached.
- durable tier: route durable ("always-surface") hits through the relevance
judge by default (kg_durable_relevance_judge_enabled) so off-topic curated
facts stop surfacing for unrelated queries; runs even on a term-less turn.
Falls back to always-surface when no judge is wired.
- health: expose a KG capability snapshot on /health (backend, durable,
embeddings, semantic recall, durable tier, process-reuse, warnings) so silent
degradation is observable instead of requiring boot-log archaeology.
Tests cover dual-active self-heal (incl. vault-DSN guard), embeddings reconcile,
verdict caching, durable opt-in/out and term-less judging, and the health
snapshot. Full middleware gate green.
…hannel Wires plan-resume into the only path that reaches the model prompt. Turn hooks are observer-only (they cannot inject system context), so resume is realized in the cross-session plan-recall the context-assembler already injects. - RecalledPlan gains completedStepGoals + resumeFromInProgress/Sideeffecting. - loadPlanHits collects the done-step goals and the resume-from step's status. - An interrupted plan (some steps done AND some open) renders as an explicit resume hint: completed steps marked "do not redo", an explicit resume point, and the remaining steps — so a follow-up continues the plan instead of restarting it. - Side-effect safety: when the resume-from step was in_progress (possibly mid-execution) it carries a caveat to verify before re-running, stronger when the step is side-effecting, mirroring buildResumePlan's ambiguous-side-effect guard that the recall channel would otherwise drop. Tests assert the framing reaches the assembled text, the side-effect caveat, and that fully-done / fully-pending plans do NOT render resume framing.
…access consolidation
Operator-UI overhaul developed across the rename effort:
- Rename "Agents" -> "Orchestrators" throughout the operator UI (en + de),
including nav, agent picker, and the operator agents page.
- Dashboard as the landing page: the root route now renders a dashboard
(system-health strip, quick-access, role onboarding); chat moves to /chat.
- Business-case-first onboarding (businessCases.ts + dashboard component)
replacing the technical bootstrap-profile flow.
- Admin consolidation: providers and subscription-CLI pages refactored into
reusable panels (ProvidersPanel, SubscriptionClisPanel, LlmAccessTabs) with
inline LLM key entry and in-page tab switching.
- Standard ("fallback") orchestrator protection: the operator agents API
refuses to disable or delete it (409 fallback_protected); its seeded default
name becomes "Standard Orchestrator".
- Store page copy + nav cleanup; chat turn-usage telemetry badge removed;
header subtitle removed.
- CI: bundle the official Claude CLI into the published middleware image
(INSTALL_SUBSCRIPTION_CLIS=true) so subscription-CLI agents work on self-host.
web-ui `next build` passes; middleware typecheck/lint/tests green.
…delete)
The fallback orchestrator is now protected from disable/delete in the operator
agents API, but that load-bearing invariant had no CI coverage. Add regression
tests: PATCH {status:'disabled'} and DELETE on the fallback return 409
fallback_protected; enabling/renaming it still works; and the protection also
covers whatever the platform fallbackAgentId points at (id-based, non-"fallback"
slug).
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
1 participant
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.
Summary
Bundles the operator-UI rename/dashboard work with a round of knowledge-graph recall stabilization and a plan-resume feature. All four commits are deployed and verified on the local docker test stack, and each was reviewed with an independent (GPT-5.4) pass.
What's in here
Knowledge-graph recall stabilization (
fix(kg))ollama_base_urlfrom env on every boot, not just first install. Enabling the Ollama overlay on an existing deployment used to be a silent no-op that left semantic recall, the durable tier and process-reuse inactive (FTS-only)./health(backend, durable, embeddings, semantic recall, durable tier, process-reuse, warnings) so silent degradation is observable instead of requiring boot-log archaeology.Plan resume (
feat(kg))Operator UI: Orchestrators rename + dashboard (
feat(web-ui))/chat.409 fallback_protected); seeded default name becomes "Standard Orchestrator". Backed by regression tests (test(operator)).Testing
next buildclean./healthreports backend neon, embeddings on, all capabilities true), dual-active self-heal fired on a real pre-existing conflict, web-ui serving.Reviewer note — action required before publishing the public image
This branch sets
INSTALL_SUBSCRIPTION_CLIS=truefor the publicghcr.io/byte5ai/omadia-middlewareimage inpublish-images.yml, which bundles the proprietary Claude CLI. The Dockerfile defaults this OFF for public images pending legal review. It is branch-only (no publish is triggered by this PR), but redistribution of the vendor CLI in the public image needs legal sign-off before the next release.Follow-ups (not in this PR)
buildResumePlanas a plan-runner service the context-assembler consumes, instead of the current parallel realization.