Changelog

2026-10-04 — Re-pin the e2e runtime to v0.8.8

docker-compose.e2e.yml now defaults to ghcr.io/multiagentcoordinationprotocol/macp-runtime:0.8.8 (was the v0.8.0 SHA tag f97fd15). Release builds now publish semver image tags, so the SHA workaround is gone. macp-control-plane / macp-playground pin 0.8.6; this stack is two patch releases ahead of them. Override with MACP_RUNTIME_IMAGE. This resolves the "Known drift" item in the entry below.

2026-10-04 — Docs sync: post-0.8.0 fixes, dependency majors, de-duplication against sibling docs

Covers the work merged after the 2026-09-23 entry (#39, #43, #46–#50) and trims console docs that restated sibling-repo content. No app-code change in this entry.

Since the last entry

#46 — live-run reconnect no longer targets a stale run. attemptReconnect reconnects through connectSSERef, so navigating run A → B without a remount no longer re-subscribes to A after a heartbeat timeout or SSE error. Details: api-integration.md § SSE integration.
#47 — dependency majors. mermaid 12 (the Mermaid 11 look is pinned in components/docs/mermaid-block.tsx), jsdom 30, vitest 5, jest-dom 7; CI runs on Node 22 (and #39 bumped actions/setup-node to v7); engines.node is >=22.22.2.
#48 — policy schemaVersion default is 3 at the control plane (fail-closed on an empty decisive tally) and unknown rule keys are rejected (macp-control-plane#96). No console behaviour change — the registration form already sent an explicit value.
#49 — docs loader marks its dynamic fs calls turbopackIgnore so the build stops tracing the whole project (lib/docs/loader.ts).
#43 / #50 — docs-content/macp-playground/ re-synced from macp-playground (#43: policy authoring, direct-agent-auth, scenario authoring, worker bootstrap contract; #50: deployment). Those files are automated copies; edit them upstream, not here.

De-duplication

Rule: console docs describe what the console does and link to the owning repo for everything else.

api-integration.md: runtime notes, POST /runtime/policies validation rules, the control-plane error envelopes and the SSE resume rationale are now short console-side summaries linking to macp-runtime/docs, macp-control-plane/docs/API.md and INTEGRATION.md.
backend-repo-notes.md: the credential-scoping (FORBIDDEN:) walkthrough is reduced to the console triage rule and where to read the reason; the mechanism lives in macp-control-plane/docs/INTEGRATION.md.

Known drift (not changed here)

Sibling repos have moved on since the e2e pin: macp-runtime is at v0.8.8, macp-control-plane absorbed runtime v0.8.1–v0.8.6 (#96), and both SDKs are at 0.14.x. docker-compose.e2e.yml still pins runtime f97fd15 (v0.8.0), and the docs describe that pin accurately. Re-pinning is an infra change that needs a stack run, so it is left as a follow-up.

2026-09-23 — Absorb macp-control-plane / macp-playground (Jul–Sep 2026), runtime v0.8.0

Absorbs the control-plane and playground work landed since the v0.5.0 window, and repoints the dev/e2e stack at runtime v0.8.0 (ghcr.io/multiagentcoordinationprotocol/macp-runtime:f97fd15, the tag both sibling repos pin). Mostly correctness and fidelity work in existing surfaces — no data migration and no wire-contract change on the console side.

Correctness

CommitmentAuthority wire value corrected designated_roles → designated_role. This repo was the only one in the stack with the plural; the runtime matcher, the runtime validator, the control plane's rejection check, the playground contract mirror and a shipped policy file all use the singular. The playground now rejects the mistake explicitly with designated_role authority requires a non-empty designated_roles array. The sibling array field designated_roles: string[] keeps its plural — it is a different thing.
SSE resume cursor no longer skips events. lastSeq — what goes out as ?afterSeq= on every reconnect — was derived from timeline.latestSeq, the server's head, in three places. The load-bearing one was the snapshot handler: the control plane republishes a snapshot on every commit batch, so the cursor was dragged to the head continuously during normal streaming and anything not yet delivered was skipped permanently. The cursor is now derived only from events actually received. Note afterSeq is exclusive and replay is gated on afterSeq > 0, so 0 means "snapshot plus live tail", not "send everything".
The live event rail no longer collapses to a single row. The hook reads its seed events at mount only, and the live route mounts before the history query settles, so the fetched history was shown until the first live event arrived and then discarded. The two sources are now merged by id and ordered by seq.
That merge no longer orders on a comparator that can return NaN. seq is typed number but reaches the console through JSON.parse of an SSE frame, so the type is a claim about the wire, not a guarantee — which is why the resume cursor above is Number.isFinite-guarded. a.seq - b.seq returns NaN against a missing or non-numeric seq, and a NaN-returning comparator is not a valid ordering: the engine may leave the array in any order at all, so a single malformed frame could scramble the whole rail rather than merely misplace itself. Unusable values now sort to the end, in arrival order.

Fidelity & errors

decision.current.supersedes surfaces the non-canonical/canonical distinction.
Structured errorCode on ApiError, with the control plane's three error envelopes handled distinctly: an AppException carries a real code; a Nest exception with an object body is emitted verbatim and usually has no errorCode (this is what every POST /runtime/policies validation rejection looks like); and a string body or unhandled error is rewritten with a hardcoded INTERNAL_ERROR. That last one is why a throttled 429 arrives labelled INTERNAL_ERROR — a present errorCode is not always a real classification. Read message via describeApiError rather than branching on a code.
Runtime session drift (GET /admin/runtime/sessions) surfaces on the Infrastructure tab: sessions the runtime holds with no control-plane run, and runs whose session the runtime has lost. A sweep cut short by its page or time budget reports complete: false, so a partial answer is never read as "no drift".
Policy schemaVersion constrained to {1, 2, 3} in the request type and the registration form.
historyGap — the control plane's own signal that a run's history is incomplete, set when it could not resume the runtime session because that history had been compacted away — is now surfaced as a fidelity notice in the event feed. Its contract doc had asked for exactly this and the console had ignored it. The notice offers no retry, because those envelopes were never recorded: after emitting the gap the control plane degrades to poll-only, and polling returns session state, never the missing events.
Event summaries no longer render half-built strings. formatEventSubject returned a dangling kind: when a subject carried a kind but an empty id, and the outcome suffix on policy.resolved / policy.violated / policy.commitment.evaluated carried its own separator, producing a doubled · ·. Both fixed.
policy.denied now shows its reasons. The inline denial path populates none of the fields the old summary read, so a denial rendered as the bare words "Policy denied" with the reasons unread in the payload. session.stream.gap gained a label and a /logs filter entry.
No client-side seq-gap detector, deliberately. Canonical seq values are not contiguous per run — for each runtime envelope persisted, the control plane allocates one sequence span, gives the first value to the raw row and the rest to the canonical events derived from it, and publishes only the canonical ones, so each such batch burns a seq no subscriber ever sees. When a raw envelope yields exactly one canonical event, the published seqs therefore run 2, 4, 6, 8 — but that ratio is not fixed either: normalization can emit two canonical events for one envelope, or none at all, so the stride varies as well as skipping. (Control-plane-authored lifecycle events take a different path that allocates exactly what it publishes and burns nothing — which is precisely why a client cannot infer loss from a delta: the two paths interleave in one counter.) A detector on "seq jumped, therefore an event was lost" was built and would have warned once per batch on every healthy run; it was removed, and a regression test pins the hook's return surface so it cannot return under another name.

Launch flow

A failed bootstrap now says so. Previously the spinner simply stopped; the page now surfaces the failure and does not navigate. Both launch triggers — the main submit and the quick-run button — were byte-identical duplicate mutations, so a fix to one would have left the other on the old behaviour; they now share a single mutation.
/examples/run responses carry controlPlaneRun, and the new-run page redirects to the live view using controlPlaneRun.runId. That id is not the session id: POST /runs mints a fresh run id and stores the session id separately, and the live route resolves by run id. When the field is absent the page explains what it does and does not know rather than redirecting somewhere broken.

Upstream behaviour worth knowing (no console change needed)

Event counts on /runs/:id/events can legitimately go down. The control plane now drops StreamSession envelopes whose sessionId is empty or belongs to another session; previously a short-circuit let empty ones through. Dropped envelopes are never persisted.
Runs fail faster and more honestly. A stream-loop crash now finalizes the run as failed immediately, bounded by a last-resort timeout, instead of leaving it hanging — narrowed so an intentional shutdown is not misreported as a failure.
Large envelopes now arrive. The gRPC receive limit is explicit and raised to 16 MiB; it was previously grpc-js's implicit 4 MB default, which surfaced as mysterious gaps or 503s. A send limit was added alongside it.
Proto moved 0.1.9 → 0.1.10 — audited, and it touches nothing the console reads. The whole delta across every .proto file is a four-line comment change on PolicyDescriptor.schema_version in policy.proto, documenting that RFC-MACP-0012 now defines version 3. No field was added, removed, renamed, re-nested or retyped. policy.proto is loaded only for the gRPC policy RPCs and is not among the descriptors the control plane decodes into decodedPayload, so it cannot reach the event stream at all. The console already implements the semantics that bump documents — POLICY_SCHEMA_VERSIONS = [1, 2, 3], with new registrations defaulting to 3. Worth keeping for next time: mode-payload decodedPayload is a straight pass-through of the proto decode, so a future bump touching decision/proposal/task/handoff/quorum/ multi_round does need this check — which is one command, git diff proto-v<old> proto-v<new> -- packages/proto-npm against the monorepo's tags, not the open-ended audit it was first assumed to be. Demo fixtures remain non-evidence for upstream drift, since they are this repo's own mock data.

Corrections to this repo's own docs

Several long-standing errors were found while verifying the above, and are worth calling out because each had been wrong for some time:

The proxy service identifiers were documented as example / control-plane. They are macp-playground / macp-control-plane.
The documented backend env var names — EXAMPLE_SERVICE_BASE_URL, CONTROL_PLANE_BASE_URL, EXAMPLE_SERVICE_API_KEY, CONTROL_PLANE_API_KEY — do not exist. The real names all carry the MACP_ prefix.
GET /runtime/modes returns five descriptors, not six. The control plane calls the runtime's ListModes, which returns the standards-track set only; ext.multi_round.v1 lives behind ListExtModes, for which the control plane exposes no route. Demo mode still shows six — a known divergence, now flagged in the mock data rather than silently disagreeing with production.
The runtime advertises list_changed: false for roots, so the console never watches them. The previous claim that "there is no change-notification stream" was wrong about the runtime — a WatchRoots RPC exists and parks idle — even though the console's behaviour was right.
The Commitment-terminal rule is now enforced at registration, not merely conventional.
The runtime exposes 24 gRPC RPCs; the notes said 22.
Policy management moved to /policies; docs still pointed at /settings.
Mock runtime manifest protocolVersion → 0.1.10. It tracks the proto package, not the runtime build. The previous 0.5.0 came from this changelog's own 2026-07-07 entry, which set it alongside the v0.5.0 image pin and conflated the two. A real runtime returns empty manifest metadata, so the field is illustrative — but not invisible: /modes renders the whole manifest as raw JSON, so in demo mode a wrong value here is a wrong value on screen.
Source line counts were removed from the docs rather than updated — client.ts was documented as ~460 lines against an actual 1346, and counts like these are guaranteed to rot.

Infra & docs

docker-compose.e2e.yml runtime pinned to ghcr.io/multiagentcoordinationprotocol/macp-runtime:f97fd15 (runtime v0.8.0, overridable via MACP_RUNTIME_IMAGE), matching macp-control-plane and macp-playground byte for byte. The SHA tag rather than a semver one: the runtime repo moved to per-crate tags, so macp-runtime:0.8.0 is not a published image. The three RUNTIME_LIST_SESSIONS_* knobs are present as commented-out entries, with the recipe for forcing the drift table's incomplete-sweep branch and the CircuitBreaker footgun that comes with it.
Demo timeline counters are derived from the event fixtures instead of restating them. They were hand-written literals and three of the six runs had drifted, so the demo showed "11 events" beside a 14-row rail. The cancelled run was the worst case: it claimed three events and had no event fixture at all, which also left /logs's run.cancelled filter entry matching nothing in demo mode — the default. It now has the three events its projection was already asserting, and mock-data.test.ts fails if any counter is restated rather than derived.
Updated: docs/api-integration.md, docs/architecture.md, docs/feature-matrix.md, docs/backend-repo-notes.md, README.md, and two code comments that had gone stale alongside them (lib/data/mock-data.ts, app/modes/page.tsx). Also CLAUDE.md, which is gitignored and so appears in no diff. The in-app /docs surface needs no separate edit — it renders docs/*.md directly, so there is no second copy to drift.
Not updated, by design: everything under docs-content/macp-playground/. Those files are auto-synced from the upstream playground repo, so hand edits would be clobbered by the next sync. They still carry v0.5.0-era statements, including a stale "authority": "designated_roles" that upstream has already corrected; the sync will carry it.

2026-07-07 — Absorb macp-runtime v0.5.0 / macp-proto 0.1.6

Absorbs the runtime v0.5.0 release into the console, consumed via the macp-control-plane and macp-playground (both of which shipped their own v0.5.0 absorptions). All changes are additive display / mock / config work — no data migration or wire-contract change.

Lifecycle & events

Session canonical-event vocabulary corrected to what the control plane actually emits: session.bound, session.stream.opened, session.state.changed (transitions carry data.state, e.g. SESSION_STATE_SUSPENDED). The /logs Session filter group and summarizeEvent were keyed to phantom session.opened/resolved/expired events; those are retained only for legacy exports. Added run.suspended / run.resumed summarizers.
Demo mode now covers the full lifecycle: added a suspended and a cancelled run (records, projections, metrics with SESSION_STATE_SUSPENDED / _CANCELLED) so the suspended badge, Resume action, and Suspended KPI render in demos.
Breadcrumb id detection widened for 36-char base64url session ids.

Handoff implicit accepts (RFC-MACP-0010 §5.1)

Runtime-synthesized implicit handoff accepts (silent target past the accept window) now render with an implicit badge in the live feed, /logs, and the event dialog (isImplicitAccept()), distinguishing them from accepts a participant actually sent.

Registries

/modes now displays terminalMessageTypes per mode. Mock descriptors rewritten to match runtime all_mode_descriptors(): all six modes (was three), quorum terminal set fixed to [Commitment], ext.multi_round.v1 bumped to 1.0.0.
/settings policy management detects a file-managed (read-only) runtime registry from the CP's 405 REGISTRY_READ_ONLY error, shows a persistent banner, and disables register/unregister instead of looping dead-end toasts.

Launch flow

InitiatorPayload.sessionStart gains maxSuspendMs (proto 0.1.5 max_suspend_ms, compiled by the Examples Service) — surfaced as a "Max suspend" badge in the launch review. Removed the deprecated inline SessionStart.context field; launch inputs flow only via scenarioMeta.sessionContext.

Infra & docs

docker-compose.e2e.yml runtime pinned to ghcr.io/multiagentcoordinationprotocol/macp-runtime:0.5.0 (overridable via MACP_RUNTIME_IMAGE), MACP_METRICS_ADDR + port 9464 exposed, MACP_ALLOW_INSECURE kept (now load-bearing). Mock runtime manifest protocolVersion → 0.5.0.
Runtime Prometheus metrics documented as an ops-only surface (the CP does not re-serve them, so the console's runtime-metrics UI was intentionally not built).
docs/api-integration.md, in-app /docs, and README.md updated for the above.

2026-04-22 — In-console docs surface + Examples Service docs sync

Moves the UI Console and Examples Service docs out of the marketing website and into the Console app itself at /docs. Operators now read the docs next to the product.

New

/docs — single-page landing with hand-authored themed SVG diagrams (architecture, scenario-pack pipeline, run flow) plus a doc index.
/docs/macp-ui-console/[slug] — renders this repo's docs/*.md via react-markdown + remark-gfm. Mermaid code fences render as diagrams via a lazy-loaded client component that re-renders on theme change.
/docs/macp-playground/[slug] — renders markdown from docs-content/macp-playground/, which is synced from the upstream repo.
Sidebar gains a Docs entry (BookOpen icon, immediately after Dashboard).
New sync workflow .github/workflows/sync-examples-docs.yml + scripts/sync-examples-docs.sh. Listens for repository_dispatch: docs-updated from the macp-playground repo (plus a 6-hour fallback schedule), copies the docs, and opens an auto-PR.
lib/docs/loader.ts — small server-side helper (title / first-paragraph extraction).
components/docs/markdown-renderer.tsx with link rewriting (relative .md → /docs/<collection>/<slug>, cross-repo paths → GitHub URLs), styled code / tables / blockquotes, and github-slugger heading anchors.

Removed

.github/workflows/notify-website.yml — the console is the docs host now, not a source for the website. Website-side cleanup (removing app/console/ + app/macp-playground/ + their sync blocks) lands in a separate website commit.

Dependencies

Added react-markdown, remark-gfm, mermaid, github-slugger.

Operator note

The macp-playground repo's notify-website.yml is retargeted at ui-console in a sibling commit. Once that ships, a push to macp-playground main automatically opens a sync PR here. Merging the PR redeploys the Console with fresh Examples Service docs.

2026-04-22 — Docs refresh: cross-repo references, reduced duplication

Backend docs (macp-playground, macp-control-plane, runtime, python-sdk, typescript-sdk) were all refreshed upstream. The UI console docs are realigned to reference those sources rather than mirror their endpoint tables. No UI code changes.

backend-repo-notes.md rewritten as a pointer index. Per-service endpoint inventories removed; each service now links to its canonical upstream doc plus a short note on the UI-facing role.
api-integration.md trimmed: endpoint schemas now point to macp-control-plane/docs/API.md and macp-playground/docs/api-reference.md; the UI-specific pieces (proxy, normalizers, SSE hook, demo mode, launch sequences) remain documented inline.
architecture.md gained an Upstream services section linking to each repo's architecture doc (CP, ES, runtime, SDKs). Client line counts resynced.
README.md gained an Upstream repositories table and updated doc index.
All occurrences of "Example Service" normalized to the upstream spelling Examples Service.

2026-04-22 — Observer-only CP alignment, session-interaction rework, Playwright removal

Consolidates a run of UI ↔ backend integration changes (5adcc20, 1045a59, 65433b6, fdbe99f, 6cd9da6, 4965a63, 6da8a76).

Control Plane contract alignment

Aligned the client with the observer-only Control Plane authority model: agent-originated traffic (messages, signals, context updates) now flows through the runtime SDKs. CP endpoints POST /runs/:id/messages, POST /runs/:id/signal, and POST /runs/:id/context return 410 Gone. POST /runs/:id/cancel proxies to the initiator agent's cancel callback (or calls CancelSession directly when policy-delegated).
POST /runs is the only run-creation path and accepts a whitelisted-safe compiled RunDescriptor only — CP rejects any body containing kickoff[], participants[].role, commitments[], policyHints, or initiatorParticipantId. The UI continues to compile via Examples Service, which is already whitelisted-safe.
POST /runs/:id/clone rejects non-empty context overrides.
CreateRunResponse now surfaces sessionId alongside runId (equal under observer-only; reserved for future auto-discovery).
Extended RunStateProjection with the llm projection: { calls[], totals: { callCount, promptTokens, completionTokens, totalTokens, estimatedCostUsd } } plus contextId and extensionKeys on the run block. NodeInspector surfaces the LLM tab from this data.

Runtime policy registry integration (RFC-MACP-0012 pass-through)

New client functions: listRuntimePolicies, getRuntimePolicy, registerRuntimePolicy, unregisterRuntimePolicy against CP /runtime/policies{?mode} and /runtime/policies/:policyId.
Surface: /policies (registry browser) and /modes (mode-scoped policy list).
PolicyBadge and PolicyPanel tones updated to reflect the revised governance rules and commitment-evaluation contract.
Mock dataset gains MOCK_RUNTIME_POLICIES to keep demo mode exercisable.

Session interaction simplification

Removed UI-side message/signal/context POST forms that targeted the now-410 CP endpoints. The runs workbench is now a read-only consumer of CP state, canonical events, and artifacts; agents drive envelopes via macp-sdk-python / macp-sdk-typescript.
Artifact creation (POST /runs/:id/artifacts) and projection rebuild (POST /runs/:id/projection/rebuild) remain CP-local and are kept.

Dashboard chart series

CP now emits a single decisionOutcome series (positive/negative encoded as +1/-1 per bucket) instead of the earlier split decisionOutcomePositive / decisionOutcomeNegative arrays. Client and mock data updated; any callers relying on the split fields were removed.

Response normalization cleanup

normalizeRun() no longer synthesizes archivedAt from tags — CP exposes the column directly; the client passes it through unchanged.
normalizeEvent() promoted out of the SSE hook and applied uniformly in getRunEvents and listEvents so every CP event consumer gets the nested source / subject shape.
listEvents gained a session-cached eventsEndpointMissing flag — when CP returns 404 for /events, the fallback per-run fan-out is used for the rest of the browser session without re-probing.

Test infrastructure

Playwright removed (4965a63) — the visual-regression and E2E spec suites are retired. All golden-PNG baselines deleted. The Vitest + React Testing Library suite remains canonical.
Integration tests under test/integration/ tightened for the observer-only contracts and the new 410 responses (6da8a76).

UI regression fixes

Color palette and chart/config warning cleanup that regressed during the CP rewrite work (fdbe99f).
Lint-only fixes across decision-panel, policy-panel, and run-workbench (6cd9da6, ddd8a26).
ExecutionGraph regained the four extra auto-layout passes lost in the CP rewrite (part of 6da8a76).

Docs

api-integration.md, backend-repo-notes.md, and architecture.md regenerated against the current lib/api/client.ts. Runtime policy endpoints, /admin/circuit-breaker/history, /readyz, /runs/:id/replay/state?seq=, /runs/batch/export, and the Jaeger proxy route are now documented.
README.md route map includes /modes and /policies; added the npm run local:{up,down,status} Docker full-stack section.

2026-04-14 (continued) — Phase F + Phase D remainder

Second push on 2026-04-14: completes Phase F (filters + observability) and the remaining Phase D items (D3 Prometheus parser, D4 tree waterfall, D5 /traces filters). All items below ship against the backend substantially-complete plan; no console / lint warnings; npm run build clean.

Added

API client extensions

listEvents(demoMode, query) (lib/api/client.ts) — wrapper around BE §4.1 cross-run /events endpoint. Query shape: { runId, scenarioRef, type, afterTs, beforeTs, afterSeq, limit }. Returns { data, total, nextCursor }. Demo mode fans out across MOCK_RUN_EVENTS with the same filters so /logs is exercisable without a backend.
getRunEvents accepts RunEventsQuery — new fields afterTs, beforeTs, type (csv) map to BE §4.2. Legacy (runId, demoMode, limit, afterSeq) call signature preserved for back-compat.
getDashboardOverview(demoMode, query) — accepts { window, scenarioRef, environment, from, to } per BE §5.1. Forwards as query params; demo mode filters MOCK_RUNS client-side so KPIs respond to the pickers.
Chart series extended: charts.throughput, queueDepth, latencyP50/P95/P99, cost, successRate, decisionOutcomePositive, decisionOutcomeNegative, perScenario — matches BE §5.2 output.
getCircuitBreakerHistory(demoMode, window?) — BE §5.3 endpoint. Returns CircuitBreakerHistoryEntry[] ({ state, enteredAt, reason? }).
computeDashboardKpis(runs?) accepts an optional filtered run list so demo mode can respond to scenario/environment filters.

Pure helpers

lib/utils/prometheus.ts — parsePrometheusText() with full support for counters, gauges, histograms (grouped under base name via _bucket/_sum/_count suffixes), summaries, and escaped quotes in labels. metricSummaryValue() returns the one-value-per-row summary (sum of labeled counters, _count for histograms). histogramQuantile(metric, p) interpolates percentiles from le buckets — same algorithm as Grafana's histogram_quantile. 13 unit tests.
lib/utils/traces.ts — buildSpanTree() builds depth-annotated tree from span.references[CHILD_OF], sorts siblings by startTime, treats orphans as additional roots, and counts descendants. findCriticalPath() post-order DP finds the longest root-to-leaf chain by cumulative duration. 7 unit tests.

New components

PrometheusMetricsTable (components/observability/prometheus-metrics-table.tsx) — sortable, filterable table over parsed exposition. Type chips (counter/gauge/histogram/summary), name/help search, click-row → EventDetailDialog with per-series breakdown. 4 integration tests.
CircuitBreakerTimeline (components/observability/circuit-breaker-timeline.tsx) — horizontal proportional-width state timeline + current-state badge + per-transition table. Freezes now at mount for render-purity. 3 tests.

Page rewrites

/logs (PR-F1) — rewritten against the cross-run /events endpoint:
- Shared <RunSelectorFilters> (scenario + runId + time window)
- Event type filter with EVENT_TYPE_GROUPS incl. new LLM group
- Cursor-based "Load more" pagination (cursors kept as an append-only useState array; each new cursor concatenates a dedup'd page)
- Client-side free-text search over the current page
- summarizeEvent() one-line summary in the Payload column
- EventDetailDialog opens on row click with full payload + Run / Trace / Event id meta rows
- URL state for every filter → shareable links (runId / scenario / window / type)
- Empty states + proper loading/error panels
/traces (PR-D4 + PR-D5) — tree-indented span waterfall:
- Calls buildSpanTree() + findCriticalPath()
- Per-row depth indentation with └ connector
- Critical path segments tinted accent-blue; non-critical use accent-2; errors tinted red
- Click-a-span → EventDetailDialog with Span ID / Trace ID / Start / Status meta
- Shared <RunSelectorFilters> (scenario + status + runId) replaces the single-select dropdown; URL deep-linkable
- Graceful fallback to flat render when references[] is missing (runtime-side fix still pending — BE §10)
/observability (PR-F3 / PR-F4 / PR-F5 / PR-F6 / PR-D3):
- Shared <RunSelectorFilters> (scenario + environment + time window)
- New chart grid for BE §5.2 series (throughput / queue depth / latency percentiles / cost over time) — each renders only when data is present
- Three new KPI cards for p50 / p95 / p99 latency computed client-side from the Prometheus histogram (PR-F6)
- CircuitBreakerTimeline card (BE §5.3)
- Raw <pre> dump replaced with PrometheusMetricsTable (raw link preserved as "Open raw exposition ↗")
Signal Rail (components/runs/signal-rail.tsx) — "Open in full view ↗" link to /logs?runId=...&type=signal.emitted (PR-B3, the last remaining surface).

Changed

Many pages now wrap their content in <Suspense> boundaries to satisfy Next.js 16 useSearchParams requirements.
RunWorkbench passes runId into SignalRail (for the deep-link) and LiveEventFeed.

Tests

New unit tests: prometheus.test.ts (13), traces.test.ts (7), prometheus-metrics-table.test.tsx (4), circuit-breaker-timeline.test.tsx (3).
Full suite now at 371 unit tests passing (up from 344), across 30 test files.
36 visual baselines regenerated and passing deterministically.
npm run build succeeds — all pages render without runtime errors.

Notes

Spurious posttooluse-validate hook recommendations suggesting Vercel Workflow migration were ignored — lib/api/client.ts is a browser-side fetch wrapper, not a serverless handler.

2026-04-14 — Phase A–E implementation against the enriched backend projection

This release consumes the substantially-complete backend plan (see plans/backend-changes-plan.md status table as of 2026-04-14): state projection completeness, decision/policy enrichment, canonical event contracts, cross-run query endpoints, observability enrichment, and the re-scoped llm.call.completed event are all live on the Control Plane.

Added

Shared primitives

<Tooltip> (components/ui/tooltip.tsx) — CSS-positioned, keyboard-focusable hint primitive. Used by policy panel commitment ids and ready for further adoption.
<EventDetailDialog> (components/ui/event-detail-dialog.tsx) — shared native-<dialog> modal for click-through event / span / metric details. Consumed by LiveEventFeed; ready for /traces span details and /observability metric details.
<RunSelectorFilters> (components/runs/run-selector-filters.tsx) — shared filter bar (scenario / runId / status / environment / time-window). Prop-driven rendering so each consumer shows only the controls it needs. Q24 decision: build once, use three times.
<EmptyState> (components/ui/empty-state.tsx) — standardized icon + title + description + action.
<KpiCard> (components/ui/kpi-card.tsx) — Syne-displayed KPI tile with optional colored top-stripe accent.
<Panel> (components/ui/panel.tsx) — replaces ad-hoc Card + CardHeader + section-actions composition.
<Breadcrumbs> (components/layout/breadcrumbs.tsx) — derives a navigation trail from the current route, truncating UUID-like segments.

`summarizeEvent()` helper (`lib/utils/events.ts`)

Pure, type-aware helper that produces a one-line semantic summary for a CanonicalEvent. Covers the high-volume types (run., signal., proposal., decision., policy., message., tool.*, llm.call.completed) with a generic fallback for everything else. Unit-tested with 11 assertions.

LLM interaction visibility (finding #12)

New LlmCallCompletedData type mirrors the CP-synthesized event payload.
NodeInspector gains a conditional "LLM (N)" tab that lists llm.call.completed events for the selected participant — each row shows model, token counts, latency, and expandable prompt + response. Honors redactedPrompt when the RedactionService applied.
Mock data includes llm.call.completed events for the completed fraud run so demo mode exercises the tab.

Enriched projection consumption (BE-3 / BE-5 / BE-6 / BE-7 / BE-8 / BE-9)

DecisionProjection.current type extended with proposals[], resolvedAt, resolvedBy, prompt. outcomePositive widened to boolean | null per BE-3.
PolicyProjection type extended with expectedCommitments[], voteTally[], quorumStatus.
Decision panel (components/runs/decision-panel.tsx) rewritten to render the action header with action→tone mapping (approve/step_up/decline → success/warning/danger, neutral fallback), bulleted reasons, contributors table (from proposals[]), resolvedBy/at sub-line with deep-link to originating event seq, and the scenario prompt labeled distinctly from reasons. Outcome badge branches on run.status first, then outcomePositive: boolean | null. Removes the literal 'mode' / 'unknown' badges (finding #6a).
Policy panel (components/runs/policy-panel.tsx) rewritten to show expected commitments with Tooltip-on-hover revealing title + description + required roles, inline tally (allow/deny/quorum), per-commitment evaluation decisions, and quorumStatus badge. Falls back to the legacy commitmentEvaluations[] table when new fields aren't yet populated. Policy version badge links to /policies/[policyVersion].
Run overview card (components/runs/run-overview-card.tsx) — Q1/Q2/Q3 KPI unification. Events + Messages counters now source from the SSE projection during live statuses and from the metrics aggregate once terminal. Tokens/Cost removed from the top strip (moved to the observability summary card).

Event feed redesign (PR-B1)

LiveEventFeed rewritten:

Drops the truncated JSON.stringify() snippet per row.
Each row: type badge + seq + semantic summary (summarizeEvent) + meta line (timestamp, source).
Whole row is a button — click opens <EventDetailDialog> with the full payload, copy-JSON action, and metadata rows (subject, source, trace id, event id).
Size selector chips (100 / 500, default 100 per Q13).
Type filter chips auto-populated from observed event types.
"Open in full view ↗" link to /logs?runId=....

Changed

RunWorkbench now passes state + events + runId into RunOverviewCard, DecisionPanel, and LiveEventFeed to support the new projection-consuming behaviour.
ExecutionGraph outcomePositive handling widened to accept boolean | null | undefined from the enriched decision projection; null maps to undefined accent.
policy-panel.tsx removes the policyHints.type !== 'none' gate — full rules now render whenever a descriptor is reachable.

Tests

New unit tests: events.test.ts (11 cases), run-selector-filters.test.tsx (5), event-detail-dialog.test.tsx (3), tooltip.test.tsx (6), empty-state.test.tsx (3), panel.test.tsx (2), kpi-card.test.tsx (4).
New integration-style tests: decision-panel.test.tsx (9 cases covering outcome branching, prompt/reasons/contributors), policy-panel.test.tsx (5 cases covering enriched + legacy fallback paths).
Full suite: 344 unit tests passing (up from 311), 36 visual baselines passing deterministically.

New PolicyHints type with governance fields: type, threshold, vetoEnabled, vetoRoles, vetoThreshold, minimumConfidence, designatedRoles
PolicyProjection and CommitmentEvaluation types for runtime policy resolution
PolicyBadge component displays policy type with color coding (none/majority/supermajority/unanimous)
PolicyPanel component shows policy version, resolution status, and commitment evaluation table
Scenario catalog cards now show policy type badge
Scenario detail page shows full policy governance section (thresholds, veto config, designated roles)
Launch form displays policy version and description when template is selected
Run workbench shows PolicyPanel with commitment evaluations when policy data is available

Token and cost tracking

MetricsSummary now includes promptTokens, completionTokens, totalTokens, estimatedCostUsd
Run workbench observability summary shows token count and estimated cost when available

Mock data updates

Fraud scenario templates expanded to include majority-veto and unanimous variants
All mock scenarios enriched with policyVersion and policyHints
Mock run state projections include policy section with commitment evaluations
Mock metrics include token usage and cost
Policy canonical events (policy.resolved, policy.commitment.evaluated) added to mock events

Changed

API client

normalizeRun() simplified: archivedAt now passed through directly from CP instead of synthesized from tags
listRuns() always sends limit and offset defaults (required by CP validation)

Deep analysis fixes: gaps, stubs, bugs, and test coverage

Critical bug fixes

Delete run redirect: deleteMutation in RunWorkbench now redirects to /runs after successful deletion instead of leaving user on an error page
Form reset after send: SendMessageForm and SendSignalForm in SessionInteractionPanel now reset fields on successful mutation

Type safety and error handling

14 API functions now have explicit return type annotations with shared interfaces (MutationAck, BatchOperationResult, RebuildProjectionResult, CircuitBreakerResult, RunExampleResult, CreateArtifactResult, AgentMetricsEntry) defined in lib/types.ts
ApiError class added to lib/api/fetcher.ts — preserves HTTP status, service, and path for error discrimination
getAgentProfile() now only returns undefined for 404; re-throws other errors
getDashboardOverview() returns degraded: true flag when overview endpoint fails; logs warning
getAgentMetrics() logs warning on failure instead of silently returning []
normalizeRun() validates required fields (id, status, runtimeKind) before processing

Agent metrics improvements

averageLatencyMs is now optional — displays "N/A" instead of misleading "0ms" when unavailable
messages field from CP response is now mapped through to AgentMetricsEntry
Demo mode returns representative mock metrics instead of empty array

Pagination and configuration

getRunEvents() accepts optional limit parameter (default 500)
getAuditLogs() accepts optional limit and offset parameters
integrations.ts throws in production if base URLs are not configured; warns on empty auth tokens
Preset IDs now use crypto.randomUUID() instead of Date.now()

UI and component improvements

Mutation error messages now include backend error details
ExecutionGraph uses data-driven auto-layout by node kind instead of hardcoded positions
formatDateTime() and formatChartLabel() accept optional timeZone parameter
Tabs component: defaultTab prop deprecated in favor of defaultValue
Removed unused imports (Plus, createArtifact, updateRunContext) from RunWorkbench

Documentation

Added 10 missing API endpoints to docs/api-integration.md: session interaction, batch operations, context updates, artifact creation, projection rebuild, run deletion

Test coverage

47 new unit tests (226 total): real-mode API client tests, SSE hook tests, agent metrics tests
6 new integration tests (85 total): error scenarios for getRun/cancelRun/sendRunMessage (500/422), dashboard degradation, empty results
New test files: lib/api/client.real-mode.test.ts, lib/hooks/use-live-run.test.ts

getAgentMetrics() now normalizes CP response: participantId mapped to agentRef, averageLatencyMs passed through when available (optional)
Previously the agent catalog page silently failed to enrich profiles because the CP returns participantId while the UI looked up by agentRef
Updated integration test fixtures to match actual CP response shape (participantId, messages fields)

Dynamic environment filter

Runs page environment dropdown now dynamically populated from actual run metadata
Removed hardcoded local-dev | stage | prod options that didn't match real environments
Supports any environment value including development (the default set by Example Service)

LineChartCard and BarChartCard now use formatChartLabel() to auto-detect ISO timestamp labels and format them as short date strings (e.g. "Apr 1, 3 PM")
added formatChartLabel utility in lib/utils/format.ts

Token and cost KPIs from control plane

getDashboardOverview() now extracts totalTokens and totalCostUsd from CP KPIs when available, instead of hardcoded zeros

Server-side run filtering

runs page now passes status, environment, and search params to GET /runs for server-side filtering
client-side filtering retained as fallback in demo mode

Agent metrics enrichment

added getAgentMetrics() API function calling GET /dashboard/agents/metrics on the control plane
agents page fetches real per-agent metrics (runs, signals, latency, confidence) and merges into profiles

Clone with overrides

cloneRun() now accepts optional { tags, context } overrides parameter
added clone UI form in run workbench with tag and context JSON inputs

Batch export

added batchExportRuns() API function calling POST /runs/batch/export
added "Export" button to batch toolbar in runs table

Webhook delivery stats

WebhookSubscription type now includes optional deliveryStats with total, succeeded, failed, lastDeliveredAt
settings page displays delivery counts and last delivery timestamp per webhook

Testing

added unit tests for formatChartLabel, getAgentMetrics, cloneRun with overrides, batchExportRuns
added integration tests for batch export, clone with overrides, server-side filters, agent metrics
updated dashboard integration tests: token/cost KPI extraction from CP, fallback to zero when omitted
updated backend-response fixtures: dashboardOverview includes token/cost KPIs, added agentMetrics() and batchExportResponse() fixtures

added normalizeRun() helper in lib/api/client.ts that maps Control Plane response shapes to UI types
listRuns() now unwraps the paginated { data, total } response from the Control Plane
getRun() now normalizes the raw response through normalizeRun()
flat sourceKind/sourceRef fields are mapped to nested source: { kind, ref } on all run records
archivedAt is synthesized from tags.includes('archived') as a bridge until the Control Plane adds a dedicated column

Validate, cancel, and archive response mapping

validateRun() now maps the Control Plane's { valid, errors, warnings, runtime } response to a typed ValidateRunResponse with ok field
cancelRun() now extracts { ok, runId, status } from the full run record returned by the Control Plane
archiveRun() now extracts { ok, runId, archived } from the full run record returned by the Control Plane

Agent profiles via Example Service

getAgentProfiles() now calls Example Service GET /agents directly instead of computing profiles client-side via N+1 cascading calls to packs/scenarios/launch-schemas
getAgentProfile() now calls Example Service GET /agents/:agentRef directly with fallback to undefined

Dashboard overview with Control Plane aggregation

getDashboardOverview() now fetches from Control Plane GET /dashboard/overview for KPIs and chart data
falls back gracefully to client-side computation if the endpoint is unavailable
Control Plane chart data ({ labels, data } arrays) is converted to UI ChartPoint[] format

Types

added ValidateRunResponse interface to lib/types.ts

created a new Next.js App Router project scaffold
added TypeScript, Tailwind-free custom CSS shell, React Query, Zustand, React Flow, and Recharts integration points
added route-handler proxy for Example Service and Control Plane

app shell with sidebar navigation
sticky topbar with demo-mode and theme toggles
command palette with route shortcuts

Scenario surfaces

scenario catalog page
scenario detail page with version/template switching
launch schema and defaults inspection

Run launch flow

new-run page with pack/scenario/template selection
schema-driven input editor
raw JSON editor
compile launch flow against Example Service
validate execution request against Control Plane
submit run flow
optional Example Service one-shot bootstrap flow

Live and historical run analysis

live-runs page
shared run workbench for live and historical views
execution graph built with React Flow
node inspector with overview, payloads, signals, logs, traces, and metrics tabs
live event rail
signal rail
final decision panel
observability summary panel
artifacts and messages panel
replay descriptor request panel

History and comparison

run history page with filtering
run comparison page

Agent and observability surfaces

agent catalog and detail pages
logs explorer
traces explorer
observability dashboard
settings page with webhook and audit surfaces

Demo mode

rich mock data for packs, scenarios, runs, states, metrics, traces, and audit/webhooks
simulated live run streaming frames

Integration alignment

The UI was aligned to the uploaded repositories by using their exposed contracts and endpoints:

Example Service for scenario/launch compilation
Control Plane for run lifecycle, state, streaming, metrics, traces, audit, and webhooks
runtime metadata as surfaced by Control Plane runtime endpoints

Intentionally not changed

no modifications were made to the uploaded backend repositories
no assumptions were added that require backend code changes to render the demo mode