Host generic Flow Agents Kit contributions and expose a feedback workspace

[briananderson1222]

Outcome

Console gives an operator a cross-run view of Kit value, friction, grounded explanations, and proposed improvements without hardcoding Builder or Knowledge semantics. Builder and Knowledge are the first conformance fixtures for a generic Flow Agents Kit contribution host.

Boundary

• Consume the versioned Flow Agents Kit observability contribution contract; do not inspect Kit-private artifacts or branch on Kit IDs.
• Console owns authenticated ingest, tenancy, storage lifecycle, correlation, generic rendering, cohort selection, queues, and operator decisions.
• Flow/Surface/Flow Agents authority projections are rendered with their labels and references; Console never re-evaluates gates, claims, or Kit outcomes.
• Observational real-run trends and controlled Evals results are visibly separate. Console must not turn correlation into a causal claim.
• Kit-generated proposals are advisory. Actions route back to the owning provider/ledger; Console does not silently mutate a Kit or routing policy.
• Headless Kits and Flow Agents remain fully usable without Console.

First vertical slice

1. Register and validate Kit contribution descriptors/records by version and tenant.
2. Render generic surfaces for run summaries, funnel/gate progress, metric series, queues, narratives, and learning proposals/decisions.
3. Add cohort filters for Kit/version, treatment revision, model/runtime, task class, completion/completeness class, time window, and observational versus controlled source.
4. Show outcome, defect/acceptance references, route-backs, retries/no-progress, human corrections, evidence completeness, tokens, latency, and cost without collapsing them into one score.
5. Drill from aggregate -> run -> grounded narrative/evidence -> exact raw source or typed unavailable reason.
6. Link proposals to their evidence window, owner repository/issue, ratify/reject/defer decision, and later effect report.
7. Preserve historical lifecycle semantics and treatment revisions rather than rewriting old cohorts.

Acceptance

• Builder and Knowledge fixtures render through one registry with no Kit-ID conditionals.
• A synthetic third-party Kit contribution appears without modifying Console core beyond registration/configuration.
• Invalid/future versions are quarantined with diagnostics while valid records continue.
• Tenant isolation, cross-tenant source refs, hostile content, redaction canaries, and large-run behavior have tests.
• CONFIRMED, FAIL, NOT_VERIFIED, timeout, partial usage, and workflow-not-engaged remain filterable and distinct.
• Observational and controlled panels use different provenance labels and cannot be combined into a causal lift number.
• Every aggregate is traceable to run IDs and producer-owned source refs.
• Narrative drill-down composes Human execution narrative: summary, evidence argument, and raw provenance drill-down #207 rather than creating a second explanation format.
• Learning decisions compose the Flow Agents decision ledger projection rather than becoming a Console-owned authority.
• Local fixture verification covers responsive, keyboard, and screen-reader use.

Dependencies and relations

• Flow Agents: Kit observability contribution contract
• Flow Agents: retrospective Kit observation compiler
• Flow Agents #463: proposal/decision ledger mirror
• Flow Agents #611 and Console Human execution narrative: summary, evidence argument, and raw provenance drill-down #207: grounded narrative
• Console Define lifecycle and stale-state semantics for historical operating records #247: historical lifecycle/stale-state semantics
• Evals: treatment/effect reports and independent grading
• Ops feat(console): authenticated ApiSink + route flow-bridge through the Sink layer (#73) #80/feat(telemetry): token/cost aggregation + cost-by-model dashboard #84 and Flow Agents #410: cross-suite direction and dogfood

[briananderson1222]

Owner clarification (2026-07-23): Console must be extendable by a Kit, not only generically aware of several record shapes.

The intended symmetry is:

• a Kit extends Flow Agents' execution behavior through the Kit contract;
• that same Kit may extend Console's inspectability experience through an optional contribution declared by the Kit;
• Console hosts the contribution through a governed capability API and shared UI primitives;
• neither Console nor Flow Agents core imports the Kit's private semantics.

Acceptance therefore requires install/discovery/version/capability lifecycle, a no-core-change third-party Kit proof, and no Builder/Knowledge identifier branches. A purely hardcoded generic dashboard over Builder-shaped records does not satisfy this issue.

[briananderson1222]

Station convergence link: kontourai/station#732 is the peer host for the same Flow Agents #911 Kit experience contribution.

Console should support the portable MCP App view where its hosted security/tenant boundary can do so, and always support the declarative standard-view fallback. It must not consume Station plugin React bundles or server modules. Shared conformance must prove the same third-party fixture, view identity, navigation, provenance, and capability-degradation behavior in both hosts.

[briananderson1222]

2026-07-30 concrete Knowledge child: #269 now defines the source/store/projection/retrieval-quality/migration health projection needed by the cross-suite Knowledge program. Compose it through this generic Kit contribution path where the contracts fit; do not fork a Knowledge-only authority or telemetry envelope.

[briananderson1222]

Implemented and pushed the first generic host slice in draft PR #271.

What is now covered:

• public Flow Agents contract/conformance consumption only
• generic descriptor/record registry with Builder, Knowledge, and synthetic third-party fixtures using the same path
• tenant/source-authority isolation and safe quarantine behavior
• historical descriptor revision and lifecycle-at-ingest preservation
• separate observational/controlled cohorts with no causal-lift computation
• aggregate -> run -> producer-owned source reference traceability
• distinct CONFIRMED, FAIL, and NOT_VERIFIED values
• hostile-content, redaction-canary, future-version, cross-tenant, and 1,000-record tests
• authenticated contribution, record, workspace, and run APIs

Verification is green for clean install, focused tests (8/8), typecheck, build, content boundary, lock integrity, and the full repository suite. The full server suite is 482/482 and browser coverage is 63 pass with 1 intentional skip.

Still intentionally open for later slices: durable storage/replay, the responsive/accessible Console UI, MCP Apps execution, Station parity, and controlled treatment-effect computation. This PR remains draft and stacked on the unreleased Flow Agents contract at adff692d2aa72bcf8b600b30089b3c8cc6effb76.