# Harness view and evals

The /harness pages: registry with editable approval policies, run history, run detail with replay diff, and evals with a score trend.

The harness is the part of the demo a support lead takes over: which tools the agent may use, under which approval policy, what each run did, and how both engines score on frozen cases. Every page reads through the `harness` and `admin.harness` procedures; nothing is hard-coded in components. Only the support Workspace reaches these routes (`_auth/harness/route.tsx`).

## Routes [#routes]

| Route                                                      | Page                        | Shows                                                                                                                                                                                                |
| ---------------------------------------------------------- | --------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `/harness?tab=werkzeuge\|skills\|kontext\|richtlinien`     | `harness.registry.page.tsx` | Counts, a tablist (arrow keys move between tabs), tools table with a sheet (description, context sources, skills, input JSON Schema), skill cards, context sources with PII level, approval policies |
| `/harness/laeufe?surface=&engine=&status=&seite=&mandant=` | `harness.runs.page.tsx`     | Run history with engine and fallback reason, status, steps, duration and start; filters and pagination live in the URL                                                                               |
| `/harness/laeufe/$runId`                                   | `harness.run.page.tsx`      | Input, engine per phase, findings with evidence links, full trace with raw JSON per step, actions with previews, audit timeline, replay with engine choice and a diff to the original                |
| `/harness/evals?lauf=`                                     | `harness.evals.page.tsx`    | "Evals ausführen" for rules and AI, progress, results matrix (case × check), run history as an SVG score trend plus a table, the eval cases with their expectations                                  |

## Policies [#policies]

Policy edits go through `admin.harness.policies.update`, a platform-admin procedure. The page enables editing only for a Workspace lead who is also platform admin (`workspace.me()`: `isLead && isPlatformAdmin`); everyone else sees the same table read-only with the reason. Write tools offer only `confirm` and `four_eyes` (`auto` is for reads). Every change is confirmed in a dialog, then the newest `policy.updated` audit event (sequence, short hash, actor, time) is shown next to the table.

## Engine per phase [#engine-per-phase]

`harness.run-model.ts` reads the engine of each phase from the trace: a `model` step means the AI answered (the model id comes from the step title), a blocked `guardrail` step means the AI output was dropped and the rules engine re-ran the phase, anything else is the rules engine. The run header shows this next to the run's overall `EngineBadge` and its fallback reason.

## Replay [#replay]

Replay runs the original input again on current data, read-only (`harness.runs.replay`, then `execute`); its actions get status `replay` and can never be executed. The panel compares root cause, actions, step count, engine and status and says whether the original cause is gone.

## Evals chart [#evals-chart]

`harness.evals-trend.tsx` is hand-written SVG (`role="img"`, `<title>`, `<desc>` listing every point): score per completed eval run, oldest left, one line per engine. Rules are circles and AI squares, so the engine is not told by colour alone. The history table below it carries the same data and links each run to its results.

## Shared agent components [#shared-agent-components]

`apps/web/src/platform/agent/` holds the presentational pieces the support console and the Elefant assistant reuse: `AgentTrace` (optional `phaseAside` and `highlightSeqs`), `AgentTraceStep`, `EngineBadge` (optional `stacked`), `ActionCard`, `DiffView`, `AuditTimeline`, `DomainChip` and the chip specs in `agent.labels.ts` (categories, priorities, statuses, approval, risk, surface, run kind, PII, context kind, step kind and status; label keys for root causes, guardrails and eval checks).

## Known gaps [#known-gaps]

- `RunSummary` carries no "started by"; the detail page reads it from the `run.started` audit event, the list does not show it.
- There is no organization list for platform admins; the Mandant filter lists the tenants seen in the newest 100 runs (`admin.harness.runs.list`), and that procedure filters by tenant and surface only.
- `RunDetail` has no `usage` field, so token usage is not shown.