Organization Dashboard evidence depth

Traceability

• Spec ID: 2026-09-02-organization-dashboard-evidence-depth
• Status: Implemented
• Extends: 2026-09-01-organization-harness-dashboard, 2026-09-01-task-evidence-upload-end-to-end

Intent

A review of the Dashboard against the scripts behind it found three classes of gap: values the collector gathered and the page discarded, values the page rendered that were not interpretable, and signals the scripts already produce that the collector never asked for. This spec closes them without adding a metric that has no source field.

The correctness problems came first, because the page's largest numbers were the wrong ones:

• Asset inventory double-counted. runAgentLint was called once per host and the per-host summaries were summed. Eleven hosts read the same four .agents/skills/*/SKILL.md and the same AGENTS.md/DESIGN.md, so the page reported 28 Skills and 18 Rules for a repository holding 4 and 2.
• The Model chart implied the full response population. It plotted the 1,474 responses that carried a model beside a headline of 17,163 responses, with no statement that 15,712 were unattributed.
• Token lanes summed incompatible counters. codex reports included-in-input (cache reads already inside inputTokens) and claude reports separate-input-lane. Adding them produced an Input total that is not a quantity of anything, and the summary's own accountingMode: "mixed" was never shown.

The organization angle had one structural hole: the upload record carries destination.organization, receipt.acceptedAt, receipt.state, and the packet digest, and the collector projected only plan.packet, so an organization Dashboard could not group by organization.

Acceptance Scenarios

• AC-1 — Distinct assets, with instances preserved: agent-lint emits a host-stable assets[] identity per configured asset — the workspace-relative path for a contained file, otherwise a scope-qualified name that carries no absolute path. The Dashboard reports distinct assets as the headline and the summed configured instances beside them. A report without identities marks the distinct total incomplete rather than silently undercounting.
• AC-2 — Model attribution stays visible: the Model section states how many responses carry a model out of the analyzed total, and names the unattributed remainder that the chart cannot show. The overview's response card reports the attribution rate rather than an unqualified count.
• AC-3 — Cache relationships travel with the totals:session-efficiency records the distinct cacheAccountingModes behind its retained counters, usage-summary carries them, and the Token section states the observed relationship. Hosts that disagree produce an explicit non-comparability note and overlapping lane labels.
• AC-4 — Organization identity survives collection: the collector reads upload records, not bare packets, so organization, acceptance time, receipt state, and digest reach the page. A bounded read reports how many records exist, so a capped list never reads as the whole population.
• AC-5 — Delivery behavior is reported: post-edit validation status, edit and later-validation counts, Task Episode closure, execution friction, observed tools, and observed hooks are projected from the insight pack the analyzer already builds. Evidence refs stay behind: they carry absolute session-file paths.
• AC-6 — Delivered change and project shape: commit-session-link supplies session-attributed commits and attributed line counts, and workspace-topology supplies member routes, tracked files, and instruction scope activation. Only explicit, high, and medium matches attribute a commit; a low match is a bare time overlap and attributes nothing.
• AC-7 — Host axis: one row per session provider reports that host's own sessions, active minutes, responses, model attribution, edits, episodes, and cache relationship. The table scrolls inside its own container.
• AC-8 — Analyzed window is stated: the page leads with the workspace label, the UTC date range and day count the series cover, the host counts, and the collection time.

Non-goals

Unchanged from 2026-09-01-organization-harness-dashboard, and specifically:

• No cost, first-pass-success, autonomy, or quality score is introduced. Commit attribution is a correlation with a stated confidence floor, not a claim that an agent authored a commit.
• Episode closure and post-edit validation are reported as observed counts, not converted into a readiness grade.
• No authentication, remote ingestion, or multi-repository aggregation.

Test and Review Evidence

• AC-1, AC-2, AC-3, AC-4, AC-5, AC-6, AC-7: npm test -w @qoder-ai/harness-ui passes 34 tests. test/collector-signals.test.mjs drives the real buildInsightPack, runAgentLint, createUploadPlan, and storeUploadPlan, and asserts that delivery projections contain no evidenceRef or absolute path, that a low commit match attributes nothing, that a bounded delivery read still reports its total, and that a workspace file yields the same identity for every host. test/dashboard-model.test.mjs proves one shared file counts once while its configured instances stay visible, that an inventory without identities marks the distinct total incomplete, and that mixed cache modes are reported as non-comparable.

• AC-8 and the rendered contract: npm run build -w @qoder-ai/harness-ui produces the dynamic route, and the page was driven manually against that production server with the browser tooling: the header reports the workspace, the Jul 28 - Sep 2 UTC / 37 days window and the collection time; assets read Skills 4 with 28 configured instances / 4 observed invocations; the model caption reads 1,650 of 17,362 responses carry a model beside the explicit unattributed remainder; the Token section reports mixed accounting with the cache-disagreement note and (overlapping) lane labels; the host table lists only the five hosts with sessions and names the seven scanned-empty ones; and the accepted evidence row shows acme-engineering, its acceptance time, and a 12-character digest. document.documentElement.scrollWidth equals the client width at both 1440 and 390, and the host table keeps overflow-x: auto.

  The Playwright suite (npm run test:browser -w @qoder-ai/harness-ui) was updated for this contract but was not executed: this machine cannot fetch the browser build the pinned @playwright/test@1.55.0 requires (chromium_headless_shell-1187), so the run fails before reaching the page. Run it where that download succeeds before treating its assertions as passing.

  Follow-up: CI hit the same missing browser build, because the repository-level playwright install chromium step provisions only the version the other workspaces share. The Dashboard now depends on that same @playwright/test@^1.62.1, and the suite passes under it.

• No regression in the owners: npx vitest run at the repository root passes 1,639 tests, covering session-efficiency, usage-summary, and agent-lint after the contract additions.