[pelikhan]

Why

Routing choices, response credits, and missing telemetry are difficult to investigate from aggregate workflow results. This adds a Model routing page and per-run drill-down based on the analytics specification in github/gh-aw#69507, under the Experimental navigation section.

Approach

• Project metadata-only routing evidence from usage/aw_session.jsonl into the existing canonical Audit collection and typed PostgreSQL columns, without creating a parallel telemetry source. Prompts, bodies, free-form errors/reasons, and content-derived hashes are excluded.
• Declare routing state precedence, run-level filters, adoption, classifier dimensions, pick stability, response attribution, complete-cost distributions, observational comparisons, deviations, and agent outcomes in Dashboard Language. Extend shared quantile/window operations and parameter resolution consistently across the worker and hosted query engine.
• Add responsive run navigation, accounting/collection coverage, and timelines. Unknown costs remain unknown; partial costs do not enter complete-cost comparisons. Exact and correlated sub-agent attribution remain distinct.

Scope and deployment notes

This is a substantial implementation, not full conformance with every upstream metric. Candidate ranking, inventory/router change context, alias moves, prioritized deviation causes, failure clustering, and full safe-output/grader/delayed-outcome comparisons remain documented gaps.

Canonical schema 31 / IndexedDB version 41 rebuild older browser projections. Hosted PostgreSQL requires a refreshed generated schema. Generated aggregate query catalogs now share the existing aggregate-dashboard size-check exception; authored fragments remain size-checked.

Validation

After merging main through 8332050ca:

• npm run check passed, including root tests, schema checks, workflow compilation, and documentation build.
• Full dashboard unit suite: 3,466 passed, one skipped. Dashboard lint, typecheck, and document validation passed.
• Full browser suite: 270 passed, one skipped, including desktop/mobile routing, reactive filters, Experimental navigation, drill-down, and the upstream ingestion fixes. The previously failing ingestion smoke assertion now passes.
• Native PostgreSQL query tests, collection tests, and Go vet passed. Earlier PostgreSQL storage tests and the isolated deployed-data worker-memory budget run also passed.
• Populated desktop/mobile screenshots are attached in a comment.

Outstanding performance check: the earlier Lighthouse run reported CLS 0.137 against the 0.1 budget in populated CFO mobile scenarios; other measured scenarios passed. Lighthouse was not rerun after the base merge, so this is not reported as green.

[pelikhan]

Populated Model routing page under Experimental, at desktop and mobile widths. These are full-site screenshots rendered through the SQLite worker/query boundary from unified-session fixture evidence, not mocked page payloads.

[github-actions]

Dashboard dead views

Dead element views (0):
Dead dashboard pages (0):
Analyzed 8 element renderers and 65 dashboard pages across 1 dashboard.json file(s) from dashboard/site/src/components/ui-elements.js.

⚠️ This reachability report is informational and does not block the pull request.
🔎 View the workflow run and report artifact.

[pelikhan]

@copilot merge main and recompile

[Copilot]

@copilot merge main and recompile

Merged main at 4df4baeee in b7f87576b and recompiled all 72 workflows; no lock-file changes. Follow-up fixes through fb771ec66 address Go lint and preserve numeric routing status/HTTP evidence.

npm run check, 309 focused dashboard tests, dashboard lint/typecheck, and Go lint/query/Postgres tests passed. Full dashboard suite was stopped after timeouts; affected tests passed in isolation. Worker-memory validation was blocked by deployed-site DNS resolution. Go CodeQL found no alerts; remaining language analysis timed out.