Routing Benchmarks Phase 3

Date: June 20, 2026.

What Changed

• Added dogfood routing goals with --routing-goal <label=goal>.
• Added explicit route expectations with --expect-route <label=task-id[,task-id]>.
• Recorded routing results in summary.json and compact routing columns in summary.csv.
• Preserved report-only routing by default. Runs fail only when an explicit route expectation misses.
• Added an offline fixture assertion that verifies the quickstart goal routes to the quickstart task pack.

Why This Matters

Readiness and search quality show whether generated context exists and looks useful. Routing benchmarks check a more agent-specific question: when a user asks for a concrete goal, does AgentDocs select the expected task pack or fall back to generic source evidence?

Routing Classifications

Classification	Meaning
matched_exact	Selected task pack matches the expected task-pack ID list.
matched_related	A task pack was selected, but no expectation was declared or it was not expected.
fallback	No task pack was selected; AgentDocs used source search and goal-bundle evidence.
unsafe_mixed_context	Verification reported mixed task or search context.

Commands To Run

corepack pnpm build
corepack pnpm regression:fixtures
corepack pnpm docs:build
corepack pnpm typecheck
corepack pnpm test

Known Limitations

• Routing accuracy is a beta metric and should be read alongside source coverage, readiness, search captures, and agent_task_passed.
• Historical dogfood rows may not have routing metrics. Treat those as historical_metric_not_captured, not zero.
• This phase adds measurement. It does not add new task-pack families or an LLM router.