fix(router): use the LLM to pick from the full roster on low-confidence routes (v0.292.5)

[vikasprogrammer]

Fixes the "wrong list" behind Cockpit's "Route to an agent instead" — and the underlying routing weakness it exposed.

Root cause

Routing "build a feature" returned app-packager, marketer, newsletter-writer — not engineer. Without an embedder, the keyword scorer matches word overlap, and "build a feature" doesn't lexically hit engineer's description ("writes code, debugs, ships fixes"). So it disambiguated to the wrong few. The LLM step existed but only re-ranked the near-tie shortlist — it never reconsidered the full fleet, so it couldn't surface engineer.

Fix

On any low-confidence keyword result (disambiguate/none), if an LLM is configured, it now picks the single best-fit agent from every agent's description (llmTieBreak → llmPick, over the full roster). A confident keyword route is left alone (no LLM cost); UNSURE keeps the keyword decision; no LLM configured → unchanged keyword-only behavior.

Result — the reported case and its variants now resolve correctly:

• "build a feature" → engineer
• "which agent can help me build a feature" → engineer
• "spam and abusive content" → trust-safety via keyword, 0 LLM calls (confident routes skip the LLM)

This lifts routing quality across Cockpit, Discord, and Slack (all go through chooseAgent), now that a Claude key is configured.

Verification

• typecheck + build clean; test:governance → 159/159 + tier-A 18/18 + capability 18/18
• Stubbed-model e2e: keyword-weak → full-roster LLM pick → engineer (1 call); confident keyword → trust-safety (0 calls); no key → keyword-only unchanged

After merge

Deploy, then in Cockpit hit "Route to an agent instead" on that ask answer (or just type "build a feature") → it should route to engineer, not the marketing agents.

src/edge/router.ts.

🤖 Generated with Claude Code