[xisen-w]

Installed SharedNet skills miss large execution tasks that do not mention SharedNet. The skill now automatically selects Goal mode for requested execution teams and substantial work with independent tracks and a verifiable deliverable, using the requested headcount or four workers by default.

Explicit Room operations remain available to solo agents and assigned workers. Simple/sequential tasks, conceptual requests, solo instructions, explicitly chosen alternative coordination systems and existing team assignments do not start new teams. Building software containing agents alone is distinguished from requesting an execution team. The Goal reference documents prerequisites, finite limits, real startup evidence and installed-skill updates. Runtime code and npm versions are unchanged; copied skills must be updated separately.

Measured behavior

• Paired native project-skill discovery: baseline 0/10 to final 10/10 actual skill loads for English/Chinese 5,000-contact research; both conditions 0/10 false positives for typo/solo negatives. Every final positive successfully read SKILL.md and the Goal reference.
• Prompted catalog simulation, 25 frozen bilingual requests × 5 fresh repetitions per condition: positive recall 30/60 to 60/60; strict false positives 12/65 to 0/65. Nine baseline false positives were debatable nested-worker guard loading; the report also gives that sensitivity.
• Full-body dry-run routing: final 50/50; separate review-driven regression routing 15/15. Default four/requested six workers each 5/5; prerequisite blockers preserved 10/10; no nested teams 35/35. Exact pressure-label disagreements and the read-only protocol confound are retained and explained rather than regraded as a perfect overall score.
• 455 scored fresh-model calls with compact raw decisions, expected labels, hashes, methods, setup exclusions and manual audits in docs/verification/2026-10-26-goal-mode-trigger-eval.{md,json}.

These are controlled observations on one model. Catalog/routing simulations are separated from native skill reads, and none of the tests proves live worker startup or population-wide trigger reliability. Independent review found two routing ambiguities; both were corrected, with the solo-plus-Room failure observed before the fix.

Validation

Skill validator, export boundary, typecheck, 471 passing unit tests (1 skipped), 16 process integration tests, installed-package smoke and Gitleaks history/evidence scans passed locally. Implementation CI passed on Node 22.18.0, 24.x and 26.x plus secrets; require the same checks on the final evidence head before merge.