Enforce skill invocation via UserPromptSubmit routing hook

[vivek7405]

Problem

Agent Skills (webjs-file-issue, webjs-start-work, webjs-list-todos, use-railway) are model-invoked: they fire only when the model judges a request to match their triggers, and that judgement can be wrong. In one session a "can we add prefetch, verify in the remix repo, let me know if we can" request was treated as pure research, so the webjs-file-issue workflow was skipped and code investigation began with no tracked issue. Skill descriptions and AGENTS.md prose are advisory; only a hook is deterministic, the same way block-prose-punctuation.sh enforces the em-dash rule.

Design / approach

Add a UserPromptSubmit hook (.claude/hooks/route-skills.sh) that runs on every prompt and, from the prompt TEXT (not model judgement):

1. Keyword routes the prompt against each skill's documented triggers and, for each hit, injects a high-priority directive naming the skill to invoke via the Skill tool BEFORE other work.
2. Always injects a generic standing policy: before acting, check the available skills and invoke any whose purpose matches, because skills encode required workflow. This is generic across ALL skills, not a rule about filing GitHub issues specifically. It also covers the research-becomes-build case keyword routing can miss.

Verified mechanics against the Claude Code hooks reference: a hook CANNOT force a Skill tool-call; the strongest deterministic lever is UserPromptSubmit additionalContext injected as a system reminder. So keyword cases are deterministic; genuinely-ambiguous prompts lean on the always-injected policy. Honest ceiling, documented as such.

Wired into .claude/settings.json alongside the existing PreToolUse / PostToolUse hooks. Lives in the repo .claude/ so every agent on webjs gets it (personal ~/.claude/skills already shadow project skills, so the skills themselves are not duplicated into the repo).

Acceptance criteria

• UserPromptSubmit hook routes file-issue / start-work / list-todos / railway trigger phrases to the right skill
• Standing skill policy is injected on every prompt, including ones with no keyword match
• Hook emits valid hookSpecificOutput.additionalContext JSON and exits 0 (never blocks a prompt)
• Registered in .claude/settings.json; merged config is valid JSON
• AGENTS.md documents the hook in the agent-config table
• Tests cover the routing decisions (each skill's triggers + the no-match standing-only case)