[4nrry]

Why the hook: skill-creator's run_eval on 20 Portuguese prompts (3 runs each, claude -p on Opus, installed plugin disabled in a scratch folder) shows the description fired on only 3 of 30 should-trigger runs, with 0 of 30 false. That matches the English eval of 2026-08-06, so the UserPromptSubmit hook carries the trigger.

hooks/inject.sh

• Fires on fan-out words (ultracode, deep research, dynamic workflow, /dwf, /fleet), or on a form of "pesquis" together with a sign of several sources: fontes, links, artigos, estudos, compara, changelogs, releases, Reddit, comunidade, "o que estão falando", no X, concorrentes, agentes, a fundo, de mercado, DOI.
• Portable: the patterns are POSIX ERE (no \b, no \s), so GNU grep, BSD grep and ugrep agree. Lint: shellcheck -S warning.
• New X rule in the injected header: search X only when the question is about posts, announcements, an account or what people say, or when the user asks.

Measured with the real hook (bench/plugins/agent-fleet/results/2026-10-26-pesquisa-trigger/, check_hook.py + logs/):

Set	New hook	Old hook
dev, written before the rule (28)	14/15 fired, 0/13 false	2/15, 0/13
blind 1, Haiku; its misses shaped the rule (24)	12/12, 0/12	1/12, 0/12
blind 2, Haiku, written after the rule (30)	12/15, 0/15	0/15, 0/15
2026-08-06 eval queries (20)	4/10, 0/10	1/10, 0/10

• Blind 2 is the honest score. Two of its three misses never say "pesquisa".
• Why the rule is two-sided: "pesquisa" alone fired on 9 of 13 near-misses (code search, a survey form, a search box…).

Other agents: the same script runs through adapters/claude_hooks.py (Kimi, ZCode, Muse), the opencode and Hermes bridges, and natively in Grok Build and Cursor. Through the dispatcher:

• a market-research prompt and a /dwf prompt each injected about 6 KB;
• "pesquisa no repo onde calcularFrete() é chamada" injected nothing.

Codex and Antigravity run none of these hooks.

The README row is updated. agent-fleet goes from 1.5.1 to 1.6.0. just check: 108 checks, 0 failed.

🤖 Generated with Claude Code