pipelinecpc got an agent planning bridge

pipelinecpc can talk to agents now.

not as an ad launcher. that line matters. the useful version is safer: codex, claude, or hermes can read saved keyword research, group it into ad groups, suggest negatives, and export a draft google ads structure. fresh keyword fetches are still gated behind explicit key permissions and credit caps.

the first test exposed the real problem: saved keyword lists can be noisy. an "openai api alternative" list tried to turn consumer ai junk into ad groups. so the planning layer now infers the likely campaign goal first, then splits keywords into included, needs review, and excluded before exporting anything. not perfect, but much harder to blindly turn bad inputs into a bad campaign.