Run per-lead web research on a candidates CSV, write one bundle.md per lead, and emit assignments.json splitting leads into A/B arms for experiments. Uses parallel sub-agents for batches of 10. Spawned by cold-experiment when the experiment design has personalization as the variable, or by the user saying "research these leads", "personalize this list", "build per-lead bundles".
Import a CSV of leads, score each against icp.md, drop misaligned ones, optionally enrich with per-lead web research, then push the survivors into FoxReach as a campaign. Replaces the v0 stub. Use when the user says "import leads", "score these prospects", "filter to my ICP", "build a campaign from this CSV".
Read the active experiment's results from FoxReach, run the statistical test, and propose a cold.md diff (or auto-commit if trust earned). Computes interested-reply rate per variant via foxreach inbox categorize-stats, applies a two-proportion z-test, checks guards (bounce rate, not-interested rate), and either writes .cold/proposed-diff.patch or applies it directly. Use 7+ days after /cold send when an experiment is active. Updates beliefs.md and decisions.md. Use when the user says "read the experiment", "/cold learn", "what did we learn", or "is the test done yet".
Push the active campaign to FoxReach — creates sequences, registers variants from cold-draft output, runs pre-flight checks (banned-phrase lint, audience check, plan-limit math, bounce-rate guard), and starts the campaign. Use after /cold leads + /cold draft. Replaces v0 stub.
Pull unread inbox replies from FoxReach, surface interested ones, draft responses per cold.md
Generate a digest of cold outreach performance - deliverability, replies, booked calls, domain health. Use when the user asks for a report, digest, summary, weekly recap, performance check, or asks "how did my campaigns do". Reads live metrics from the FoxReach API. Designed to be called by cron (daily or weekly) and post results to Slack or email. Part of the cold.md suite.
Print the current state of the autoresearch loop — what we've learned, what's being tested, what's pending review, what's next. Reads everything in .cold/ and outputs a one-screen dashboard. Use anytime to orient. No FoxReach calls — purely local read.
Design the next A/B experiment for the autoresearch loop. Reads the current tier from .cold/config.json, picks the variable to test, drafts arm specs (e.g. statement vs. question subject lines), declares sample size + success criteria, and writes a protocol to .cold/experiments/. Use after /cold offer and before /cold draft. The experiment protocol is what cold-draft reads to generate variant pairs. Use when the user says "design the next experiment", "what should we test next", "set up the A/B", or "/cold experiment".