Status and known gaps¶

BotsHub Agent is alpha, built on AAA, which is itself 0.x.

Verified¶

Discovery run against live GitHub, Hacker News and SearXNG (266 signals → 221 candidates).
Three enrichments against a live gateway (Firecrawl, claude-mem, Browser Use): all published, and an independent audit confirmed 79 of 79 stored facts have their quote on the cited page.
28 offline tests; MLX critic and embeddings smoke-tested.

Lightly live-tested. No repo-only candidate enriched yet, BOTSHUB_RUN_MAX_COST tripping not verified, and cost per listing not measured.
Run-to-run variance. A re-run can drop a fact the previous run found. Grounding guarantees what is kept is on the page, not that everything on the page is kept.
Weak-signal facts. Integrations can come from "works with X" prompts, "free trial" from a one-time credit, and plan names from table text.
Local models are limited. No tool calling; a 2B critic is a soft check.
Reasoning models need config (BOTSHUB_LLM_PARAMS=reasoning_effort=none).
Screenshots are viewport captures of a few pages. They need Playwright plus a Chromium or system Chrome; sites that block headless browsers or gate content behind logins yield thin captures. Videos found by search are matched by title/channel name only, so an unusual product name can still pick a tangential video.
Semantic merge is conservative and merged nothing in the first run.
Prompt-injection defence is line filtering, data-only prompting, grounding and own-site-only research tools. AAA tool_result guardrails are not used because page text is not returned through tools.
Auto-publish has no human review. Audit a sample; keep unpublish handy.

MIT. See LICENSE.