| name | browser-harness |
| description | Browser Harness is a self-healing browser-automation harness from browser-use that lets LLMs complete arbitrary web tasks. Registered as an external reference for OctoCode's exploratory web-automation pathway. |
| source | https://github.com/browser-use/browser-harness |
| integration | external-browser |
| status | descriptor-only |
browser-harness
Descriptor-only registration. Sits next to the existing
agent-browser,
browser-use,
lightpanda-headless-browser, and
playwright-webui-qa skills.
Division of responsibility
| Concern | Tool | Scope |
|---|
| Liveness probe | lightpanda-headless-browser | ultra-fast Zig/Rust binary |
| Self-healing flows | browser-harness | LLM retries on selector/UI drift |
| Exploratory navigation | agent-browser | LLM-driven semantic navigation |
| Hard assertions | playwright-webui-qa | deterministic E2E |
Integration plan (deferred to a later iteration)
- Add
integrations.browser_harness.enabled (default off).
- Expose
browser-harness-run <goal> tool routed through
WorkspaceToolExecutor, gated by enforce_approval.
- Reuse the
skills-install HTTPS allowlist (see
crates/octocode-cli/src/skills_install.rs) plus a strict page-size
cap before any model context is hydrated.
- Emit one
event_feed_json entry per step for operator audit.
Safety notes
- Refuse forms whose fields contain
password, card, ssn, or otp.
- Snapshots and downloads must traverse
runtime/file_guard.rs.
- Treat LLM-suggested URLs as untrusted; deny RFC1918 and metadata IPs.
