ADR 0021 — Interactive browser automation: deferred, with named triggers (v37-F5)
Date: 2026-07-11 · Status: accepted
Context
Hermes and OpenClaw both ship browser automation, and the project-comparison
review (2026-07-10) listed it as an open gap. skep already ships the read
half: v29's network.read capability gives workers governed page reading —
fetches only through the per-task domain allowlist, stdlib HTML-to-text, full
event evidence. What skep does not have is interactive browsing: clicking,
forms, sessions. Building it means a real browser dependency, a much larger
attack surface, and a new class of side effects to govern.
Decision
No interactive browser automation now. network.read covers the
research/read use case that field tests have actually surfaced; nothing
observed so far needs page interaction.
Revisit only when a named trigger fires:
- A Hermes-replacement field test (the v36 Stage F acceptance ritual) records a concrete inventoried action that cannot be governed without page interaction; or
- a recurring worker task class fails that a read-only fetch cannot complete.
A trigger reopens the question; it does not pre-approve the feature.
Constraints binding any future implementation
- A worker capability behind the domain allowlist and the policy engine, with full event evidence — never a free tool.
- The browser process runs inside the OS sandbox with the same egress enforcement as the worker itself.
- Never available to the chat Queen directly; only to dispatched workers under a contract.
Consequences
The comparison page keeps an honest "no" in the browser-automation row, and the recurring "should skep browse?" debate has a recorded answer with exit conditions instead of a vibe.