Hardened the web agents. Search, fetch, and verify now carry an explicit
WebFetch/WebSearch-only instruction and are told to skip a source rather
than fetch it another way, so they do not fall back to a shell curl on an
untrusted URL when WebFetch fails. The reasoning is in the README and in
the fetcher shouldn’t have a shell.
Shipped web-fetcher.md: the opt-in shell-less agent type for
hard enforcement (create the agent, add agentType to the three web calls).
Off by default — the plain install is the skill without the lock.
0.0.2 — 2026-08-30
First published version (this repo).
The verify-phase comment’s issue reference made resolvable:
anthropics/claude-code#69883 instead of an internal shorthand.
SKILL.md stripped to runtime-only content; history, provenance, and
install moved to this folder’s README and the skills index
(a skill is paid context).
0.0.1 — 2026-07-19
Initial fork from the built-in deep-research (base: the shipped
script as of 2026-07-13), private. Per-stage model pinning — search
on Haiku, fetch and verify on Sonnet; verify cap restored to 25;
three-outcome verdicts (unverified vs refuted); scope-schema
robustness fix.