$ the-wire · showcase
ClawSweeper passes security fix for agent skill MCP boundary drift
By RepoJournal · Filed · About OpenClaw
OpenClaw merged a security fix that flags agent skill MCP boundary drift, and removed the direct skill scaffold helper to keep durable skill work inside workshop proposals.
OpenClaw merged a security fix [1] that adds a warning when agent skill MCP boundary drift is detected. The change passed ClawSweeper review, and no repair was needed after automerge opt-in. This is a security-relevant safeguard for anyone running agent skills with MCP boundaries.
OpenClaw also removed the direct `init_skill.py` scaffold helper [2] to prevent skill-creator from bypassing workshop proposals. Current main had `init_skill.py` writing a live `SKILL.md` directly; the branch removes that helper and its test, routing durable work through Skill Workshop proposals. The diff is Docs +6, Other -429, for a net -423 across 3 files.
The in-loop context-overflow guard now estimates harness role sizes [3]. Previously, `installToolResultContextGuard` sized only `user`, `assistant`, and `toolResult` roles; harness roles like `bashExecution`, `compactionSummary`, `branchSummary`, and `custom` all returned 256 characters, undercounting summary- and bash-dominated context. The new estimator accounts for the full-text expansion `convertToLlm` performs.
Doctor added `core/doctor/memory-search` [4] as a structured lint finding source, reusing the existing memory-search note logic through an injected note sink so legacy `openclaw doctor` output stays unchanged. The check is opt-in, running only via `--only core/doctor/memory-search` or `--all`, not plain `doctor --lint`.
In acpx, development dependencies bumped across 7 packages [5], including `@vitejs/plugin-react` from 6.0.2 to 6.0.3, `@xyflow/react` from 12.11.0 to 12.11.1, and lint-staged from 17.0.7 to 17.0.8; @types/node moved from 25.9.3 to 26.0.1 [6].
Action items
- → Verify agent skills using MCP boundaries are on a build that includes the security warning openclaw/openclaw [immediate]
- → Remove reliance on init_skill.py for durable skill creation openclaw/openclaw [immediate]
- → Review openclaw doctor warns if context guard undercounts harness roles in your workloads openclaw/openclaw [plan]
- → Run openclaw doctor with --all to test memory-search lint findings openclaw/openclaw [monitor]
References
- [1] fix(security): warn on agent skill MCP boundary drift (#98352) ↗ openclaw/openclaw
- [2] fix: prevent skill-creator from bypassing workshop proposals (#98346) ↗ openclaw/openclaw
- [3] fix(agents): estimate harness role sizes in context guard char estimator (fixes #97927) ↗ openclaw/openclaw
- [4] doctor: add memory search lint findings ↗ openclaw/openclaw
- [5] chore(deps-dev): bump the development group across 1 directory with 7 updates ↗ openclaw/acpx
- [6] chore(deps-dev): bump @types/node from 25.9.3 to 26.0.1 ↗ openclaw/acpx
