ShieldCortex 5.8.0: Recall That Answers What You Asked, Action Guard for Claude SDK Browser and Computer Use, Recall Repetition Check
ShieldCortex 5.8.0 was released on 12 October 2026. Prompt recall now uses what you asked, rather than the OpenClaw 2026.9.x wrapper around your message. A new Claude SDK toolsets export can observe browser and computer use, and doctor can flag repeated recall results. This is a minor release on 5.7.0.
Recall starts with your message
The Claude Code UserPromptSubmit hook previously included the OpenClaw 2026.9.x message envelope in its search query. That could keep returning memories matched to the wrapper and earlier replies instead of the owner's new question. The hook now strips the envelope before querying. It strips only a complete envelope; a message that merely mentions its marker, has an unfinished block, or has no envelope remains unchanged.
There is an optional vector recall plane. It is off by default; set "recallVector": true in the config or SHIELDCORTEX_RECALL_VECTOR=1 to turn it on. It re-ranks and tops up full-text search candidates, but never admits a memory that would fail the same full-text search match and admission rules. It cannot retrieve a memory that shares no search term with the prompt.
shieldcortex doctor now has a read-only Recall repetition row. It warns when recent interactive recalls repeatedly return the same candidate set. This is evidence of repetition, not a diagnosis of its cause; repeated questions can produce it too. The check does not change recall or turn on vector search.
Observe Claude SDK browser and computer calls
The new shieldcortex/toolsets export puts Action Guard in front of the Claude SDK browser_toolset_20260801 and computer_toolset_20260801 toolsets. It classifies calls, scans page text and emits values-free audit events. Its default mode: 'observe' changes nothing about what runs: the host's answer and the driver's result pass through unchanged. This entry was mistakenly listed under 5.7.0; the export first ships in 5.8.0.
Opt in with mode: 'enforce' to refuse the deny set and hold approval-required calls. Enforce mode needs the host to supply its own approver, the inner callback: without one, every held call is refused. When the host does supply one, a held call is passed to it with a readable card without typed values, page content, URL paths or queries, and runs only if that approver answers yes in time. Approval is single-use and has a deadline; a second identical call must ask again. The approval binds to the host in the SDK's last state report, not the live page. Before the first read, a typed newline in the browser toolset is still allowed.
Action Guard stays off by default in the shared ShieldCortex config. The toolsets export has its own observe-only default and requires an explicit enforce mode to hold calls.
Clearer setup and threat-graph errors
status, doctor, help and the README now point to shieldcortex setup for onboarding and shieldcortex setup claude for Claude Code repair. This is wording only; install and quickstart keep working.
Threat-graph realtime projection now reports an unreadable listed day file and holds its cursor there. A malformed stored cursor halts projection with a fixed error instead of stalling or replaying. This changes behavior only for broken state. Recovery is an operator action: restore a canonical cursor or run shieldcortex threat-graph rebuild, which replays only files still present.
Release tooling: the ClawHub sync step now reports a version still in ClawHub's security scans as submitted, which proves neither a failed upload nor public availability. A failed ClawHub read is unknown rather than absent, and only a version ClawHub actually serves counts as success.
Upgrade to 5.8.0
The supported Node range stays ^22.14.0 || >=24.0.0. No existing export, CLI command, config default, Node floor or Cloud pin changes.
npm install -g shieldcortex@5.8.0Built by Drakon Systems.