Skip to content

fix: enforce tool approval and isolate legacy ownership - #133

Merged
alimobrem merged 6 commits into
mainfrom
codex/agent-deep-safety
Oct 3, 2026
Merged

alimobrem merged 6 commits into
mainfrom
codex/agent-deep-safety

Conversation

@alimobrem

@alimobrem alimobrem commented Oct 3, 2026 •

Copy link
Copy Markdown
Collaborator

Enforce mutation approval consistently across chat, dynamically registered MCP tools, investigation plans, and Temporal activities. Registry write classification cannot be removed by a caller's empty write set. Unknown MCP tools require confirmation, read-only skills exclude them, and refresh/disconnect removes only tools owned by that connection. MCP and durable workflow administration require administrator identity; approval payloads must contain actual JSON booleans.

Legacy view migration now moves only the owner derived from the current caller's forwarded token. It never claims another user's hash-based views, and commits zero-row updates to release database locks. Plans preserve configured write gates and caller context; unattended investigations cannot silently run write tools. Temporal phases gain write permission only through their explicit approval gate and server trust, with workflow-history compatibility preserved.

Refresh repository Markdown against current source, including setup, database, API, security, testing, architecture, release and model-facing guidance. Label historical plans and distinguish dry-run validation from live behavioral gates.

Validation: final focused safety regressions cover real chat/monitor WebSocket malformed approvals, PostgreSQL transaction release, ownership isolation, MCP refresh/collisions and classification, and plan/Temporal write gates. Full suite: 3,441 tests passed with exit 0 using an isolated PostgreSQL database and inert kubeconfig; 34 focused safety cases and 49/49 replay dry runs passed. Rebase onto merged main preserved the tested file tree. Ruff, Mypy and documentation checks passed. Real OpenShift acceptance remains pending.

CI also fetches full history and compares the explicit PR base SHA for changes under sre_agent, including model-facing references. Git errors fail detection instead of silently skipping live evaluations.

Merge gate: commit4eb46ef is published and run37101198048 is validating the live judged replay and SRE-Bench integrity checks. Await these results before merging. SRE-Bench resolution scores remain advisory under the existing workflow; passing integrity checks does not establish every scenario was resolved.

@alimobrem
alimobrem merged commit d3025cb into main Oct 3, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants