[neekolas]

Summary

Add a shared, explicitly invoked security-hardening skill for full-repo audits. It guides a host agent from a reviewed prompt matrix through sandboxed discovery, deduplication, validation, user triage, and a remediation plan. Fixes remain in separate sessions.

The skill uses the existing dev/sbx VM. A host helper reads selected API keys from a permission-restricted .env and registers them through Docker's credential proxy without putting keys in command arguments or guest files. A guest helper launches reviewed Claude and Codex CLI lanes in separate worktrees with bounded concurrency and durable status files.

There are optional user pauses after validation and during document triage for cleanup or adversarial review by other models. The prompt pack and output records are included for review.

Validation

• dev/nix-shell 'python3.11 -B dev/sbx/test_sandbox.py && python3.11 -B dev/sbx/test_provision.py && python3.11 -B dev/sbx/test_guest_lanes.py'
• Ruff on the new Python files
• Skill quick validation and Markdown lint

The guest runner was tested with fake Claude and Codex commands in temporary worktrees. No live audit or credential registration was run.

Note

Add shared security-hardening audit skill with lane runner and credential provisioning

• Adds an explicit-invocation skill that runs a user-reviewed, sandboxed multi-model security audit in the dev/sbx VM: parallel discovery, deduplication, validation, triage, and a remediation plan. Fixes are not implemented in the same run.
• Adds guest_lanes.py: validates lane manifests, creates one detached Git worktree per lane, and runs Claude or Codex lanes up to a configured concurrency limit. A lane is marked done only when the process exits successfully and the expected final CLI event is present.
• Adds provision.py: loads an owner-only dotenv file and registers selected keys with the sandbox proxy via standard input, so key values never appear in command arguments.
• Adds reference guides for lane running, prompts, and run records (lanes.md, prompts-and-records.md), plus tests for both scripts and a rule in AGENTS.md routing full-repository audit requests to the skill.
• Behavioral Change: the Claude lane uses bypass-permissions mode and the Codex lane uses workspace-write with network access, per the runner in run_lane.

^{Macroscope summarized 6b438cc. (Automatic summaries will resume when PR exits draft mode or review begins).}