Hermes Agent is a self-improving AI agent runtime with tools, persistent knowledge, reusable skills, scheduled automation, and support for multiple model providers.
MiniCISO builds on that runtime as an agentic security staff, distributed as a reproducible overlay of profiles, prompts, templates, and operating policies.
This repository is not a Hermes fork. MiniCISO is a public overlay installed on top of a Hermes runtime pinned by version and commit in
config/hermes-version.env.
MiniCISO v0.7.0 adds a cost and context governance layer with:
disabled,observe, andenforcemodes;- budgets for each agent and delegated task, with a protected QA reserve;
- a
reserve → persisted dispatch → reconcile/finalizelifecycle and an accounting latch that fails closed when accounting cannot be trusted; - limited compaction per pressure episode, with re-arming after recovery;
- identity tracking from
profile → engagement → task → logical call → attempt → request; - human authorization before additional LLM calls. Telegram approval callbacks are opaque, single-use, and do not call the LLM to process approval or denial;
- artifact minimization and profile-scoped authorization.
Headroom Phase 1's output optimizer remains in shadow mode. Headroom Phase 2 is not part of v0.7.0. See the governance design and the governance skill.
MiniCISO v0.7.0 uses the icidade/hermes-agent fork at commit 489c6f2103ccca0ac1fc4f6249c71924ec8f024c. This SHA, not a floating branch, is the authority for the runtime. MiniCISO remains an overlay: there is no local patchset and no vendored Hermes code.
MiniCISO packages a reusable security operating model around Hermes:
- a
chief-of-staffcoordinator for intake, routing, synthesis, QA enforcement, and institutional retrieval of lessons learned; - specialized security SMEs for threat modeling, architecture, code review, AppSec, compliance, offensive validation, recon, and QA;
- bundled MiniCISO skills for KAG finding gating, Headroom Phase 1 selective retrieval, and institutional learning;
- evidence-driven workflows with explicit gates for finding validation, lessons-learned reuse, and post-submission follow-up.
It is designed to help a human operator run structured security engagements more consistently. It does not replace human authorization, judgment, or accountable decision-making.
git clone https://github.com/icidade/miniCISO.git
cd miniCISO
.\scripts\bootstrap.ps1git clone https://github.com/icidade/miniCISO.git
cd miniCISO
./scripts/bootstrap.shThe bootstrap restores the pinned Hermes runtime, creates the MiniCISO profile set, installs the overlay prompts, bundled chief-of-staff skills, and templates, prepares the shared workspace, and runs structural checks.
To update an existing installation to v0.7.0, use the platform command in INSTALL.md without --skip-hermes-install / -SkipHermesInstall, so the pinned runtime is installed and verified. The bootstrap checks provenance and fails closed when the runtime diverges. See INSTALL.md for provider-setup options and rollback. After deployment, restart existing sessions and gateways; a new session is recommended after the restart.
git pull --ff-only
.\scripts\bootstrap.ps1 -SkipProviderSetupgit pull --ff-only
./scripts/bootstrap.sh --skip-provider-setupCredentials requested by hermes setup stay in the user's Hermes environment. They are never copied into this repository.
MiniCISO ships the following public profile set:
chief-of-staffsecurity-threat-modelingsecurity-architecturesecurity-code-reviewsecurity-appsec-assessmentsecurity-compliance-mappersecurity-offensive-securitysecurity-recon-attack-surface-strategistsecurity-qa
For responsibilities, handoffs, and usage examples, see the wiki staff guide and the canonical profile contract in docs/profile-setup.md.
Bundled chief-of-staff skills installed by bootstrap:
miniciso-kag-finding-gateminiciso-headroom-phase1miniciso-institutional-learning
Offline repository validation:
.\\scripts\\validate-repo.ps1./scripts/validate-repo.shThe validation pipeline now includes a bundled-capability contract check so future changes cannot silently drop the distributed chief-of-staff skills, their SOUL references, or their bootstrap/sync wiring.
Runtime smoke test after bootstrap:
.\scripts\smoke-test.ps1./scripts/smoke-test.shUse -Online / --online only when you want to send a real question to each profile using the configured provider.
This repository contains only non-secret configuration and sanitized public content. Do not publish .env files, tokens, sessions, memories, logs, real reports, or unsanitized evidence. See SECURITY.md and .env.example.
For external vulnerability or bug bounty work, report drafting starts only after a GO decision. RESEARCH means produce an impact-validation plan, and NO-GO means block submission and capture the lesson learned.
MiniCISO can also evaluate prior operational experience through institutional retrieval of lessons learned before and during analysis, so repeated methodological errors are less likely to survive into a final claim.
Canonical sources:
docs/kag-finding-validation.mddocs/institutional-learning-capability.mdtemplates/finding-decision-template.mddocs/submission-followup.mdtemplates/submission-followup-template.md
MiniCISO can maintain a public, sanitized copy of selected runtime-safe artifacts without publishing private state. See docs/self-update-capability.md.
- Dependencies and configuration
- Cost & Context Governance design
- Safe self-update capability
- GitHub PR access from the VPS
MIT. Hermes Agent is a separate project and keeps its own license.