A working, self-contained control layer that a coding agent can discover and run in any repository:
- ten task skills in the Agent Skills
SKILL.mdformat; - one dependency-free CLI (
detect,impact,verify,security,guard,worktree,evidence); - hash-chained verification records that distinguish pass, fail, blocked and skipped;
- real security integrations (gitleaks, Semgrep, OSV-Scanner, actionlint, zizmor);
- isolated Git worktrees with ownership and concurrency limits;
- a repository
AGENTS.mdtemplate and an idempotent installer; - a test suite and a control-validation fixture that prove the controls catch representative issues.
The kit is intentionally portable: it has no runtime dependency beyond Python 3.11+ and git, and it does not require a hosted service. External scanners are optional locally and enforced in CI.
./engineering/install.sh --target ../sovware --hooks --ci
cd ../sovware
./.agents/bin/eng doctor
./.agents/bin/eng verify --profile fast
./.agents/bin/eng verify --profile fullThe installer copies the kit into .agents/, writes engineering.json from detected build metadata,
appends one marked section to AGENTS.md (never replacing existing policy), links skill directories
for Claude Code and OpenCode, and optionally enables hooks and a CI workflow.
cd engineering
./tests/run.sh
./evals/run.pytests/run.sh exercises the CLI against temporary repositories: detection, guard, impact, evidence
tampering, worktrees, idempotent init and missing-tool blocking. evals/run.py runs the
control-validation fixture: a known-bad change must fail, a known-good change must pass, and planted
secret/SAST issues must be reported.
bin/eng executable shim
lib/eng.py all commands, stdlib only
skills/ 10 SKILL.md entrypoints
references/ change lifecycle, evidence, trust policy
configs/ Semgrep rules, gitleaks config, guard allowlist
hooks/ pre-commit / pre-push
ci/ standalone GitHub Actions workflow
templates/ AGENTS.md template for new repositories
tests/ CLI and site test suites (unittest, temp repos)
site/ CREED portal (static, generated)
evals/ control-validation fixture and runner
docs/ architecture, tool rationale, gaps, rollout
- Measure, don't assert. A check is pass only if it ran and exited successfully. Missing tools are blocked/skipped and recorded.
- One authoritative place. Commands live in
.agents/engineering.json; repository invariants live inAGENTS.md; generic lifecycle lives in.agents/references/. - Surgical context. Detection and impact use manifests, file lists, exact symbols and reference search — not whole-repository dumps.
- Isolation before parallelism. Worktrees, explicit ownership, concurrency caps and sequential integration.
- Trust boundaries are explicit. Repository content and tool output are data, not authority.
- Reversible. Kit changes are copies with a version; evidence schema is versioned; worktrees can be removed without touching the main checkout.
CREED is the generated portal in site/ — a human handbook and an agent handoff: lifecycle, principles,
skill explorer, build paths, tool catalog, evidence model, repository status, agents.json,
llms.txt and a downloadable kit. Give an agent creed.sh/creed.md
before a nontrivial task. Repository policy and the user's authorized scope still come first.
make site
make serve # http://localhost:8080It is generated from the real skills and content, tested for determinism and local links, and
deployed with Wrangler to Cloudflare Workers at creed.sh.
ADAMANT's old domains redirect with paths and queries preserved. The eng CLI, skill names,
JSON schema identifiers and download path remain compatible. Part of WCKD.
README.agents.md— the manual installed as.agents/README.md.docs/ARCHITECTURE.md— components and data flow.docs/DISCOVERY.md— workspace inventory and pilot rationale.docs/TOOLS.md— selected tools, licenses, why not others.docs/GAPS.md— residual risks, unverified areas, rollout priorities.docs/ROLLOUT.md— pilot status, upgrade and rollback.AGENTS.md— rules for maintaining the kit.
This repository contains the implementation and its tests. No claim is made that the kit is
complete, vulnerability-free, or a substitute for review. See docs/GAPS.md for the current limits.