secretzero-agent

name	secretzero-agent
description	Use for agentic and operational SecretZero workflows including unified `agent sync`, CLI/API/MCP parity, secure human-in-the-loop vectors, and automation-safe run loops.
license	Apache-2.0
compatibility	Requires SecretZero CLI (Python 3.12+). Set SZ_AGENT_MODE=true on agent hosts. MCP extra for sz_* tools.
metadata	{"author":"secretzero","version":"1.1"}
allowed-tools	Bash(secretzero:*) Read

SecretZero Agent Skill

Use this skill when running SecretZero in agentic workflows (manual-assisted, web-assisted, or fully automated), or when guiding users through runtime usage scenarios.

Install / Verify

Preferred install:

uv tool install -U "secretzero[all]"

Lean or provider-specific installs:

uv tool install -U secretzero
uv tool install -U "secretzero[aws]"
uv tool install -U "secretzero[azure]"

Alternative:

pip install -U "secretzero[all]"

Verify CLI:

secretzero --help
secretzero agent sync --help

MCP host (optional)

When the agent host exposes SecretZero MCP tools (secretzero-mcp):

uv tool install -U "secretzero[mcp]"
secretzero mcp config generate --workspace /path/to/repo --format cursor

See docs/mcp-setup.md. The MCP server sets SZ_AGENT_MODE=true by default (same spill contract as skills/secretzero-handle/SKILL.md). Prefer secretzero mcp serve; secretzero-mcp remains a deprecated wrapper.

Core Agent Contract

Absolute rule: skills/secretzero/SKILL.md — never consume secrets in context; use SecretZero for all handling (agent sync, --web, validate, status). Never request, receive, or print plaintext secret values.

• Prefer JSON output for machine handling.
• When running in automation or spill-sensitive hosts, set SZ_AGENT_MODE=true so the CLI blocks or redacts commands that would dump secret-bearing config (see skills/secretzero-handle/SKILL.md).
• Use the unified entrypoint:

secretzero agent sync --json [--web] [--dry-run] [--verbose]

Three Usage Vectors

1. Human-instructed (Vector 1)

   • Run secretzero agent sync --json.
   • Relay pending_secrets[].summary and ordered steps exactly.
   • Re-run until clean.

2. Secure local web capture (Vector 2)

   • Run secretzero agent sync --web (add --json only when you also need structured stdout after the run finishes).
   • Do not block the whole agent session on the foreground TTY if your environment allows it: start the command in a background terminal (or equivalent job control) on the operator’s machine so the process can stay alive until the form is submitted, and capture stdout if you need to quote the URL line back reliably.
   • When the CLI prints the localhost URL, relay the complete URL verbatim to the user (scheme, host, port, and path — no truncation). Vector 2 binds to 127.0.0.1; the printed URL is the handoff surface (this is not the same as secretzero web, which uses a separate bootstrap token for the network dashboard — if you use that command, pass through its full login URL including the token per that command’s output).
   • Remind the operator not to paste secret values into chat; only the URL and UI instructions belong in the agent thread.
   • Remind the operator to submit the form once and close the browser tab when done: the CLI helper stops the temporary localhost server after a successful submit; if they walk away, Ctrl+C (or waiting for timeout) ends the wait.
   • API Vector 2: POST /agent/sync with web: true returns web_url and web_session_id — give the operator the exact web_url, poll GET /agent/sync/web/{web_session_id} until done, and treat the session id as sensitive correlation data. The API process may keep that localhost port open until restart; do not forward web_url beyond the trusted operator.
   • Poll/re-run agent sync --json until pending_secrets clears.

3. Fully automated (Vector 3)

   • Ensure provider auth is available (and optionally SZ_AGENT=true).
   • Run secretzero agent sync --json.
   • Handle failed_secrets with manifest fixes and retry.

Agent runtime integration

When bootstrapping or restoring Hermes/OpenClaw installs:

secretzero agent list --format json
secretzero agent adopt --dry-run --format json
secretzero agent adopt --preseed-lockfile --format json

Load skills/secretzero-agent-adopt/SKILL.md for the full adopt/restore playbook. agent backup is an alias of agent adopt (not secretzero backup create).

Standard Agent Loop

1. Run secretzero agent sync --json first; escalate to --web when you need the localhost form (see Vector 2 handoff above).
2. Parse status and pending_secrets/failed_secrets.
3. Execute the appropriate vector behavior.
4. Re-run command until both arrays are empty.
5. Continue downstream only after clean completion.

API Parity

Use API when running remote orchestration:

• POST /agent/sync with { dry_run, web, lockfile?, sz_agent? }
• For Vector 2 polling: GET /agent/sync/web/{session_id}

Treat API payload semantics the same as CLI semantics.

MCP operational tools

When MCP tools are available, prefer them for metadata-only maintenance (hashes, counts, drift — never values):

MCP tool	CLI parity	Use for
sz_status	secretzero status --format json	Preflight, lockfile hashes, per-target sync state
sz_sync	secretzero sync --format json	Reconcile targets (non-interactive; prompts disabled)
sz_rotate	secretzero rotate --format json	Rotation policy execution
sz_drift_check	secretzero drift --format json	External target drift
sz_discover	secretzero discover --format json	Credential inventory during authoring (see secretzero-author)

Pass environment and var_files on MCP tools for multi-environment lanes (same semantics as CLI --environment / --var-file). Set SZ_WORKSPACE (or tool workspace) to the repo root.

MCP gap — bootstrap / human seeding: MCP does not expose agent sync (Vectors 1–3) or --web. For pending manual secrets, first-time bootstrap, or Vector 2 localhost forms, use:

secretzero agent sync --json [--web]

or API POST /agent/sync with { "web": true } when remote.

Agent loop with MCP: run agent sync --json until pending_secrets / failed_secrets are empty; then use sz_status / sz_sync / sz_drift_check for ongoing operations.

Operational Playbooks

• Agent adopt (Hermes/OpenClaw): see skills/secretzero-agent-adopt/SKILL.md — agent list → agent adopt → agent sync
• Bootstrap: validate -> init --install -> test -> agent sync --json
• Preflight: sz_status or secretzero status --format json; dry-run via sz_sync (dry_run: true) or secretzero sync --dry-run
• Maintenance: sz_rotate / sz_drift_check or CLI secretzero rotate, secretzero drift

Common Failure Handling

• Missing extras/provider dependencies -> run install/secretzero init --install
• Auth missing/expired -> fix provider auth, re-run secretzero test
• Manual-seeded secret lacks instructions -> add agent_instructions, retry
• Variable interpolation issues -> check vars and --var-file usage

Definition of Done

• Secret bootstrap/sync flow reaches no pending or failed secrets.
• Workflow used secure vector handling with no secret leakage.
• CLI/API/MCP behavior stays consistent for the selected scenario.