Sponsio Agent Safety Setup

Sponsio — Agent Safety Lifecycle Companion

Sponsio is a Python/TypeScript runtime safety layer for LLM agents: it evaluates deterministic contracts against each tool call and can block (enforce) or just log (observe) violations. The engine is deterministic-only. This skill covers the full lifecycle — first-time setup, contract authoring/review, observe-mode tuning, and flipping to enforce — by orchestrating Sponsio's CLI and explaining its output in plain language.

This skill does NOT reimplement Sponsio's logic; it calls the CLI and interprets results.

When to use this skill

Dispatch by what the user is trying to do. Pick ONE workflow and follow it; do not run multiple workflows in one turn.

Do NOT trigger for: general LLM-safety discussions not tied to a specific codebase; non-agent code review (linting, correctness).

Prerequisites (run silently before any workflow)

sponsio --version

• Not found → install: pip install sponsio (or pip install -e ".[all]" from a local clone).
• For --llm inference, check: OPENAI_API_KEY / ANTHROPIC_API_KEY / GEMINI_API_KEY / GOOGLE_API_KEY. Absent → still proceed; AST-based extraction and all of W3/W4/W5 work with zero keys.

Editing contract YAML — write rules by file

Sponsio contract YAMLs split into two trust zones with different write rules. Pick the right zone before any edit; the runtime's self-modify pack enforces the host-zone rules, but cross-zone slips are still a config-correctness bug we'd rather avoid up-front.

Zone A — project YAML (you may add additively)

Path: <project>/sponsio.yaml — the file sponsio onboard writes into the user's repo. This evolves through every onboard / scan cycle. Adding new contracts via Edit (extending old_string) is the supported workflow.

Three legal write modes:

1. Add a new contract — Edit with new_string extending old_string (the invariant: old_string ⊆ new_string):

   old_string  = the verbatim tail of the existing file ending at
                 the last contract entry (or the last `contracts:`
                 line if the list is empty)
   new_string  = old_string + "\n      - <new contract YAML block>"

2. Tune an existing pack-shipped rule — never edit the rule directly; append a customized: entry:

   customized:
     - match: { desc: "<the shipped rule's desc, exact>" }
       A: "<extra assumption that narrows when it fires>"   # to relax
       # or args: [...]                                     # to retune thresholds
       # or disabled: true                                  # to silence (last resort)

3. Run sponsio scan for bulk additions from code / policy, merges additively and writes atomically:

   sponsio scan <paths> -o ./sponsio.yaml --append

Zone B — host bucket + plugin bundle YAMLs (user-only — never write directly)

Paths:

• ~/.sponsio/plugins/{{HOST_BUCKET}}/sponsio.yaml (this host's runtime library)
• ~/.sponsio/plugins/{{HOST_BUCKET_SUBAGENT}}/sponsio.yaml (this host's subagent library)
• ~/.sponsio/plugins/<plugin-id>/sponsio.yaml (per-plugin / per-MCP-server bundle — github, filesystem, my-plugin, …)

These files govern your own future tool calls. The runtime self-modify pack blocks every Edit / Write / MultiEdit you'd attempt against them — for the host bucket because rewriting your own rules is privilege escalation, and for the per-plugin bundles for the same reason (they constrain the plugin tools you'll call later). The {{HOST_BUCKET}} placeholders above are baked in at skill install time (_host_cursor, _host_claude_code, _host_openclaw, …) so each host only loses write access to its own bucket.

You must NOT use Edit, Write, MultiEdit, or Bash with shell redirects (>, >>, tee, sed -i, cp, mv, rm, dd) on any path under ~/.sponsio/plugins/. The legitimate update paths are:

• CLI (you can run via Bash):

  sponsio plugin install <name>            # copy a fresh bundled starter
  sponsio plugin scan --apply              # regenerate per-plugin bundle
  sponsio plugin show <name>               # surface what's loaded

• Hand-edit by the user in their text editor.

For customizations the user agreed to during a tuning conversation, do NOT ghostwrite the YAML they should paste. Contract content in Zone B has exactly four legitimate sources: bundle libraries (sponsio plugin install), CLI extraction (sponsio plugin scan, sponsio scan, sponsio onboard), the user's own keystrokes, and customized: blocks the user authors themselves. An LLM-composed snippet has none of those provenances — it's configuration with no audit trail.

The flow:

1. Restate the user's intent in plain English and identify the shipped rule it affects. sponsio plugin show <id> prints the desc: of every loaded rule; quote a desc verbatim from that output if you need to refer to one — do NOT compose YAML around it.
2. Tell the user the file path and describe the change in words ("add a customized: entry beside contracts: whose match.desc matches the rule you want to silence, with disabled: true"). Point them at the existing pack's syntax for reference. Let them write the YAML themselves.
3. When they say they're done, run sponsio validate --config ~/.sponsio/plugins/<id>/sponsio.yaml and help debug if it doesn't parse.

Forbidden write modes (universal)

• Write on any contract YAML above — overwrites the whole file in one go; bypasses the additive evidence even when the result happens to be a superset.
• Edit on a Zone-B path — denied by the self-modify pack regardless of additive intent.
• Edit on a Zone-A path where new_string does NOT contain old_string — a modification or deletion masquerading as an edit; treat as forbidden.
• MultiEdit on any contract YAML — same shape as Write.
• Bash with shell write operators (>, >>, tee, sed -i, cp, mv, rm, dd) targeting any of these paths — the runtime blocks Zone B; treat Zone A the same way.

Why this matters

The user's invariant: adding contracts is always allowed; modifying or deleting existing ones is not. Following this protocol makes that invariant easy to see at the diff level (old_string ⊆ new_string). When the user audits your edits later, "was anything removed or changed?" reduces to a string-containment check.

If you genuinely need to remove a rule

You don't. Use customized: ... disabled: true (an additive edit that silences the rule) or ask the user to delete it by hand. The agent never has authority to remove its own contracts.

Pattern generality — match operator intent, not demo data

When the operator's NL says "block public gists" or "cap files at 3", write the rule against the operator's literal intent. Do not infer file extensions, content types, naming conventions, or path structures from sample data, demo fixtures, or examples you've seen — unless the operator explicitly named them.

Concrete: a regex like (\.md"\s*:.*?){4,} matches only keys ending in .md, which means a 4-file gist of .json / .txt / .csv / extension-less keys passes freely. If the operator said "cap files at 3", the right form is (\".+?\"\s*:\s*\{){4,} — any 4+ keys. Same principle for path globs (/work/notes/.*\.md vs /work/notes/.+), arg_field_has value patterns, and match: selectors.

If the operator's intent IS demo-specific ("block any .md dump from this notes plugin"), keep the narrow pattern but record the assumption in the contract desc: so a later reviewer sees the constraint instead of inferring a bug.

The flip side — don't broaden past what the operator named

The operator's NL also has a verb scope. When a policy says "no destructive verbs against AWS" the agent must not emit \baws\s+ (matches every aws invocation including aws s3 ls / aws sts get-caller-identity). The regex must require both the provider and a destructive-verb token from the policy.

Concrete:

The shipped packs already follow this convention (see sponsio:incident/cursor-railway-wipe). When you extract from a policy doc, mirror the pack style: every command-shape regex is (host or tool prefix) AND (destructive verb the operator named). Bare provider prefixes are false-positive factories — every routine aws s3 ls in a CI script becomes a violation.

If the operator named providers exhaustively in prose ("Railway, Fly, Render, Supabase, Vercel, Cloudflare, Heroku, AWS, GCP"), keep all named providers, each anchored to its own verb set. Do not collapse them into a single broad alternation that drops the verb anchor.

W1 — Initial setup

Goal: from "project has no Sponsio" to "agent runs under observe mode with a sane contract file".

The CLI now has a one-shot wizard that covers the common path — detect framework + IDEs, ask which to install, dispatch sponsio onboard / sponsio host install / sponsio skill install accordingly, then verify. W1 is just orchestrating that wizard, not reimplementing it.

Steps

1. Run the wizard.

   sponsio init                                                    # interactive TTY
   sponsio init --apply 'framework=<name>;ides=<ide>:<level>,...;mode=observe'
                                                                   # non-interactive

   Picks string format:

   • framework=<one of langgraph / langchain / crewai / openai / anthropic / claude_agent / openai_agents / google_adk / vercel_ai / mcp / none>. none is "bare loop, generic guard.guard_before/after wiring"; an empty framework= skips the onboard step entirely.
   • ides=<ide>:<level>,... per-IDE pick. <level> is none / skill (drop SKILL.md only) / full (host hooks + SKILL.md, the canonical "protect this IDE" pick).
   • mode=observe|enforce. Default observe; enforce flips are W4's job.

   Use sponsio init --plan '<picks>' first to surface the exact commands the wizard will run, especially when you're non-interactive. Surface sponsio init's output to the user verbatim — the panel + preview + recap blocks are designed to be readable as-is.

2. Patch the agent entry file (only when framework ≠ "").

   sponsio onboard . --emit-context > /tmp/sponsio-onboard-context.json

   Read the JSON. Pick the entry file in this priority:

   1. entry_file_candidates has exactly one strong match.
   2. Else dedupe tool_inventory[*].filepath — one file, use it.
   3. Else stop and ask the user.

   If the file already imports from sponsio.<adapter> (Python) or from "@sponsio/sdk" (TS), it's already wrapped — skip. Otherwise splice wrap_snippet from the JSON: imports + guard construction at the top, wrap site adapted to the file's actual idiom (canonical create_react_agent(model, guard.wrap(tools)) if present, else adapt — e.g. tools_by_name = guard.wrap(TOOLS).tools_by_name for a name-keyed dispatch loop). Show the diff before writing.

3. Verify.

   sponsio doctor
   sponsio validate --config sponsio.yaml

   Surface every warning / fail line verbatim. If sponsio doctor FAILED (not warned, failed), stop here — don't let the user run their agent thinking the install is healthy when it isn't.

4. Explain observe mode explicitly: "Nothing is blocked on day 1. Every contract is still evaluated; violations are logged to ~/.sponsio/sessions/<agent_id>/*.jsonl. Use sponsio report --since 24h after a day of real traffic to see what would have been blocked. When you're ready to flip, that's W4."

Beyond W1 — pointers

Onboarding is just install + wire + verify. These are SEPARATE workflows, not extensions of W1:

• Author tighter contracts from a policy doc / threat model / actually-scanned codebase → W2 (audit & refine).
• Tune host-plugin libraries (Bash / Read / Write / MCP gating in this Claude Code or Cursor session) → invoke the sponsio-claude-code:configure skill (or the cursor analog). That skill owns sponsio plugin scan / sponsio plugin append / per-MCP-server library generation. W1 doesn't duplicate it.
• Move from observe to enforce → W4.
• Something doesn't fire / fires wrong → W5.

Choosing the write target — Zone A vs Zone B

sponsio init's axes already encode the destination decision:

• Axis 1 (framework wrap, picked at non-empty / non-none) → writes <project>/sponsio.yaml. This is Zone A — governs the LLM agent the user is wiring Sponsio INTO. You may Edit/Write this file; sponsio validate --config <path> after.
• Axis 2 (per-IDE level=full) → writes ~/.sponsio/plugins/_host_<ide>/sponsio.yaml via sponsio host install. This is Zone B — governs the HOST agent (i.e. you, Claude Code / Cursor). Edit/Write/ MultiEdit are denied by the self-modify pack; the only way to extend it from inside the host is sponsio plugin append.

If a user later asks you to add rules and the destination is ambiguous (project rules vs host-IDE rules), ask:

"These rules — should they govern (a) the LLM agent your project deploys (Zone A: <project>/sponsio.yaml), or (b) my own tool calls inside this IDE (Zone B: ~/.sponsio/plugins/_host_<ide>/ sponsio.yaml)?"

Two destinations need two files; conflating them silently breaks one or both.

Adding rules to Zone B (host bucket) — sponsio plugin append

When a Zone-B addition is the right call, you can't Edit/Write the host yaml directly. Use the append CLI:

# 1. Write the new rules to a staging file OUTSIDE Zone B (project
#    root is fine).  Never name it sponsio.yaml — collides with
#    project-config.
sponsio plugin append --from <staging-path> --target <bucket> --dry-run
sponsio plugin append --from <staging-path> --target <bucket>
rm <staging-path>     # delete after success

plugin append is structurally additive: it rejects anything beyond contracts: entries — no customized:, no disabled:, no desc collisions, no top-level keys. Tell the user up front you'll be running this on their behalf so they're not surprised when it appears.

Auto-selected packs

sponsio onboard uses simple, conservative heuristics:

sponsio packs lists all shipped packs with live rule counts and include specs.

Do NOT

• Do NOT edit sponsio/contracts/*.yaml inside the installed package — those are the shipped packs; they're read-only. Adjustments go in the user's sponsio.yaml via customized: or contracts:.
• Do NOT flip mode: enforce during W1. The whole point of observe mode is to find false positives before they break production.

W2 — Audit & refine (from scratch, or deepen an existing yaml)

Goal: produce or improve a sponsio.yaml from code / policy docs / traces, and explain every contract in plain language.

Routing rules between layers

A policy document can mix two kinds of rules that land in different YAMLs. Misrouting them (writing host rules to the project file, or agent rules to the host bucket) means the rule loads but never fires.

Classify each rule before writing:

Process:

1. Score each rule by the signals above.
2. Write each rule to its target YAML using the additive-only protocol in the previous section.
3. For genuinely ambiguous rules ("PII must not leak"), ask the user which agent the rule is about before picking a target.
4. After writing per-plugin rules, run sponsio plugin show <plugin-id> and surface the digest verbatim so the user sees what's now active.

The default failure mode is to dump everything into the file you were already editing. Cross-layer leakage is a worse user error than the extra clarification round.

Decide which sources to use

Sponsio contracts come from four sources, mixable in one yaml:

Match the user's input to the source(s):

• "Explain / review my sponsio.yaml" → source 1 and/or others already applied; jump to "Explain contracts" below.
• "Scan my agent code" → source 2, code-only.
• "We have a security policy document" → source 2, add --policy <doc> --llm.
• "I know the pattern I want but not the syntax" → source 4, then show them the yaml entry.

If ambiguous: ask ONE question — "(a) scan your code, or (b) extract from a policy document?"

Run scan (when extraction is needed)

You ARE the LLM. Use the agent-mediated path — Sponsio collects deterministic inputs, you do the cognition, Sponsio validates. No --llm / API key needed.

# 1. Sponsio dumps the deterministic inputs (AST tool inventory, policy
#    docs, existing yaml) as JSON:
sponsio scan <PATHS> --agent <AGENT> [--policy <doc>] --emit-context

# 2. Sponsio prints the contract-authoring prompt template:
sponsio prompt scan

Read both, apply the prompt to the JSON in your own context, and produce contract YAML entries. Source-tag every entry you author with source: agent-extracted so later tooling can distinguish them from heuristic rules.

Decide the write target by intent BEFORE writing, using the W1 step-4 (A) vs (B) test. Write path differs by destination:

If destination is project YAML — <project>/sponsio.yaml

Write via Edit/Write, then validate:

sponsio validate --config ./sponsio.yaml

If destination is host bucket — ~/.sponsio/plugins/{{HOST_BUCKET}}/sponsio.yaml

Edit / Write / MultiEdit on this path are denied by Zone B's self-modify pack. Use sponsio plugin append — a structurally-additive CLI that writes through validated Python:

1. Write the YAML you produced (above) to a transient staging file outside Zone B (e.g. <project>/.sponsio.staging.yaml or /tmp/sponsio-host-<timestamp>.yaml). Never name it sponsio.yaml — that collides with project-config and triggers the wrong-destination anti-pattern.

2. Dry-run the merge to confirm what would land:

   sponsio plugin append --from <staging-path> \
     --target {{HOST_BUCKET}} --dry-run

3. Apply. The CLI rejects anything beyond additive contracts: entries — no customized:, no disabled:, no desc collisions with existing rules, no top-level runtime: / include: smuggling. It validates the merged file via the loader before declaring success.

   sponsio plugin append --from <staging-path> --target {{HOST_BUCKET}}

4. Delete the staging file once plugin append reports success — keeping it around invites accidental re-apply.

This is the only way an in-host agent can extend its own governance file from inside the host. Direct shell redirects (cat >> ...) are blocked by the self-modify pack; the append CLI is the explicit, audited bypass.

Bare-CLI fallback (only when no host agent is driving — you're being run from a CI shell or one-shot script):

sponsio scan <PATHS> --policy <doc> --llm \
  -o ~/.sponsio/plugins/{{HOST_BUCKET}}/sponsio.yaml --append

--llm uses the configured API key. Skip this when you (the host agent) are already in the loop — plugin append is faster and free.

Anti-pattern (do NOT do this). Writing the contracts to <project>/sponsio.yaml and adding a comment block that says "Cursor's hooks read ~/.sponsio/plugins/_host_cursor/sponsio.yaml; copy this file there". That comment is an admission you wrote to the wrong place. Use the staging file + plugin append, not a fake project yaml.

Validate existing yaml (explain-only path)

sponsio validate --config ./sponsio.yaml --json

JSON shape per entry: {nl, ok, type: "det"|"unknown", pattern, formula, agent, section}. ok: false means the NL didn't match any pattern — surface those as "ambiguous — needs refinement".

Explain contracts (use this as the report template)

## Sponsio contract summary — <agent>

**Setup**: <N> contracts total, <sources breakdown>. Mode: <observe|enforce>.

### Active packs
- `sponsio:core/universal` (empty stub, kept so existing `include:` lines don't error)
- `sponsio:capability/shell` (11 det) — guards on <tool name after rename>
<...>

### Contracts explained
Each contract is an assumption → enforcement pair. State the assumption explicitly even when absent ("unconditional"), because the assumption defines **when** the enforcement applies.

- **When**: the agent has called `modify_order`
  **Then**: `get_order_details` must have happened earlier.
  **Runtime effect**: if unmet, the call is blocked (or logged, in observe) and the agent is told why.

- **When**: unconditional
  **Then**: `send_email` is called at most 5 times per session.
  **Runtime effect**: the 6th call is blocked.

### Ambiguous contracts (optional — only if validate flagged any)
- `"..."` — didn't match any pattern; rephrase or rewrite as a structured dict.

### Next steps
1. Review; prune anything that doesn't fit the agent's real job.
2. Run in observe mode for 1–2 days.
3. `sponsio report --since 24h` to see what would have been blocked (W3).
4. Prune false positives, then flip `mode: enforce` (W4).

Keep it dense. Don't pad with "consult an expert" disclaimers — the user is already using the expert.

W2b — Author plugin / MCP-server contracts

Goal: write a Sponsio contract library for the tools a Claude Code plugin or OpenClaw plugin (or a bare MCP server) exposes. This is the analogue of W2 but the input is a plugin manifest + MCP tools/list introspection rather than your own agent's source code.

You ARE the LLM — same agent-mediated pattern as W1 / W2.