agents-shipgate

agents-shipgate skill

agents-shipgate is the deterministic merge gate for AI-generated agent capability changes — a local-first, static Tool-Use Readiness review. It analyzes shipgate.yaml plus tool sources (MCP exports, OpenAPI specs, OpenAI Agents SDK Python files, Anthropic Messages API artifacts, Google ADK files, LangChain/LangGraph files, CrewAI files, OpenAI API artifacts, Codex repo config, Codex plugin packages and marketplaces, n8n workflow JSON, Conductor OSS workflow JSON) and emits deterministic verifier artifacts, findings, Markdown, JSON, and SARIF.

It does not run agents, call tools, invoke LLMs, connect to MCP servers, or send telemetry by default. Static analysis only; audited exceptions are pinned in tests/test_adapter_static_only.py::ALLOWED_EXCEPTIONS.

The skill name is intentionally agents-shipgate (not shipgate) so it does not collide with the /shipgate slash command shipped at .claude/commands/shipgate.md — Claude Code lets a skill with the same name preempt a command, which would bypass the bootstrap flow the slash command is meant to deliver.

When to use this skill

• The current diff adds or modifies MCP server or tool definitions, @tool/@function_tool decorators, OpenAPI specs describing agent tools, agent prompts, permission scopes, approval/confirmation policies, or agent CI workflows — run the verifier before reporting the change complete or creating a PR.
• The user asks to add Tool-Use Readiness or pre-merge checks to an agent project.
• The user asks whether an AI-generated agent PR can merge.
• The repo already has shipgate.yaml, agents-shipgate-reports/verifier.json, or agents-shipgate-reports/report.json.
• The user asks to fix, triage, suppress, or explain a Shipgate finding.
• The user wants to add Shipgate to CI (GitHub Actions, GitLab CI, CircleCI).

When NOT to use this skill

• Generic linting / type checking — use the project's existing tooling.
• Runtime monitoring, evals, or behavioral testing — Shipgate is static-only.
• LLM output quality assessment — out of scope.
• Editing agents-shipgate's own check implementations — that's upstream-repo work, not user-repo work.

How to act

Pick the matching task and follow the linked recipe verbatim. Recipes are bundled inside this skill so behavior is pinned to the installed version and works offline. Each prompt is self-contained: install commands, exit codes, and AGENTS_SHIPGATE_AGENT_MODE=1 error handling are in the prompt itself.

Always:

1. Set AGENTS_SHIPGATE_AGENT_MODE=1 so errors emit a next_action JSON line on stderr (auto-enabled inside Claude Code via the harness's CLAUDECODE=1 env var, and Cursor via CURSOR_TRACE_ID).
2. For local agent control, run shipgate check --agent claude-code --workspace . --format agent-boundary-json and read the stdout shipgate.agent_boundary_result/v1 object. Switch on control.state; follow only control.next_action, control.allowed_next_commands, and control.human_review. Treat decision as diagnostic context only.
3. For verifier runs, validate agents-shipgate-reports/verification-receipt.json first, then parse agent-handoff.json, verifier.json, and verify-run.json: control.state, merge_verdict, can_merge_without_human, control.next_action, fix_task, and capability_review.top_changes. Then parse agents-shipgate-reports/report.json.release_decision.decision; it is the release gate.
4. Before editing shipgate.yaml, Shipgate CI, AGENTS/CLAUDE/Cursor rules, policy packs, baselines, waivers, suppressions, Codex hooks/config, Codex plugin manifests, .mcp.json, .app.json, or SKILL.md, plan to run agents-shipgate verify before completion and route trust-root review to a human when the verifier requires it.
5. Before finishing an agent-related diff, run shipgate check --agent claude-code --workspace . --format agent-boundary-json. For committed PR/CI verification, run agents-shipgate verify --workspace . --config shipgate.yaml --base origin/main --head HEAD --ci-mode advisory --format json after making the base ref available. verify never fetches. For host grants, run shipgate audit --host --json --out agents-shipgate-reports/host-grants.json.
6. Do not bypass the verifier by suppressing findings, lowering severity, expanding baselines or waivers, removing Shipgate CI, or weakening agent instructions; verify-mode SHIP-VERIFY-* checks make those trust-root edits release-visible.
7. Confirm with the user before any command that writes files (init --write, baseline save).

First-time CI setup (advisory)

If the user has no Shipgate CI yet, default to advisory mode — never strict, never with a baseline. The promotion path comes later, only after findings have been reviewed.

1. Confirm the repo has shipgate.yaml and a clean local scan (agents-shipgate scan -c shipgate.yaml --ci-mode advisory exits 0). If not, run the bootstrap recipe first.
2. Create .github/workflows/agents-shipgate.yml from ci-recipes/advisory-pr-comment.yml. It runs on every pull request, posts a summary comment, uploads the report as an artifact, and never fails the job.
3. Confirm permissions: pull-requests: write is acceptable to the user before committing — required for the PR comment.
4. Push and open a test PR. Verify the agents-shipgate comment appears.
5. Stop here. Promotion to strict mode is a separate task — only run prompts/stabilize-strict-mode.md after the user has reviewed the advisory output and decided which findings they accept.

For non-GitHub CI (GitLab, CircleCI, Jenkins, Azure Pipelines, Buildkite, Bitbucket, pre-commit) refer to https://github.com/ThreeMoonsLab/agents-shipgate/tree/main/examples or docs/integrations.md in the upstream repo. Always start in advisory mode.

Stable contracts (rely on these)

• CLI surface follows the current 0.x contract line — see https://github.com/ThreeMoonsLab/agents-shipgate/blob/main/STABILITY.md.
• Installed CLI contract: when available, run agents-shipgate contract --json to verify local schema versions, capability/research surfaces, release_decision.decision, and manual-review signal fields. Older installs should use docs/agent-contract-current.md or upgrade before automating against the local contract command.
• Verifier JSON: verifier_schema_version: "0.5". Switch on control.state, then read merge_verdict, can_merge_without_human, control.next_action, fix_task, capability_review.top_changes, trust_root_touched, and policy_weakened before summarizing an AI-generated PR. merge_verdict is a deterministic projection; the gate remains report.json.release_decision.decision.
• Verification receipt: verification-receipt.json uses schema_version: "shipgate.verification_receipt/v1" and is written last. Validate it before trusting any projected verdict; it content-addresses the request, executor, unit result, decision, and complete artifact set.
• Verify run JSON: verify-run.json uses schema_version: "shipgate.verify_run/v3", embeds the content-addressed plan and executor, and binds unit-result and decision IDs. run_id is an exact compatibility alias of request_id; do not treat the run projection as a second gate.
• Report JSON: report_schema_version: "0.34". Read release_decision.decision first. A passed decision requires a complete root-reachable static binding graph plus complete, conflict-free identity, effect, and authority evidence for every reachable action; it does not prove runtime behavior. Preserve release_decision.static_analysis_only=true, runtime_behavior_verified=false, and static_verdict_disclaimer in summaries. Read release_decision.evidence_coverage.binding_coverage, semantic_coverage, identity_coverage, and policy_gap_count, then work every evidence_gaps[].next_action in order. Binding, semantic, and policy-applicability gaps are not Findings and cannot be suppressed, baselined, severity-overridden, cleared by --no-heuristics, or satisfied by human_ack; binding, effect, and authority declarations are human assertions and must never be auto-written. Use tool_catalog[] for diagnostics and tool_inventory[] for the proven reachable surface. The current schema is docs/report-schema.v0.34.json; v0.33 is a frozen compatibility reference. See the current agent contract, verification identity contract, and evidence-backed passed contract. These binding-backed fields require runtime contract v13; the unambiguous AgentControl projection requires runtime contract v14. A contract-v12 CLI emits the frozen v0.30 report and must not be described using the v0.31 binding claim.
• Release Evidence Packet: agents-shipgate-reports/packet.{md,json,html} (and packet.pdf with the [pdf] extras) is a supporting/provisional reviewer artifact. Packet v0.12 projects the verification request and decision IDs plus binding and semantic coverage; it never creates a second gate. See docs/packet-schema.v0.12.json and STABILITY.md §Release Evidence Packet.
• Capability standard: agents-shipgate capability export emits a stable static capability lock (capability_lock_schema_version: "0.6") and agents-shipgate capability diff emits a stable semantic diff (capability_lock_diff_schema_version: "0.7"). These artifacts implement capability standard v0.5, are supporting/provisional and non-gating, exclude runtime trace evidence, and are documented in docs/capability-standard.md.
• Governance benchmark: benchmark/agent-pr-governance/cases.yaml and scripts/run_governance_benchmark.py are the stable research benchmark substrate (governance_benchmark_result_schema_version: "0.2"), not a release gate. See docs/governance-benchmark.md.
• Single source of truth for the contract: docs/agent-contract-current.md. When the schema bumps, that file updates first.
• Exit codes: 0 pass, 2 config error, 3 parse error, 4 other error, 20 strict-mode gate failure.
• Check IDs (e.g. SHIP-POLICY-APPROVAL-MISSING) are stable; new ones may be added but existing ones will not be renamed or repurposed.

Boundaries (do not violate)

• Do not claim a finding is fixed without re-running agents-shipgate verify and reporting the new merge verdict and release decision.
• Do not silently suppress findings — checks.ignore requires a reason and the manifest validator rejects empty reasons.
• Do not commit agents-shipgate-reports/ — it's regenerated each run; add it to .gitignore.
• Do not run agents-shipgate baseline save until the user has reviewed the initial findings; baselining ratchets in noise.
• Do not enable strict CI as the first CI step. Always start advisory.
• Do not modify checks in agents-shipgate's own source — that's upstream repo work.
• Do not weaken Shipgate trust roots to make a verifier run pass. Policy, baseline, waiver, CI, trigger-catalog, and agent-instruction changes require human review.

If something errors out

Set AGENTS_SHIPGATE_AGENT_MODE=1 and re-run. The CLI appends a JSON line to stderr with {error, message, next_action}. Follow the next_action. The error kinds emitted by the current CLI:

For deeper troubleshooting see https://github.com/ThreeMoonsLab/agents-shipgate/blob/main/docs/troubleshooting.md.