TITMAS Agent Action Gate

Evidence-verified, deterministic authorization boundaries for AgentTeams workflows.

中文：面向 AgentTeams 多智能体工作流的证据验证与确定性行动闸门。

TITMAS Agent Action Gate is a production-oriented reference architecture and competition demo candidate for the GOAI 2026 Agent Infra track. It separates uncertain agent analysis from deterministic authorization, evidence verification, policy evaluation, and human approval.

Current milestone: M4_RUNTIME_BLOCKER_CLOSURE_IN_PROGRESS; the bounded Alibaba Cloud official Skill sub-milestone is COMPLETE, while full M4 remains INCOMPLETE.

The deterministic Action Gate, append-only state store, pinned agent-evidence adapter, authenticated native MCP boundary, six-role target topology, and allowlisted in-memory GitHub provider adapter are implemented in the current worktree. A bounded public sandbox run previously created a branch and Draft PR. On 2026-08-02, a historical isolated deployment of official AgentTeams v1.2.0 started one Manager and five Workers; it remains operator-supervised negative evidence. A separate disposable cloud-context-inspector ran one native Qwen Worker turn: it resolved the externally installed official Alibaba Cloud Resource Center Skill, verified the source-lock digest, invoked the frozen typed read-only adapter, and returned EMPTY_RESULT as NOT_ASSESSED_NO_VISIBLE_RESOURCE. The retained chains and canonical agent-evidence receipt validate, while Worker decision records and Resource Center write calls remain zero. This proves only the bounded specialist turn, not a complete cloud inventory, broader autonomous M4 completion, or deployment authorization.

This repository is not submitted to, endorsed by, or affiliated with GOAI, and it makes no certification, compliance, production-readiness, or security guarantee.

Why this exists

Agents are useful at interpreting ambiguous requests, decomposing work, and explaining uncertainty. They should not be the component that silently grants their own authority. This project uses:

• AgentTeams v1.2.0 for transparent Manager/Leader/Worker collaboration;

• agent-evidence 0.6.0 as the canonical evidence packaging and verification dependency;

• versioned JSON contracts and a deterministic Action Gate for ALLOW, BLOCK, and REQUIRE_APPROVAL;

• a human approval record for scoped, high-risk actions;

• provider MCP servers, such as GitHub MCP, only after an ALLOW decision.

Related MCP server: gov-mcp

AgentTeams team

Agent identities and intended tool boundaries are machine-readable in agents/registry.json. The reviewable deployment template is deploy/agentteams/team.v1.2.0.yaml; the non-idempotent macOS Docker Desktop smoke profile is deploy/agentteams/team.native-smoke.v1.2.0.yaml.

Native local smoke boundary

The retained machine-readable evidence is demo/evidence/agentteams-native-20260802.json. It records both the verified chain and the failures that prevent a stronger claim:

• Qwen qwen3.8-max-preview specialist Workers invoked the real six-tool MCP endpoint; preview model availability is not a stable runtime contract;

• agent-evidence 0.6.0 returned VALID, after which the deterministic gate returned a five-minute ALLOW that expired without execution;

• the leader did not complete the workflow autonomously, one unrelated request entered the global store during concurrent prompts, and github-operator called a tool outside its declared registry allowlist;

• all Workers shared the same MCP endpoint, so prompts described role boundaries but the smoke did not enforce per-Worker tool ACLs;

• repository Skill names were declared in resources, but the run did not independently prove that those Skill packages were materialized inside the Workers.

This is native local orchestration evidence, not a persistent deployment, autonomous-workflow proof, least-privilege proof, or production-readiness evidence.

Deterministic decisions

The decision contract and precedence rules are in specs/action-gate-decision-v0.1.md.

GitHub demo path

Agent request
  -> request analysis
  -> pre-action evidence verification
  -> deterministic Action Gate
  -> exact GitHub action after ALLOW
  -> official Alibaba Cloud read-only context preflight before a deployment-related release request is evaluated
  -> post-action evidence generation
  -> agent-evidence verification
  -> deterministic release decision
  -> human approval when required

The end-to-end scenario and retained public evidence are documented in docs/GITHUB-WORKFLOW-DEMO.md. The repository contains four reproducible runtime cases: valid execution, missing evidence, tampered evidence, and a high-risk release action requiring approval.

The historical adapter-only Alibaba Cloud evidence is demo/evidence/alibabacloud-resourcecenter-preflight-20260802.json. The later native Worker-turn evidence is demo/evidence/agentteams-native-alibabacloud-skill-20260802.json. The frozen four-file evidence set is demo/evidence/alibabacloud-evidence-set-freeze-20260802.json. Together they retain exact external Skill source verification, pinned CLI/plugin digests, same-profile live STS identity binding, the complete one-policy RAM attachment set, sanitized invocation trace, CLI exit 0, native AgentTeams Worker and Matrix receipts, replayable VALID agent-evidence, and scoped zero-write accounting. The Worker ZIP contains reference metadata but no upstream Skill bytes.

Run and validate

Python 3.11 or newer is required.

python3 -m pip install -e '.[dev]'
python3 scripts/validate_milestone.py
python3 scripts/validate_governance.py
python3 scripts/validate_alibabacloud_runtime_evidence.py
python3 scripts/validate_alibabacloud_evidence_set.py
python3 scripts/validate_native_agentteams_cloud_skill_evidence.py \
  demo/evidence/agentteams-native-alibabacloud-skill-20260802.json
python3 -m unittest discover -s tests -v
python3 -m titmas_action_gate.cli evaluate-fixtures
python3 -m titmas_action_gate.cli demo --state-dir artifacts/runtime/local-demo
python3 -m titmas_action_gate.cli validate-install

Run a future real Alibaba Cloud preflight only through the runner's internal same-run RAM readback. It generates an unpredictable run ID and atomically reserves the evidence path before any provider call. It does not accept an external observation or credential bytes:

python3 scripts/run_alibabacloud_skill_evaluation.py \
  --control-profile '<RAM-readback-profile-label>' \
  --profile '<read-only-profile-label>' \
  --role-name '<read-only-role-label>' \
  --output '<new-evidence-path>' \
  --confirmation-ref '<explicit-user-confirmation-reference>'

The tests execute the deterministic engine, pinned agent-evidence validator, append-only chain, MCP stdio protocol, all six tools, AgentTeams-compatible local handoffs, in-memory provider workflow, native-smoke manifest/evidence checks, and negative boundaries. They do not prove persistent AgentTeams deployment, autonomous orchestration, production security, or operational readiness.

Start the MCP server over stdio:

TITMAS_ACTION_GATE_STATE_DIR='artifacts/runtime/mcp' \
TITMAS_ACTION_GATE_CALLER_TOKEN='replace-with-agent-token' \
TITMAS_ACTION_GATE_APPROVER_TOKEN='replace-with-distinct-approver-token' \
TITMAS_ACTION_GATE_DEMO_MODE='true' \
TITMAS_ACTION_GATE_MCP_TRANSPORT='stdio' \
  titmas-action-gate-mcp

The real GitHub runner requires a separately provisioned sandbox repository and exact local worktree. It is intentionally not part of default CI. See docs/RUNBOOK.md.

Repository map

• architecture/README.md: system boundaries and data flow;

• docs/AGENTTEAMS-INTEGRATION-PLAN.md: AgentTeams v1.2.0 integration;

• docs/SKILL-SPECIFICATION-v0.1.md: reusable Skill package contract;

• docs/MCP-TOOL-SPECIFICATION-v0.1.md: MCP server and tool contract;

• docs/THREAT-MODEL-v0.1.md: trust boundaries and bypass threats;

• docs/ALIBABA-CLOUD-SKILL-INTEGRATION.md: exact official Skill source, typed read-only boundary, credential isolation, evidence path, and current exit status;

• docs/EXECUTION-ROADMAP.md: implementation and evaluation gates;

• docs/RUNBOOK.md: reproducible local and sandbox commands;

• SECURITY.md: security model, reporting, and non-production limitations;

• evaluations/: deterministic contract cases;

• governance/: DBA source lock, existence declaration, recommendation, and management boundary.

Truth boundaries

AGENTTEAMS_ORCHESTRATION_NE_ACTION_AUTHORITY=true
AGENT_ANALYSIS_NE_POLICY_DECISION=true
EVIDENCE_NE_TRUTH=true
EVIDENCE_VERIFICATION_NE_ACTION_AUTHORIZATION=true
ALLOW_NE_EXECUTION_SUCCESS=true
MCP_TOOL_AVAILABILITY_NE_PERMISSION=true
SPECIFICATION_NE_IMPLEMENTATION=true
TEST_PASS_NE_PRODUCTION_READINESS=true
LOCAL_HANDOFF_HARNESS_NE_NATIVE_AGENTTEAMS_RUNTIME=true
NATIVE_LOCAL_SMOKE_NE_PERSISTENT_OR_PRODUCTION_DEPLOYMENT=true
OPERATOR_SUPERVISED_NE_AUTONOMOUS_END_TO_END=true
HASH_CHAIN_VALID_NE_SEMANTIC_ORCHESTRATION_CLEAN=true
PROMPT_ROLE_BOUNDARY_NE_ENFORCED_PER_WORKER_ACL=true
CLOUD_CONTEXT_NE_DEPLOYMENT_AUTHORIZATION=true
CLOUD_READ_SUCCESS_NE_COMPLETE_INVENTORY_OR_READ_ONLY_POLICY_PROOF=true
GITHUB_PR_CREATED_NE_GITHUB_PR_MERGED=true
COMPETITION_REPOSITORY_NE_COMPETITION_SUBMISSION=true
TITMAS_CORE_PROTOCOLS_CHANGED=false

License

Original project code is Apache-2.0. See LICENSE.

The externally installed third-party alibabacloud-resourcecenter-search subtree has SPDX-License-Identifier: NOASSERTION in this repository's source lock because upstream has no applicable license file and its README contains conflicting Apache-2.0 and MIT statements. Installation and byte provenance do not establish redistribution, derivative-work, commercial-use, or compatibility clearance. The subtree is absent from the repository, Worker packages, wheel, and sdist; it must not be included in a public commit or release. See governance/alibabacloud-resourcecenter-search-source-lock.json.