agentctrl 0.2.0__tar.gz → 0.2.2__tar.gz

agentctrl-0.2.2/CHANGELOG.md

@@ -0,0 +1,74 @@
+# Changelog
+
+All notable changes to `agentctrl` will be documented in this file.
+
+Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/). This project uses [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+---
+
+## [0.2.2] — 2026-04-12
+
+### Added
+- **Advisory context on early exit.** When a pipeline stage short-circuits with ESCALATE or BLOCK, remaining stages (risk, conflict) still run as `ADVISORY` — their results are appended to the decision record for reviewer visibility but do not change the decision.
+- New `ADVISORY` status for `PipelineStageResult` (non-decision, informational).
+- 3 new tests: `test_advisory_context_on_autonomy_escalate`, `test_advisory_context_on_early_exit`, `test_allow_has_no_advisory_stages`.
+
+### Changed
+- `_run_pipeline()` early-exit paths now call `_collect_advisory_context()` instead of returning immediately.
+- Test count updated to 8 (was 5 before advisory context tests).
+
+---
+
+## [0.2.1] — 2026-04-11
+
+### Added
+- **Bidirectional trust calibration.** New agents (< 5 governed actions) receive a +0.35 risk surcharge, pushing routine actions into ESCALATE territory. Proven agents (50+ actions, >90% success rate) receive up to 15% risk discount.
+- `trust_context` parameter on the `@governed` decorator — pass `{"total_actions": N, "success_rate": R}` to influence trust calibration.
+- 2 new tests: `test_new_agent_premium`, `test_new_agent_premium_bypassed_after_threshold`.
+- Apache-2.0 license header on `examples/inbound_governance.py`.
+- `CONTRIBUTING.md` — library-scoped contribution guide.
+- `SECURITY.md` — vulnerability reporting and security model.
+- This `CHANGELOG.md`.
+
+### Changed
+- Risk scoring dimensions documented as 13 (was incorrectly stated as 9 in README).
+- Test count updated to 76 (was 74 before trust calibration tests).
+
+### Fixed
+- README: "nine factors" corrected to "13 dimensions" to match actual `score()` implementation.
+
+---
+
+## [0.2.0] — 2026-04-10
+
+### Added
+- CLI: `agentctrl demo`, `agentctrl validate`, `agentctrl init`.
+- JSONL audit logging via `PipelineHooks`.
+- `RuntimeDecisionRecord` — subscriptable (`record["decision"]`) and attribute-accessible (`record.decision`).
+- Inbound governance example (`examples/inbound_governance.py`).
+- Instance isolation — multiple `RuntimeGateway` instances with independent config.
+- 76 tests (up from initial release).
+
+### Changed
+- Core pipeline tightened: fail-closed invariant enforced at three levels.
+- Trust calibration discount for proven agents (50+ actions, >90% success).
+- Consequence class floors (irreversible actions never score LOW).
+- Factor interaction multiplier (3+ concurrent factors trigger compounding).
+
+---
+
+## [0.1.0] — 2026-04-07
+
+### Added
+- Initial release.
+- 5-stage governance pipeline: Kill Switch → Rate Limiter → Policy Engine → Authority Graph → Risk Engine.
+- `RuntimeGateway` — the main entry point.
+- `PolicyEngine` — AND/OR groups, 14 operators, temporal conditions.
+- `AuthorityGraphEngine` — NetworkX delegation, SoD, decay, time-bound edges.
+- `RiskEngine` — factor-based scoring with configurable weights.
+- `ConflictDetector` — resource contention checking.
+- `@governed` decorator for enforcement.
+- SDK adapters: LangChain, OpenAI Agents SDK, CrewAI.
+- 4 runnable examples.
+- Zero required dependencies.
+- Apache-2.0 license.

agentctrl-0.2.2/CONTRIBUTING.md

@@ -0,0 +1,91 @@
+# Contributing to agentctrl
+
+Thanks for your interest in contributing to `agentctrl` — the institutional governance layer for AI agents.
+
+---
+
+## Quick Start
+
+```bash
+git clone https://github.com/moeintel/AgentCTRL.git
+cd AgentCTRL
+pip install -e ".[dev,all]"
+python -m pytest tests/ -v
+```
+
+---
+
+## Development Setup
+
+**Requirements:** Python 3.11+
+
+```bash
+pip install -e ".[dev,all]"
+```
+
+This installs the library in editable mode with all optional dependencies (networkx, langchain-core, openai-agents, crewai) and dev tools (pytest, pytest-asyncio, ruff).
+
+---
+
+## Running Tests
+
+```bash
+python -m pytest tests/ -v
+```
+
+79 tests total (78 pass, 1 skipped). No external services required — everything runs in-process.
+
+---
+
+## Linting
+
+```bash
+ruff check src/ tests/
+```
+
+---
+
+## Making Changes
+
+### Before you start
+
+1. Check existing issues and discussions to avoid duplicate work.
+2. For larger changes, open an issue first to discuss the approach.
+
+### Guidelines
+
+- **Keep the library self-contained.** Zero required dependencies. No imports from external packages in the core library (adapters are the exception — they use lazy imports).
+- **Write tests.** Every new feature or bug fix should include tests.
+- **Preserve the fail-closed invariant.** Any error in the governance pipeline must produce BLOCK, never silent ALLOW.
+- **Type hints everywhere.** The library is PEP 561 typed.
+
+### Pull request process
+
+1. Fork the repository
+2. Create a feature branch (`git checkout -b my-feature`)
+3. Make your changes
+4. Ensure all tests pass
+5. Ensure linting passes
+6. Submit a pull request with a clear description of what and why
+
+---
+
+## Areas Where Contributions Are Welcome
+
+- **More tests** — edge cases for policy engine, authority graph, risk scoring
+- **Integration examples** — additional `examples/` scripts showing `agentctrl` with different agent frameworks
+- **Documentation** — usage guides, tutorials, integration walkthroughs
+- **Bug reports** — especially around edge cases in policy evaluation or authority resolution
+- **Adapter coverage** — new framework adapters in `src/agentctrl/adapters/`
+
+---
+
+## Code of Conduct
+
+Be respectful, constructive, and professional. We're building governance infrastructure — the bar for quality and honesty is high.
+
+---
+
+## Questions?
+
+Open an issue on [GitHub](https://github.com/moeintel/AgentCTRL/issues).

{agentctrl-0.2.0 → agentctrl-0.2.2}/PKG-INFO

@@ -1,8 +1,8 @@
 Metadata-Version: 2.4
 Name: agentctrl
-Version: 0.2.0
+Version: 0.2.2
 Summary: Institutional control layer for AI agent actions — authority, policy, risk, and audit before execution.
-Project-URL: Homepage, https://moeintel.com
+Project-URL: Homepage, https://moeintel.ai
 Project-URL: Repository, https://github.com/moeintel/AgentCTRL
 Project-URL: Issues, https://github.com/moeintel/AgentCTRL/issues
 Author: MoeIntel
@@ -116,7 +116,7 @@ Those are institutional controls. They existed for human employees. They need to
 - **Fail-closed.** Any pipeline error produces BLOCK, never silent approval.
 - **Structural enforcement.** Policies are operator-based rule matching, not prompt instructions. Authority is graph traversal. Risk is weighted factor scoring. None of this is prompt engineering.
 
-> **Status:** 74 tests passing. Published on [PyPI](https://pypi.org/project/agentctrl/).
+> **Status:** 79 tests passing. Published on [PyPI](https://pypi.org/project/agentctrl/).
 
 ---
 
@@ -333,7 +333,7 @@ agent = Agent(role="analyst", tools=[governed_tool])
 
 ## Decision Pipeline
 
-Every action passes through 5 stages in order. Each can short-circuit.
+Every action passes through 5 stages in order. Each can short-circuit with BLOCK or ESCALATE. When a stage short-circuits, remaining stages still run as **ADVISORY** — their results are appended to the decision record for reviewer visibility but do not change the decision.
 
 ```
 Agent proposes action
@@ -345,6 +345,7 @@ Agent proposes action
     → Risk Scoring (how risky is this action in context?)
     → Conflict Detection (does this clash with other active workflows?)
     → Decision: ALLOW / ESCALATE / BLOCK
+         (+ ADVISORY stages from remaining pipeline on early exit)
 ```
 
 ### Policy Engine
@@ -359,7 +360,7 @@ Authority is opt-in. When no graph is configured, the authority check passes. Co
 
 ### Risk Scoring
 
-Deterministic factor-based scoring. Nine factors: high-value transaction, novel vendor, off-hours activity, data sensitivity, rate pressure, velocity, behavioral anomaly, cumulative exposure, input confidence. Plus consequence class floors (irreversible actions never score LOW) and trust calibration discounts (agents with 50+ governed actions and >90% success rate earn up to 15% reduction).
+Deterministic factor-based scoring across 13 dimensions: base action risk, high-value transaction, novel vendor, off-hours activity, data sensitivity, rate pressure, velocity, behavioral anomaly, cumulative exposure, input confidence, trust calibration (agents with 50+ governed actions and >90% success rate earn up to 15% reduction), factor interaction (3+ concurrent factors trigger compounding), and consequence class floors (irreversible actions never score LOW).
 
 ### Conflict Detection
 
@@ -440,7 +441,7 @@ result = await gateway.validate(proposal)
 python -m pytest tests/ -v
 ```
 
-74 tests covering: pipeline stages, fail-closed behavior, policy evaluation (AND/OR groups, 14 operators, temporal conditions), authority graph (delegation, SoD, limits, decay), risk scoring (9 factors, trust calibration, consequence class), conflict detection, `@governed` decorator, CLI, demo, audit logging, subscriptable record, empty authority default, instance isolation, and library boundary.
+79 tests covering: pipeline stages, advisory context, fail-closed behavior, policy evaluation (AND/OR groups, 14 operators, temporal conditions), authority graph (delegation, SoD, limits, decay), risk scoring (13 dimensions, trust calibration, consequence class), conflict detection, `@governed` decorator, CLI, demo, audit logging, subscriptable record, empty authority default, instance isolation, and library boundary.
 
 ## Requirements
 
@@ -455,4 +456,4 @@ python -m pytest tests/ -v
 
 ---
 
-Built by [MoeIntel](https://moeintel.com). Created by [Mohammad Abu Jafar](https://github.com/moeadnan). [GitHub](https://github.com/moeintel/AgentCTRL)
+Built by [MoeIntel](https://moeintel.ai). Created by [Mohammad Abu Jafar](https://github.com/moeadnan). [GitHub](https://github.com/moeintel/AgentCTRL)

{agentctrl-0.2.0 → agentctrl-0.2.2}/README.md

@@ -76,7 +76,7 @@ Those are institutional controls. They existed for human employees. They need to
 - **Fail-closed.** Any pipeline error produces BLOCK, never silent approval.
 - **Structural enforcement.** Policies are operator-based rule matching, not prompt instructions. Authority is graph traversal. Risk is weighted factor scoring. None of this is prompt engineering.
 
-> **Status:** 74 tests passing. Published on [PyPI](https://pypi.org/project/agentctrl/).
+> **Status:** 79 tests passing. Published on [PyPI](https://pypi.org/project/agentctrl/).
 
 ---
 
@@ -293,7 +293,7 @@ agent = Agent(role="analyst", tools=[governed_tool])
 
 ## Decision Pipeline
 
-Every action passes through 5 stages in order. Each can short-circuit.
+Every action passes through 5 stages in order. Each can short-circuit with BLOCK or ESCALATE. When a stage short-circuits, remaining stages still run as **ADVISORY** — their results are appended to the decision record for reviewer visibility but do not change the decision.
 
 ```
 Agent proposes action
@@ -305,6 +305,7 @@ Agent proposes action
     → Risk Scoring (how risky is this action in context?)
     → Conflict Detection (does this clash with other active workflows?)
     → Decision: ALLOW / ESCALATE / BLOCK
+         (+ ADVISORY stages from remaining pipeline on early exit)
 ```
 
 ### Policy Engine
@@ -319,7 +320,7 @@ Authority is opt-in. When no graph is configured, the authority check passes. Co
 
 ### Risk Scoring
 
-Deterministic factor-based scoring. Nine factors: high-value transaction, novel vendor, off-hours activity, data sensitivity, rate pressure, velocity, behavioral anomaly, cumulative exposure, input confidence. Plus consequence class floors (irreversible actions never score LOW) and trust calibration discounts (agents with 50+ governed actions and >90% success rate earn up to 15% reduction).
+Deterministic factor-based scoring across 13 dimensions: base action risk, high-value transaction, novel vendor, off-hours activity, data sensitivity, rate pressure, velocity, behavioral anomaly, cumulative exposure, input confidence, trust calibration (agents with 50+ governed actions and >90% success rate earn up to 15% reduction), factor interaction (3+ concurrent factors trigger compounding), and consequence class floors (irreversible actions never score LOW).
 
 ### Conflict Detection
 
@@ -400,7 +401,7 @@ result = await gateway.validate(proposal)
 python -m pytest tests/ -v
 ```
 
-74 tests covering: pipeline stages, fail-closed behavior, policy evaluation (AND/OR groups, 14 operators, temporal conditions), authority graph (delegation, SoD, limits, decay), risk scoring (9 factors, trust calibration, consequence class), conflict detection, `@governed` decorator, CLI, demo, audit logging, subscriptable record, empty authority default, instance isolation, and library boundary.
+79 tests covering: pipeline stages, advisory context, fail-closed behavior, policy evaluation (AND/OR groups, 14 operators, temporal conditions), authority graph (delegation, SoD, limits, decay), risk scoring (13 dimensions, trust calibration, consequence class), conflict detection, `@governed` decorator, CLI, demo, audit logging, subscriptable record, empty authority default, instance isolation, and library boundary.
 
 ## Requirements
 
@@ -415,4 +416,4 @@ python -m pytest tests/ -v
 
 ---
 
-Built by [MoeIntel](https://moeintel.com). Created by [Mohammad Abu Jafar](https://github.com/moeadnan). [GitHub](https://github.com/moeintel/AgentCTRL)
+Built by [MoeIntel](https://moeintel.ai). Created by [Mohammad Abu Jafar](https://github.com/moeadnan). [GitHub](https://github.com/moeintel/AgentCTRL)

agentctrl-0.2.2/SECURITY.md

@@ -0,0 +1,46 @@
+# Security Policy
+
+## Reporting a Vulnerability
+
+If you discover a security vulnerability in `agentctrl`, please report it responsibly.
+
+**Do not open a public GitHub issue for security vulnerabilities.**
+
+Instead, use [GitHub's private security advisory feature](https://github.com/moeintel/AgentCTRL/security/advisories/new) or email **security@moeintel.ai**.
+
+Include:
+- Description of the vulnerability
+- Steps to reproduce
+- Potential impact
+- Suggested fix (if you have one)
+
+We will acknowledge receipt within 48 hours and provide a fix or mitigation plan within 7 days for critical issues.
+
+---
+
+## Security Model
+
+`agentctrl` is a governance enforcement library. It evaluates agent actions against policies, authority graphs, and risk scores, then returns ALLOW / ESCALATE / BLOCK decisions.
+
+### What agentctrl enforces
+
+- **Fail-closed design.** Any error in the governance pipeline produces BLOCK, never ALLOW. Three independent layers enforce this (gateway catch, stage-level catch, top-level catch).
+- **Deterministic evaluation.** Policy matching, authority resolution, and risk scoring are all deterministic — no LLM calls, no prompt engineering, no probabilistic behavior.
+- **Structural enforcement.** Policies use operator-based rule matching. Authority is graph traversal. Risk is weighted factor scoring. None of this is prompt-based.
+
+### What agentctrl does NOT enforce
+
+- **Caller identity verification.** `agent_id` is a self-declared string. The library does not verify that the caller actually is that agent. Your application is responsible for identity.
+- **Bypass prevention.** If a tool is called directly without going through `RuntimeGateway` or `@governed`, agentctrl has no visibility. Governance only covers actions routed through the library.
+- **Persistence.** The library is stateless by default. Rate limiting uses in-memory counters that reset on restart. For durable state, integrate with your own storage.
+
+---
+
+## Supported Versions
+
+| Version | Supported |
+|---------|-----------|
+| 0.2.x   | Yes       |
+| < 0.2   | No        |
+
+Security fixes will be applied to the latest release.

{agentctrl-0.2.0 → agentctrl-0.2.2}/examples/inbound_governance.py

@@ -1,3 +1,17 @@
+# Copyright 2026 MoeIntel
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 """Inbound governance — controlling what external agents can do in YOUR system.
 
 This example shows how to use agentctrl to govern actions initiated by external

{agentctrl-0.2.0 → agentctrl-0.2.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "agentctrl"
-version = "0.2.0"
+version = "0.2.2"
 description = "Institutional control layer for AI agent actions — authority, policy, risk, and audit before execution."
 readme = "README.md"
 license = "Apache-2.0"
@@ -40,7 +40,7 @@ classifiers = [
 agentctrl = "agentctrl.cli:main"
 
 [project.urls]
-Homepage = "https://moeintel.com"
+Homepage = "https://moeintel.ai"
 Repository = "https://github.com/moeintel/AgentCTRL"
 Issues = "https://github.com/moeintel/AgentCTRL/issues"
 

{agentctrl-0.2.0 → agentctrl-0.2.2}/src/agentctrl/__main__.py

@@ -233,7 +233,7 @@ def _print_footer(results: list):
     print(f"  {DIM}5 stages × {len(results)} proposals = every action evaluated through:{RESET}")
     print(f"  {DIM}autonomy → policy → authority → risk → conflict{RESET}")
     print()
-    print(f"  {WHITE}pip install agentctrl{RESET}  {DIM}·{RESET}  {WHITE}github.com/moeintel/agentctrl{RESET}")
+    print(f"  {WHITE}pip install agentctrl{RESET}  {DIM}·{RESET}  {WHITE}github.com/moeintel/AgentCTRL{RESET}")
     print()
 
 

{agentctrl-0.2.0 → agentctrl-0.2.2}/src/agentctrl/cli.py

@@ -155,7 +155,7 @@ def cmd_init(args):
         print()
         print("Quick start:")
         print(f'  agentctrl validate --policies {target / "policies.json"} \\')
-        print(f'    \'{{"agent_id": "analyst", "action_type": "invoice.approve", "action_params": {{"amount": 6000}}}}\'')
+        print('    \'{"agent_id": "analyst", "action_type": "invoice.approve", "action_params": {"amount": 6000}}\'')
     else:
         print("No files created (all already exist).")
 

{agentctrl-0.2.0 → agentctrl-0.2.2}/src/agentctrl/decorator.py

@@ -27,6 +27,7 @@ def governed(
     agent_id: str,
     autonomy_level: int = 2,
     action_type: str | None = None,
+    trust_context: dict | None = None,
 ):
     """Decorator that wraps an async function with governance evaluation.
 
@@ -53,6 +54,7 @@ def governed(
                 action_type=resolved_action_type,
                 action_params=action_params,
                 autonomy_level=autonomy_level,
+                trust_context=trust_context,
             )
 
             result = await gateway.validate(proposal)

{agentctrl-0.2.0 → agentctrl-0.2.2}/src/agentctrl/risk_engine.py

@@ -198,11 +198,22 @@ class RiskEngine:
                 "value": f"Daily exposure: ${daily_exposure:,.0f} (threshold: ${exp_threshold:,.0f})",
             })
 
-        # Trust calibration — agents with demonstrated reliability get a risk discount.
+        # Trust calibration — bidirectional risk adjustment based on agent track record.
+        # New agents (< new_agent_threshold actions) receive a risk surcharge.
+        # Proven agents (50+ actions, >90% success) receive a risk discount.
         trust_ctx = getattr(proposal, "trust_context", None) or {}
         trust_total_actions = trust_ctx.get("total_actions", 0)
         trust_success_rate = trust_ctx.get("success_rate", 0.0)
-        if trust_total_actions >= 50 and trust_success_rate > 0.90:
+        new_agent_threshold = self._factors.get("new_agent_premium", {}).get("threshold", 5)
+        new_agent_weight = self._factors.get("new_agent_premium", {}).get("weight", 0.35)
+        if trust_total_actions < new_agent_threshold:
+            total += new_agent_weight
+            factors.append({
+                "factor": "trust_calibration",
+                "contribution": round(new_agent_weight, 3),
+                "value": f"New agent — {trust_total_actions} prior actions (threshold: {new_agent_threshold})",
+            })
+        elif trust_total_actions >= 50 and trust_success_rate > 0.90:
             trust_discount = min(0.15, (trust_success_rate - 0.90) * 1.5)
             total = max(0.01, total - trust_discount)
             factors.append({

{agentctrl-0.2.0 → agentctrl-0.2.2}/src/agentctrl/runtime_gateway.py

@@ -19,7 +19,9 @@ Runtime Governance Model — Layer 1
 The pipeline evaluates every ActionProposal through 5 sequential decision
 stages (autonomy → policy → authority → risk → conflict).  Each stage can
 short-circuit with BLOCK or ESCALATE.  If all stages pass, the decision
-is ALLOW.
+is ALLOW.  On early ESCALATE/BLOCK, remaining stages still run as ADVISORY
+context — their results are appended to the decision record for reviewer
+visibility but do not change the decision.
 
 The kill switch is an optional pre-gate callback (`kill_switch_fn`) so
 the library works without platform dependencies.  Fail-closed: any
@@ -154,9 +156,8 @@ class RuntimeGateway:
         if stage1.status == "BLOCK":
             return self._make_decision(proposal, stages, "BLOCK", stage1.reason, 0.0, "LOW")
         if stage1.status == "ESCALATE":
-            risk = await self.risk_engine.score(proposal)
-            stages.append(PipelineStageResult("risk_scoring", "INFO",
-                                               {"risk_score": risk.score, "risk_level": risk.level}))
+            risk, advisory = await self._collect_advisory_context(proposal, from_stage=1)
+            stages.extend(advisory)
             return self._make_decision(proposal, stages, "ESCALATE", stage1.reason,
                                        risk.score, risk.level, escalated_to="approver_required")
 
@@ -164,7 +165,8 @@ class RuntimeGateway:
         stage2 = await self.policy_engine.validate(proposal)
         stages.append(stage2)
         if stage2.status in ("BLOCK", "ESCALATE"):
-            risk = await self.risk_engine.score(proposal)
+            risk, advisory = await self._collect_advisory_context(proposal, from_stage=2)
+            stages.extend(advisory)
             return self._make_decision(proposal, stages, stage2.status, stage2.reason,
                                        risk.score, risk.level)
 
@@ -172,7 +174,8 @@ class RuntimeGateway:
         stage3 = await self.authority_engine.resolve(proposal)
         stages.append(stage3)
         if stage3.status in ("BLOCK", "ESCALATE"):
-            risk = await self.risk_engine.score(proposal)
+            risk, advisory = await self._collect_advisory_context(proposal, from_stage=3)
+            stages.extend(advisory)
             escalated_to = stage3.details.get("escalate_to")
             return self._make_decision(proposal, stages, stage3.status, stage3.reason,
                                        risk.score, risk.level, escalated_to=escalated_to)
@@ -187,6 +190,8 @@ class RuntimeGateway:
         )
         stages.append(stage4)
         if stage4.status == "ESCALATE":
+            _risk, advisory = await self._collect_advisory_context(proposal, from_stage=4)
+            stages.extend(advisory)
             return self._make_decision(proposal, stages, "ESCALATE", stage4.reason,
                                        risk.score, risk.level)
 
@@ -202,6 +207,40 @@ class RuntimeGateway:
         reason = f"All validation stages passed. Action '{proposal.action_type}' approved for execution."
         return self._make_decision(proposal, stages, "ALLOW", reason, risk.score, risk.level)
 
+    async def _collect_advisory_context(
+        self, proposal: ActionProposal, from_stage: int,
+    ) -> tuple:
+        """Run remaining pipeline stages as non-decision ADVISORY context.
+
+        Gives the human reviewer visibility into what risk and conflict
+        would have said, even though an earlier stage already decided.
+        Returns (risk_result, list_of_advisory_stages).
+        ``from_stage`` is the stage number that triggered the early exit
+        (1=autonomy, 2=policy, 3=authority, 4=risk).
+        """
+        advisory_stages: list[PipelineStageResult] = []
+        if from_stage < 4:
+            risk = await self.risk_engine.score(proposal)
+            advisory_stages.append(
+                PipelineStageResult(
+                    "risk_scoring", "ADVISORY",
+                    {"risk_score": risk.score, "risk_level": risk.level, "factors": risk.factors},
+                    f"Advisory: Risk level {risk.level} (score: {risk.score:.2f})",
+                )
+            )
+        else:
+            risk = None
+        if from_stage < 5:
+            conflict = await self.conflict_detector.check(proposal)
+            advisory_stages.append(
+                PipelineStageResult(
+                    "conflict_detection", "ADVISORY",
+                    {"original_status": conflict.status, **(conflict.details or {})},
+                    f"Advisory: {conflict.reason}",
+                )
+            )
+        return risk, advisory_stages
+
     async def _check_autonomy(self, proposal: ActionProposal) -> PipelineStageResult:
         level = proposal.autonomy_level
         action = proposal.action_type.split(".")[0] if "." in proposal.action_type else proposal.action_type

{agentctrl-0.2.0 → agentctrl-0.2.2}/src/agentctrl/types.py

@@ -51,7 +51,7 @@ class ActionProposal:
 class PipelineStageResult:
     """Result from a single pipeline stage."""
     stage: str
-    status: str  # PASS | FAIL | ESCALATE | BLOCK
+    status: str  # PASS | FAIL | ESCALATE | BLOCK | ADVISORY
     details: dict = field(default_factory=dict)
     reason: str = ""
 

{agentctrl-0.2.0 → agentctrl-0.2.2}/tests/test_decorator.py

@@ -26,7 +26,9 @@ async def test_governed_execute_happy_path():
 
     gateway = RuntimeGateway()
 
-    @governed(gateway=gateway, agent_id="ap_analyst", autonomy_level=2, action_type="invoice.approve")
+    @governed(gateway=gateway, agent_id="ap_analyst", autonomy_level=2,
+             action_type="invoice.approve",
+             trust_context={"total_actions": 10, "success_rate": 0.95})
     async def approve_invoice(amount: float):
         return {"ok": True, "amount": amount}
 

{agentctrl-0.2.0 → agentctrl-0.2.2}/tests/test_parity_features.py

@@ -91,7 +91,8 @@ class TestRiskEngineNewFactors:
         from agentctrl import RiskEngine, ActionProposal
         engine = RiskEngine()
         p = ActionProposal(agent_id="a", action_type="search",
-                           action_params={}, consequence_class="reversible")
+                           action_params={}, consequence_class="reversible",
+                           trust_context={"total_actions": 10, "success_rate": 0.95})
         score = await engine.score(p)
         assert score.score < 0.30
 
@@ -458,12 +459,14 @@ class TestGatewayNewFeatures:
         for _ in range(2):
             r = await gw.validate(ActionProposal(
                 agent_id="ap_analyst", action_type="invoice.approve",
-                action_params={"amount": 100}, autonomy_level=2))
+                action_params={"amount": 100}, autonomy_level=2,
+                trust_context={"total_actions": 10, "success_rate": 0.95}))
             assert r["decision"] == "ALLOW"
 
         r = await gw.validate(ActionProposal(
             agent_id="ap_analyst", action_type="invoice.approve",
-            action_params={"amount": 100}, autonomy_level=2))
+            action_params={"amount": 100}, autonomy_level=2,
+            trust_context={"total_actions": 10, "success_rate": 0.95}))
         assert r["decision"] == "BLOCK"
         assert "rate_limit" in r["pipeline"][0]["stage"]
 
@@ -476,7 +479,8 @@ class TestGatewayNewFeatures:
         gw = RuntimeGateway(autonomy_scopes=scopes)
         r = await gw.validate(ActionProposal(
             agent_id="ap_analyst", action_type="invoice.approve",
-            action_params={"amount": 3000}, autonomy_level=2))
+            action_params={"amount": 3000}, autonomy_level=2,
+            trust_context={"total_actions": 10, "success_rate": 0.95}))
         assert r["decision"] == "ALLOW"
 
     @pytest.mark.asyncio

{agentctrl-0.2.0 → agentctrl-0.2.2}/tests/test_pipeline.py

@@ -30,6 +30,7 @@ async def test_low_risk_action_executes():
         action_type="invoice.approve",
         action_params={"amount": 1000},
         autonomy_level=2,
+        trust_context={"total_actions": 10, "success_rate": 0.95},
     ))
     assert result["decision"] == "ALLOW"
 
@@ -83,7 +84,7 @@ async def test_hooks_called():
 
     decisions = []
     hooks = PipelineHooks(
-        on_decision=lambda d, p, s, l: decisions.append(d),
+        on_decision=lambda d, p, s, log: decisions.append(d),
     )
     gateway = RuntimeGateway(hooks=hooks)
     await gateway.validate(ActionProposal(
@@ -93,3 +94,63 @@ async def test_hooks_called():
         autonomy_level=2,
     ))
     assert len(decisions) == 1
+
+
+@pytest.mark.asyncio
+async def test_advisory_context_on_autonomy_escalate():
+    """Level 1 agent escalates at autonomy; risk + conflict run as ADVISORY."""
+    from agentctrl import RuntimeGateway, ActionProposal
+
+    gateway = RuntimeGateway()
+    result = await gateway.validate(ActionProposal(
+        agent_id="junior-agent",
+        action_type="email.send",
+        action_params={"to": "user@example.com"},
+        autonomy_level=1,
+    ))
+    assert result["decision"] == "ESCALATE"
+    stages = result["pipeline"]
+    advisory_stages = [s for s in stages if s["status"] == "ADVISORY"]
+    assert len(advisory_stages) == 2, f"Expected 2 advisory stages, got {len(advisory_stages)}"
+    advisory_names = {s["stage"] for s in advisory_stages}
+    assert "risk_scoring" in advisory_names
+    assert "conflict_detection" in advisory_names
+
+
+@pytest.mark.asyncio
+async def test_advisory_context_on_early_exit():
+    """Early BLOCK/ESCALATE still collects advisory risk + conflict stages."""
+    from agentctrl import RuntimeGateway, ActionProposal
+    from agentctrl.policy_engine import PolicyEngine
+
+    policies = [{"action_type": "delete.*", "effect": "BLOCK", "reason": "Deletes are forbidden"}]
+    gateway = RuntimeGateway(policy_engine=PolicyEngine(policies=policies))
+    result = await gateway.validate(ActionProposal(
+        agent_id="analyst",
+        action_type="delete.records",
+        action_params={},
+        autonomy_level=3,
+    ))
+    assert result["decision"] in ("BLOCK", "ESCALATE")
+    advisory_stages = [s for s in result["pipeline"] if s["status"] == "ADVISORY"]
+    assert len(advisory_stages) >= 1, "At least one ADVISORY stage expected"
+    advisory_names = {s["stage"] for s in advisory_stages}
+    assert "risk_scoring" in advisory_names or "conflict_detection" in advisory_names
+
+
+@pytest.mark.asyncio
+async def test_allow_has_no_advisory_stages():
+    """Normal ALLOW path has no ADVISORY stages — all stages run as real decisions."""
+    from agentctrl import RuntimeGateway, ActionProposal
+
+    gateway = RuntimeGateway()
+    result = await gateway.validate(ActionProposal(
+        agent_id="ap_analyst",
+        action_type="invoice.approve",
+        action_params={"amount": 1000},
+        autonomy_level=2,
+        trust_context={"total_actions": 10, "success_rate": 0.95},
+    ))
+    assert result["decision"] == "ALLOW"
+    advisory_stages = [s for s in result["pipeline"] if s["status"] == "ADVISORY"]
+    assert len(advisory_stages) == 0

{agentctrl-0.2.0 → agentctrl-0.2.2}/tests/test_risk_engine.py

@@ -141,3 +141,42 @@ async def test_configurable_base_risks():
     score = await engine.score(proposal)
     base_factors = [f for f in score.factors if f.get("factor") == "base_action_risk"]
     assert base_factors[0]["contribution"] == 0.99
+
+
+@pytest.mark.asyncio
+async def test_new_agent_premium():
+    """New agents (< 5 actions) receive a risk surcharge via trust_calibration."""
+    from agentctrl import RiskEngine, ActionProposal
+
+    engine = RiskEngine()
+    proposal = ActionProposal(
+        agent_id="brand-new-agent",
+        action_type="invoice.approve",
+        action_params={"amount": 1000},
+        autonomy_level=2,
+        trust_context={"total_actions": 0, "success_rate": 0.0},
+    )
+    score = await engine.score(proposal)
+    trust = [f for f in score.factors if f.get("factor") == "trust_calibration"]
+    assert trust
+    assert trust[0]["contribution"] > 0
+    assert score.level in ("HIGH", "CRITICAL")
+
+
+@pytest.mark.asyncio
+async def test_new_agent_premium_bypassed_after_threshold():
+    """Agents above the new-agent threshold do not receive the surcharge."""
+    from agentctrl import RiskEngine, ActionProposal
+
+    engine = RiskEngine()
+    proposal = ActionProposal(
+        agent_id="established-agent",
+        action_type="invoice.approve",
+        action_params={"amount": 1000},
+        autonomy_level=2,
+        trust_context={"total_actions": 10, "success_rate": 0.90},
+    )
+    score = await engine.score(proposal)
+    trust = [f for f in score.factors if f.get("factor") == "trust_calibration"]
+    assert not trust
+    assert score.level in ("LOW", "MEDIUM")

{agentctrl-0.2.0 → agentctrl-0.2.2}/tests/test_v02_features.py

@@ -132,7 +132,8 @@ class TestEmptyAuthorityDefault:
         gw = RuntimeGateway()
         r = await gw.validate(ActionProposal(
             agent_id="agent-1", action_type="invoice.approve",
-            action_params={"amount": 100}, autonomy_level=2))
+            action_params={"amount": 100}, autonomy_level=2,
+            trust_context={"total_actions": 10, "success_rate": 0.95}))
         assert r.decision == "ALLOW"
 
 

{agentctrl-0.2.0 → agentctrl-0.2.2}/uv.lock

@@ -9,7 +9,7 @@ resolution-markers = [
 
 [[package]]
 name = "agentctrl"
-version = "0.2.0"
+version = "0.2.2"
 source = { editable = "." }
 
 [package.optional-dependencies]