docs: add Aegis to third-party capabilities by Acacian · Pull Request #4888 · pydantic/pydantic-ai

Acacian · 2026-03-29T01:06:30Z

Summary

Add agent-aegis to the "Governance & Safety" subsection of third-party capabilities.

What it adds

One bullet in the existing list — AegisCapability (from aegis.contrib.pydantic_ai) is a native AbstractCapability that wraps a guardrail engine into before_model_request / after_model_request hooks. Features exposed through the capability:

Prompt injection detection (13 categories, multi-language, encoding evasion)
PII masking with Luhn/IBAN validation
Toxicity filtering, hallucination pattern detection, prompt-leak detection
Keyword/regex matching and YAML policy-as-code via GuardrailEngine.from_pack()
on_block="raise"/"warn" and per-agent check_input/check_output control

Import: from aegis.contrib.pydantic_ai import AegisCapability

devin-ai-integration

Devin Review found 1 new potential issue.

View 1 additional finding in Devin Review.

devin-ai-integration · 2026-03-29T03:34:06Z


+### Governance & Safety
+
+* [`agent-aegis`](https://github.com/Acacian/aegis) -- Auto-instruments Pydantic AI agents with governance guardrails: prompt injection detection, PII masking, policy-as-code (YAML), and audit trail. Setup: `aegis.auto_instrument()`.


🚩 Package name in link text does not match GitHub repo name

All existing third-party capability entries use a link text that matches the GitHub repository name (e.g., pydantic-ai-shields → vstorm-co/pydantic-ai-shields, pydantic-ai-skills → DougTrajano/pydantic-ai-skills). The new entry uses agent-aegis as the link text but links to https://github.com/Acacian/aegis where the repo is named aegis, not agent-aegis. This could be intentional if the PyPI package name is agent-aegis while the repo is aegis, but it's a departure from the pattern of other entries and could confuse users trying to find the package on GitHub vs PyPI. Worth verifying whether the pip-installable name is indeed agent-aegis.

Was this helpful? React with 👍 or 👎 to provide feedback.

DouweM · 2026-04-01T17:16:32Z

@Acacian It doesn't look like those features are actually exposed as Pydantic AI capabilities?

Acacian · 2026-04-01T23:13:26Z

@DouweM Good point — you're right, the previous version used monkey-patching (auto_instrument()), not the native capability API.

I've updated this: Aegis now ships AegisCapability, a proper AbstractCapability subclass that uses before_model_request / after_model_request lifecycle hooks to enforce guardrails. The listing text has been updated accordingly:

from pydantic_ai import Agent
from aegis.contrib.pydantic_ai import AegisCapability
from aegis.guardrails import GuardrailEngine, InjectionGuardrail

engine = GuardrailEngine()
engine.add(InjectionGuardrail())

agent = Agent(
    "openai:gpt-4o-mini",
    capabilities=[AegisCapability(engine)],
)

Also supports from_spec() for YAML/JSON agent specs.

Implements pydantic_ai.capabilities.AbstractCapability so Aegis guardrails run as a first-class Pydantic AI capability via before_model_request / after_model_request lifecycle hooks — no monkey-patching required. Addresses pydantic/pydantic-ai#4888 reviewer feedback. Co-Authored-By: Claude Opus 4.6 <[email protected]>

devin-ai-integration

Devin Review found 2 new potential issues.

View 1 additional finding in Devin Review.

devin-ai-integration · 2026-04-01T23:15:07Z


+### Governance & Safety
+
+* [`agent-aegis`](https://pypi.org/project/agent-aegis/) ([source](https://github.com/Acacian/aegis)) -- `AegisCapability` implements `AbstractCapability` to enforce governance guardrails (prompt injection detection, PII masking, policy-as-code) via `before_model_request` / `after_model_request` lifecycle hooks. Setup: `Agent(..., capabilities=[AegisCapability(engine)])`.


🚩 Third-party package legitimacy not verified

The listing references agent-aegis on PyPI and links to https://github.com/Acacian/aegis as the source. The description claims it implements AbstractCapability with before_model_request / after_model_request hooks. Since this is a third-party package listing, the maintainers may want to verify that the package actually implements AbstractCapability as claimed (rather than, e.g., monkey-patching or using the older auto_instrument() pattern referenced in an earlier commit a96aef92). The commit history shows the description was updated from aegis.auto_instrument() to AegisCapability(engine), suggesting the package API may have recently changed.

Was this helpful? React with 👍 or 👎 to provide feedback.

Acacian · 2026-04-05T15:23:27Z

@DouweM Just checking if the updated implementation addresses your concern — Aegis now exposes AegisCapability as a proper AbstractCapability subclass rather than relying on monkey-patching. Happy to adjust further if needed.

Implements pydantic_ai.capabilities.AbstractCapability so Aegis guardrails run as a first-class Pydantic AI capability via before_model_request / after_model_request lifecycle hooks — no monkey-patching required. Addresses pydantic/pydantic-ai#4888 reviewer feedback. Co-Authored-By: Claude Opus 4.6 <[email protected]>

Kludex · 2026-04-06T21:00:01Z

@Acacian Please move this to the right section in the docs.

Acacian · 2026-04-06T23:09:52Z

@Kludex Thanks for the review! Moved Aegis under the existing "Guardrails & Safety" section and rebased on latest main.

Implements pydantic_ai.capabilities.AbstractCapability so Aegis guardrails run as a first-class Pydantic AI capability via before_model_request / after_model_request lifecycle hooks — no monkey-patching required. Addresses pydantic/pydantic-ai#4888 reviewer feedback. Co-Authored-By: Claude Opus 4.6 <[email protected]>

github-actions · 2026-04-21T14:30:04Z

This PR is stale, and will be closed in 7 days if no reply is received.

DouweM · 2026-04-21T23:34:44Z

 Capabilities for cost control, input/output filtering, and tool permissions help keep agents safe and within budget:

 * [`pydantic-ai-shields`](https://github.com/vstorm-co/pydantic-ai-shields) - Ready-to-use guardrail capabilities: `CostTracking` (tracks token usage and USD cost per run, raises `BudgetExceededError` on budget overrun); `ToolGuard` (block or require approval for specific tools); `InputGuard` and `OutputGuard` (custom sync or async validation functions); `PromptInjection`, `PiiDetector`, `SecretRedaction`, `BlockedKeywords`, and `NoRefusals` content shields.
+* [`agent-aegis`](https://pypi.org/project/agent-aegis/) ([source](https://github.com/Acacian/aegis)) - `AegisCapability` implements [`AbstractCapability`][pydantic_ai.capabilities.AbstractCapability] to enforce governance guardrails (prompt injection detection, PII masking, policy-as-code) via lifecycle hooks. Setup: [`Agent`][pydantic_ai.Agent]`(..., capabilities=[AegisCapability(engine)])`.


It's not clear where AegisCapability should even be imported from; can you link to a Pydantic AI specific page in your docs/repo? If there isn't one with full examples, it doesn't warrant a mention here -- sorry.

Updated the entry and the aegis repo to address your feedback:

Import path is now shown inline: from aegis.contrib.pydantic_ai

Linked to the Pydantic AI integration guide with full runnable examples (native AbstractCapability usage)

Dropped the auto_instrument mention to keep it focused on the capability interface

Also added pydantic-ai as an optional dependency and integration tests with TestModel on the aegis side.

…antic-ai Add import path, full example, and comparison table for the native AbstractCapability approach alongside the existing auto_instrument docs. Addresses pydantic/pydantic-ai#4888 review feedback. Co-Authored-By: Claude Opus 4.6 <[email protected]>

Add agent-aegis to the Governance & Safety section of third-party capabilities. AegisCapability wraps a guardrail engine into capability lifecycle hooks, providing prompt injection detection, PII masking, toxicity filtering, and YAML policy-as-code rules. Co-Authored-By: Claude Opus 4.6 <[email protected]>

Acacian · 2026-05-04T08:50:43Z

@DouweM Gentle ping — I addressed your feedback on 4/22 (inline import path, linked full integration guide with examples). Could you take another look when you get a chance? Happy to make further changes if needed. Thanks!

github-actions Bot added size: S Small PR (≤100 weighted lines) docs Improvements or additions to documentation labels Mar 29, 2026

This comment was marked as resolved.

Sign in to view

Acacian force-pushed the docs/aegis-capability branch from acb4d42 to 4ccb564 Compare March 29, 2026 03:32

devin-ai-integration Bot reviewed Mar 29, 2026

View reviewed changes

devin-ai-integration Bot reviewed Apr 1, 2026

View reviewed changes

Kludex added the awaiting author revision label Apr 6, 2026

Acacian force-pushed the docs/aegis-capability branch from 516109a to 394906e Compare April 6, 2026 22:58

This comment was marked as resolved.

Sign in to view

github-actions Bot added the Stale label Apr 21, 2026

DouweM requested changes Apr 21, 2026

View reviewed changes

Acacian force-pushed the docs/aegis-capability branch from 6ed34fd to 2685920 Compare April 22, 2026 01:22

github-actions Bot removed the Stale label Apr 22, 2026

DouweM merged commit fe50b13 into pydantic:main May 4, 2026
45 checks passed

adtyavrdhn mentioned this pull request May 5, 2026

chore: Revert "docs: add Aegis to third-party capabilities" #5305

Merged

3 tasks


		### Governance & Safety

		* [`agent-aegis`](https://github.com/Acacian/aegis) -- Auto-instruments Pydantic AI agents with governance guardrails: prompt injection detection, PII masking, policy-as-code (YAML), and audit trail. Setup: `aegis.auto_instrument()`.

Conversation

Acacian commented Mar 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

What it adds

Uh oh!

This comment was marked as resolved.

Uh oh!

devin-ai-integration Bot left a comment

Choose a reason for hiding this comment

Uh oh!

devin-ai-integration Bot Mar 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

DouweM commented Apr 1, 2026

Uh oh!

Acacian commented Apr 1, 2026

Uh oh!

devin-ai-integration Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

devin-ai-integration Bot Apr 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Acacian commented Apr 5, 2026

Uh oh!

Kludex commented Apr 6, 2026

Uh oh!

This comment was marked as resolved.

Uh oh!

Acacian commented Apr 6, 2026

Uh oh!

This comment was marked as resolved.

Uh oh!

github-actions Bot commented Apr 21, 2026

Uh oh!

DouweM Apr 21, 2026

Choose a reason for hiding this comment

Uh oh!

Acacian Apr 22, 2026

Choose a reason for hiding this comment

Uh oh!

Acacian commented May 4, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Acacian commented Mar 29, 2026 •

edited

Loading

devin-ai-integration Bot Mar 29, 2026 •

edited

Loading

devin-ai-integration Bot Apr 1, 2026 •

edited

Loading