-
A signed webhook does not make a pull request trustworthy
Webhook authentication, repository admission, prompt-injection resistance, and execution isolation solve different problems. A valid signature does not make PR content safe to follow or run.
-
Agent artifacts need an expiration date
Retention is part of an agent workflow’s design: preserve useful evidence, bound sensitive output, and avoid deleting the state that makes recovery and deduplication work.
-
When should an agent be allowed to fix code?
A finding is not permission to edit. A practical policy for bounded automatic fixes, grounded in the PR reviewer’s disabled-by-default fixer primitives.
-
Operating an agent workflow: status, evidence, and failure visibility
A practical look at the PR reviewer’s operations CLI, why a successful process is not a successful workflow, and which failures a simple health summary can miss.
-
Testing agent workflows without calling a model
Use deterministic fixtures to test validation, orchestration, and failure policy. Save live model calls for the questions those tests cannot answer.
-
A webhook is not a durable queue
What happens when a PR review worker is busy or crashes? A look at the gap between accepting a webhook and reliably completing the work it triggers.
-
Memory, skills, or files? Where agent context belongs
A practical policy for agent context: stable preferences in memory, reusable procedures in skills, task evidence in files, and earlier decisions in session history.
-
Structured agent outputs: reviewer, fixer, arbiter
How the PR reviewer validates agent outputs and reduces them into decisions, and why valid JSON is still a long way from verified correctness.
-
Agentic workflows need boring guardrails
A practical approach to agent guardrails: narrow permissions, revision-specific reviews, validated outputs, visible failures, and explicit stopping conditions.
-
Building a Hermes-backed GitHub PR reviewer
Inside a small Hermes-backed PR reviewer: signed webhooks, per-commit job identity, worker dispatch, visible failures, and an important queue limitation.