From past chats

Build one agent-portability fixture

Recent interest in AGENTS.md, plugins, and cross-tool agent formats suggests a useful test: make one tiny repository whose expected result is unambiguous, then run the same task through two coding agents. Include a root instruction, one nested override, a local skill, and a deliberately irrelevant instruction. Compare whether each agent discovers the right scope, runs the required check, and leaves the same minimal diff. The result would turn “portable” from a format claim into a behaviour you can periodically verify. A good follow-up question is which parts of the workflow genuinely need a shared standard, and which should remain tool-specific.

• AGENTS.md open format
• Model Context Protocol specification
• Agentic AI Foundation

Give agent handoffs a trace ID

The recent idea of separate agent “towns” exchanging messages becomes much easier to study if every delegated job carries a trace ID. Try a small append-only ledger with trace_id, parent_id, sender, recipient, purpose, outcome, elapsed time, and artefacts—but no prompt bodies by default. One real multi-agent task should reveal whether delays come from delegation, tool use, or rework, while also making orphaned handoffs obvious. The practical question is whether a trace is useful enough with metadata alone, before accepting the privacy and storage cost of recording content.

• OpenTelemetry agent-service example
• OpenTelemetry GenAI semantic conventions
• W3C Trace Context

From the lmika.org front page

Make fake addresses unmistakably synthetic

The surprise that “123 Fake St” exists is a good reason to stop treating memorable prose as safe test data. Build a small seeded fixture generator that produces locally valid field shapes while stamping every record with an obvious marker such as TEST-<case-id>; use reserved domains for email and URLs, and stub the address-lookup provider in automated tests. Keep a separate, explicitly labelled manual sandbox case for testing the real lookup UI. This makes failures reproducible and reduces the chance that a test record points at a real person or triggers a downstream action.

• Gofakeit: seeded fake data for Go
• RFC 2606 reserved test and example domains
• Google libaddressinput

Put a smoke test around the Hugo mount trick

Mounting static into assets is a neat way to bring old files into Hugo’s image pipeline, but custom mounts replace the default mount for that component. Add one tiny build fixture that checks both sides of the contract: a mounted image can be fetched and resized as a resource, while an ordinary static file still appears unchanged in the published output. Run it after Hugo upgrades. If the experiment proves stable, a follow-up is to convert only images that benefit from thumbnails or fingerprints instead of reorganising the whole archive.

• Hugo module mount configuration
• Hugo resources.Get
• Hugo image resizing