How a tiny naming quirk hid seventy tests in Google Mug

Aditya Gaurav — Sun, 23 Aug 2026 18:28:13 +0000

I spent my morning digging into google mug and came away with a fun debugging story.

While reading through their dot parse regex engine I noticed malformed escape sequences like incomplete unicode or empty character names were quietly passing through. A broad fallback rule was catching them and turning them into plain letters.

Looking deeper into the test folder I found an entire suite of error tests that had never been running in ci simply because the file name had a plural s at the end.

I renamed the test file so maven would pick it up, tightened the escape parser rules, and opened a pull request.

The lead maintainer reviewed it within hours, suggested adding a few more characters to the exclusion set to catch property edge cases, and merged the whole thing.

Fixing a silent parsing bug and bringing seventy tests back to life in one morning felt pretty sweet.

You can check out the merged pull request here: https://github.com/google/mug/pull/130

When Your AI Skill Calls a Ghost: Retiring Zombie Models in Anthropic Skills

Aditya Gaurav — Sun, 23 Aug 2026 13:51:13 +0000

Imagine this scenario. You build an autonomous agent pipeline. You wire in standard skills, configure your API keys, grab a cup of coffee, and watch it tackle complex tasks. Everything feels futuristic.

Then suddenly, a cold splash of reality hits your logs:

Error: 404 Not Found - model 'claude-opus-4-20250514' has been retired.

No timeout. No rate limit. Just an abrupt HTTP 404 from the Anthropic API. Your autonomous agent just spent five minutes trying to summon a ghost.

Here is the story of how dead model identifiers linger inside official agent skills, how an autonomous AI babysitter bot reviewed my pull request, and why keeping documentation in sync across files is harder than it looks.

Act 1: The Zombie Model Problem

Large Language Model development moves at breakneck speed. New checkpoints drop, older snapshots get deprecated, and eventually, older model IDs get decommissioned completely.

In the official anthropics/skills repository, the claude-api skill is supposed to be the source of truth for how agent frameworks talk to Claude. It tells agents which model strings to use, which models support adaptive thinking, and which models have reached end-of-life.

While inspecting skills/claude-api/shared/models.md and skills/claude-api/shared/model-migration.md, I noticed several prominent models were still marked as "Deprecated (retiring soon)" even though their retirement dates had already passed:

claude-opus-4-20250514 (retired June 15, 2026)
claude-sonnet-4-20250514 (retired June 15, 2026)
claude-3-haiku-20240307 (retired April 19, 2026)
claude-opus-4-1-20250805 (retired August 5, 2026)

If an agent reads this skill table to resolve friendly names (such as "opus 4" or "haiku 3"), it would attempt to call a decommissioned model ID instead of routing to modern replacements like claude-opus-4-8, claude-sonnet-5, or claude-haiku-4-5.

The fix looked simple enough: remove the outdated "Deprecated" table, move all four models into the "Retired" table, and point alias resolutions to active successors.

I submitted PR #1607 feeling satisfied. Simple documentation cleanup. Easy victory. Right?

Wrong.

Act 2: Enter the Cron Babysitter

A few hours after opening the PR, a GitHub notification popped up. A review had arrived.

Except it was not from a human maintainer. It came from an autonomous agent calling itself:

Reviewed by Hermes Agent (cron babysitter)
Verdict: Comment: correct and useful model-status update; two cross-file consistency nits.

There is something wonderfully surreal about writing open-source code for AI skills, only to have an automated AI agent run a cron job, read your git diff, and evaluate your work like a sharp-eyed proofreader.

And Hermes was not giving generic praise. It found two very real, very subtle cross-file discrepancies that I had missed:

Nit 1: The One-Day Time Slip

In model-migration.md, the old deprecated table had listed claude-3-haiku-20240307 retiring on April 19, 2026. But in my new table rows in models.md, I had typed April 20, 2026.

A single day difference. To a casual reader, April 19 versus April 20 is trivia. But to an API consumer writing automated compliance checks, date drift between sibling markdown files creates confusion.

Nit 2: The Missing Opus Row

In models.md, I had added claude-opus-4-1-20250805 to the Retired table. But in model-migration.md, I had forgotten to add claude-opus-4-1-20250805 to its corresponding migration table.

The two documentation files inside the same skill were out of sync.

Act 3: The Surgical Reconciliation

The feedback was spot on.

I checked Anthropic's official release timeline to verify the canonical retirement date: April 19, 2026 was the exact date.

Then I made two surgical updates:

skills/claude-api/shared/model-migration.md: Added claude-opus-4-1-20250805 pointing to drop-in replacement claude-opus-4-8, and standardized claude-3-haiku-20240307 to April 19, 2026:

| Retired model                 | Retired       | Drop-in replacement  |
| ----------------------------- | ------------- | -------------------- |
| `claude-opus-4-1-20250805`    | Aug 5, 2026   | `claude-opus-4-8`    |
| `claude-opus-4-20250514`      | Jun 15, 2026  | `claude-opus-4-8`    |
| `claude-sonnet-4-20250514`    | Jun 15, 2026  | `claude-sonnet-5`    |
| `claude-3-haiku-20240307`     | Apr 19, 2026  | `claude-haiku-4-5`   |

skills/claude-api/shared/models.md: Synchronized the retirement date for Haiku 3 to April 19, 2026, ensuring complete cross-file consistency.

Pushed commit bdff74f to the branch and replied to the review thread.

Takeaways from the Trenches

Agent skills are software, not static prose: In traditional documentation, a typo is a visual annoyance. In an agent skill, a stale model ID or wrong parameter definition directly breaks autonomous execution loops in production.
Watch out for multi-file documentation drift: When a repository maintains multiple guides describing the same underlying domain (like models.md and model-migration.md), always audit every sibling file when updating a status.
Embrace autonomous babysitters: Automated PR review bots running static cross-reference checks can catch real human oversights long before core maintainers sit down to triage the queue.

DEV Community: Aditya Gaurav