Claude Code shipped five developer-facing changes between 1 and 6 October 2026, and a solo developer should switch on two of them today. Version 2.1.287 on 1 October added Claude Mods and a built-in mod called You should know. Version 2.1.292 on 6 October added an effort parameter to the Agent tool, a --marketplace flag on plugin install and a new retry-delay environment variable for 529 overload errors. Versions 2.1.293 and 2.1.294 followed on 7 and 8 October.
Short answer: turn on the Agent-tool effort habit immediately, because it costs nothing and cuts the bill on mechanical sub-agent work. Trial the You should know side agent on one repository for a week. Set the 529 backoff variable only if you actually see overload errors. Ignore --marketplace until you distribute plugins to a team, and read about Mods before you write one. Everything below comes from the official changelog at code.claude.com/docs/en/changelog unless a line says "reported".
What Claude Mods change
Until 2.1.287, a plugin could add commands, skills, hooks and agents around Claude Code. The changelog entry for that release reads: "Added Claude Mods: plugins may now modify deeper behavior." The word that matters is deeper. A mod is a plugin that reaches into parts of the product that used to be fixed.
The proof is in what already shipped as one. Some built-in features now arrive as mods, and the changelog uses /diff as the example: you can disable it in /plugin or replace it with your own. That is a change of ownership. A built-in command stops being a fixed feature and becomes a default you can swap, the same way you swap an editor extension.
Two mod APIs landed in 2.1.292 and show the direction. prompt.autocomplete lets a mod shape what the prompt box suggests. $.tool.register registers a tool, and it accepts an isDeferred option, so a mod can add a tool without paying its schema cost in every request until Claude needs it. If you have watched your context fill with tool definitions you never call, deferred registration is the part to care about.
The catch: Mods raise the trust bar. A plugin that only adds a slash command can do limited damage. A plugin that modifies core behaviour sits in the path of everything you type and every tool call Claude makes. Install mods from sources you can read, pin versions, and keep your plugin list short. The same rule applies to the new --marketplace flag, covered below.
For a solo developer the practical takeaway is narrow. You do not need to write a mod this month. You do need to know that "Claude Code behaves like this" is now sometimes "my installed plugins make it behave like this", which matters the next time a teammate's session acts differently from yours. When you debug odd behaviour, open /plugin first.
The You should know side agent
You should know is the first built-in mod, and it is the one most people will try. The 2.1.287 changelog describes it as a side agent that watches your back and flags things you or Claude might miss. It runs alongside the main session rather than inside it, so its job is the opposite of the main agent's job: it is not trying to finish your task, it is looking for what the task is about to break.
The changelog gives the exact enable command, and notes it applies to first-party sessions with telemetry on:
/plugin enable cc-plugin-you-should-know@builtin
That condition is easy to miss. If you run Claude Code through a third-party cloud provider or you have telemetry switched off, the mod may not be available to you. Check before you assume it is broken.
Version 2.1.288 on 2 October changed the wording of its notes so they say "we", "the main agent" or "you" depending on who is responsible for the thing flagged. That sounds cosmetic and is not. A flag that says "you" asks you to act. A flag that says "the main agent" tells you Claude should fix it and you can leave it alone. Version 2.1.290 on 5 October fixed a bug where it wrote English regardless of your configured language setting. Two fixes in four days tells you the feature is young.
How I would trial it
Pick one repository with real tests and a CLAUDE.md you trust. Enable the mod, then work normally for five working days. Keep a plain text file open and add a line each time a flag was useful, ignorable or wrong. At the end of the week you have three numbers and a decision. If useful flags outnumber wrong ones, keep it. If it nags you about things your hooks already catch, switch it off with /plugin.
One cost question stays open. The changelog does not state how many tokens the side agent consumes, and a second agent watching a session is not free. Watch your usage numbers during the trial week rather than guessing. If you are running agents at volume, the Prompt Bloat Analyzer, new today, shows how much of each request your standing instructions consume, which is the part every extra agent pays again.
The effort parameter on the Agent tool
Version 2.1.292 added an effort parameter to the Agent tool. The changelog line: "so Claude runs a sub-agent at the effort level you ask for". Before this, a sub-agent inherited the session's effort. If you ran your main session at a high effort setting for a hard refactor, every file-search and test-run sub-agent it spawned ran hot too.
The Agent tool exposes the levels low, medium, high, xhigh and max. You do not call the tool yourself. You tell Claude which level to use, either in the moment or as a standing rule in CLAUDE.md. A rule that has worked on my own repositories looks like this:
## Sub-agent effort
- Mechanical work (grep and summarise, run tests, rename, format): effort low
- Single-file edits with a clear spec: effort medium
- Cross-file refactors, security review, migrations: effort high
- Never use max unless I ask for it
The reasoning is plain. Mechanical sub-agents spend most of their tokens reading and reporting, and extra reasoning effort buys almost nothing on a task whose answer is a list of grep hits. Hard review work is where effort pays. Matching the two is the cheapest quality-per-token improvement in this release, and it needs no new infrastructure.
The catch: a low-effort sub-agent can miss a subtle case it would have caught at high. Keep the final review step on the main agent at your normal setting, and make the sub-agents report evidence, not conclusions. For defining "done" in a way a low-effort agent cannot fudge, the Gauntlet Prompt Generator builds the success-criteria and verification block you paste into the task. The Claude MD Production Rules product includes ready-made rule blocks of the kind above, and the CLAUDE.md Starter Templates Pack gives you a full file to edit instead of a blank one.
If you are weighing when a skill file beats spawning several sub-agents at all, read Agent Skills Beat Agent Crews. The effort parameter makes crews cheaper, and a skill file still makes many tasks need no crew.
The marketplace flag and the 529 backoff variable
Both of these are small and both solve a real problem.
--marketplace <source> is now accepted by claude plugin install. The changelog says it was added to that command and nothing more, so treat the details as unverified until you run claude plugin install --help on your machine. The use case is deterministic installs: a script or a new teammate's setup command can say which marketplace a plugin comes from instead of relying on whatever is already registered. If you work alone and install by hand, you will not notice it.
CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS sets a longer base delay before Claude Code retries after a 529 overloaded response. A 529 means the API is shedding load. Retrying fast during a shed adds to the problem and burns your patience. A longer base delay spaces the retries out. The name tells you the unit is milliseconds, and the changelog gives no default, so choose a value deliberately:
export CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS=5000
Five seconds is my starting point, not an official recommendation. Set it in your shell profile, run for a few days and see whether 529 errors resolve on their own more often. If you rarely see 529s, leave the variable unset. A setting you do not need is a setting you will forget the reason for.
One more item, labelled as reported. Third-party coverage describes a Projects beta in which one project coordinates parallel threads with shared memory. It does not appear in the visible changelog, so I am not including it in the verdict. Check the changelog before you plan around it.
Which to turn on for a solo developer
| Change | Version and date | Turn it on? | Why |
|---|
| Claude Mods (concept) | 2.1.287, 1 Oct 2026 | Read, do not build yet | It changes who owns core behaviour; the benefit arrives when mods you trust exist |
| You should know side agent | 2.1.287, 1 Oct; fixes 2.1.288, 2.1.290 | Trial one repo for a week | Catches what you and the main agent miss; token cost is not stated, so measure it |
| Agent tool effort parameter | 2.1.292, 6 Oct | Yes, today | Free; low for mechanical sub-agents, high for review and migrations |
| --marketplace on plugin install | 2.1.292, 6 Oct | Later | Matters for scripted or team installs, not hand installs |
| 529 retry base delay variable | 2.1.292, 6 Oct | Only if you see 529s | One line in your shell profile; no downside if you leave it unset |
If you only do one thing from this post, paste the sub-agent effort rule into your CLAUDE.md and keep it there for a month. Compare your usage before and after. If you want to sanity-check that file first, run it through the Prompt Bloat Analyzer and trim anything that no longer earns its tokens.
Quick answers
How do I enable the You should know mod?
Run /plugin enable cc-plugin-you-should-know@builtin inside Claude Code. The 2.1.287 changelog says this is for first-party sessions with telemetry on, so it may not appear on third-party cloud setups or with telemetry disabled.
Can I replace the built-in /diff command?
Yes. Because some built-in features now ship as mods, /diff can be disabled in /plugin or replaced by a plugin of your own. Open /plugin to see which built-ins are listed as mods in your version.
What does the effort parameter on the Agent tool do?
It lets Claude run a sub-agent at the effort level you ask for, instead of inheriting the main session setting. Use low for search, test runs and renames, and high for cross-file refactors and security review.
What does the 529 environment variable change?
CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS lengthens the base wait before Claude Code retries after a 529 overloaded response. The unit is milliseconds and the changelog does not publish a default value.
Every product mentioned is available at wowhow.cloud — pay once, ship forever.
Originally published at wowhow.cloud
Top comments (0)