Maya: Three days before Anthropic relaunched Claude Code Projects on September seventeenth, the company quietly swapped a temporary fifty percent boost to the weekly usage limit for a permanent twenty-five percent increase. Against what users had all summer, that lands as a seventeen percent cut.

James: And the timing bites, Maya. Projects splits one goal into parallel threads, each a full cloud session with its own git branch and repo copy. Each thread draws a full session's allowance, so running six threads burns six sessions' worth.

Maya: So it is a usage multiplier arriving the same week the ceiling shrank. Anthropic capped it at two hundred new threads per day per project, and the beta opened to select Pro and Max subscribers, with repos required on GitHub specifically.

James: There is no separate Projects price; it rides existing tiers. When you hit the limit, usage credits switch you to pay-as-you-go at API rates, capped at two thousand dollars a day. That is what catches the overflow.

Maya: That overflow problem connects to something else people keep hitting: Claude Code losing details mid-task. Long sessions fill the context window, so it compacts, summarizing everything so far and discarding the original messages to keep going.

James: The summary holds the broad shape but drops specifics, like why retry logic moved to the client. A developer on Medium notes those fall out silently, and you cannot see what got cut. Claude later re-adds something you already removed.

Maya: Two fixes are circulating. ContextForge ships a hook on the compact event: when compaction fires, a recall command pulls the ten most recent saved decisions and up to five pending tasks, prints them as text, and Claude appends them straight into context.

James: The other skips chat entirely. A Reddit workflow called wb-flow tracks task state in a real plan file on disk, with dependencies, status, and completion reports. Validation crosses a provider boundary, since same-family review waved bugs straight through that cross-provider checks caught.

Maya: Keeping state outside the transcript is a theme. Agit, short for Agent Git, applies Git's save-and-branch logic to conversations with Claude Code and Codex. Instead of storing code, an Agit repo stores the back-and-forth itself, plus memory and skills.

James: So every finished exchange becomes a commit, and each session gets its own branch. That log is a rewind button too, Maya. A revert command rolls the conversation back a set number of turns, editing only what the agent sees, not code already written.

Maya: A fork command goes further, converting a whole Claude Code conversation into the format Codex reads and opening it there. And because agents read secrets, Agit scans before saving and swaps recognized keys for placeholders.

James: Plain-English passwords are the gap, though. Without a fixed pattern, Agit can miss one even when a message names it directly. AI Labs built a companion skill to register passwords in advance, but it is sold separately, not built in.

Maya: From conversation memory to code correctness. A developer in a ten-year-old ASP.NET Core and Vue app found five recurring failures and fixed each with an Agent Skill, a markdown file Claude Code loads when a task matches.

James: The mistakes are so specific. Claude defaults to textbook patterns like a generic repository even when the codebase uses hand-written mappers, and it happily generates a migration that drops a column when you rename a property.

Maya: The ef-migration skill bans applying to shared databases, requires opening every generated migration to check for drops and non-nullable adds, and asks for a plain-English risk rating for the database administrator. There is also a Vue skill that checks package versions.

James: And the underlying move works without any specific rule: open every skill by telling Claude to find and read the most similar existing code first. The report calls that single sentence the highest-value line in any skill.

Maya: Once code is written, a new tool called assay checks it. It runs as a Claude Code Stop hook, opens every generated page in a real browser, finds the controls, drives them, and reports what broke, with no AI model involved.

James: It deliberately checks one narrow thing: moments where a page contradicts itself, like an Undo button one press behind. On a hand-tested set of two hundred twenty-five pages, it caught fifteen of twenty real bugs and never flagged a working page.

Maya: That is our look at what is quietly reshaping Claude Code this week, here at Claude Code Review. Keep those decisions somewhere compaction cannot reach, and we will see you next time.