You return to a Claude Code task with claude --continue after a tool call failed, only to find that part of the earlier conversation is missing. On a long coding task, an agent may repeat work or miss a constraint you already gave it. Claude Code 2.1.318, released on November 4, 2026, fixes a specific failure path that could cause that loss.

Anthropic's official release notes are precise: --continue sometimes lost the earlier conversation after a malformed tool call exhausted its retries. The regression started in 2.1.314. That does not mean every damaged, deleted, or shortened session can now be recovered. The fix addresses the named sequence of events. If you never use the CLI continuation flag, this particular regression is unlikely to change your daily workflow.

What to check in your coding workflow

Before resuming an important task, run claude --version. If your installation is on an affected version, update Claude Code using the installation method for your environment and check the version again. With 2.1.318 or later, resume a low-risk session using claude --continue. Ask it to summarize prior decisions, then compare that answer with the visible transcript, diff, and actual files. SoyJB has not reproduced the bug or measured the fix in its own installation; this is a verification plan based on the release note.

Consider an AI-assisted database migration. You had already decided to keep an old column until a compatible application version was deployed. If the agent resumes without that constraint, it might suggest dropping the column too early. This update reduces the risk in the documented failure case. It does not replace reviewing the plan and repository state before a change that is hard to undo.

A safe continuity check

1. Record the result of claude --version and open a disposable session in a test repository.
2. State two concrete decisions, such as a file name and an implementation constraint. Exit the session normally.
3. Resume with claude --continue and ask Claude to list those decisions before editing anything.
4. Compare its answer with the transcript, diff, and files. If it misses context, provide the missing facts explicitly; a confident answer alone does not prove a complete memory.

This check does not reproduce the exact regression trigger, which involved a malformed tool call and exhausted retries. It tests your ordinary resumption path. There is no reason to force a tool failure in production just to test a patch. If a previous session appears incomplete, preserve the available session identifier or transcript and review the written work before asking the agent to continue.

One more change for unattended jobs

The same release adds loginExpiresAt to claude auth status for stored claude.ai sign-ins and credential_expires_at to the JSON output of claude -p. Text output from claude -p now warns on stderr starting ten days before that stored login expires. For unattended workflows, that is a signal to renew credentials before the next job fails. The release contains many other macOS, MCP, and remote-session fixes, but no official incidence rate for this particular context-loss bug and no promise to restore sessions that were already lost.