A long Claude Code coding session can lose the details you need while the task is still open. Claude Code can show what occupies its current context and summarize the conversation when space runs short. The useful check comes afterward: compare what the session remembers with the files and work that actually exist. This guide uses Anthropic's current context documentation and describes a procedure to follow in your own session; BIG CHANGE has not run it for you.

The big change

  • What changed: Claude Code exposes a live context breakdown and lets you focus a conversation summary on the task you are continuing. The summary carries working context forward while some files and instructions reload from disk.
  • Why it matters: A developer can make room in a long session without abandoning the authorized coding task, then check the resumed state against the workspace before making another change.
  • What to watch: The context display describes capacity, not the accuracy of a summary. After compaction, the next decision depends on the real files, applicable instructions, running work and the task's existing acceptance checks.

Before you change the context

You need an interactive Claude Code session with a coding task you are authorized to continue, access to its workspace, and a way to identify the task's existing acceptance checks. Anthropic's setup guide says claude --version reports the installed CLI version; /status shows the session's version, model, account and connectivity. Check the commands offered by your installed client's / menu before using those below. The documentation was checked on 19 October 2026 and does not assign these pages a fixed version. Menu support and screen details can change.

  1. Inspect the session. Run /context; use /context all for the expanded item breakdown when the display offers it. Look at the capacity warning and the categories taking space, including loaded CLAUDE.md and memory files. Open /memory to inspect the instruction and auto-memory entries relevant to this task. These commands show what the client has loaded; a CLAUDE.md instruction guides the model and does not enforce access or security. Context window, command reference and memory documentation describe those views.
  2. Preserve the task before summarizing. In a durable task note or the conversation, identify the requested outcome, relevant files, constraints, decisions already made, outstanding work, running background work and the next existing acceptance check. Confirm those details against the current workspace and task record. Do not assume a prior conversation message reflects later file changes or completed background commands. The note is your checkable reference when the session resumes.
  3. Compact only when the context needs it. Claude Code compacts automatically near its configured limit. When you need room sooner, /compact accepts focus instructions. Specify the authorized task and the state that must survive: relevant file paths, constraints, unresolved failures, running work and the next check. The context guide says root CLAUDE.md and unscoped rules reload from disk, while nested and path-specific instructions reload on demand. It rereads up to five recently modified files; long files can return as references. Background commands and subagents keep running. A compacted summary therefore needs verification, even when it sounds complete.
  4. Reconcile and resume. Ask Claude to state its current task, the relevant files, constraints, known failure or unfinished step, and what it intends to do next. Compare that account with the actual task note, applicable instructions and workspace files. Read any missing file or path-specific rule directly, inspect the status of existing background work, and correct a mistaken summary before proceeding. Continue the same authorized task. Run its actual existing acceptance checks and review the resulting diff against the request; investigate any mismatch rather than treating a successful compaction as task completion.

If the session or result does not line up

If /context or /compact is absent, use /status, the installed command menu and claude --version to establish what this client supports. If a summary omits a constraint or file, supply the missing task record and reread the current source file before further edits. If a background job is still running, inspect that job's state before starting the same operation again. Anthropic documents /resume for returning to a conversation, including a running background session; resuming does not roll back file changes or external effects. /clear starts an empty conversation and is inappropriate when you need the current task's history. See the command reference.

Context capacity and spending are separate questions. /context shows how full this conversation's model window is. /usage can show session token statistics, plan allowance and activity, depending on account and client version. Anthropic says the displayed session dollar figure is a local estimate; Pro and Max subscribers have included usage, while authoritative API billing belongs in the Claude Console. Compaction can cause a prompt-cache rebuild and consume additional usage. Do not infer a token saving, remaining subscription allowance or invoice amount from a larger free-context area. Check your own account and the cost documentation before making a billing decision.

For a different task, BIG CHANGE has covered building a Claude Code context display mod and auditing a completed coding-agent transcript with AgentsView.

Sources & further reading