Compaction And Context

Compaction and Context

This page describes what fills the context of Claude Code, when it compacts the context, and what a compaction keeps. Read it if you tune the context settings of claudetto. It covers Claude Code 2.1.296. Each fact from the binary comes from a search of its strings, or from the official docs. The page names the source of each fact. Each inference is marked as inference.

Window Sizes

• The default window is 200,000 tokens. Source: the default window constant in the 2.1.296 strings.
• A model ID with the [1m] suffix selects a window of 1,000,000 tokens. Source: the 2.1.296 strings match the suffix, and they offer sonnet[1m] for a 1M window.
• CLAUDE_CODE_DISABLE_1M_CONTEXT turns off the 1M choice. Source: the 2.1.296 strings.
• claude-sonnet-5-5 had a window of 1,000,000 tokens in the measurement of 2026-10-12. Source: Handoff Cost.
• The comment on CONTEXT_WINDOW in src/kernel/budget.ts says that Claude Code 2.1.302 warns against a 200K window for claude-opus-5-5. It warns unless autoCompactWindow is set. I did not check this comment in the 2.1.296 strings. The strings have a warning for an unknown model id. That warning names CLAUDE_CODE_MAX_CONTEXT_TOKENS and CLAUDE_CODE_DISABLE_UNKNOWN_MODEL_WINDOW_ENFORCEMENT.
• The sizes of the other models are not settled, so this page gives no others.

What Fills the Context

The official docs list these items, in the order that a session loads them.

• The system prompt, which the user does not see.
• The CLAUDE.md files of the project and the user.
• Auto memory, which is the first 200 lines or 25 KB of MEMORY.md.
• Environment information, such as the working directory and the platform.
• The names of MCP tools. By default, the schemas stay deferred until a tool search loads them.
• The one-line descriptions of the skills. The body of a skill loads only when Claude uses it.
• The messages of the conversation, the tool calls, the tool output, and the file reads.

The docs give a token count for some items as an illustration. This page does not repeat those counts, because they are not measured.

When Auto-Compaction Starts

The threshold is the effective window minus 13,000 tokens. The effective window is the window minus the smaller of the output limit and 20,000 tokens. Source: the threshold function and the 20,000 constant in the 2.1.296 strings. The threshold is an inference from these two formulas, and I did not measure it.

• With a 200,000-token window and an output limit of at least 20,000 tokens, the threshold is 167,000 tokens (inference).
• With the 300,000-token window of the claudetto snippet, the threshold is 267,000 tokens (inference). The snippet also sets CLAUDE_AUTOCOMPACT_PCT_OVERRIDE, which gives a lower threshold (see the section "Context in claudetto").
• CLAUDE_AUTOCOMPACT_PCT_OVERRIDE sets a percentage of the effective window from 0 to 100. The threshold is then the lower of that value and the default threshold.
• The 2.1.296 strings also have a second threshold, from a field named precomputeBufferFraction. It is the window minus a fraction of the window. The trigger takes the lower of that value and the threshold above, so it can fire earlier. The fraction comes from a model table. This page does not examine that second threshold.
• The next request is blocked at the effective window minus 3,000 tokens. With the default output limit, that is the window minus 23,000 tokens (inference). CONTEXT_BLOCK_MARGIN in src/kernel/budget.ts has the same value, 23,000. Its source is 20,000 plus 3,000.

Settings and Env Vars

Name	Kind	Effect	Source
autoCompactEnabled	settings key, boolean, default true	Turns auto-compaction on or off.	Settings schema, dossier/cc-2.1.296/settings.md
autoCompactWindow	settings key, 100000 to 1000000 or "auto", per model	Sets the window for auto-compaction of one model.	Settings schema description
CLAUDE_CODE_AUTO_COMPACT_WINDOW	env var	Sets the window for auto-compaction, and takes precedence over autoCompactWindow.	2.1.296 strings, dossier/cc-2.1.296/env.md
CLAUDE_AUTOCOMPACT_PCT_OVERRIDE	env var	Sets the threshold as a percentage of the effective window.	2.1.296 strings
DISABLE_AUTO_COMPACT	env var	Turns off auto-compaction.	2.1.296 strings
DISABLE_COMPACT	env var	Turns off auto-compaction and /compact.	2.1.296 strings, dossier/cc-2.1.296/env.md
CLAUDE_CODE_MAX_CONTEXT_TOKENS	env var	Sets the window of the model, with DISABLE_COMPACT or as a fallback for an unrecognized model id.	2.1.296 strings, dossier/cc-2.1.296/env.md

The 2.1.296 strings read CLAUDE_CODE_MAX_CONTEXT_TOKENS in two places. A search of external/cc-2.1.296/strings.txt for the name and its nearby code found both. One read sets the window only when DISABLE_COMPACT is set. The other read is a fallback for a model id. The name check does not take that id as a claude- id. The fallback runs after the model table and the 1M checks. It does not need DISABLE_COMPACT. So the comment in src/kernel/budget.ts is incomplete for 2.1.296. That comment says that Claude Code honors the variable only with DISABLE_COMPACT.

What a Compaction Keeps and Drops

A compaction asks the model for a summary of the conversation. The summary replaces the earlier messages. The summary prompt in the 2.1.296 strings asks for these sections.

1. The primary request and intent.
2. The key technical concepts.
3. The files and code sections.
4. The errors and fixes.
5. The problem solving.
6. All user messages that are not tool results.
7. The pending tasks.
8. The current work.
9. The next step.

The binary has two variants of this prompt.

• A partial summary covers only the recent messages. The earlier messages stay intact.
• A shorter prompt asks for a summary of 2,000 words or fewer. The strings do not say when Claude Code uses it.

The official docs list what the context holds after a compaction.

• The system prompt, CLAUDE.md, auto memory, and the names of the MCP tools reload.
• Claude Code re-reads up to five of the files that were modified most recently.
• Claude Code re-injects the body of each invoked skill, up to 5,000 tokens for each skill.
• The skill descriptions do not reload.

The messages, the tool output, and the file contents that are not re-read leave the context (inference from the docs).

/compact and /clear

• /compact starts a compaction now. The strings describe the command as a way to free context by summarizing the conversation so far. DISABLE_COMPACT removes the command.
• /clear ends the conversation and starts a new one with no earlier messages. The strings suggest /clear for a fresh start when a file is too large for the window.
• /compact keeps a summary of the conversation. /clear keeps nothing of it (inference from the two commands).

Context in claudetto

• The settings snippet sets autoCompactWindow to 300000. It sets autoCompactWindow to 144444 for claude-haiku-5-5 in modelSettings. It sets CLAUDE_AUTOCOMPACT_PCT_OVERRIDE to 80.3571. Source: the settings snippet, and the constants in src/kernel/budget.ts.
• The percent and the window put the start of the auto compaction at 225,000 tokens for a 300,000-token window. The comment on AUTOCOMPACT_PCT in src/kernel/budget.ts gives the formula (Claude Code 2.1.302 bundle). The snippet does not set DISABLE_COMPACT.
• The prompt text of the context_management section says that the conversation is summarized. The prompt slice replaces that text with a text that agrees with the compaction state of the session. The states are handoff, built-in, and off (DISABLE_COMPACT is on). Source: src/plugins/claudetto/slices/prompt/domain/compaction.ts, src/plugins/claudetto/slices/prompt/transforms/context_management.ts, and the Prompt page.
• The handoff is the alternative to the summary of a compaction. The handoff slice writes the state of the work to a file when the compaction starts. The handoff replaces the summary, and the next turns read the file. See Handoff.
• One measured run on 2026-10-12 compared the two, with the old /dotclaude:handoff of dotclaude 1.0.0. A handoff cost $0.0630. A compaction with its follow-up turn cost $0.2288. Both costs are at the API list price. The run used one session of about 81,000 tokens, so it has no spread. See Handoff Cost.
• The bounds of the context of a subagent are in Handoff.

Sources

• The 2.1.296 strings, searched with grep for the names and the constants above.
• The extracts of the compaction prompts in external/cc-2.1.299/cc-extract-2.1.299/compaction/.
• The dossier of the binary, external/dossier/cc-2.1.296/env.md and settings.md.
• The official docs page "Explore the context window" at code.claude.com/docs/en/context-window, read on 2026-10-13.
• The kernel constants in src/kernel/budget.ts, and the settings snippet.

Not Settled

• The window of each model, except the three cases above.
• The threshold for a model whose output limit is below 20,000 tokens.
• The exact content that a compaction keeps, which only the docs describe.

Related Pages

• Handoff
• Auto Mode