Context Bound

Context bound

Compaction is off in the settings snippet, so a session that reaches its limit stops and loses its state. This page says how dotclaude bounds the context of each subagent. The decision is D22 in Decisions. The measures are in Usage evidence.

Why

• Investigator runs, and not the start of a session, used most of the quota in the measured sessions. Each run paged files into its context and ended at 110K to 223K tokens (measured, research/subagent-usage.md).
• The main session grows too. In four main sessions, 48 of 86 file reads (56%) read a file again, and almost all were slices through Bash (measured, research/session-growth.md).
• A bound in prose did not hold. The 0.19 output style already told the model to read only the part of a file that it needs.

Bounds

All bounds are in plugins/dotclaude/features/context/limits.mjs, because only the context feature reads them. Tests pin each copy of a bound in code and config.

Bound	Value	Use
SUBAGENT_CONTEXT_MAX	100,000 tokens	A subagent gets the report-now message from this context.
HAIKU_CONTEXT_MAX	85,000 tokens	The same for a Haiku 5.5 subagent.
REPORT_MAX_WORDS	400	The most words in the final report of an agent.

Haiku 5.5 costs 5 times as much for each request above 100K prompt tokens. HAIKU_CONTEXT_MAX leaves room for the report turn below that step.

File reads through Read

• Each agent that reads files tells the model to search first. The output style of 0.28.0 has no read rule, and the tools shared section and the reads hook below hold it. The model then reads only the matched range with Read, and refers to an earlier result instead of a second read of the same lines.
• Read returns a short stub when the same file and the same offset and limit are read again and the file did not change. A Bash slice gets no stub.
• The PreToolUse hook features/reads/cli.mjs denies a Bash command that only prints a file. This covers cat, sed -n with a print range, head, and tail on one file, with no pipe and no other command. The deny reason gives the Read call with the same file and range. For example, sed -n '40,80p' design.md gets Read with offset 40 and limit 41.
• The snippet sets bashOutputMaxChars to 4000, the lowest value. Claude Code saves a longer output to a file and sends a preview and the path.

Subagent report at the bound

A classic PostToolUse hook (features/context/cli.mjs) reads the transcript of the subagent. When the context passes the bound of its model, the hook gives additionalContext that tells the subagent to report now.

• The hook tells and does not block, because a block loses the work of the run.
• Below the bound, and in the main session, the hook prints nothing.
• Each agent body names REPORT_MAX_WORDS, because the main session keeps each report to the end.
• A run at maxTurns returns no report, so each maxTurns holds one spare turn (Agents).
• The hook counts the turns of the subagent in its transcript and reads maxTurns from the agent file. With TURNS_LEFT_AT (2) turns left, it gives one note that tells the subagent to report now, joined with the context message when both apply.
• A resume after the limit starts a new count. The delegation text therefore tells the coordinator to take the report of a stopped run and not to resume it.

Main session

• The main session gets no context note. Compaction is off, and the context_management text of features/prompt/data.mjs tells Claude that and to keep the state of work in files.
• The status line may show the figure to the user.
• No prompt text tells the model to estimate or announce its context share. A model can report a false state, so the figure comes from a program (D27 in Decisions). A test reads the output style, the shared sections, the prompt variants, and the agent bodies for this.

Usage report

just measure reports the context of a session from its transcripts:

• For each subagent: the requests, the final context, and the total input tokens.
• For the main session: the tool-result size for each tool, and the count of repeat reads of the same file.

See Usage evidence.

Related pages

• Agents
• Usage evidence
• Prompt surface