Undocumented Reads

Undocumented reads

Claude Code documents most data that dotclaude uses. A few reads use data that Claude Code does not document. Claude Code can change or remove such data in any release, and it gives no notice. This page lists each read, its source, what shows when the read breaks, and how to find a break. The evidence is finding F4 in research/audit-plugins.md of the change dotclaude-0-28-0.

What a break looks like

Each read fails closed. A parser that does not know the shape returns null, and the part that needs the data does not show. Most breaks are silent: the status line has no segment, or the figure is less exact. Only the drift check is visible, and then only for a change that it knows to look for. No read here blocks a tool call or changes a file.

Summary

Read	Used by	Source	Break
Credits segment	Status line	cachedUsageUtilization in <config dir>/.claude.json, key cedar_ember	The segment disappears.
/context scrape	Status line	The Markdown table that /context writes into the transcript	The breakdown shows only the free tokens, or a wrong messages figure.
Drift section ids	Drift check	The section ids of prompt.compose in the bundle of Claude Code 2.1.296	A new id gives a line in the drift note at each session start.
Built-in plugin names	Drift check	Internal cc-plugin-*@builtin names in the bundle	A renamed plugin gives a false line or no line.
Tested version	Drift check	TESTED_VERSION in features/drift/data.mjs	Another version gives a line at each session start.
Prompt and tool texts	Surface cuts	Hand copies of Claude Code texts	A changed text makes the cut do nothing.
Context margin	Status line	CONTEXT_BLOCK_MARGIN of 23,000 tokens, from the 2.1.296 bundle	The free-token figure is off.

The credits segment

The status line shows the count of reset credits that the user can spend, and the next expiry. It also shows the weekly window of Fable from the same file.

• Source: the file <config dir>/.claude.json, key cachedUsageUtilization. The config dir is CLAUDE_CONFIG_DIR, or the home folder. The credits are in utilization.cedar_ember, and the Fable window is in utilization.limits[].
• Code: plugins/dotclaude/features/statusline/credits.mjs. The read opens the file one time for each status line run, and it keeps only these two parts.
• Evidence: the schema of the cache in Claude Code 2.1.296 has no cedar_ember key, but it passes unknown keys through. The grant fields (resets_left, paused, usable_now, starts_at, ends_at) are in the grant schema of the bundle. The first decoder of the shape was the one of CodexBar 0.69.0.
• Checks that the parser makes: eligible is true, grants is an array, and each counted grant is not paused, is usable_now, has resets_left above 0, and is inside its dates. A cache that is older than 1 hour (CREDITS_MAX_AGE_MS) counts as absent.

How it breaks at a Claude Code release:

Change in Claude Code	Result
The key cedar_ember is renamed or removed	The credits segment does not show.
A grant field is renamed or its type changes	The grant does not count, and the segment shows fewer credits or none.
The cache moves to another file	The segment does not show.
The cache is written less often than every hour	The segment does not show most of the time.
limits[] changes its kind or scope shape	The Fable window does not show.

A test pins the known shape with a fixture, and no test reads a real cache, so just check does not find a break in Claude Code. To check by hand, read cachedUsageUtilization in your .claude.json and compare the keys with the list above. The rate limits of the 5-hour and 7-day windows are not in this read. They come from rate_limits in the status JSON, which Claude Code documents.

The /context scrape

The status line shows an exact breakdown of the context (system, memory, and skills) only after the user ran /context.

• Source: /context writes its Markdown table into the transcript as a user record whose text starts with ## Context Usage. Its rows come from the count-tokens API.
• Code: plugins/dotclaude/features/statusline/context-table.mjs. Each status line run reads the last 1,000,000 bytes of the transcript (TABLE_TAIL_BYTES_MAX), keeps no state, and takes the newest table.
• The parser reads only the first table, with a regex for a row and a map of category names: System prompt, System tools, Custom agents, Memory files, and Skills. System tools (deferred) is not a part, because deferred tools are not in the context.
• A subagent record (isSidechain) is skipped.

How it breaks at a Claude Code release:

Change in Claude Code	Result
The header ## Context Usage changes	No table is found, and the breakdown shows only the free tokens.
Every category name changes	The same as above, because no row counts.
One category name changes	That part counts as 0 and the messages figure grows by the same size. The figure is wrong and nothing reports it.
The row format changes, such as the k suffix	The same as one changed name.
/context stops writing to the transcript	No table is found.

The table is as old as the last /context, so the figure ages while the session goes on. A documented source exists in the hooks module: $.session.usage({ breakdown: "summary" }), which the drift check uses. The status line is a separate process and has no $, so it cannot use that call. The audit gives this as the way to drop the scrape if the status line moves into the module.

The drift section ids

The drift check reports a section in the system prompt that dotclaude does not know.

• Source: ENGINE_SECTIONS in plugins/dotclaude/features/drift/data.mjs. It lists 33 section ids of Claude Code 2.1.296, such as intro, tools, context_management, heron_brook, tengu_silent_harbor, and willow_tern. The ids come from the section list in the bundle, because no document lists them. The ids after memory come only on some conditions, such as a flag, a model, or a deferred tool.
• Code: unknownSections in features/drift/rules.mjs takes the ids from a prompt.compose call with the bare trait. An id with a colon belongs to a plugin, so it does not count.

How it breaks at a Claude Code release:

Change in Claude Code	Result
A new section id	The note says that the system prompt has a section with an id that dotclaude does not know. The line comes at each session start until a maintainer adds the id. The model can pass it to you as noise.
An id is removed or renamed	No line for the removed id. A renamed id gives a line for the new name.
prompt.compose changes its result shape	The read fails, the fact gives no line, and the check goes on with the other facts.

The line is a hint and not an error. dotclaude keeps an unknown section as Claude Code made it.

A second list, KNOWN_SECTIONS in features/prompt/select.mjs, holds the ids that a prompt variant replaces: intro, system, doing_tasks, actions, tools, tone, and lean_body. composeSections replaces only these ids, and passes any other id on unchanged.

Change in Claude Code	Result
A replaced id is renamed	The old text stays next to the variant, and the drift note names the new id

The same data file holds two more facts of this kind:

• OFF_PLUGINS lists 9 internal cc-plugin-*@builtin names. The check reports each one that your enabledPlugins does not set to false. A renamed plugin gives a line for the old name until the list changes.
• TESTED_VERSION is 2.1.296. Any other version gives one line at each session start, and each Claude Code release needs a code change to stop it. A test pins the value to the schema file in schemas/.

Other texts that follow Claude Code

These are not reads of stored data, but they depend on text that Claude Code does not document.

Text	Code	Break
The Agent, Bash, and Read tool descriptions	features/surface/data.mjs (TOOL_TEXT)	A handler replaces a description only when its own text is shorter. A long new description from Claude Code is cut to the old copy, and a short one stays.
The agent_listing_delta sentence about parallel agents	features/surface/rules.mjs	An exact string replacement. A changed sentence makes it do nothing.
The Recent commits: block of the session_context attachment	features/surface/rules.mjs	A regex. A changed layout makes it do nothing.
The 23,000-token block margin	lib/budget.mjs (CONTEXT_BLOCK_MARGIN)	The margin is the 20,000-token output reserve plus 3,000 tokens in the 2.1.296 bundle. A change moves the free-token figure.
Model names such as haiku-5-5 and opus-5-5	lib/model.mjs, features/context/rules.mjs, features/statusline/limits.mjs	A new model name takes the default subagent context bound.

A cut that does nothing is silent by design: it passes the text on with no change. The effect is only a larger request. The lab tests in plugins/dotclaude/tests/surface.test.ts run the handlers with stubbed engine texts, so they do not see a change in the real text.

What to do after a Claude Code update

1. Read the drift note in the first answer of a new session. The note goes to the model, and it tells the model to report the differences to you and to change no setting. It names a Claude Code version that is not the tested version, the built-in plugins that are on, the synced skills, the settings that stop the agent files, and the ids of the sections that dotclaude does not know.
2. If the status line lost the credits segment or the context breakdown, read the matching table above.
3. Report a break that you cannot fix in the issues of the repository, with the Claude Code version.

Related pages

• Status line
• Core plugin
• Hooks
• Troubleshooting
• Claude mods