Probes

A probe is a small test that runs Claude Code 2.1.305 and checks what one setting changes. The probes ran against a local stub of the Messages API, so they cost nothing and sent no data to Anthropic. They show which environment variables and settings have an effect that you can see.

How We Probed

1. We ran the unmodified 2.1.305 program with claude -p "Reply with 'OK'." --output-format stream-json --verbose. Every probe used that prompt.
2. We gave each run a fresh config folder, a fresh home folder, a fake API key, and a base URL that points to the local stub.
3. We turned off telemetry and other optional traffic. We pointed the proxy settings to a closed port, so a stray request to api.anthropic.com fails.
4. The stub answered POST /v1/messages as a stream or as JSON. It sent one scripted tool call and then the plain text OK. It wrote the method, path, headers, and body of each request to a log.
5. We made two baselines. The first was a plain run. The second ran a Bash tool call with command hooks on five events and a local plugin with a skill, a hook, and an MCP server. Two repeat runs of each baseline were identical after we removed ids, paths, and times.
6. We set each of 1,506 variable names to 1 in both baselines, which gave 3,012 runs. The names came from the program itself. We listed every upper-case word in its strings that starts with ANTHROPIC_, API_, BASH_, CLAUDE_, DISABLE_, ENABLE_, IS_, MAX_, MCP_, NO_, OTEL_, or USE_ and ends in a letter or digit. That gave 1,499 words, and we added 7 cut-off prefixes that end in _ (they had no effect). Not every word is an environment variable that the program reads: the list also has shell variables such as BASH_ARGV. Names with other prefixes were not swept. Each run got an explicit environment with only the probe setup and the name under test, so no variable of the parent shell reached the program. One name, ANTHROPIC_API_KEY, was dropped, because 1 replaced the fake key. We compared headers, betas, the system prompt, tools, body fields, output events, files in the config folder, the hook log, and the environment names that hooks saw.
7. Some names got extra runs: auto mode with a Bash call, --brief, --resume, and the plugin variables with a plugin. --brief changed nothing that we compared, in either baseline. --resume with the id of an earlier run loaded that run: the request had both prompts and the first reply, and the session id stayed the same.

Every request carried the fake key. The only path called was /v1/messages?beta=true, except with CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY, which also called /v1/models?limit=1000.

Results

Of the 1,506 names, 108 had an effect that we could see: 53 are

Undocumented: Only the program shows this item. It can change without notice.

in the Environment Variables pages, 2 (CLAUDE_CODE_HOST_HTTP_PROXY_PORT and DISABLE_COLORS) have no row there, and 53 are documented there. The rest had none. We did not count 13 names whose only difference was in the environment names that hooks saw, because the program sets those names itself. The exception is CLAUDE_CODE_SUBPROCESS_ENV_SCRUB: its effect is real (hooks lost ANTHROPIC_API_KEY, CLAUDE_CONFIG_DIR, GIT_CONFIG_GLOBAL, and CLAUDE_CODE_MESSAGING_TOKEN), and the Static and Print row covers it. We also did not count ANTHROPIC_BASE_URL, NO_PROXY, and CLAUDE_CONFIG_DIR, which the probe setup sets itself. No visible effect in print mode does not prove that a variable does nothing. Many variables work only in other modes, in remote sessions, or with a login. This page lists the undocumented names with a clear effect, the two documented plugin variables, and the items that differ from the docs. The other documented names with an effect, and all names with no visible effect, are left out.

Effects and Differences

An item is

Differs: The official docs and the program disagree. The row says how.

when the result differs from the docs. The status of an environment variable is the one on its row in the Environment Variables pages. This table records only what the probe saw.

Name	What We Set	Status	What We Saw
--model X[1m]	model flag	Undocumented: Only the program shows this item. It can change without notice.	context-1m-2025-08-07 is added for sonnet[1m], opus[1m], and every claude-*[1m] tried, also for models with a native 1M window. Without [1m] it is never sent.
hook type: "script"	settings	Differs: The official docs and the program disagree. The row says how.	Rejected at load. Under PostToolUse the text is Unknown hook type "script"; entry ignored, and the other hooks in the file still ran. Under PreToolUse the text is Unknown hook type "script" — a PreToolUse/PermissionRequest hook that cannot be loaded. No script ran. A PreToolUse entry also disables the whole settings file it sits in (the SessionStart and PostToolUse command hooks in the same file did not run).
CLAUDE_AGENT_SDK_VERSION	1	Undocumented: Only the program shows this item. It can change without notice.	Adds agent-sdk/1 (the value) to the user agent.
CLAUDE_CODE_ACTION	1	Undocumented: Only the program shows this item. It can change without notice.	Sets the entrypoint to claude-code-github-action in the billing header and the user agent, and adds claude-code-guide to the agent list.
CLAUDE_CODE_ADDITIONAL_PROTECTION	1	Undocumented: Only the program shows this item. It can change without notice.	Adds the request header x-anthropic-additional-protection.
CLAUDE_CODE_AMBER_ASTROLABE	1	Undocumented: Only the program shows this item. It can change without notice.	Adds an "operating autonomously" block to the system prompt (do not ask "Shall I...?", finish the work before ending the turn).
CLAUDE_CODE_ATTACHMENTS_BY_PATH	1	Undocumented: Only the program shows this item. It can change without notice.	Adds the SendUserFile tool.
CLAUDE_CODE_BASALT_COVE	1	Undocumented: Only the program shows this item. It can change without notice.	Adds a "Communicating with the user" section and a code-comment rule to the system prompt.
CLAUDE_CODE_BISON_CAIRN	1	Undocumented: Only the program shows this item. It can change without notice.	Adds a "Delivering work" section to the system prompt.
CLAUDE_CODE_BRIDGE_CHILD_AUTO_DEFAULT	1	Undocumented: Only the program shows this item. It can change without notice.	Starts in auto permission mode (permissionMode in init). Adds the betas dangerous-tool-use-2026-09-03 and afk-mode-2026-01-31, a safeguards body field, an auto mode reminder in the first message, and a system/informational event.
CLAUDE_CODE_BRIEF	1	Undocumented: Only the program shows this item. It can change without notice. Gated: Works only behind a flag or a server gate. Works only behind a feature flag or server gate	Adds a "Talking to the user" section and the SendUserMessage tool. The reply goes through that tool, so the output has an extra tool turn.
CLAUDE_CODE_CONTAINER_ID	1	Undocumented: Only the program shows this item. It can change without notice.	Adds the request header x-claude-remote-container-id.
CLAUDE_CODE_COORDINATOR_MODE	1	Undocumented: Only the program shows this item. It can change without notice. Gated: Works only behind a flag or a server gate. Works only behind a feature flag or server gate	The tool list shrinks to Agent, ListAgents, SendMessage, Skill, TaskStop, and Workflow. The only agent is worker. The beta advanced-tool-use-2025-11-20 is not sent, and the first message lists the worker tools.
CLAUDE_CODE_DISABLE_CLAUDE_API_SKILL	1	Undocumented: Only the program shows this item. It can change without notice.	Removes the claude-api skill from the skill list and the slash commands.
CLAUDE_CODE_DISABLE_REFUSAL_RETRY	1	Undocumented: Only the program shows this item. It can change without notice.	Adds the beta fallback-credit-2026-06-01.
CLAUDE_CODE_EMIT_STARTUP_TIMING	1	Undocumented: Only the program shows this item. It can change without notice.	Adds startup_timing to the init event.
CLAUDE_CODE_EMIT_TOOL_USE_SUMMARIES	1	Undocumented: Only the program shows this item. It can change without notice.	Adds tool_use_summary events after a tool runs.
CLAUDE_CODE_ENABLE_EXPERIMENTAL_ADVISOR_TOOL	1	Undocumented: Only the program shows this item. It can change without notice. Gated: Works only behind a flag or a server gate. A feature flag or server gate decides by default. This variable overrides it Experimental: The program or the official docs call this experimental or beta.	Adds the beta advisor-tool-2026-03-01 and an advisor slash command.
CLAUDE_CODE_ENABLE_REFRESH_MCP_TOOLS	1	Undocumented: Only the program shows this item. It can change without notice.	Adds the RefreshMcpTools tool (seen with a plugin MCP server configured).
CLAUDE_CODE_ENABLE_TOKEN_USAGE_ATTACHMENT	1	Undocumented: Only the program shows this item. It can change without notice.	Adds a line Token usage: 0/980000; 980000 remaining to the first message.
CLAUDE_CODE_ENTRYPOINT	1	Undocumented: Only the program shows this item. It can change without notice.	Sets the entrypoint in the billing header and the user agent to the value (1 here), and adds claude-code-guide to the agent list.
CLAUDE_CODE_FOOTER_INDICATOR	1	Undocumented: Only the program shows this item. It can change without notice.	Adds footer_indicator to the init event.
CLAUDE_CODE_HUMBLE_HAMMOCK	1	Undocumented: Only the program shows this item. It can change without notice. Gated: Works only behind a flag or a server gate. A feature flag or server gate decides by default. This variable overrides it	Adds wire_tool_inputs (raw tool_use.input) to the assistant event when a tool runs.
CLAUDE_CODE_INCLUDE_PARTIAL_MESSAGES	1	Undocumented: Only the program shows this item. It can change without notice.	Adds stream_event events and a system/status event to the output.
CLAUDE_CODE_INLINE_TOOLS	1	Undocumented: Only the program shows this item. It can change without notice. Gated: Works only behind a flag or a server gate. Works only behind a feature flag or server gate	Adds the beta inline-tools-2026-09-15 to anthropic-beta.
CLAUDE_CODE_LANTERN_PRISM	1	Undocumented: Only the program shows this item. It can change without notice. Gated: Works only behind a flag or a server gate. Works only behind a feature flag or server gate	Adds the skill-doctor skill to the init event.
CLAUDE_CODE_LARCH_CISTERN	1	Undocumented: Only the program shows this item. It can change without notice. Gated: Works only behind a flag or a server gate. A feature flag or server gate decides by default. This variable overrides it	Adds a "Corrections" section to the system prompt (avoid excessive self-correction).
CLAUDE_CODE_PEWTER_OWL	1	Undocumented: Only the program shows this item. It can change without notice. Gated: Works only behind a flag or a server gate. A feature flag or server gate decides by default. This variable overrides it	Adds the SendUserMessage tool and a "Talking to the user" section. The reply goes through that tool, so the output has an extra tool turn.
CLAUDE_CODE_PEWTER_OWL_TOOL	1	Undocumented: Only the program shows this item. It can change without notice. Gated: Works only behind a flag or a server gate. A feature flag or server gate decides by default. This variable overrides it	Adds only the SendUserMessage tool.
CLAUDE_CODE_PLUGIN_DIRS	a plugin folder path	Documented: An official docs page names this item.	Loads the plugin folder like --plugin-dir (the plugin is in the plugins list in init). The value 1 changed nothing.
CLAUDE_CODE_PROFILE_STARTUP	1	Undocumented: Only the program shows this item. It can change without notice.	Writes a .json and a .txt file under startup-perf/ in the config folder.
CLAUDE_CODE_REMOTE_MEMORY_DIR	1	Undocumented: Only the program shows this item. It can change without notice.	Changes the memory folder in the system prompt to the value plus /projects/ and the project folder name (1/projects/... here).
CLAUDE_CODE_SDK_READS_SESSION_STATE	1	Undocumented: Only the program shows this item. It can change without notice.	Adds system/session_state_changed events (with sdk_host_only).
CLAUDE_CODE_SDK_READS_STARTUP_MARKS	1	Undocumented: Only the program shows this item. It can change without notice.	Adds keep_alive events with a startup_mark field.
CLAUDE_CODE_SIMULATE_PROXY_USAGE	1	Undocumented: Only the program shows this item. It can change without notice.	Removes the whole anthropic-beta header and the thinking.display field.
CLAUDE_CODE_SKILL_PROPOSALS	1	Undocumented: Only the program shows this item. It can change without notice.	Adds a "Saving skills" section to the system prompt (skill files are a read-only cache; point the user to claude.ai settings).
CLAUDE_CODE_SKIP_PLUGIN_LOCAL_MCP_SERVERS	1	Undocumented: Only the program shows this item. It can change without notice.	Plugin MCP servers are not started. mcp_servers is empty and the tool list shrinks.
CLAUDE_CODE_SKIP_PLUGIN_MCP_SERVERS	1	Undocumented: Only the program shows this item. It can change without notice.	Plugin MCP servers are not started. mcp_servers is empty and the tool list shrinks.
CLAUDE_CODE_SYNC_PLUGIN_INSTALL	1	Documented: An official docs page names this item.	Emits plugin_install events (started, completed) in stream-json.
CLAUDE_CODE_TETHER_LIVE	1	Undocumented: Only the program shows this item. It can change without notice. Gated: Works only behind a flag or a server gate. A feature flag or server gate decides by default. This variable overrides it	Adds the betas mid-conversation-system-clear-at-2026-08-21 and message-threads-2026-08-12 and a thread body field.
CLAUDE_CODE_TOTAL_TOKENS_REMINDER_BUDGET	1	Undocumented: Only the program shows this item. It can change without notice.	Sets the number in the total_tokens reminder (1 tokens left here).
CLAUDE_CODE_TURN_UPDATES	1	Undocumented: Only the program shows this item. It can change without notice.	Adds guidance to say what you are about to do and to end with a recap. The line "Write code that reads like the surrounding code..." is gone.
CLAUDE_CODE_USE_COWORK_PLUGINS	1	Undocumented: Only the program shows this item. It can change without notice.	User settings come from cowork_settings.json, so its hooks and env apply and those of settings.json do not. Plugins and marketplaces use cowork_plugins. claude plugin install writes there.
CLAUDE_CODE_WILLOW_TERN	1	Undocumented: Only the program shows this item. It can change without notice. Gated: Works only behind a flag or a server gate. A feature flag or server gate decides by default. This variable overrides it	Adds a "Writing for the user" section to the system prompt.
CLAUDE_COWORK_MEMORY_EXTRA_GUIDELINES	1	Undocumented: Only the program shows this item. It can change without notice.	Appends the value to the system prompt.
CLAUDE_COWORK_MEMORY_GUIDELINES	1	Undocumented: Only the program shows this item. It can change without notice.	Replaces the memory section of the system prompt with # auto memory and the value.
CLAUDE_COWORK_MEMORY_INDEX_CONTENT	1	Undocumented: Only the program shows this item. It can change without notice.	Adds a message that shows the value as the content of the MEMORY.md auto-memory file.
CLAUDE_MEMORY_STORES	1	Undocumented: Only the program shows this item. It can change without notice.	Changes the memory section of the system prompt: it adds a 4 KB limit per memory file and drops the MEMORY.md pointer rule.
CLAUDE_PROJECT_UUID	1	Undocumented: Only the program shows this item. It can change without notice. Gated: Works only behind a flag or a server gate. Works only behind a feature flag or server gate	Adds the Projects tool.

Ten undocumented names have no row in this table. Five ended the run with the value 1, because they need a real value: CLAUDE_CODE_BLOCKING_LIMIT_OVERRIDE (result Prompt is too long, exit code 1), CLAUDE_CODE_CLIENT_DATA_URL (it must be the https://downloads.claude.ai/ address), CLAUDE_CODE_CUSTOM_OAUTH_URL (exit code 1), CLAUDE_CODE_REMOTE_TOOLS_PIN_STORED_LOGIN (exit code 78), and CLAUDE_CODE_USE_ANTHROPIC_GOOGLE_CLOUD (no workspace ID). ANTHROPIC_LOG printed that 1 must be one of off, error, warn, info, or debug. CLAUDE_CODE_HOST_HTTP_PROXY_PORT, CLAUDE_CODE_SILENT_TURN_REMINDER_TURNS, and DISABLE_COLORS changed one baseline in a way that we could not tie to the name, so we did not describe them. CLAUDE_CODE_WEB_FETCH_AGENT is in Subagents and Hooks.

Request Details

• cc_version=2.1.305.b57: the suffix b57 stayed the same in all runs with the prompt Reply with 'OK'., on both models tried. Auto mode classifier requests carried a different suffix, 034.
• /v1/messages/cache_touch: never called. On 2.1.305 with a real CLAUDE_CODE_OAUTH_TOKEN, claude -p "Reply with 'OK'." got no anthropic-cache-keepalive header, the request had no cache-keepalive beta, and no cache_touch call was made (seen through a local proxy that logged only paths, and a debug file). The program sends it only when the response has anthropic-cache-keepalive: active for a request with the cache-keepalive-2026-09-03 beta. That beta needs a 1-hour cache TTL for a subscriber login, allow_cache_keepalive, and a server-side flag. It is also off when ANTHROPIC_UNIX_SOCKET is set. The body is {"request_id": ...}, and the response needs a request id. Touches repeat every 30 s by default (10 to 120). We ran the stub with that response header and a request id, with and without CLAUDE_INTERNAL_FC_OVERRIDES setting the flags. The program never sent the beta and never called the path, because the probe had no subscriber login. The path, the header name, the body field, and the 30 s default with its 10 to 120 range are in the program's code.
• context-1m-2025-08-07: added by a [1m] suffix, also on models with a native 1M window. Not sent without the suffix. The default model in this setup was claude-opus-5-5 and it did not send the beta.
• Base betas in the plain run: claude-code-20250219, interleaved-thinking-2025-05-14, thinking-token-count-2026-05-13, context-management-2025-06-27, per-turn-control-2026-07-01, mid-conversation-tool-changes-2026-07-01, advanced-tool-use-2025-11-20, effort-2025-11-24, thinking-display-updates-2026-08-18. Up to 2.1.306 the list also had prompt-caching-scope-2026-01-05. 2.1.307 no longer sends it behind a custom ANTHROPIC_BASE_URL, such as the stub. A first-party connection adds mid-conversation-system-clear-at-2026-08-21, thinking-binding-controls-2026-08-01, and cache-diagnosis-2026-04-07, and a diagnostics body field. We tested that with the base URL http://api.anthropic.com and ANTHROPIC_UNIX_SOCKET pointing to a local socket, so nothing left the machine.

Auto Mode

With --permission-mode auto and a Bash call, the stub got classifier requests with xml_s1 and xml_s2 stages. There were five requests of each stage, because the stub's OK reply makes the program retry. The debug log says server_no_result and falls back to billed local classifier requests. CLAUDE_CODE_TWO_STAGE_CLASSIFIER=1 and CLAUDE_CODE_AUTO_MODE_SERVER=1 changed nothing. CLAUDE_CODE_AUTO_MODE_SERVER=0 removed the server_no_result line from the debug log, and the stub still got the same ten requests. These four runs gave the same results on 2.1.306, 2.1.307, 2.1.308, and 2.1.309.

autoMode.classifyAllShell

We ran claude -p --permission-mode auto against the stub with the command curl -s http://127.0.0.1:9/x and a user allow rule Bash(curl -s http://127.0.0.1:9/x). The table shows whether the stub got requests that carried the classifier system prompt: 1 means the classifier ran, 0 means it did not. A classifier run sent about 10 requests per Bash call, because the stub's canned reply makes the program retry. A managed folder needed sudo, so the user ran the managed cases on 2.1.306 (managed cases classify-*): managed true, managed false, and managed unset, each in all 7 settings combinations.

Settings	Classifier Requests
No allow rule	1
Allow rule, flag unset	0
Allow rule, user true or --settings true	1
Allow rule, user false or --settings false	0
User false and --settings true, or the reverse	1
Managed false, with user or --settings true	1
Managed false, with user or --settings false, or neither	0
Managed true, with user false, --settings false, or neither	1

Any true in user, --settings, or managed settings turns the flag on. false is the same as unset. The program reads no project or local source: true in project settings or in local settings gave 0 classifier requests.

Subagents and Hooks

Follow-up probes ran 2.1.305 against the same stub, with a fake key, a temp home folder, and -p "Reply with 'OK'." --output-format stream-json --verbose. The stub scripted Agent, Workflow, Bash, and Skill calls, so each subagent request could be logged. No login was used, so cloud paths were not run. The interactive screen has its own section below.

What We Ran	Status	What We Saw
Workflow script with agent(), and the model variables	Undocumented: Only the program shows this item. It can change without notice.	The subagent model comes from CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL (not inherit), then CLAUDE_CODE_SUBAGENT_MODEL, then the main model. A model option in agent() is overridden. With CLAUDE_CODE_SUBAGENT_MODEL_FORCE, the workflow variable is dropped.
CLAUDE_CODE_WEB_FETCH_AGENT=1	Undocumented: Only the program shows this item. It can change without notice.	The roster gains web-fetch, and WebFetch leaves the main session's tools. The agent has only WebFetch.
Agent with each built-in type	Undocumented: Only the program shows this item. It can change without notice.	claude, Explore, general-purpose, Plan, statusline-setup, and fork started. claude-code-guide, comment-thread-analyst, observer, and workflow-subagent gave Agent type '...' not found in -p. The interactive screen started claude-code-guide (see Interactive Probes).
--append-subagent-system-prompt	Undocumented: Only the program shows this item. It can change without notice.	The text is the last lines of the subagent system prompt. The env var alone changed nothing.
observer frontmatter	Undocumented: Only the program shows this item. It can change without notice.	Without CLAUDE_CODE_EXPERIMENTAL_OBSERVER_AGENTS=1 no observer ran. With it, extra observer requests ran (three or five, from run to run, on 2.1.306, 2.1.307, 2.1.308, and 2.1.309), with an ObserverReport tool, and observerMessage was the last line of the digest.
Subscription (OAuth) login, main and Explore requests	Undocumented: Only the program shows this item. It can change without notice.	Same main system prompt text (apart from the memory folder path), same 12 default tools, and same Explore prompt as an API key. Different: a 1-hour cache ttl, thinking.display updates, the ScheduleWakeup description, a schedule skill, and RemoteTrigger in the deferred list. The Explore request had the same 7 tools: Bash, ListAgents, Read, ReportFindings, Skill, DeferredToolPlaceholder, and ToolSearch. See Override a Prompt.
Agent frontmatter spelling	Undocumented: Only the program shows this item. It can change without notice.	Tools and disallowed-tools were ignored. tools and disallowedTools worked.
isolation: remote in a file	Undocumented: Only the program shows this item. It can change without notice.	No error. The agent ran in a local worktree, like worktree.
Command hooks with continue:false and systemMessage	Documented: An official docs page names this item.	See Hook Input and Output.
Setup hook additionalContext	Differs: The official docs and the program disagree. The row says how. Not in the model request.	It never reached the model request.

Configure and Memory

These checks ran 2.1.305 against the same stub, plus static reading of the program. Nothing was run with a login.

What We Ran	Status	What We Saw
User outputStyle plus a plugin style with force-for-plugin: true	Undocumented: Only the program shows this item. It can change without notice.	The system prompt used the forced plugin style. The init event still named the configured style.
CLAUDE_CODE_REMOTE=1, with and without CLAUDE_CODE_REMOTE_MEMORY_DIR	Undocumented: Only the program shows this item. It can change without notice.	Without the dir, the auto-memory section is missing. With it, the section returns and names the dir.
CLAUDE_MEMORY_STORES set to a JSON list	Undocumented: Only the program shows this item. It can change without notice.	A list such as [{"name":"s1","path":"/some/dir"}] kept the memory section, but it named a memory/team/<dir name> folder under the project folder in the config folder, marked "(shared with all users of this project)", and added a 4 KB limit note per memory file. A list of strings, [], and text that was not JSON changed nothing. No memory_* tool was added. 2.1.304 gave the same result.
claude auto-mode config with autoMode.allow in each source	Documented: An official docs page names this item.	User settings and --settings applied. .claude/settings.json and .claude/settings.local.json were ignored. Managed settings applied too, see Managed Folder Probe.

Context Budget

These runs used a live subscription login on 2.1.305, not the stub, with a fresh config folder and an empty project.

What We Ran	Status	What We Saw
claude -p "/context" --output-format stream-json, empty and after claude -p "Reply with 'OK'."	Undocumented: Only the program shows this item. It can change without notice.	An empty session held 12,524 tokens and 16,277 after Reply with 'OK'. (the reply was exactly OK in every run). The first request matched /context exactly. Reruns on 2.1.306, 2.1.307, 2.1.308, and 2.1.309 (default model claude-sonnet-5-5, warm cache) had the same categories and matched again. On 2.1.307 the System tools category was 393 tokens larger than on 2.1.306, with the same warm cache. On 2.1.309 it was 106 tokens larger than on 2.1.308, which fits the longer Bash tool description. Its token counts are not quoted, because the cache was warm. See Context Budget.

Live Login Probes

These runs used a live subscription login (CLAUDE_CODE_OAUTH_TOKEN) on 2.1.306, and again on 2.1.307, 2.1.308, 2.1.309, 2.1.310, 2.1.311, and 2.1.312 with the same results, with a fresh config folder and an empty project. The request runs went through a local proxy that kept only the paths and the anthropic-beta header.

What We Ran	Status	What We Saw
claude -p "Reply with 'OK'." with the login, compared with the same run using a fake API key against the stub	Undocumented: Only the program shows this item. It can change without notice.	The main request (the first one with tools) had the same 11 body keys (context_management, diagnostics, max_tokens, messages, metadata, model, output_config, stream, system, thinking, tools), the same tool names, and 4 system blocks on both sides. On 2.1.307 the system text was also the same once model names, temp paths, and the per-run billing hash were masked. The login sent 15 betas and the key sent 13. Only the login sent oauth-2025-04-20 and extended-cache-ttl-2025-04-11. The default model was claude-sonnet-5-5 with the login and claude-opus-5-5 with the key. The login run also called GET /api/claude_code/policy_limits (200) and GET /api/claude_code/settings (404). The cache was warm, so no token counts are quoted. See Betas.
/artifacts typed in the interactive screen with the login, then Esc	Undocumented: Only the program shows this item. It can change without notice.	The screen showed Unknown command: /artifacts and no panel opened. This differs from the 2.1.305 run, where the panel opened with the same kind of login. 2.1.307, 2.1.308, 2.1.309, 2.1.310, 2.1.311, and 2.1.312 showed the same. The cause is not known. See Version 2.1.306.

Static and Print Probes

These runs used 2.1.305, the same stub, a fake key, a temp home folder, and a dead proxy, plus static reading of the program. Nothing was run with a login or with sudo. The Managed Folder Probe is the one section that used sudo.

A hook inherits the environment of the probe, so a change in the probe environment shows as a change in the hook environment. The 2.1.311 to 2.1.312 comparison reports AWS_EC2_METADATA_DISABLED as new in the environment of PreToolUse, PostToolUse, SessionStart, Stop, and UserPromptSubmit. The probe sets that variable, so this is not a change in Claude Code. Read a hook-environment difference against tools/release/lib/env.ts before you record it.

What We Ran	Status	What We Saw
Control requests on stdin with --input-format stream-json	Undocumented: Only the program shows this item. It can change without notice.	set_model, get_binary_version, set_permission_mode, mcp_status, and interrupt got replies that match Control Protocol. An unknown model ID still got success. A bad mode got error with error_code invalid_mode.
A user agent with mcpServers frontmatter	Undocumented: Only the program shows this item. It can change without notice.	The server started with the agent and stopped when it ended. The same agent in .claude/agents/ was skipped in an untrusted folder.
claude plugin validate with odd version values	Documented: An official docs page names this item.	A string of any text passed. A number failed. An empty string counted as missing. See Plugin Manifest.
CLAUDE_CODE_SUBPROCESS_ENV_SCRUB with test variables	Undocumented: Only the program shows this item. It can change without notice.	Credential and config names were removed. The four GitHub token variables stayed. The permission mode was forced to default. See Tools and Bash Variables.
Default -p "Reply with 'OK'." for every model in the catalog	Undocumented: Only the program shows this item. It can change without notice.	Short, extended short, or full prompt by model. See System Prompt.
CLAUDE_COWORK_MEMORY_* and CLAUDE_CODE_POST_TURN_MEMORY	Undocumented: Only the program shows this item. It can change without notice.	The four Cowork variables changed the memory section as described in Memory and Rules. The post-turn pass did not run locally: the log said the memory server was not mounted.
PostToolUse hook logging per tool	Undocumented: Only the program shows this item. It can change without notice.	The tool_response shapes in Hook Events.
claude -p "/context"	Undocumented: Only the program shows this item. It can change without notice.	24 count_tokens requests and no /v1/messages request.
Skill variables in the init skill list	Documented: An official docs page names this item.	CLAUDE_CODE_DISABLE_CLAUDE_API_SKILL removed claude-api. The other four had no visible effect.
CLAUDE_CODE_MANAGED_SETTINGS_PATH	Undocumented: Only the program shows this item. It can change without notice.	Ignored. The log still read the machine default managed folder.
/status size warning with a 45,000-character CLAUDE.md in -p, claude doctor, and get_status	Undocumented: Only the program shows this item. It can change without notice.	No warning in these three. It shows in the interactive screen (see Interactive Probes).

Managed Folder Probe

These runs used the real 2.1.306, 2.1.307, 2.1.308, 2.1.309, 2.1.310, and 2.1.311 binaries on macOS with the stub, a fake key, and a temp home and config folder. The managed folder /Library/Application Support/ClaudeCode was made with sudo (owner root:wheel). It held CLAUDE.md, a rule, a skill, an output style, managed-settings.json, a managed-settings.d/ drop-in, and managed-mcp.json with one stdio server. Each run also had a user server and a project .mcp.json server. A baseline ran without the folder. Runs used -p --output-format stream-json, plus claude mcp list and claude mcp get. Every row below was seen in the managed cases of 2.1.306, and again in all 55 cases of 2.1.307, of 2.1.309, of 2.1.310, and of 2.1.311: mcp-run, mcp-list, mcp-get-managed, mcp-get-user, mcp-config-flag, outputstyle, skill-default, skill-disabled, claudemd-order, excludes-any-claude-md-user and -project, excludes-managed-abs-user and -project, excludes-managed-glob-user and -project, agent-mcp, and auto-mode-config. The NotebookEdit deny is in claudemd-order. The only data that changed in 2.1.307, 2.1.310, and 2.1.311 is the order in which the baseline servers started, and no row depends on it.

What We Ran	Status	What We Saw
managed-mcp.json with a user server and a project server	Documented: An official docs page names this item.	Only the managed server started. The user and project servers were not started and not in init (the baseline started both). claude mcp list showed only the managed server. claude mcp get on a user server exited 1 with No MCP server named "user_srv". Configured servers: managed_srv.
claude mcp get on the managed server	Undocumented: Only the program shows this item. It can change without notice.	Scope: Enterprise config (managed by your organization), and the removal hint names the managed-mcp.json path.
--mcp-config <file> with managed-mcp.json present	Documented: An official docs page names this item.	Exit 1 with You cannot dynamically configure MCP servers when an enterprise MCP config is present. No init event. See Servers passed with --mcp-config.
Managed outputStyle with a style file in the managed .claude/output-styles/	Documented: An official docs page names this item.	The init event named the style and its text was in the request. It won over outputStyle in user settings, .claude/settings.json, .claude/settings.local.json, and --settings.
A skill in the managed .claude/skills/, with and without CLAUDE_CODE_DISABLE_POLICY_SKILLS=1	Documented: An official docs page names this item.	The skill was listed and its description was in the request. The variable removed it. Nothing else changed.
Managed CLAUDE.md with user and project CLAUDE.md	Documented: An official docs page names this item.	Order in the request: managed, then ~/.claude/CLAUDE.md, then project CLAUDE.md.
claudeMdExcludes of **/CLAUDE.md, in user or project settings	Documented: An official docs page names this item.	User and project CLAUDE.md were removed. The managed CLAUDE.md and the managed rule still loaded.
claudeMdExcludes of the managed absolute path, or <managed>/**, in user or project settings	Documented: An official docs page names this item.	Nothing was removed: user and project CLAUDE.md stayed, and the managed CLAUDE.md and the managed rule loaded.
Managed .claude/rules/*.md	Undocumented: Only the program shows this item. It can change without notice.	It loaded, and claudeMdExcludes with **/CLAUDE.md, the managed absolute path, or <managed>/** did not remove it.
A managed-settings.d/ drop-in that denies NotebookEdit	Documented: An official docs page names this item.	NotebookEdit left the init event tool list (present there in the baseline). It was not in the request tool list in either run (12 tools).
The prompt note "not verified as your organization's policy"	Undocumented: Only the program shows this item. It can change without notice.	It did not appear with a root-owned folder. The program text says it is added for managed files that the session's own account could write (not run).
User agent with mcpServers frontmatter, no --mcp-config	Undocumented: Only the program shows this item. It can change without notice.	The agent ran, but its server did not start (without the managed folder it did). A notice said probe-agent agent MCP server blocked by enterprise MCP config: agent_srv.
autoMode in managed-settings.json and a managed-settings.d/ drop-in, with claude auto-mode config	Documented: An official docs page names this item.	Both files applied. Lists from all sources join, with user or --settings rules first and managed rules after. $defaults expands where it sits. A list with no $defaults in any source has no defaults. See Auto Mode.

Interactive Probes

These runs used the real 2.1.305 binary in a pseudo-terminal (a Python pty driver with pyte to render the 200x50 screen, because tmux was not installed). The method is the same as tmux send-keys and capture-pane. The stub, the fake key, the temp home and config folders, the dead proxy, and CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 were as in the other probes, with env -i so no variable from the parent session leaked in (the child got only the variables the probe set). Plain prompts were Reply with 'OK'. and the stub answered OK. Onboarding and trust were skipped by seeding the config file. The stub answered a DO:{...} line in the prompt with one scripted tool call. No login and no real credentials were used.

What We Ran	Status	What We Saw
Agent with claude-code-guide	Documented: An official docs page names this item.	It started on claude-haiku-5-5 with the tools Bash, Read, WebFetch, and WebSearch. The agent call ran in the background. The request text is the same as on 2.1.304. The program leaves it out when CLAUDE_CODE_ENTRYPOINT is sdk-ts, sdk-py, or sdk-cli, which includes -p. See Subagent, Claude Code Guide.
Agent with other types	Undocumented: Only the program shows this item. It can change without notice.	Unknown types return Agent type '<x>' not found. Available agents: claude, claude-code-guide, Explore, general-purpose, Plan, statusline-setup. comment-thread-analyst, workflow-subagent, and web-fetch (default settings) gave that error. fork returned Fork started. See Subagent Runtime.
PermissionRequest, PermissionDenied, Notification, and PreCompact hooks	Undocumented: Only the program shows this item. It can change without notice.	All four fired. See Hook Input and Output.
Notification types	Documented: An official docs page names this item.	Seven values seen live: permission_prompt, idle_prompt, elicitation_dialog, elicitation_response, elicitation_url_dialog, elicitation_complete, and agent_needs_input. See Hook Events.
Elicitation and ElicitationResult hooks with an MCP server	Undocumented: Only the program shows this item. It can change without notice.	Both fired. Inputs had mcp_server_name, message, mode, and requested_schema, and for the result action and content.
statusLine and subagentStatusLine commands that log their input	Undocumented: Only the program shows this item. It can change without notice.	subagentStatusLine ran on a fixed 5-second tick and got no COLUMNS or LINES. statusLine got transcript_path from the first run and prompt_id after the first prompt. Runs paused under dialogs. See Status Line.
Instruction file size warning with 45.6k, 156.0k, and four 34.4k files	Undocumented: Only the program shows this item. It can change without notice.	A banner under the header at start and a line in /status. The 40k, 120k total, and 150k (Opus 5.5) limits held. See Memory and Rules.
PostToolUse hook logging for Read, Write, Edit, Bash, and an MCP tool	Undocumented: Only the program shows this item. It can change without notice.	Same shapes as the -p probe (no contentNotInModelContext on an Edit after a Read). On 2.1.305 an Edit of a file that was never read worked, and its PostToolUse input had contentNotInModelContext: true; no PostToolUseFailure fired. See Hook Events.
/usage when the usage endpoint answers 429, with fake subscriber credentials and a local TLS stub behind a proxy (without CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC)	Undocumented: Only the program shows this item. It can change without notice.	Rerun on 2.1.305 (fake credentials, local stub): a 429 with no Retry-After gave Error: Usage endpoint is rate limited until <time>. with the time 5 minutes away, and Retry-After: 1800 gave a time 30 minutes away. A 200 with an error envelope gave Error: Usage endpoint is rate limited. Please try again in a moment. When rate-limit headers from a model reply had filled the first rows, the panel showed Per-model breakdown unavailable (rate limited until <time>) (or (rate limited — try again in a moment) for the envelope). After one good fetch and a later 429, it showed Showing last-known usage as of 2m ago (rate limited until <time>). These texts match the 2.1.304 run. With the real CLAUDE_CODE_OAUTH_TOKEN on 2.1.305 (a proxy that answered 429 for the usage endpoint and passed everything else on) the usage endpoint was never called, so the texts could not be checked: the panel showed the Current session and Current week rows from reply headers, and Refreshing… stayed up. A claude.ai login was not tried. See Slash Commands. With CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC set, the panel showed Failed to load usage data and sent no request. Not seen on 2.1.305: Partial usage data, the (could not refresh) variants, the form without as of, and the 403 path. On 2.1.306 the same real-token run was repeated with three proxy faults: a 429, a 429 with Retry-After: 1800, and a 200 error envelope. The usage endpoint was again never called, so the texts are still seen only against the stub. On 2.1.307, 2.1.308, 2.1.309, 2.1.310, and 2.1.311 the same three runs gave the same result, and the program holds the same text fragments as 2.1.306. On 2.1.312 the same three runs gave the same result: the usage endpoint was never called. The 2.1.312 program has new text, so it differs from 2.1.311. The 200 error envelope now reads Claude is temporarily limiting requests to the usage page (not your usage limit). Try again in a moment. The panel suffix now reads (requests temporarily limited — try again in a moment). Only the program text shows this, not a run.
A mod that calls $.ui.notify	Undocumented: Only the program shows this item. It can change without notice.	The hook got message, title, and notification_type (type plugin_notification), even with no notification channel. See Hook Events.

Notification Types

Follow-up probes ran the interactive screen with a hook that logs each Notification input. The same stub and fake key were used, with a temp home and config folder. They were rerun on 2.1.305 with env -i.

What We Ran	Status	What We Saw
PushNotification tool call {message, status: "proactive"}	Undocumented: Only the program shows this item. It can change without notice. Gated: Works only behind a flag or a server gate. The tool is behind a server-side flag that is off by default	The hook got session_id, transcript_path, cwd, prompt_id, hook_event_name, message (probe push), and notification_type push_notification. No title. The tool result was Terminal notification sent. The flag is tengu_kairos_push_notifications. The probe forced it through cached feature flags in the temp config, with CLAUDE_CODE_GB_DISK_CACHE_WHEN_TELEMETRY_OFF=1, ENABLE_TOOL_SEARCH=false (the tool is deferred), and CLAUDE_CODE_DISABLE_NOTIFICATION_PRESENCE_CHECK=1. A run without the presence-check variable gave the tool result Not sent because you're active in this terminal. and no hook call.
Agent teams (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1), one named background teammate, and a hand-written permission_request in <config>/teams/<team>/inboxes/team-lead.json	Undocumented: Only the program shows this item. It can change without notice. Gated: Works only behind a flag or a server gate. Set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 or use --agent-teams	The hook got worker_permission_prompt with the message A subagent needs permission for Bash (the name part was A subagent, not the teammate name) and no title. A normal permission_prompt (Claude needs your permission) followed when the dialog showed Bash command · from a subagent. The lead's poller starts only when a teammate exists, and it sends the notification only when the lead is idle with no dialog shown. The network form, <name> needs network access to <host>, was not run (it needs a sandbox and a teammate in another process). On 2.1.305 the same hand-written permission_request did not work: the lead showed Message from @mate (unverified) as a plain message, and no hook ran. The program reads an inbox message as verified only when the same process wrote it (it has a msg_id that the process recorded), so a hand-written file is treated as forged or unverifiable (the program). The worker_permission_prompt text was not seen again. A real in-process teammate whose scripted Bash call needed permission showed the dialog Bash command · from the mate agent and fired only permission_prompt (Claude needs your permission). The lead's inbox path, and so worker_permission_prompt and its network form, needs a teammate in another process (a pane backend), which was not available here.
Built-in computer-use MCP server, enabled in /mcp	Undocumented: Only the program shows this item. It can change without notice. Gated: Works only behind a flag or a server gate. Needs a Pro or Max plan, a server flag, and not HIPAA	It connected with its tools. list_granted_applications and wait calls did not fire the hook. The program sends computer_use_enter when it takes the computer-use lock, so the probe stopped before any action that controls the screen. Not seen. Not rerun on 2.1.305.
Search of the program for model_refusal_fallback	Undocumented: Only the program shows this item. It can change without notice.	The name is in the notification_type list, but no code sends a Notification with it. The only other use is a system message subtype. This is a reading of the 2.1.305 program, not a live run.
auth_storage_failure	Undocumented: Only the program shows this item. It can change without notice.	Not run. The token URL cannot point at localhost (CLAUDE_CODE_CUSTOM_OAUTH_URL allows only three Anthropic hosts), and on macOS the save goes to the real keychain. The 2.1.305 program still sends it when the login save fails (reading only).

Not run, because they need a login, a cloud or remote session, or another host:

• callback hook forwarding for cloud sessions.
• The Notification types auth_success, agent_completed, quota_auto_resume_fired, quota_auto_resume_stale, and quota_auto_resume_disabled. The other types are in Notification Types.
• Cloud and Cowork memory variables.
• Glob, Grep, NotebookEdit, WebFetch, and WebSearch in the PostToolUse shape probe.

What We Did Not Probe

• Cloud hook forwarding (callback) and the post-turn memory pass need a cloud session.
• Chrome variables need the extension.
• Remote, CCR, and cloud variables need a remote session.
• Subscriber-only paths such as cache_touch need a login.

Example

Repeat the [1m] probe by hand against any stub or proxy that logs request headers.

Check the 1M beta

ANTHROPIC_BASE_URL=http://127.0.0.1:8787 \
ANTHROPIC_API_KEY=sk-ant-fake-probe \
claude -p "Reply with 'OK'." --model "sonnet[1m]" --output-format stream-json --verbose

Look for context-1m-2025-08-07 in the anthropic-beta header. Run it again without [1m] and the value is gone.

How It Works

The harness reads an env var at start or at the point of use. A probe sets one variable, runs the same short session, and compares every output to the baseline. A change in a header, the system prompt, the tool list, or an event is an effect. A run with no change shows nothing in print mode, but the variable can still matter in another mode.

Related Pages

• API Betas
• API Endpoints and Headers
• Environment Variables
• Hooks
• Request and Cache