Claude Code Probes

Claude Code Probes

A probe is a small test that runs Claude Code 2.1.303 and checks what one setting changes. The probes in this wiki 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 2.1.303 program with claude -p "hi" --output-format stream-json --verbose.
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, and 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 plain text. 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 166 variables to 1 in both baselines, which gave 332 runs. 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.

Every request carried the fake key, and the only path called was /v1/messages?beta=true.

Results

Mark	Meaning
✅	An effect was seen
❌	No visible effect
⚠️	The result differs from the docs

Of the 166 variables, 11 had an effect and 155 had none that we could see. 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.

Effects and Differences

Name	What We Set	Mark	What We Saw
--model X[1m]	model flag	✅	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	⚠️	Rejected at load: Unknown hook type "script"; entry ignored. No script ran. A PreToolUse entry also disables the whole settings file it sits in.
CLAUDE_CODE_AMBER_ASTROLABE	1	✅	Adds an "operating autonomously" block to the system prompt (do not ask "Shall I...?", finish the work before ending the turn).
CLAUDE_CODE_BASALT_COVE	1	✅	Adds a "Communicating with the user" section and a code-comment rule to the system prompt.
CLAUDE_CODE_HUMBLE_HAMMOCK	1	✅	Adds wire_tool_inputs (raw tool_use.input) to the assistant event when a tool runs.
CLAUDE_CODE_INLINE_TOOLS	1	✅	Adds the beta inline-tools-2026-09-15 to anthropic-beta.
CLAUDE_CODE_LANTERN_PRISM	1	✅	Adds the skill-doctor skill to the init event.
CLAUDE_CODE_LARCH_CISTERN	1	✅	Adds a "Corrections" section to the system prompt (avoid excessive self-correction).
CLAUDE_CODE_PLUGIN_DIRS	1	✅	Loads the plugin dir like --plugin-dir (same plugins list in init).
CLAUDE_CODE_SIMULATE_PROXY_USAGE	1	✅	Removes the whole anthropic-beta header and the thinking.display field.
CLAUDE_CODE_SKIP_PLUGIN_MCP_SERVERS	1	✅	Plugin MCP servers are not started: mcp_servers is empty and the tool list shrinks.
CLAUDE_CODE_SYNC_PLUGIN_INSTALL	1	✅	Emits plugin_install events (started, completed) in stream-json.
CLAUDE_CODE_USE_COWORK_PLUGINS	1	✅	User settings come from cowork_settings.json and plugins from cowork_plugins, so the hooks in settings.json did not run.

No Visible Effect

156 variables had no visible effect. Part 2 lists the first 78 rows, from CLAUDE_INTERNAL_FC_OVERRIDES to CLAUDE_CODE_MOCK_TRIAL. Part 3 lists the last 78 rows, from CLAUDE_CODE_NANKEEN_KESTREL to VOICE_STREAM_BASE_URL.

Request Details

• cc_version=2.1.303.914: the suffix 914 stayed the same in all runs.
• /v1/messages/cache_touch: never called. The bundle 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. The body is {"request_id": ...}. Touches repeat every 30 s by default (10 to 120). The stub with the response header and a flag override did not start it, because we have no subscriber login.
• 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, prompt-caching-scope-2026-01-05, per-turn-control-2026-07-01, mid-conversation-tool-changes-2026-07-01, effort-2025-11-24, thinking-display-updates-2026-08-18. First party adds advanced-tool-use-2025-11-20, thinking-binding-controls-2026-08-01, and cache-diagnosis-2026-04-07.

Auto Mode

With --permission-mode auto and a Bash call the stub got classifier requests with xml_s1 and xml_s2 stages. The debug log says server_no_result and falls back to billed local classifier requests. Neither classifier var changed that.

What We Did Not Probe

• Chrome vars need the extension.
• Remote, CCR, and cloud vars need a remote session.
• Subscriber-only paths (cache_touch) need a login.

See Also

• Probes (Part 2)
• Probes (Part 3)
• Environment Variables
• Hooks
• API Betas
• API Endpoints
• Request Build