Eliza Fingerprint Surface — extracted via code-read

Source: nyx container (nyx:v2.9.0-rc.1), package @elizaos/native-reasoning at /app/eliza/plugins/plugin-discord/typescript/node_modules/@elizaos/native-reasoning/dist/

Date: 2026-05-15 Profile method: docker exec nyx cat ... (read source dist files; no runtime capture needed — code-read gives the full enumeration).

Why code-read over runtime profile

We considered inserting a logger HTTP proxy between nyx and her tunnel, but two reasons made code-read strictly better:

1. Completeness. Runtime profile gives a sample (5–10 messages); code gives the full enumeration of every tool name, system prompt section, and property the framework can ever emit.
2. Safety. Code-read is docker exec ... cat — read-only, no env change, no recreate, no risk of breaking nyx.

Enumerated outbound surface

Tool names (18 total)

The native-reasoning tool registry registers these name fields on the tools[] array (Anthropic shape: {type:"custom", name, description, input_schema}):

eliza tool	source file	rough purpose
bash	tools/bash.js	shell exec
read_file	tools/file_ops.js	read text file
write_file	tools/file_ops.js	overwrite/create file
edit_file	tools/file_ops.js	unique-string replace
glob	tools/file_ops.js	glob list
grep	tools/file_ops.js	regex search
web_fetch	tools/web.js	URL fetch
web_search	tools/web.js	brave search
recall	tools/memory.js	semantic memory query
remember	tools/memory.js	persist a fact
ignore	tools/ignore.js	silent skip
journal	tools/journal_tools.js	private journal append
note_thread	tools/journal_tools.js	open-threads.md add
close_thread	tools/journal_tools.js	open-threads.md remove
update_project	tools/journal_tools.js	projects.md upsert
spawn_codex	tools/spawn_codex.js	codex subagent
spawn_agent	tools/acp_agent.js	ACPX subagent
create_task	tools/acp_agent.js	ACPX task creation

sessions_spawn is a name collision with OpenClaw's sessions_spawn. Different semantics but same wire token — important for dictionary design.

System-prompt fingerprint markers

Strings injected by system-prompt.js and loop.js that scream "eliza":

• HARD RULE: If a human in this channel told you to be quiet ... (CHANNEL_GAG_HARD_RULE — appears before every conversation)
• ## Your Identity, ## Your Soul, ## About Your Human, ## Recent Context (IDENTITY_FILES headers)
• ## Recent Conversation (room context section)
• ## Current Moment, ## Active Projects, ## Today's Journal (your own private thoughts from earlier), ## Open Threads (things you want to follow up on), ## Relevant Past Conversations (nyxDynamicContext / nyxRelevantMemories)
• nyx stay quiet, nyx be quiet, nyx shut up, stay silent, nyx you can speak, nyx unmute — literal example phrases inside the CHANNEL_GAG_HARD_RULE
• Path strings: /workspace/projects.md, /workspace/open-threads.md, /workspace/journal/, IDENTITY.md, SOUL.md, USER.md, MEMORY.md

Tool description giveaways

Phrases that make tool-listing fingerprintable as eliza:

• "Execute a shell command inside the agent's allowed workspace" (bash)
• "agent's allowed workspace", "the allowed workspace" (file_ops)
• "agent's persistent memory" (recall)
• "agent's long-term memory" (remember)
• "PREFER THIS for any multi-step coding task" (spawn_agent — distinctive)
• "acpx-compatible agent" (spawn_agent)
• "Spawn a codex subagent to do focused multi-step work" (spawn_codex)

Property names on the wire

Eliza emits these field names in tool inputs and tool_results:

• roomId, entityId, agentId, messageId, tableName
• tool_use_id (Anthropic native; not framework-specific)

Backend headers

AnthropicBackend calls client.beta.messages.create(...) with the betas: ["advanced-tool-use-2025-11-20"] flag. Default model: claude-opus-4-7. Model is configurable via ANTHROPIC_LARGE_MODEL.

Framework-named env vars

These appear in stack traces and may leak into logs or error contexts:

• NATIVE_REASONING_* (HOOKS_ENABLED, VISION_ENABLED, EVALUATORS_ENABLED, ENABLED_PROVIDERS, PROVIDER_CACHE_TTL_MS, MAX_TURNS, etc.)
• ELIZA_REASONING_MODE
• NYX_HYBRID_EVALUATORS
• NATIVE_REASONING_HOOK_SURFACE_MARKER = "native-reasoning-hooks-v1"

Module-name leakage in error stacks

@elizaos/native-reasoning, @elizaos/core, plugin-discord — appear in any stack trace included in tool_results. Should be stripped or generalized.

Dictionary design implications

1. String replacements must catch: eliza, Eliza, nyx, native-reasoning, framework-specific section headers. These should map to identity values (find === replace) initially — they're presence detectors, not transformations. Future tuning can replace with neutral synonyms.
2. Tool renames must remap eliza tools to CC-shaped tools. Mapping choices below.
3. Property renames are minimal — eliza already uses fairly generic names. roomId → thread_id mirrors the OC mapping cleanly.
4. System-prompt paraphrase for eliza needs a different anchor. Eliza's prompt structure starts with character.system (free text), then CHANNEL_GAG_HARD_RULE. The strip should target the CHANNEL_GAG_HARD_RULE block, which is the most distinctive marker.

Tool rename mapping (eliza → CC names)

Designed to make eliza's tool surface look like a Claude Code session:

eliza	→ CC name	rationale
bash	Bash	direct shape match
read_file	Read	CC's Read
write_file	Write	CC's Write
edit_file	Edit	CC's Edit
glob	Glob	CC's Glob
grep	Grep	CC's Grep
web_fetch	WebFetch	CC's WebFetch
web_search	WebSearch	CC's WebSearch
recall	KnowledgeSearch	same shape as OC mapping
remember	KnowledgeStore	sibling to KnowledgeSearch
ignore	SkipResponse	non-CC; needs neutral name
journal	NotebookEdit	CC has NotebookEdit
note_thread	TodoWrite	CC has TodoWrite
close_thread	TodoComplete	CC-adjacent
update_project	ProjectUpdate	neutral
spawn_codex	Task	CC has Task (subagent)
spawn_agent	Agent	CC stub uses Agent
sessions_spawn	TaskCreate	distinct from Task