Harness Integration

Claude Code, Codex, OpenClaw, and Hermes — how each connects, and how TELOS detects which harness a request belongs to.

A harness is the upstream agent runtime. TELOS ships first-class support for four, and is adapter-driven so you can add more without rewriting your agent logic.

Supported harnesses

Harness	Typical usage	telos init auto-connect	Status
Claude Code	Anthropic-native coding agent workflow	✅	🟢 First-class
OpenClaw	Open-source agent runtime with TELOS parser integration	✅	🟢 First-class
Hermes	Multi-agent orchestration with independent sub-IR handling	✅	🟢 First-class
Codex	OpenAI-style coding workflow via local gateway injection	✅	🟢 Supported

How each connects

Claude Code

telos init --harness claude-code writes ANTHROPIC_BASE_URL into the env field of ~/.claude/settings.json (atomic, backed up to .telos.bak, marked with __telos_installed). Survives npm update, never touches PATH or the npm package.

Codex

telos init --harness codex adds a model_provider = "telos" custom provider to ~/.codex/config.toml, pointing Codex at http://127.0.0.1:7171/upstreams/openai/v1. The OpenAI Responses API path is passthrough; ChatCompletions-compatible traffic is TELOS-processed.

OpenClaw / Hermes

telos init patches each harness’s own provider config to point at the local gateway.

Generic

telos init --harness generic prints manual export instructions for any client that respects ANTHROPIC_BASE_URL or OPENAI_BASE_URL.

How a harness is parsed

Each harness plugin is a stateless parse(raw) → TelosIR. They differ in wire shape and identity:

	OpenClaw	Hermes	Telos
Wire shape	Anthropic /v1/messages	Anthropic /v1/messages	OpenAI ChatCompletions
Tool classification	_classify_anthropic_tool (type prefix + server field + metadata)	reuses _classify_anthropic_tool	_classify_openai_tool (metadata only)
<file> into ref-pool	no (whole oversized system item → slug system-doc-{i})	yes (<file path=…>, dotted slug)	yes (dash slug)
thinking block	not handled	FOLD kind=thinking	n/a
tool_result source	embedded in the user message	embedded in the user message	standalone role=tool → wrapped into a user message
system extraction	the system field	the system field	leading consecutive role=system in messages[]
source_tag prefix	openclaw/*	hermes/*	telos/*

Tool classification writes source (builtin / mcp / user) into TelosBlock.extra, feeding the bridge’s _tool_sort_key to protect PIN prefix stability across multi-MCP startup races.

Harness detection (_detect_harness)

For Anthropic-shaped traffic, TELOS picks between hermes and openclaw. Detection runs in order and returns on the first hit:

1. HTTP header fingerprint

   User-Agent containing claude-cli, or x-app: cli → hermes. Only the proxy path sees request headers. This reliably catches Claude Code’s tool-less Haiku auxiliary requests (title generation / topic detection), which steps 1–3 below all miss.

2. Envelope markers

   Paired open/close regexes (system-reminder / command-message / command-name) in the system text and in every user text block → hermes. Paired matching avoids misclassifying a user who is merely discussing tags in prose.

3. thinking block

   Any assistant message containing a thinking content block → hermes.

4. Tool fingerprint

   The intersection of tool names with {Bash, Edit, Read, Write, Grep, Glob, TodoWrite, Task, WebFetch, WebSearch, NotebookEdit} ≥ 3 → hermes. Catches the first-turn request before the reminder is injected.

5. Fallback

   Otherwise → openclaw.

Detection is sticky. The proxy remembers the confident signal per client (ProxyApp._client_harness); the SDK transport writes it into BridgeSessionState.sticky_harness. This avoids re-probing per call, which would flip the harness and destabilize the prefix.

Support Matrix

The full harness, model, and inference-framework support tables.