halo

RLM-based analyzer for AI agent execution traces

TLDR

Analyze a trace file with a diagnostic prompt

$ halo [traces.jsonl] -p "[Diagnose errors and suggest fixes]"

Use a specific model for the analysis run

$ halo [traces.jsonl] -p "[prompt]" --model [gpt-5.4-mini]

Point at an OpenAI-compatible provider

$ halo [traces.jsonl] -p "[prompt]" --base-url [https://openrouter.ai/api/v1]

Emit telemetry for the HALO run itself

$ halo [traces.jsonl] -p "[prompt]" --telemetry

SYNOPSIS

halo TRACEPATH **--prompt** TEXT [options_]halo TRACEPATH **-p** TEXT [options_]

halo is the command-line front end for the HALO engine, a recursively structured trace analyzer built around RLM (Recursive Language Models). It ingests OpenTelemetry-compatible JSONL execution traces from agent harnesses, decomposes failures across many runs, and produces a report of systemic issues rather than overfitting to a single bad trace.The CLI sends a user prompt plus the trace file to a root agent that can spawn subagents, summarize long contexts, and synthesize findings. Output is intended to feed back into coding agents such as claude or cursor so harness prompts, tools, and guardrails can be improved iteratively. A separate desktop app and hosted service exist, but halo is the local, scriptable entry point installed from the halo-engine Python package.

PARAMETERS

--prompt, -p TEXT

Required user prompt describing what to look for in the traces.

--model, -m MODEL

Model used for root and subagent calls (default `gpt-5.4-mini`).

--synthesis-model MODEL

Cheaper model for trace summarization (defaults to --model).

--compaction-model MODEL

Cheaper model for context compaction (defaults to --model).

--max-depth N

Maximum subagent recursion depth (default 2).

--max-turns N

Maximum turns per agent (default 20).

--max-parallel N

Maximum concurrent subagents (default 10).

--base-url URL

OpenAI-compatible API base URL.

--api-key KEY

Provider API key (falls back to OPENAI_API_KEY).

--header, -H NAME:VALUE

Extra provider header; repeatable.

--temperature N

Sampling temperature forwarded to the model.

--max-output-tokens N

Maximum output tokens forwarded to the model.

--telemetry

Emit OpenInference telemetry for HALO's own LLM and tool activity.

CONFIGURATION

OPENAI_API_KEY

Default credentials for the configured provider.

OPENAI_BASE_URL

Default API base URL when --base-url is not set.

CATALYST_OTLP_TOKEN

When --telemetry is enabled, upload spans to inference.net Catalyst over OTLP.

CAVEATS

halo requires a valid API key and provider access for the chosen model. Trace files must be JSONL in an OpenTelemetry-compatible format produced by an integrated harness. Analysis cost scales with trace size, recursion depth, and model choice; using smaller models for synthesis and compaction is recommended on large runs.

HISTORY

HALO was developed by Context Labs (inference.net) in 2026 as a methodology and engine for improving agent harnesses from production traces. The halo CLI ships with the open-source halo-engine package on PyPI.

RESOURCES

Source code · Homepage · Documentation