Oh My Pi Session Tracing with Future AGI

Oh My Pi (omp) is a coding agent for the terminal, a fork of the Pi coding agent. This page adds a small omp hook that subscribes to omp’s documented hook events and sends them to Future AGI as OTLP traces. Each omp session shows up as one trace, with every model call and tool call inside it.

Use this guide with the omp.sh fork of Oh My Pi, not the upstream Pi binary.

This is not the same as:

• Upstream Pi (pi, pi.dev). Pi has its own extension API and its own page in this Coding agents section. Loading this file into pi is not supported: the hook only handles events whose context carries omp’s agent identity (ctx.agent), which Pi’s does not, so under pi it exports nothing.
• omp’s built-in OpenTelemetry export (OTEL_EXPORTER_OTLP_*). omp can send its own traces, logs, and metrics without any hook. This page uses a hook instead; see omp’s built-in OpenTelemetry export for the comparison and for how to avoid exporting twice.
• Other forks or embedded SDKs. This hook targets only the omp.sh CLI hook API.

This recipe requires omp 18.11.0. The hook has no npm dependencies. See the hook source.

What you get

omp.session                 AGENT   one per omp session
  omp.prompt                CHAIN   one per prompt you submit
    omp.turn                CHAIN   one per model/tool round trip
      chat gpt-6-luna       LLM     the model call, with tokens and cost
      execute_tool glob     TOOL    tools the model ran in that turn
      execute_tool read     TOOL
    omp.turn                CHAIN
      chat gpt-6-luna       LLM     the call that wrote the answer

A session that switches to another session inside the same omp process ends its trace and starts a new one. That covers /new, /resume, and /fork (omp’s session_switch event), and a branch omp writes to a new session file: a promoted /btw answer, or a branch or entry fork requested by an extension or over RPC (omp’s session_branch event). /branch and /rewind in the TUI move within the same session file and keep the same session id, so the trace continues. A session you continue later with omp -c or omp -r gets a new trace with the same session.id, so all of its runs are grouped under Sessions. A session in which you never send a prompt (you open omp and quit, or leave it right away with /new) exports nothing.

Set up

curl -fsSL https://omp.sh/install | sh -s -- --binary --ref v18.11.0
export PATH="$HOME/.local/bin:$PATH"
omp --version

mkdir -p .omp/hooks/post
curl -fsSL https://raw.githubusercontent.com/future-agi/docs/main/examples/coding-agents/oh-my-pi/futureagi-otlp.ts \
  -o .omp/hooks/post/futureagi-otlp.ts

export FI_API_KEY="YOUR_API_KEY"
export FI_SECRET_KEY="YOUR_SECRET_KEY"
export FI_PROJECT_NAME="my-omp-project"

omp
# or, one prompt without the TUI:
omp -p "List the files in this folder and read README.md"

Configuration

The hook posts OTLP/HTTP JSON with traces only. It never returns a value from a handler, so it never blocks or rewrites a tool call, a tool result, the context, or a message. It does not subscribe to tool_call, tool_result, context, input, or before_provider_request at all. It never runs shell commands and never reads omp’s session files. If an export fails, its spans are retried once with the next batch and then dropped. omp keeps working either way.

What is captured

Every span carries session.id, so traces also appear under Sessions. A failed tool sets its span to ERROR with error.type=ToolError. A model call that ends with stop reason error sets the model, turn, and prompt spans to ERROR with error.type=ProviderError. By default all of these use the message error details omitted by default. With content capture on, the tool span’s message is the tool’s error output and the model span’s message is the provider’s error text; the turn and prompt spans keep the fixed message. The session span stays OK; its omp.tool.error.count shows how many tools failed.

What is not available

• The full model request. The hook does not read before_provider_request, so the system prompt, history, and tool schemas are never exported. With content capture on, you get your prompt text and the assistant’s answer text, not the full request.
• Thinking. Reasoning blocks are never exported, even with content capture on. Only the reasoning token count is.
• Streaming detail. Token deltas (message_update) are not exported. Time to first token is exported as omp.llm.ttft_ms when omp reports it.
• Cancellation as an error. omp has no separate cancellation event. A model call you abort ends with stop reason aborted, which is exported in gen_ai.response.finish_reasons with an OK status. If omp shuts down in the middle of a turn, open spans are closed with omp.incomplete=true and an unset status, not ERROR.
• Cost for models omp cannot price. gen_ai.cost.* is omp’s own estimate from its model catalog or the cost you set in models.yml. Calls omp reports as free get no cost attributes.
• Subagents nested under the parent. omp gives each task-tool subagent its own copy of the hook, and each subagent session has its own session id, so every subagent becomes a separate trace carrying omp.agent.kind, omp.agent.parent_id, and omp.agent.depth. It is not nested under the parent trace.
• Hard kills. A turn’s spans are sent about a second after the turn ends. If omp is killed before that, that turn is lost. Earlier batches already carry a snapshot of the session and prompt spans, so the trace keeps its root.

Privacy

By default the hook exports ids, model and provider names, token counts, omp’s cost estimate, timings, tool names, status, and byte sizes. It does not export:

• your prompt text, the assistant’s answers, or thinking;
• tool arguments, tool results, or the tool intent text the model writes;
• error messages (only the error type and, for model calls, the HTTP status);
• local paths: the session’s working directory (omp.cwd) is not exported, because it usually contains your OS user name.

To export content, set FI_CAPTURE_CONTENT=true before starting omp. The prompt and final answer go to input.value and output.value on the prompt span (and the first prompt and last answer on the session span, which fills the trace list’s Input and Output columns), each model call’s text to its output.value, tool arguments and result text to the tool span, tool and provider error text to the status message of the tool or model span, and the working directory to omp.cwd on the session span. Each field is cut at 4,096 characters, and strings that look like sk-... keys, Bearer tokens, or api_key=... and "api_key": "..." values are replaced with [REDACTED].

The hook runs inside omp with your user’s permissions and can see everything omp sees, including prompts and tool output. It is written to export only what the tables above list.

omp’s built-in OpenTelemetry export

omp also has its own OpenTelemetry export, separate from hooks and on by default. It covers three signals, sends OTLP/HTTP protobuf only, and starts each signal that has an endpoint (omp environment variables, OpenTelemetry export):

So OTEL_EXPORTER_OTLP_ENDPOINT on its own turns on all three. The exporters also read OTEL_RESOURCE_ATTRIBUTES and the standard OTEL_EXPORTER_OTLP_HEADERS, which is how the Future AGI keys are passed. Future AGI accepts omp’s traces with no hook installed. Set only the traces endpoint, so omp’s logs and metrics stay off:

export OTEL_EXPORTER_OTLP_TRACES_ENDPOINT="https://api.futureagi.com/tracer/v1/traces"
export OTEL_EXPORTER_OTLP_HEADERS="X-Api-Key=YOUR_API_KEY,X-Secret-Key=YOUR_SECRET_KEY"
export OTEL_RESOURCE_ATTRIBUTES="project_name=my-omp-project,project_type=observe"
omp

The built-in export creates one trace per prompt:

invoke_agent             no Future AGI span kind; gen_ai.conversation.id = omp session id
  chat gpt-6-luna        LLM    model, provider, tokens
  execute_tool glob      TOOL
  execute_tool read      TOOL
  chat gpt-6-luna        LLM

If you prefer the built-in export, do not install the hook. Do not use both.

Each switch turns off a different part of omp’s own export. The hook reads only the FI_* variables, so none of them affects it.

A variable you set on the omp command line does not change your shell, but the commands omp runs for you inherit it. A program that uses OpenTelemetry and that you run through omp’s bash tool would see OTEL_SDK_DISABLED=true too. The /settings switch does not change the environment.

Troubleshooting

Related guides