Laminar Pi extension

Laminar observability for the Pi coding agent.

This extension subscribes to Pi's lifecycle events and emits, live and in-process, one Laminar trace per agent run — with LLM and tool spans — grouped into a Laminar session by Pi's session id. Tracing, export, and debugger sessions are all handled by the @lmnr-ai/lmnr SDK.

Installation

npx lmnr-cli@latest plugin add pi

The CLI logs you in, lets you pick the Laminar project that should receive your Pi traces, mints a project API key, writes it to ~/.config/lmnr/pi-extension.json, and runs pi install npm:@lmnr-ai/pi-extension. Restart Pi and every run is traced.

Prefer to wire it up by hand? Install the package and set the key in your environment:

pi install npm:@lmnr-ai/pi-extension
export LMNR_PROJECT_API_KEY="..."

Without a key from either source, the extension disables itself and Pi runs untouched (fail-open).

What you get

One Laminar trace = one Pi agent run (before_agent_start → agent_end). Spans open on start events and close on end events — fully granular and realtime — under the run root:

pi agent run            (DEFAULT)   ← before_agent_start / agent_end
├─ LLM call (turn 0)    (LLM)       ← message_start / message_end
├─ bash                 (TOOL)      ← tool_execution_start / tool_execution_end
├─ LLM call (turn 1)    (LLM)
└─ …

Each span carries the Laminar-native attributes: token usage and cost, model/provider, per-turn input/output messages (gen_ai.*), the active tool definitions, and the run's session id — so runs are grouped and cost-attributed in the Laminar UI without any extra setup.

What each LLM span reports

The messages Pi actually sent. gen_ai.input.messages comes from Pi's context event, which fires before every LLM call and carries the list Pi assembled — full session history, compaction already applied, extension-injected messages included. We report that list rather than reconstruct one from the event stream.

The system prompt, as the role: "system" message heading every turn — the prompt Pi actually sent, after each before_agent_start handler has had its say. It runs about 12k chars, so it gets its own LMNR_MAX_CHARS budget rather than sharing the conversation's; otherwise it would clip the turns.

The active tool definitions, as gen_ai.tool.definitions — name, description and JSON schema for each tool in the run, filtered to Pi's active set. Laminar reads these off LLM spans into its canonical tool-definitions column and content-addresses them, so repeating the same set on every turn costs one stored row.

Messages are {role, content: [blocks]} with Anthropic-style content blocks — the same shape every other Laminar instrumentation reports, so pi traces read like Claude Code's. Pi's content blocks map onto Anthropic blocks one for one:

| Pi | Laminar | | --- | --- | | {type: "text", text} | {type: "text", text} | | {type: "thinking", thinking} | {type: "thinking", thinking} | | {type: "toolCall", id, name, arguments} | {type: "tool_use", id, name, input} | | role toolResult | role user + {type: "tool_result", tool_use_id, content} |

Reasoning stays a structured block rather than being flattened into the reply text, so a reader can still tell the model's thinking from its answer. Images are reported as [image <mime type>] rather than re-embedded: Anthropic's image block requires the base64 payload, and Pi keeps an image in context for the rest of the run, so carrying it would repeat hundreds of KB on every later turn.

Configuration

Configuration comes from the environment, falling back to ~/.config/lmnr/pi-extension.json ({ "projectApiKey": "...", "baseUrl": "..." }, mode 0600) written by lmnr-cli plugin add pi. The environment wins, so you can redirect a single run without editing the file.

| Variable | Required | Default | Purpose | | --- | --- | --- | --- | | LMNR_PROJECT_API_KEY | yes | config file | Laminar project API key. Absent from both ⇒ tracing disabled (fail-open). | | LMNR_BASE_URL | no | config file, else https://api.lmnr.ai | Laminar API base URL (self-hosted deployments). | | LMNR_USER_ID | no | lmnr-cli login identity | Associates every trace with a user id. Falls back to the email (then user id) that lmnr-cli login stored in credentials.json, so a logged-in user is attributed with no config. | | LMNR_MAX_CHARS | no | 20000 | Truncation cap for span input/output values. | | LMNR_DEBUG | no | — | Enables debugger sessions + file logging to ~/.pi/agent/lmnr-pi-extension.log. | | LMNR_DEBUG_SESSION_ID | no | — | Explicit debugger (rollout) session id. |

Debugger sessions

With LMNR_DEBUG set to a truthy value, each run is associated with a Laminar debugger session so it shows up in the debugger UI. The session id is resolved by the SDK: LMNR_DEBUG_SESSION_ID → the nearest .lmnr/debug-session.json (written by lmnr-cli debug session new) → a freshly minted id. The SDK registers the session and stamps it on every span.

Development

npm install
npm run typecheck    # tsc --noEmit
npm test             # node:test via tsx — unit + end-to-end span-tree tests

The end-to-end test (tests/extension.test.ts) drives the extension with synthetic Pi events against an in-memory span exporter and asserts the resulting span tree — no network or Laminar account required.

License

Apache-2.0