agent-otel-bridge 0.3.0

Overview

agent-otel-bridge is a high-performance, native OpenTelemetry sidecar bridge designed to solve the hot-path lifecycle hook latency bottleneck in autonomous AI CLI agent harnesses (Google Antigravity, Claude Code, OpenAI Codex CLI, xAI Grok, Inflection Pi, and custom LLM developer agents).

Modern AI coding agents invoke synchronous lifecycle hooks (PostToolUse, PreInvocation, PostInvocation, Stop) dozens to hundreds of times per session. Traditional script-based hooks (Python, Node.js, PowerShell) impose 150ms to 1,200ms of cold-start latency per tool call, accumulating up to several minutes of wasted developer time in a single pairing session.

agent-otel-bridge decouples hot-path event emission into:

1. agent-hook.exe: An ultra-fast, microscopic static PE32 binary (242 KB) that executes in < 1ms, drops the event into a local Win32 Named Pipe via Overlapped I/O, and enforces a 3ms watchdog fail-open guarantee (never blocks or breaks the agent's workflow).
2. agent-otel-bridge daemon: A Tokio async background daemon maintaining connection pooling with the OpenTelemetry Collector (localhost:4318), performing micro-batching (50 spans / 200ms), and serializing canonical OTLP Protobuf traces and quota metrics without requiring protoc build-time dependencies.

Architecture

flowchart TD
    subgraph AgentHarness ["AI CLI Agent Harnesses (Antigravity / Claude / Codex / Grok / Pi)"]
        HookEvent["Synchronous Lifecycle Hook (e.g. PostToolUse)"]
    end

    subgraph ClientPath ["Ultra-Fast Client (< 1ms execution)"]
        HookBin["agent-hook.exe (242 KB Native PE32)"]
        Watchdog["Hard Watchdog Thread (3ms Deadline)"]
        PipeClient["Win32 Overlapped Client (\\.\\pipe\\agent-otel)"]
    end

    subgraph DaemonProcess ["agent-otel-bridge daemon (Tokio Background Process)"]
        PipeServer["Named Pipe Server (64KB Ring Buffer)"]
        Parser["ProtoJSON Parser & SemConv Mapper"]
        Batcher["Micro-Batcher (Configurable Buffer)"]
        QuotaEngine["Quota & Heartbeat Engine"]
        OtlpExporter["OTLP HTTP/Protobuf Exporter (Keep-Alive Pool)"]
    end

    subgraph OTelPipeline ["OpenTelemetry Collector Pipeline"]
        Collector["OTel Collector (:4318 / :4317)"]
        Processor["Batch / Transform Processor (OTTL)"]
    end

    subgraph ObservabilityBackends ["Pluggable Observability Ecosystem"]
        Traces["Distributed Traces (Jaeger / Tempo / SigNoz / Datadog)"]
        Metrics["Metrics & Gauges (Prometheus / Mimir / VictoriaMetrics)"]
        Dashboards["Dashboards (Grafana / SigNoz / Honeycomb)"]
    end

    HookEvent -->|"stdin (ProtoJSON)"| HookBin
    HookBin -->|"Spawn watchdog"| Watchdog
    HookBin -->|"Non-blocking write"| PipeClient
    HookBin -->|"stdout: {} & exit 0"| HookEvent
    PipeClient -->|"Binary Wire Framing [AG:v1]"| PipeServer

    PipeServer -->|"Tokio mpsc channel"| Parser
    Parser --> Batcher
    QuotaEngine -->|"Metrics"| OtlpExporter
    Batcher -->|"Traces"| OtlpExporter
    OtlpExporter -->|"POST /v1/traces (Protobuf)"| Collector
    OtlpExporter -->|"POST /v1/metrics (Protobuf)"| Collector
    Collector --> Processor
    Processor --> Traces
    Processor --> Metrics
    Processor --> Dashboards

Benchmarks

Measured on Windows 11 (AMD Ryzen / NVMe) via the dedicated benchmark suite (agent-otel-bench):

• OTLP Span Processing Throughput: > 546,000 spans/second.
• Hook Cost across 200 Tool Invocations:
  • PowerShell Hook Script: 90s – 240s lost.
  • Python Hook Script: 30s – 56s lost.
  • agent-hook.exe (Rust): 0.7s total accumulated time.
• Full methodology and data in docs/BENCHMARKS.md.

OpenTelemetry Semantic Conventions

Fully aligned with canonical CNCF OpenTelemetry GenAI & Agent Semantic Conventions (v1.28+ / v1.36):

Spans & Canonical Attributes

• Span Names:
  • execute_tool {gen_ai.tool.name} (SpanKind::Internal)
  • invoke_agent {gen_ai.agent.name} (SpanKind::Internal)
  • agent.stop (SpanKind::Internal)
• GenAI Attributes:
  • gen_ai.operation.name: execute_tool, invoke_agent
  • gen_ai.provider.name: dynamically inferred ("anthropic", "google", "openai", "xai", "inflection")
  • gen_ai.agent.name: dynamically inferred ("antigravity", "claude-code", "codex", "grok", "pi", "ai-agent")
  • gen_ai.conversation.id: correlated session identifier (mapped from conversationId or Claude Code session_id)
  • gen_ai.tool.name & gen_ai.tool.call.id
  • gen_ai.request.model: e.g. claude-3-5-sonnet, gemini-2.5-pro, gpt-4o, grok-2
• Canonical Agent Attributes:
  • agent.hook.event: PostToolUse, PreToolUse, PostInvocation, PreInvocation, Stop
  • agent.step.index & gen_ai.agent.step_index: integer turn index
  • agent.execution.num: execution sequence number
  • agent.fully_idle: boolean quiescence indicator
  • agent.termination_reason: stop / completion reason (e.g. NO_TOOL_CALL, model_stop)
• Agent Quota Gauges:
  • agent.quota.remaining_fraction: Gauge (unit "1", 0.0..=1.0).
  • agent.quota.seconds_to_reset: Gauge (unit "s", >= 0.0).
  • Attributes: bucket, group.
• Universal Agent Intelligence & Context (v0.3):
  • Behavioral CLI Archetypes (Preference-Agnostic): Commands are categorized into 6 functional archetypes (filter_compressor, structured_parser, inspector_diff, search_retrieval, build_test_verify, generic_exec) instead of hardcoding developer-specific tools (rtk, jq, bat). Tracks compression ratio, tokens saved, and pipeline depth.
  • Universal MCP & Skills Taxonomy + Waste Tracking: Standardizes tool calls across mcp servers, skill bundles, and native tools. Quantifies Schema Tax (dead-weight prompt tokens spent on inactive tools), retry thrashing, and response payload bloat.
  • Multi-Tier Zero-Subprocess Context Harvester (< 150 µs): Extracts workspace path, project root, project type (rust, node, python, go), and direct file-based Git metadata (.git/HEAD, sanitized remote origin, worktrees) with zero external process calls (git.exe is never spawned).
  • Cross-Agent Distributed Tracing (W3C traceparent): Injects and propagates W3C trace context via process environment variables ($env:TRACEPARENT). Heterogeneous subagents (e.g. Antigravity spawning Claude Code, which delegates to Codex) are unified into a single distributed trace DAG in SigNoz.
• OpenTelemetry Layering & Opt-In Compatibility:
  • Emits purely canonical, vendor-neutral attributes by default so all AI agents (Antigravity, Claude Code, Codex, Grok, Pi) share a clean data model without vendor pollution.
  • Optional AGENT_OTEL_LEGACY_ATTRIBUTES=true enables legacy agy.* aliases for older dashboards, or use an OpenTelemetry Collector transformprocessor (OTTL) to alias attributes in the collection tier.

Crates in the Workspace

Installation & Setup

1. Install via Cargo or Pre-built Binaries

From crates.io:

cargo install agent-otel-bridge
cargo install agent-otel-client

Or download pre-compiled release archives directly from GitHub Releases.

Or build locally from repository:

cargo install --path crates/agent-otel-client --force
cargo install --path crates/agent-otel-cli --force

2. Automated Hook Installation (install-hooks)

agent-otel-bridge automatically detects and configures lifecycle hooks for your AI agents:

# Automatically detects installed agents (Antigravity, Claude Code, Codex, Grok, Pi) and registers hooks
agent-otel-bridge install-hooks

# Or target a specific client
agent-otel-bridge install-hooks --client antigravity
agent-otel-bridge install-hooks --client claude
agent-otel-bridge install-hooks --client codex
agent-otel-bridge install-hooks --client grok
agent-otel-bridge install-hooks --client pi

# Or register hooks for all supported clients
agent-otel-bridge install-hooks --client all

# Or install at project level (.claude/settings.json or .gemini/hooks.json, etc.)
agent-otel-bridge install-hooks --client claude --project

Check hook status anytime:

agent-otel-bridge hooks status

3. Verify Installation (doctor)

agent-otel-bridge doctor

Output:

=== agent-otel-bridge doctor ===
Diagnosing station telemetry pipeline & OpenTelemetry invariants

[1/5] Checking environment contract variables...
  [ok] OTEL_EXPORTER_OTLP_ENDPOINT = http://127.0.0.1:4318
  [info] OTEL_SERVICE_NAME is unset (using default 'agent-otel-bridge')
  [ok] OTEL_RESOURCE_ATTRIBUTES = deployment.environment=homelab

[2/5] Checking named pipe IPC (\\.\pipe\agent-otel)...
  [ok] Daemon is RUNNING and responding on named pipe!

[3/5] Checking OTLP Collector HTTP endpoint...
  [ok] OTLP Collector reachable at http://127.0.0.1:4318/v1/traces

[4/5] Checking Observability UI reachability (http://localhost:8080)...
  [ok] Observability UI reachable at http://localhost:8080

[5/5] Checking Client Hook Registrations...
  agent-hook in PATH: [ok] present
  Google Antigravity: [ok] registered (C:\Users\samue\.gemini\config\hooks.json)
  Claude Code:        [ok] registered (C:\Users\samue\.claude\settings.json)
  OpenAI Codex:       [info] not registered (C:\Users\samue\.codex\hooks.json)
  xAI Grok:           [info] not registered (C:\Users\samue\.grok\hooks.json)
  Inflection Pi:      [info] not registered (C:\Users\samue\.pi\hooks.json)

4. Run CLI Commands

# Start background daemon
agent-otel-bridge start

# Send manual quota probe to collector
agent-otel-bridge emit-quota --ping

# Gracefully stop daemon
agent-otel-bridge stop

# Run benchmarks and submit results to community leaderboard
cargo run --release -p agent-otel-bench -- --submit --open-browser
# Or via unified CLI:
agent-otel-bridge benchmark --submit --open-browser

Documentation

License

Copyright (c) 2026 Samuel Mota. Licensed under the Apache License, Version 2.0.