---

name: agtrace-provider-normalization description: Investigate AI agent provider tool schemas (Claude Code, Codex, Gemini) and design unified domain abstractions in agtrace architecture.

Agtrace Provider Normalization Expert

This skill provides deep knowledge of how agtrace normalizes diverse AI agent log formats into unified domain types. Use this when working with provider implementations, tool normalization, or understanding the schema-on-read architecture.

Three-Tier Provider Architecture

Tier 1: Trait-Based Adapter Pattern

Every provider implements three core traits bundled in a ProviderAdapter:

CopyLogDiscovery  → File discovery and session location
SessionParser → Raw log parsing to AgentEvent timeline
ToolMapper    → Tool call normalization and classification

Key Files:

• crates/agtrace-providers/src/traits.rs - trait definitions
• crates/agtrace-providers/src/registry.rs - adapter factory

Tier 2: Provider-Specific Implementation

Each provider has this structure:

Copyprovider/
├── discovery.rs      # LogDiscovery trait impl
├── parser.rs         # SessionParser trait impl
├── mapper.rs         # ToolMapper trait impl
├── io.rs             # Raw file I/O
├── schema.rs         # Provider-specific schemas
├── tools.rs          # Provider-specific tool args
├── tool_mapping.rs   # Tool classification
└── models.rs         # Data structures

Provider directories:

• crates/agtrace-providers/src/claude/
• crates/agtrace-providers/src/codex/
• crates/agtrace-providers/src/gemini/

Tier 3: Unified Domain Types

All providers normalize to common types in agtrace-types:

Schema-on-Read Architecture

The architecture enforces Schema-on-Read where:

1. Raw logs are source of truth - Original files never modified
2. Lazy parsing - Files parsed on demand
3. Type-safe conversion - Raw JSON → Provider Args → Domain ToolCallPayload

Provider-Specific Args Pattern

Each provider defines schema-specific argument types:

Copy// Claude: crates/agtrace-providers/src/claude/tools.rs
ClaudeReadArgs, ClaudeGlobArgs, ClaudeEditArgs, ClaudeWriteArgs,
ClaudeBashArgs, ClaudeGrepArgs, ClaudeTodoWriteArgs

// Codex: crates/agtrace-providers/src/codex/tools.rs
ShellArgs, ApplyPatchArgs, ReadMcpResourceArgs

// Gemini: crates/agtrace-providers/src/gemini/tools.rs
GeminiReadFileArgs, GeminiWriteFileArgs, GeminiReplaceArgs,
GeminiRunShellCommandArgs, GeminiGoogleWebSearchArgs

Each has conversion methods:

Copyimpl ClaudeReadArgs {
    fn to_file_read_args(&self) -> FileReadArgs { ... }
}

Tool Call Normalization Flow

CopyRaw JSON Arguments
    ↓
Provider-specific Args deserialization (strict)
    ↓
Optional: Semantic reclassification (e.g., shell → Read/Write/Search)
    ↓
Convert to typed ToolCallPayload variant
    ↓
If parse fails → Generic variant (safe fallback)

Semantic Reclassification Example

Codex shell commands are intelligently classified:

Copy// File: codex/mapper.rs
match tool_name {
    "shell" => {
        if let Ok(shell_args) = parse::<ShellArgs>(arguments) {
            match classify_execute_command(&command) {
                Some(ToolKind::Search) => return Search variant
                Some(ToolKind::Read) => return FileRead variant
                _ => return Execute variant
            }
        }
    }
}

MCP Tool Handling

Different providers use different MCP naming conventions:

All normalize to ToolCallPayload::Mcp with McpArgs:

Copypub struct McpArgs {
    pub server: Option<String>,
    pub tool: Option<String>,
    pub inner: Value,
}

Event Building

The EventBuilder in builder.rs provides deterministic event construction:

Copypub struct EventBuilder {
    session_id: Uuid,
    stream_tips: HashMap<StreamId, Uuid>,  // Per-stream parent tracking
    tool_map: HashMap<String, Uuid>,        // Provider ID → UUID mapping
}

Key features:

• Deterministic UUIDs: v5 UUID from session_id + "base_id:suffix"
• Multi-stream support: Main, Sidechain, Subagent independent chains
• Tool result linking: O(1) lookup from provider tool ID to UUID

Adding a New Provider

1. Create provider module: crates/agtrace-providers/src/<provider_name>/

2. Implement Discovery:

Copypub struct NewProviderDiscovery;

impl LogDiscovery for NewProviderDiscovery {
    fn id(&self) -> &'static str { "new_provider" }
    fn probe(&self, path: &Path) -> ProbeResult { ... }
    fn scan_sessions(&self, log_root: &Path) -> Result<Vec<SessionIndex>> { ... }
}

1. Implement Parser:

Copypub struct NewProviderParser;

impl SessionParser for NewProviderParser {
    fn parse_file(&self, path: &Path) -> Result<Vec<AgentEvent>> { ... }
    fn parse_record(&self, content: &str) -> Result<Option<AgentEvent>> { ... }
}

1. Implement ToolMapper:

Copypub struct NewProviderToolMapper;

impl ToolMapper for NewProviderToolMapper {
    fn classify(&self, tool_name: &str) -> (ToolOrigin, ToolKind) { ... }
    fn normalize_call(&self, name: &str, args: Value, call_id: Option<String>)
        -> ToolCallPayload { ... }
    fn summarize(&self, kind: ToolKind, args: &Value) -> String { ... }
}

1. Register in registry.rs:

Copypub fn create_adapter(name: &str) -> Result<ProviderAdapter> {
    match name {
        "new_provider" => Ok(ProviderAdapter::new(
            Box::new(crate::new_provider::NewProviderDiscovery),
            Box::new(crate::new_provider::NewProviderParser),
            Box::new(crate::new_provider::NewProviderToolMapper),
        )),
        // ...
    }
}

Tool Classification Pattern

Provider-Specific Classification

Each provider defines in tool_mapping.rs:

Copyfn classify_tool(tool_name: &str) -> Option<(ToolOrigin, ToolKind)>

Returns Some if recognized, None for fallback.

Common Heuristic Fallback

When provider doesn't recognize a tool (tool_analyzer.rs):

Copypub fn classify_common(tool_name: &str) -> (ToolOrigin, ToolKind) {
    // name.contains("search") → ToolKind::Search
    // name.contains("read") → ToolKind::Read
    // name.starts_with("mcp__") → ToolOrigin::Mcp
}

Key Architectural Decisions

Testing Patterns

Each provider has tests for:

• Discovery: probe, scan, session extraction
• Parser: record parsing, content blocks
• Mapper: tool normalization per tool type
• Edge cases: malformed data, unknown tools

Copy#[test]
fn test_normalize_read() {
    let payload = normalize_claude_tool_call(
        "Read".to_string(),
        serde_json::json!({"file_path": "src/main.rs"}),
        Some("call_123".to_string()),
    );

    match payload {
        ToolCallPayload::FileRead { arguments, .. } => {
            assert_eq!(arguments.file_path, Some("src/main.rs".to_string()));
        }
        _ => panic!("Expected FileRead"),
    }
}

When to Use This Skill

This skill should be activated when:

• Adding a new AI agent provider to agtrace
• Modifying tool normalization logic
• Understanding how provider logs become unified events
• Debugging provider-specific parsing issues
• Designing new tool classification strategies
• Working with MCP tool handling
• Understanding the schema-on-read architecture

Key Files Reference