Tool Call Trace

简体中文

Open the browser tool

Download v0.2.1 and verification assets

A browser-local waterfall viewer and contract checker for timestamped AI Agent tool calls. It imports Generic JSON, OpenAI run steps, OpenAI Agents SDK spans, LangChain Runs, and PydanticAI/Logfire spans without uploading the trace. Opt-in redaction removes common credentials and selected fields before analysis or display.

Status: pre-release maturity. v0.2.1 hardens JSON-encoded and free-text credential redaction and fixes WASM startup readiness. No package or Agent transport is published.

Current capabilities

• Auto-detect or explicitly parse five timestamped trace contracts.
• Normalize absolute or exporter timestamps to milliseconds from trace start.
• Show total, average, maximum, error-rate, frequency, duplicate, and slow-call findings in a keyboard-accessible waterfall.
• Inspect untrusted input and output through text-only DOM rendering.
• Explicitly redact common authorization, API-key, token, password, secret, and private-key fields; credential assignments and authorization headers in free text; HTTP(S) user-info, query, and fragment components; and exact JSON Pointer paths.
• Preserve trace and call IDs so investigations remain searchable.
• Validate and normalize the same contracts from files or stdin with tool-call-trace check.
• Reject input above 5 MiB or 100,000 lines and traces above 2,000 calls.

Supported inputs

Format	Accepted timestamped export
Generic JSON	Strict flat array with millisecond timestamps and status
OpenAI run steps	data list of function tool-call run steps
OpenAI Agents SDK	Exported trace.span function spans with RFC 3339 times
LangChain	Tool Run objects, arrays, runs wrappers, and child_runs
PydanticAI / Logfire	OTel tool spans from exported_spans_as_dict()

Fixture provenance and pinned upstream revisions are documented in fixture sources.

Anthropic message blocks are intentionally unsupported because they do not contain start and end timestamps. Tool Call Trace never invents latency.

Local redaction

Redaction is off by default. Enable Redact common secrets in the browser or pass --redact to the CLI. Additional paths are exact JSON Pointers relative to each normalized call, such as /input/customer/email or /output/session/token.

cargo run -p tool_call_trace_cli -- \
  check --redact --redact-path /input/customer/email trace.json

The result reports the number of replaced values and uses [REDACTED] as the marker. Redaction is a bounded safety aid, not a complete secret or personal data detector. Review output before sharing it.

Browser quick start

Prerequisites: Rust 1.95.0, the wasm32-unknown-unknown target, wasm-pack 0.15.0, and Python 3 or another static file server.

git clone https://github.com/Tinkora/tool_call_trace.git
cd tool_call_trace
rustup target add wasm32-unknown-unknown
wasm-pack build --target web --out-dir static/pkg crates/tool_call_trace_web -- --locked
python3 -m http.server 4174 --bind 127.0.0.1 --directory crates/tool_call_trace_web

Open http://127.0.0.1:4174/static/.

CLI quick start

Use - or omit the path to read stdin. --format accepts auto, generic, openai-run-steps, openai-agents, langchain, and pydantic-ai.

cargo run -p tool_call_trace_cli -- check --format auto trace.json
cargo run -p tool_call_trace_cli -- check --redact - < trace.json

Successful normalized JSON is written to stdout; diagnostics and redaction counts are written to stderr. The command exits with code 1 for invalid trace contracts and 2 for invalid command usage.

Generic input

Each array item must contain id, name, input, start_time_ms, end_time_ms, and status. IDs must be unique, names and IDs must be non-empty, and the end time cannot precede the start time. Optional fields are output and error. Supported statuses are success, error, cancelled, and pending; the aliases completed, failed, and in_progress are also accepted.

[
  {
    "id": "call_1",
    "name": "search",
    "input": { "query": "WASM" },
    "output": { "matches": 3 },
    "start_time_ms": 1700000000000,
    "end_time_ms": 1700000000250,
    "status": "success"
  }
]

Deliberate limits

• The tool analyzes static logs; it is not a live tracer or an APM replacement.
• The repository contains no hosted API, MCP server, executable Agent tool, account system, storage service, or telemetry.
• Exporter import is contract-based. It does not install or instrument the upstream SDKs.
• Redaction is explicit and best-effort. It intentionally preserves trace and call IDs and does not claim exhaustive PII detection.

Development

cargo fmt --all -- --check
cargo test --workspace --locked
cargo clippy --workspace --all-targets --locked -- -D warnings
cargo check -p tool_call_trace_web --target wasm32-unknown-unknown --locked

wasm-pack build --target web --out-dir static/pkg crates/tool_call_trace_web -- --locked
cd crates/tool_call_trace_web
npm ci --ignore-scripts
npm run test:wasm-smoke

The browser suite runs Chromium at 375, 768, 1024, and 1440 pixel widths and checks the real WASM boundary, parsing and redaction workflows, keyboard dialog behavior, secret-free errors, reduced motion, console output, external traffic, and horizontal overflow.

Documentation

• Product specification
• Maturity evidence
• Contributing
• Security
• Support
• Changelog

License

MIT Copyright Tinkora contributors.