Extract Agent Trajectories with atifact#

Convert agent session recordings (HAR files, Claude Code JSONL logs, Copilot CLI JSONL logs) into structured ATIF v1.6 trajectory JSON using the atifact CLI.

Prerequisites#

Verify the CLI is available:

bash

command -v atifact

If not installed, install globally:

bash

npm install -g atifact

Supported input formats#

FormatFile typeDescriptionhar``.harHAR files with OpenAI (Chat Completions, Responses API) or Anthropic (Messages API) requestsclaude-code-jsonl``.jsonlClaude Code CLI session logscopilot-cli-jsonl``.jsonlCopilot CLI session logs

Format is auto-detected from file contents (not extension). Use --format / -f to force a specific format when auto-detection fails.

Usage#

Basic conversion#

The --output / -o option takes a prefix, not a filename. Output files are derived from the prefix:

• Main trajectory: <prefix>.trajectory.json
• Subagent trajectories: <prefix>.trajectory.<name>.json

Default prefix is the input file path:

bash

atifact session.har
# Writes: session.har.trajectory.json

Specify output prefix#

bash

atifact session.har -o out
# Writes: out.trajectory.json
# If subagents exist: out.trajectory.<name>.json

Force input format#

Use when auto-detection picks the wrong format for .jsonl files:

bash

atifact session.jsonl -f claude-code-jsonl
atifact session.jsonl -f copilot-cli-jsonl

Pipe JSON to stdout#

Use --json with --quiet to suppress diagnostics and get clean JSON on stdout. Output is a JSON array of all trajectories (main first, then subagents). No files are written.

bash

atifact session.har --json --quiet

Combine with other tools:

bash

atifact session.har --json --quiet | jq '.[0].steps | length'

CLI options#

OptionAliasDescription<input-file>Path to the input file (required)--output``-oOutput path prefix. Main: <prefix>.trajectory.json, subagents: <prefix>.trajectory.<name>.json (default: input file path)--format``-fForce input format: har, claude-code-jsonl, copilot-cli-jsonl``--jsonWrite JSON array of all trajectories to stdout (no files written). First element is main, rest are subagents.--quiet``-qSuppress progress messages (stderr only)

Exit codes#

CodeMeaning0Success1Runtime error (parse failure, I/O error)2Invalid usage (bad arguments, missing file)

Workflow#

1. Identify the input file and its format (HAR or JSONL).
2. For .jsonl files, determine the source (Claude Code or Copilot CLI) to use the correct --format if auto-detection fails.
3. Run atifact with the input file. Use -o to set the output prefix (e.g., atifact /path/to/session.har -o /path/to/session). The main trajectory is written to <prefix>.trajectory.json.
4. Report the output file path(s) and key metrics (total steps, total cost) from the generated trajectory.

Notes#

• HAR files may contain multiple API formats (OpenAI + Anthropic); all are parsed.
• Multi-turn HAR conversations are deduplicated (each request carries full history).
• Utility calls (e.g., gpt-4o-mini title generation) are excluded from the trajectory.
• Tool results from request N are attached as observations to the agent step from request N-1.
• Copilot CLI logs with subagent task tool calls produce separate trajectory files per subagent. The main trajectory references them via subagent_trajectory_ref with trajectory_path pointing to the sibling file.
• All timestamps are preserved from source data as-is (ISO 8601).