atifact

Convert agent logs to ATIF trajectories. One command. Zero dependencies.

Turn HAR files, Vally trajectories, Claude Code logs, Copilot CLI logs, and Codex CLI logs into standardized ATIF v1.8 trajectory JSON — ready for debugging, visualization, fine-tuning, and RL pipelines.

Give your AI coding agent the atifact skill so it can extract trajectories on your behalf:

npx skills add waldekmastykarz/atifact

Once installed, ask your agent to "extract the trajectory from this HAR file", "convert Claude Code logs to ATIF", "convert Copilot CLI logs", or "convert Codex CLI logs" and it will handle the rest.

npm install -g atifact

# Convert a HAR file (auto-detected)
atifact session.har

# Convert Claude Code logs
atifact claude-log.jsonl

# Convert Copilot CLI logs
atifact copilot-session.jsonl

# Convert Codex CLI logs
atifact codex-session.jsonl

# Convert a standalone Vally Trajectory object
atifact vally-trajectory.json

# Pipe to stdout (returns JSON array of trajectories)
atifact session.har --json | jq '.steps | length'

Output: <input>.trajectory.json in ATIF v1.8 format. Claude Code, Copilot CLI, and Codex CLI logs with subagents produce additional <input>.trajectory.<name>.json files.

--json mode outputs a single trajectory with subagents embedded in the subagent_trajectories array to stdout with no files written.

Install atifact as a project dependency:

npm install atifact

Convert serialized log content directly from memory:

import { convert } from "atifact";

const { trajectory, subagentTrajectories } = await convert(jsonlContent);

Strings are treated as content, not file paths. Uint8Array input is also supported. Use convertFile to read from a path:

import { convertFile } from "atifact";

const result = await convertFile("session.jsonl");

Both functions detect the input format by default. The format can be forced and a source name can be included in the trajectory notes:

const result = await convert(content, {
	format: "har",
	utilityModels: ["gpt-4o-mini"],
	sourceName: "captured request",
});

Use detectFormat(content) for in-memory input or await detectFileFormat(path) for a file. The package exports the ATIF trajectory and parser result TypeScript types alongside these functions.

Format is auto-detected from file contents (not extension). Force it with -f:

atifact myfile.log -f claude-code-jsonl
atifact myfile.log -f copilot-cli-jsonl
atifact myfile.log -f codex-cli-jsonl
atifact trajectory.json -f vally-json

atifact <input-file> [options]

atifact produces ATIF v1.8 JSON with:

• Steps — user messages, agent responses, tool calls, and observations
• Metrics — token counts, costs, cached tokens per step
• Tool calls — structured function name + arguments with observation results
• Subagent trajectories — Claude Code, Copilot CLI, and Codex CLI subagents produce separate trajectory files linked via subagent_trajectory_ref with trajectory_id resolution; --json mode embeds them in subagent_trajectories
• Final metrics — aggregated totals across the trajectory
• All timestamps preserved as ISO 8601 from source data
• Null/undefined fields excluded for compact output

ATIF v1.8 supports audio content parts. atifact's current input parsers do not extract audio from source logs.

atifact recording.har -o my-trajectory
# Writes: my-trajectory.trajectory.json

atifact ~/.claude/projects/*/sessions/*.jsonl --json --quiet > trajectory.json

atifact copilot-session.jsonl
# Writes: copilot-session.jsonl.trajectory.json
# Writes: copilot-session.jsonl.trajectory.<subagent-name>.json (per subagent)

atifact codex-session.jsonl
# Writes: codex-session.jsonl.trajectory.json

atifact vally-trajectory.json
# Writes: vally-trajectory.json.trajectory.json

atifact session.har --json | jq '[.steps[] | select(.source == "agent")] | length'

• Node.js 22+

MIT