Normalize agent transcripts from different runtimes into one validated, model-ready record format.

Agent tools represent the same concepts—messages, reasoning, tool calls, and tool results—in incompatible native formats. trajectory provides one TypeScript API that turns those formats into deterministic, structured records for training, evaluation, analysis, and inference.

The caller supplies a transcript string and its source. The one exception is Deep Agents, whose sessions normalizeCheckpoint reads from its local LangGraph SQLite store by thread ID; see src/adapters/deepagents/.

The TypeScript package is published as @letta-ai/trajectory:

npm install @letta-ai/trajectory

The Python wrapper is published as agent-trajectory and imports as trajectory:

pip install agent-trajectory

import { normalizeTranscript } from "@letta-ai/trajectory";

const { records, diagnostics } = normalizeTranscript({
  source: "codex",
  transcript: rawJsonl,
});

records contains the normalized trajectory. diagnostics is always present and is empty when the transcript required no recoverable cleanup.

{
  "records": [
    { "role": "meta", "source": "codex" },
    {
      "role": "user",
      "content": "Check the current directory.",
      "timestamp": "2026-07-10T12:00:00.000Z"
    },
    {
      "role": "assistant",
      "content": null,
      "tool_calls": [
        {
          "id": "call_1",
          "name": "exec_command",
          "args": "{\"cmd\":\"pwd\"}"
        }
      ],
      "timestamp": "2026-07-10T12:00:01.000Z"
    },
    {
      "role": "tool",
      "tool_call_id": "call_1",
      "content": "/workspace",
      "timestamp": "2026-07-10T12:00:02.000Z"
    }
  ],
  "diagnostics": []
}

Tool result records may include ok: boolean when the source exposes an authoritative structured outcome, such as Pi/OpenClaw isError, Claude Code is_error, Letta Code resultOk, OpenHands/Cursor is_error, OpenCode/Gemini terminal state, or Copilot CLI success. The field is omitted when the source does not expose a reliable status; result text is never interpreted as success or failure.

Each adapter lives in its own folder under src/adapters/ with a README documenting the exact input contract, decoding behavior, and what the adapter drops.

listTrajectories() enumerates the sessions in a source's standard local store, newest first, with cursor pagination. It is a discovery layer beside normalization — normalizeTranscript() itself never touches the filesystem. Copilot CLI, Cursor, Gemini CLI, and OpenCode are export-only input contracts and intentionally return listing_unavailable; callers locate and read the exports themselves.

import { listTrajectories } from "@letta-ai/trajectory";

let cursor: string | undefined;
do {
  const page = await listTrajectories({ source: "claude-code", limit: 100, cursor });
  for (const item of page.items) {
    // item.id, item.path, item.updatedAt?, item.title?, item.sizeBytes?
  }
  cursor = page.nextCursor;
} while (cursor);

A trajectory is an ordered array containing:

• One leading meta record identifying the source and available session metadata.
• user and assistant prose records.
• Optional reasoning records when the source exposes reasoning.
• Assistant tool-call records with stable IDs and stringified JSON-object arguments.
• tool records linked to earlier calls by tool_call_id.

Every conversational record has an ISO timestamp. The complete contract is available as both runtime validation and schema/trajectory-v1.schema.json.

The public function is:

normalizeTranscript(input: NormalizeInput): NormalizeResult

Each native format is implemented as a focused adapter that decodes source events into the shared internal message/tool contract. Common validation, linking, repair, timestamp handling, and bounds remain in the normalization core.

Use prompts/add-source.md with a coding agent to add a source from a local transcript corpus. The prompt covers privacy-safe corpus inspection, sanitized fixtures, compatibility checks, and the transcript-only API boundary.

Requires Node.js 20+ and Bun for development:

bun install
bun run check

bun run check runs typechecking, the complete test suite, and the package build. It also regenerates the JavaScript runtime embedded in the Python wheel and fails if the committed bundle was stale. Run the Python parity suite with:

PYTHONPATH=python/src python3 -m unittest discover -s python/tests -v

See PARITY.md for compatibility checks performed against real transcript corpora and production source adapters. See SOURCE_VERSION_AUDIT.md for the privacy-safe source-version inventory, observed format families, and current decoder gaps.

Apache-2.0. See LICENSE.