Agent Transcript Parser

agent-transcript-parser

An evidence-driven transcript engine for agent providers.

The package converts each provider into a shared conversation document and projects that document into any registered provider. It does not contain a Claude-to-Codex translator or a Codex-to-Claude translator:

Claude ─┐
Codex  ─┼─> ConversationDocument ─> Claude / Codex / future provider
Future ─┘

That hub shape is the main architectural contract. Supporting another provider requires one decoder plus that provider's output projectors; it does not require pairwise translators to every provider already supported.

Profiles

Archive and native resume are intentionally different APIs:

• projectArchive favors preservation. Unknown records can survive as bounded, non-recursive provenance.
• projectNativeResume emits only the provider-native subset supported by a named evidence profile. Unsupported data is dropped or repaired and reported.

Every projection returns a structured report describing preservation, loss, demotion, repair, synthesis, and identity changes. A green archive test is not a claim that a provider CLI can resume the result.

Basic use

import {
  classifyClaudeDocument,
  codexNativeResumeProjector,
  decodeClaudeConversation,
  decodeJsonl,
} from 'agent-transcript-parser'

const raw = decodeJsonl(sourceText)
const classified = classifyClaudeDocument(raw)
const conversation = decodeClaudeConversation(classified.records)
const projected = codexNativeResumeProjector.projectNativeResume(conversation, {
  targetSessionId: crypto.randomUUID(),
  now: new Date().toISOString(),
  cwd: '/project',
  cliVersion: '0.144.6',
  modelProvider: 'openai',
  model: configuredModelId,
})

console.log(projected.values)
console.log(projected.report)

Context and compaction planning

The parser owns decisions that can be derived from a `ConversationDocument`; the host application owns live process I/O, clocks, cancellation, and UI. Use the typed p