@codexview/adapters

Stateless adapters that convert AI coding agent transcript logs into the @codexview/react ChatStreamEvent schema.

Zero runtime dependencies, ESM only, Node ≥ 20.

Install

pnpm add @codexview/adapters

Use

Auto-detect (umbrella entry)

import { readFileSync } from 'node:fs';
import { adapt, parseJsonl } from '@codexview/adapters';

const lines = parseJsonl(readFileSync('rollout-xyz.jsonl', 'utf8'));
const { format, events } = adapt(lines);
// format: 'rollout' | 'codex-team' | 'claude-code' | 'unknown'

Per-source (tree-shakeable)

import { adaptClaudeCode } from '@codexview/adapters/claude-code';
import { adaptRollout }    from '@codexview/adapters/rollout';
import { adaptCodexTeam }  from '@codexview/adapters/codex-team';

const events = adaptClaudeCode(lines);

Supported formats

Live-host adapter: AgentWeb transcript

Apps that store chat history in a database and stream new turns live (rather than reading a JSONL log) need to merge two sources into one ChatStreamEvent stream. adaptAgentWebTranscript does that bridge for hosts shaped like AgentWeb — persisted ChatMessage[] plus an in-flight StreamingState.

import { adaptAgentWebTranscript } from '@codexview/adapters/agentweb-transcript';
import { CodexTranscript } from '@codexview/react';

const { events, status, error } = adaptAgentWebTranscript({
  sessionId,           // → thread_started.threadId
  messages,            // ChatMessage[] from your DB
  streaming,           // StreamingState atom (or null)
  // now,              // optional clock override for tests
});

<CodexTranscript events={events} status={status} />

Input shapes (AgentWebMessage, AgentWebStreamingState, AgentWebToolCall) are structural — adapters has no AgentWeb dependency. The adapter:

• emits thread_started for sessionId and synthesizes turn_started / turn_completed boundaries around each turnId group;
• preserves the same turnId across user message → live partial assistant text → final persisted assistant row so they render in one turn;
• maps run_command tool calls to exec_command_begin + exec_command_end, and everything else to function_call + function_call_output;
• maps AgentWeb { input, cachedInput, output } to CodexView TokenUsage;
• surfaces streaming.status = 'failed' | 'disconnected' | 'gaveUp' as turn_failed / turn_aborted events plus a TranscriptStatus (working, failed, stopped, completed, idle) on the result.

In-flight tool calls (completed: false) get their _begin event but no _end — the open call renders with a running indicator until the next adapter call sees the completed shape. Approval / reconnect / gave-up UI stays host-owned and renders next to <CodexTranscript>.

adapt() options

adapt(lines, {
  format?: 'rollout' | 'codex-team' | 'claude-code',  // skip detection
  patchMode?: 'function_call' | 'patch_apply_end',    // Claude Code only
  subagents?: SubagentInput[],                        // Claude Code only
  closeOpenTurn?: boolean,                            // Codex rollout only
});

• format — skip detectFormat() and use the given format directly.
• patchMode — controls how Claude Code's Edit / Write / MultiEdit tool calls are rendered.
  • 'function_call' (default) — emit as opaque function_call entries. Matches the cli's compact output.
  • 'patch_apply_end' — defer until the matching tool_result, then emit a patch_apply_end with a synthesized PatchFile[] containing diff text. Use when downstream rendering (e.g. <PatchBlock>) needs the diff.
• subagents — list of Claude Code subagent transcripts associated with this session. When supplied, each Agent tool's tool_result is rewritten to embed a Markdown summary of the matching subagent (description, agent type, tool counts, token totals, final reply). Pairing strategy: primary by toolUseResult.agentId on the parent's user-type tool_result line; FIFO fallback over unconsumed entries.
• closeOpenTurn — controls Codex rollout EOF behavior. Defaults to true for static files, preserving the historical behavior of synthesizing turn_completed if the log ends mid-turn. Set to false when tailing a currently-written rollout so the final turn remains working until a real task_complete / abort event appears.

interface SubagentInput {
  agentId: string;
  meta?: { agentType?: string; description?: string } | null;
  lines: unknown[];   // raw JSONL rows from agent-<agentId>.jsonl
}

Versioning

Independent semver from @codexview/react. The two packages share no runtime dependencies; the ChatStreamEvent types are duplicated and kept in sync by a CI guard (planned).

License

MIT.