@codexview/cli

CLI that renders AI coding agent session logs (Codex CLI rollouts, AgentWeb codex-team status logs, Claude Code sessions, OpenCode exports) as compact plaintext markdown — suitable as compressed context to hand to another LLM, or as a quick human-readable session summary.

Install

npm i -g @codexview/cli
# or
pnpm add -g @codexview/cli

Usage

codexview-md path/to/session.jsonl                       # → stdout
codexview-md path/to/session.jsonl -o out.md             # → file
cat session.jsonl | codexview-md -                       # → stdin
codexview-md path/to/session.jsonl --format rollout      # force a format
codexview-md parent.json --subagent child.json           # embed an OpenCode subagent

Supported input formats

• Codex CLI rollouts — ~/.codex/sessions/.../rollout-*.jsonl
• AgentWeb codex-team status logs — .codex-team/runs/*/events.jsonl
• Claude Code sessions — ~/.claude/projects/<repo>/<sessionId>.jsonl
• OpenCode session exports — single JSON document, output of opencode export <sessionID>

Auto-detected by content; override with --format rollout|codex-team|claude-code|opencode.

Embedding OpenCode subagents

OpenCode parent sessions invoke children via the task tool. Pass each child's export with --subagent (repeatable) to inline its summary:

opencode export ses_parent > parent.json
opencode export ses_child_a > child-a.json
opencode export ses_child_b > child-b.json
codexview-md parent.json --subagent child-a.json --subagent child-b.json

Pairing is deterministic: parent's task state.metadata.sessionId must equal the child's info.id. If a child can't be matched it's reported on stderr and the parent's task line falls back to a one-line placeholder.

Claude Code subagent summaries (the Agent tool) are rendered inline automatically when present in the parent jsonl — no flag needed.

What's in the output

• # Session <threadId> header.
• One block per turn, separated by ---.
• Role-tagged sections: ## User, ## Assistant, ## Assistant (reasoning).
• Plaintext reasoning rendered as a blockquote.
• Tool calls render as one-line placeholders, e.g.:

  🔧 `Bash` npm test
  🔧 `Edit` src/foo.ts
  🔧 `mcp__github.create_issue`
  🔧 `TodoWrite` (6 todos)

What's dropped

The point is to compress, so the output omits:

• Tool outputs (stdout, stderr, diffs, MCP results, function outputs).
• Encrypted reasoning blobs (Codex Fernet content, Claude Code empty thinking blocks).
• Token usage, timestamps, durations, raw / unknown events.

Subagent summaries (Claude Code Agent, OpenCode task) are kept and inlined — see "Embedding OpenCode subagents" above. The cli inlines a tool output only when it begins with a ### markdown header (the format used by subagent summaries); other tool outputs are still dropped.

Exit codes

| Code | Meaning | |------|---------| | 0 | Success | | 1 | Unrecognised input format | | 2 | File I/O error | | 3 | Bad argument |

Related

• @codexview/react — React components that render the same kinds of sessions interactively, with full tool output visible.

License

MIT.