Codexview CLI

codexview-cli — render agent session logs to compact markdown

codexview-md is a one-shot CLI that converts an AI coding agent session log into plaintext markdown. The output is small (typically ~5% of the raw input) and structured for both human skim-reading and feeding to another LLM as compressed context.

Accepts both line-delimited jsonl (Codex CLI, codex-team, Claude Code) and single-document JSON (OpenCode exports, GitHub Copilot Chat sessions) — auto-detected.

When you should reach for this skill

• The user references an agent session by file path (~/.codex/sessions/..., ~/.claude/projects/..., .codex-team/runs/..., ~/.local/share/opencode/...) and wants to see what happened in it.
• The user pipes opencode export <id> output and wants it as markdown.
• The user has a giant log (often hundreds of KB to many MB) and any direct Read / cat would blow context.
• The user wants a session summary to paste into another conversation, into a doc, or to compare two runs.
• The user is debugging "what did the agent do at step N" and needs to read the turn-by-turn flow.

If the user wants to see the raw tool outputs (stdout, diffs, MCP results) — this CLI drops those by design. Tell them so, and point at @codexview/react for an interactive viewer that keeps the detail.

Invocation

Always invoke via npx so the latest version is used and no global install is required:

npx -y @codexview/cli@latest <input.jsonl>

-y skips the install confirmation. The first run downloads the package (~20 KB tarball); subsequent runs are cached.

Common patterns

View a session, head-limited (the default for unknown-size files):

npx -y @codexview/cli@latest ~/.claude/projects/<repo>/<sessionId>.jsonl | head -80

Save to a markdown file, then read it back in chunks:

npx -y @codexview/cli@latest <jsonl> -o /tmp/session.md
wc -l /tmp/session.md
# then Read /tmp/session.md with offset/limit as needed

Stream from stdin:

cat <jsonl> | npx -y @codexview/cli@latest -

Force a format if auto-detect fails (rare):

npx -y @codexview/cli@latest <jsonl> --format rollout

Valid --format values: rollout (Codex CLI), codex-team (AgentWeb status log), claude-code (Claude Code session), opencode (OpenCode export), github-copilot (GitHub Copilot Chat session).

OpenCode (live, no saved file):

opencode export <sessionID> | npx -y @codexview/cli@latest -

Embed OpenCode subagents (parent's task tool invocations) by exporting each child and passing --subagent (repeatable):

opencode export ses_parent > /tmp/parent.json
opencode export ses_child_a > /tmp/child-a.json
opencode export ses_child_b > /tmp/child-b.json
npx -y @codexview/cli@latest /tmp/parent.json \
  --subagent /tmp/child-a.json \
  --subagent /tmp/child-b.json

Pairing key: child info.id must equal parent's task state.metadata.sessionId. Mismatched / missing children print a stderr note and the task line falls back to a one-line placeholder.

Other flags

Exit codes

Output format

# Session <threadId>

## User
<user text>

## Assistant (reasoning)
> plaintext reasoning rendered as blockquote

## Assistant
<assistant text>

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

<continued assistant text after tools>

---

## User
<next turn>

Rules:

• # Session header carries the thread/session id from the original log.
• Turns separated by ---.
• Three role headings — ## User, ## Assistant, ## Assistant (reasoning). Reasoning is a separate block from the assistant text.
• Tool calls render as one-line placeholders (no stdout, no stderr, no diff). The short summary shown depends on the tool — Bash shows the command, Edit/Write/Read show the file_path, mcp__server.tool shows the qualified name, Agent shows the description, TodoWrite shows the count.
• Consecutive tool calls group without blank lines; an assistant text segment that follows tool calls starts a fresh ## Assistant heading so the reader sees that work happened in between.
• turn_failed / turn_aborted append a single italic line at the end of the turn (e.g. _(turn failed: rate limited)_).

What's dropped (so you don't promise it to the user)

• Tool outputs: stdout, stderr, exit codes, file diffs, MCP results, function call outputs.
• Encrypted reasoning (Codex Fernet blobs, Claude Code empty thinking blocks).
• Token usage, timestamps, durations, raw / unknown events.
• Claude Code system noise: attachment, system, last-prompt, queue-operation lines.

Kept (inlined) since cli 0.4.0: subagent summaries — Claude Code Agent tool outputs appear automatically when present in the parent jsonl; OpenCode task tool outputs appear when the matching child export is passed via --subagent. Both render as ### {description} blocks under the parent's tool line.

If the user needs raw tool output (stdout, diffs, MCP results), the CLI is the wrong tool — recommend @codexview/react.

Where to find session files on a user's machine

• Codex CLI rollouts: ~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl
• Claude Code sessions: ~/.claude/projects/<encoded-cwd>/<sessionId>.jsonl (the directory is the project's working-directory path with / replaced by -)
• AgentWeb codex-team: <project>/.codex-team/runs/<runId>/events.jsonl
• OpenCode: real data lives at ~/.local/share/opencode/ (the ~/.opencode/ dir is the install, not the data). List sessions with opencode session list --format json (project-scoped to cwd; cd /tmp first for cross-project listing); export one with opencode export <sessionID> (single JSON document). Subagent children share the same store; pair via parent.task.state.metadata.sessionId == child.info.id.
• GitHub Copilot Chat: single JSON files at ~/Library/Application Support/Code/User/workspaceStorage/<hash>/chatSessions/<uuid>.json (macOS); ~/.config/Code/User/workspaceStorage/... (Linux); use Code - Insiders instead of Code for Insiders builds. Filter to Copilot sessions by checking agent.extensionId.value === 'GitHub.copilot-chat' or sniff for "copilot" in the first 2 KB. GitHub Copilot has no subagent model; --subagent is not applicable.

When the user gestures vaguely at "my last session" without a path, ls -t ~/.claude/projects/<...>/ to find the most recently modified file is usually the right move (Claude Code). For OpenCode, opencode session list --format json | jq '.[0]' from /tmp gets the latest across all projects.

Worked examples

"Summarize my last Claude Code session in this project"

LATEST=$(ls -t ~/.claude/projects/<encoded-cwd>/*.jsonl | head -1)
echo "Raw: $(wc -c < "$LATEST") bytes"
npx -y @codexview/cli@latest "$LATEST" -o /tmp/last-session.md
echo "Markdown: $(wc -c < /tmp/last-session.md) bytes"

Then read /tmp/last-session.md to answer the user's question.

"I want to give Claude.ai context about what I did yesterday in Codex CLI"

F=~/.codex/sessions/2026/05/16/rollout-<...>.jsonl
npx -y @codexview/cli@latest "$F" -o /tmp/yesterday.md
# then attach /tmp/yesterday.md or paste its contents

"Summarize my last OpenCode session"

LATEST_ID=$(cd /tmp && opencode session list --format json | jq -r '.[0].id')
opencode export "$LATEST_ID" | npx -y @codexview/cli@latest - -o /tmp/opencode.md

If the session dispatched subagents (parent has task tool calls), the cli will print a stderr hint like note: 3 subagent task call(s) detected. Re-run with --subagent .... Export each child and re-run with --subagent to embed them.

"What tool calls did this run make?"

npx -y @codexview/cli@latest <jsonl> | grep "^🔧" | sort -u

"Did the agent ever read foo.ts in this session?"

npx -y @codexview/cli@latest <jsonl> | grep -E "🔧 \`(Read|Edit|Write)\` .*foo\.ts"

When NOT to use this skill

• The user wants the full transcript with tool outputs visible (use @codexview/react interactive viewer instead).
• The user wants to modify the jsonl, not read it (parse it directly).
• The file is not an agent session log (e.g. it's some other jsonl data file). The CLI will exit 1 with "could not detect input format" — that's the signal it's not the right tool.
• The file is a Claude Code subagent file (path includes /subagents/agent-*.jsonl). Those are sidechain-only and the CLI will produce empty output; you want the parent session jsonl one level up.

Source: codexview/codexview — distributed by TomeVault.