codex-dsh-trace
Reads and displays recorded Codex (OpenAI) request traces from local rollout recording bundles, letting you inspect nested tool calls, subtool declarations and results, compaction items, reconstructed context, request/response JSON, and execution timing alongside the conversation.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
Codex Trace
Read recorded Codex requests, nested tool calls, compaction, and execution timing alongside your conversation.
简体中文 · Contributing · Architecture · Security

The screenshot uses synthetic data in the local preview. It does not contain a private conversation or demonstrate a native Codex installation.
Features
Open the current chat's recording automatically and follow new events.
Explore the DSH timeline and execution ledger with folding, search, and a resizable inspector.
Compare reconstructed context with the incremental request and original response JSON.
Follow nested calls, recorded subtool declarations, and tool results, including recorded images.
Keep earlier requests visible across resumed recordings, fork history, and compaction.
Read large recordings through incremental indexing, virtualized rows, and on-demand details.
Follow the host's theme and language. Upstream controls support eight locales; Codex-specific text currently falls back to Chinese or English.
Related MCP server: memex
Requirements
A Codex client that supports local plugins, MCP Apps, and official
rollout-tracerecording. Compatibility depends on the client and recording schema.Python 3.9+ available as
python3. No Python packages are required.Node.js 20.19+ and npm to build or develop from source. An already-built plugin does not need Node.js at runtime.
macOS arm64 has been checked locally. The original integration used Codex CLI 0.163.0; that is a historical test baseline, not a minimum-version guarantee. CI checks the reader and build on Ubuntu; Linux and Windows native Codex integrations remain unverified. The reader supports recording schemas 1 and 2.
Install
codex plugin marketplace add xz1220/dsh-trace-plugin
codex plugin add codex-dsh-trace@rollout-localThe repository includes the built frontend. You do not need Node.js, npm, or a source checkout to install it. If the repository is private, your Git credentials must have access.
Open Codex Trace from the conversation's right-panel tools menu. Open a new chat after installation if the entry does not appear.
For the macOS setup helper, including configuration checks and uninstall, see the installer guide.
Update
codex plugin marketplace upgrade rollout-local
codex plugin add codex-dsh-trace@rollout-localIf you previously registered a local checkout, run the two Install commands first to switch the marketplace to the Git source.
Fully quit and reopen Codex after updating so the backend and UI resource reload. Reopening only the panel can retain an older build. The marketplace identifier remains rollout-local for compatibility.
Build from source
For development or a local marketplace installation:
git clone https://github.com/xz1220/dsh-trace-plugin.git
cd dsh-trace-plugin
npm ci
npm run build
codex plugin marketplace add "$PWD"
codex plugin add codex-dsh-trace@rollout-localOpen Codex Trace from the conversation's right-panel tools menu. Reopen the panel or restart Codex after installation if the entry does not appear. Client menus and plugin commands can differ between Codex versions.
The repository is dsh-trace-plugin; the plugin ID is codex-dsh-trace and the displayed product name is Codex Trace. Existing rollout-viewer MCP identifiers and configuration remain supported for compatibility. See Architecture.
Enable recording
Recording must be enabled before Codex starts. Existing conversations cannot be captured retroactively.
For a new CLI process:
CODEX_ROLLOUT_TRACE_ROOT="$HOME/.codex/rollout-traces" codexOr record one scripted task:
CODEX_ROLLOUT_TRACE_ROOT="$HOME/.codex/rollout-traces" codex exec --ephemeral "your task"On macOS, configure subsequently launched desktop apps for the current login session:
mkdir -p -m 700 "$HOME/.codex/rollout-traces"
launchctl setenv CODEX_ROLLOUT_TRACE_ROOT "$HOME/.codex/rollout-traces"Fully quit and reopen Codex afterwards. Repeat the command after logging out. To disable desktop recording, run launchctl unsetenv CODEX_ROLLOUT_TRACE_ROOT and restart Codex. These commands do not close the application. Source checkouts also include scripts/record.py as a recording helper.
From a source checkout, register another recording directory:
python3 plugins/codex-dsh-trace/server/main.py --register /absolute/path/to/recordingsPrivacy and data
The plugin reads official local trace bundles from the default recording root, CODEX_ROLLOUT_TRACE_ROOT, and directories you register. It does not substitute ordinary session logs. Registration writes a small configuration file under $CODEX_HOME/rollout-viewer/ (default ~/.codex/rollout-viewer/); it does not modify recordings.
The reader has no upload, analytics, or remote-service integration. Recording data is delivered to the local Codex app for display. The optional preview serves the same data on authenticated loopback HTTP. Follow external links and use copy or export actions with care: recordings can include credentials, private code, prompts, and images. Read Security before sharing a trace or screenshot.
Troubleshooting
The viewer is empty. Confirm recording was enabled before the Codex process started and that its root matches the current chat. Opening the viewer does not start recording. An empty resumed recording waits for new data automatically.
The plugin still looks like an older version. Installed files and an open Codex connection can differ. Fully quit and reopen the client after updating. Installing a package alone does not prove that a running panel has reloaded.
A recording is unsupported or too large. Schemas 1 and 2 are supported. The current limits are 256 MiB of event logs and 64 MiB per payload. Limits produce an explicit error. A current-chat aggregate over the limit can still be inspected as individual captures through the API.
Some historical text or timing is missing. The viewer shows what was recorded. Opaque compaction items, uncaptured images, missing continuation prefixes, and unrecorded generation timings cannot be reconstructed reliably.
Development
npm ci
npm run build
npm testRun npm run demo for an isolated synthetic recording without access to personal Codex data.
Start with this README for installation and use, then follow the contribution guide for setup, checks, and the repository map. Read the architecture when changing the reader or UI; product behavior and the roadmap provide the detailed constraints and planned work.
Attribution and license
Independently maintained; not an official OpenAI or DeepSeek product. The interface includes MIT-licensed components from DeepSeek Harness, pinned and documented in vendor/dsh/PATCHES.md. DSH refers to that upstream project.
This project's code is MIT licensed; see LICENSE. Built plugin packages include LICENSE and THIRD_PARTY_NOTICES.txt with the bundled dependencies' notices.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Traceo bug recordings (console, network, user actions, environment) as context for coding agents.
Ingest and search LogsLoom logs from coding agents.
Experiment tracking for coding agents: read your team's ML runs, metrics and lineage. Read-only.
Read and score Opik traces, experiments, datasets and prompts from your AI assistant.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceA terminal live-tail and a browser dashboard — one process, one event stream, served from localhost. Unified timeline across Claude Code, Codex, Gemini CLI, Cursor, Hermes, and OpenClaw. Token + cost accounting, compaction + anomaly detection, hybrid search, SVG call graphs, monaco-style diff attribution, agent-aware replay ("what would the agent say if I edited the prompt?"), policy editor, MCP s8 npm16MIT
- AlicenseNot gradedqualityBmaintenanceEnables local-first personal knowledge management for Codex by turning conversations into a searchable archive, extracting durable facts, building a knowledge graph, and injecting relevant context into later sessions.9 npmMIT
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to search and retrieve past coding-tool conversations across Claude Code, Claude Desktop, Codex, and Cursor through a local, read-only index.2MIT
- AlicenseNot gradedqualityBmaintenanceEnables ChatGPT conversations to directly inspect local workspaces and delegate coding tasks to a local Codex agent, with controlled concurrency, reusable sessions, and task/usage tracking.603 npmMIT