Codebase Analysis

Codebase Analysis via Codex

Use Codex to get a second-opinion architectural read of the current project, with the sandbox locked to read-only. Codex consults; Claude writes.

Transport

Always use the MCP tool. The plugin runs codex mcp-server on stdio via .mcp.json. Tool name: mcp__plugin_codex_cli__codex. If the example below errors with an unknown-tool error, run /mcp and substitute the actual prefix (e.g., mcp__codex_cli__codex). Shell fallback is a last resort (see ../references/commands.md).

Model

Pin model: "gpt-5.6-sol" with config: { "model_reasoning_effort": "medium" } on the opening call. Codebase analysis benefits from the flagship's reasoning across many files — don't downgrade to gpt-5.6-luna here. Set both on the first codex call only; codex-reply inherits them. Honor an explicit user-named model if given. See ../references/patterns.md → Models and Reasoning effort.

Basic call

mcp__plugin_codex_cli__codex({
  "prompt": "Analyze this project's architecture: entry points, major modules, component relationships, and notable dependencies.",
  "sandbox": "read-only",
  "model": "gpt-5.6-sol",
  "config": { "model_reasoning_effort": "medium" }
})

The response includes a threadId. Use mcp__plugin_codex_cli__codex-reply with that id to drill in without re-establishing context.

When to Use

• Onboarding to an unfamiliar codebase
• Understanding legacy systems
• Mapping component relationships
• Finding hidden dependencies
• Architecture documentation
• Technical debt assessment

Examples

Full project analysis:

mcp__plugin_codex_cli__codex({
  "prompt": "Analyze this project. Report on:\n- Overall architecture\n- Key dependencies\n- Component relationships\n- Potential issues",
  "sandbox": "read-only",
  "model": "gpt-5.6-sol",
  "config": { "model_reasoning_effort": "medium" }
})

Flow mapping:

mcp__plugin_codex_cli__codex({
  "prompt": "Map the authentication flow. Identify every component involved from request to session creation.",
  "sandbox": "read-only",
  "model": "gpt-5.6-sol",
  "config": { "model_reasoning_effort": "medium" }
})

Dependency analysis:

mcp__plugin_codex_cli__codex({
  "prompt": "Analyze dependencies: direct vs transitive, outdated packages, circular dependencies, bundle-size impact.",
  "sandbox": "read-only",
  "model": "gpt-5.6-sol",
  "config": { "model_reasoning_effort": "medium" }
})

Iterative workflow (prefer codex-reply)

When you're still working on the same area of the codebase, continue the existing thread rather than starting a new codex call. Codex retains context between rounds; fresh calls force it to re-read files and drift from its prior reasoning.

Typical loop:

1. Initial consult → save the threadId from the response.
2. Claude reads related files / runs a query / makes a change.
3. codex-reply with new findings or a follow-up question.
4. Repeat — but cap at 3–4 rounds total. If the thread isn't converging, stop and bring the current state back to the user.

threadId is an MCP argument — pass it as the threadId field of codex-reply, not in the prompt text. See ../references/mcp-schema.md for wrong-vs-right examples.

Example — three rounds on the same architecture thread:

# Round 1 — initial map (opening call pins model + effort)
mcp__plugin_codex_cli__codex({
  "prompt": "Map the auth flow end-to-end.",
  "sandbox": "read-only",
  "model": "gpt-5.6-sol",
  "config": { "model_reasoning_effort": "medium" }
})
# → threadId: "019da14b-..."  /  flags: uncertainty about session rotation

# Round 2 — Claude reads src/session/ and reports back
mcp__plugin_codex_cli__codex-reply({
  "threadId": "019da14b-...",
  "prompt": "src/session/rotate.ts shows a 15m rotation window, not the 1h you assumed. Does that change anything in your flow map?"
})

# Round 3 — drill into a specific layer
mcp__plugin_codex_cli__codex-reply({
  "threadId": "019da14b-...",
  "prompt": "Focus on the data layer. What invariants does this flow depend on and where are they enforced?"
})

Start a fresh thread when: the user switches topic, the threadId is no longer in context, or Claude has made substantial code changes that would be cleaner to re-prime than to patch incrementally. See ../references/patterns.md.

After the analysis

Codex's read is a second opinion, not authoritative — it can misread structure or miss context it never saw.

• Relay the findings to the user and attribute them to Codex, rather than presenting them as verified fact.
• Spot-check claims against the actual code before acting on them (see ../references/patterns.md → Validation) — especially dependency, impact, and "nothing else uses this" claims.
• Surface uncertainty or disagreement to the user instead of smoothing it over into a confident-sounding summary.

Safety

• Always sandbox: "read-only". Codex must not modify files.
• Never use workspace-write or danger-full-access.
• Never use --dangerously-bypass-approvals-and-sandbox.

Fallback (rare)

If the MCP server is unavailable (plugin disabled, server crashed), see ../references/commands.md for the Bash equivalent. Requires dangerouslyDisableSandbox: true because Codex writes its own session state.