Docs › Recommended Plugins

Recommended Plugins

Third-party Claude Code plugins that pair well with Propel. Two of them power Propel's bridge skills (/c-codex, /codex-lead, /c-review) and are required for those skills to function. The rest are complementary — they don't change Propel's pipeline, but they make the surrounding workflow more pleasant.

Plugins vs. Propel Skills

Propel does not bundle these plugins. The bridge skills (/c-codex, /c-review) are interaction contracts that wrap a separately-installed plugin and merge its output with Propel's auditors. If a bridge skill's plugin isn't installed, the skill pauses and points you here.

Plugins That Power Propel Skills

These two plugins are what the Plugin Skills sidebar section calls out to. Installing them unlocks the corresponding Propel bridge skill. Without them, the bridge skill refuses to run and tells you where to install from.

codex-plugin-cc — OpenAI Codex inside Claude Code

Repo: github.com/openai/codex-plugin-cc

Brings OpenAI Codex into a Claude Code session as a callable tool. Propel uses this as the substrate for cross-model critique — Claude and Codex have different training biases, and pulling Codex in at planning and code-review time catches blind spots a single model would miss.

Powers these Propel skills:

/c-codex — ask Codex for a second opinion during planning or code review. Claude frames the question, Codex critiques, Claude synthesizes a comparison card for you.
/codex-lead — ephemeral flag that flips the contract for one exchange so Codex leads and Claude becomes the adversarial checker.

When you need it: at Gate 1 (plan review) or Gate 3 (post-implementation review), when a divergent-model perspective is worth the round-trip. think-deeply is the internal equivalent; /c-codex is the external one.

Anthropic `code-review` plugin — the official general-purpose review rubric

Repo: github.com/anthropics/claude-code — plugins/code-review

The official Anthropic code-review plugin ships a mature, general-purpose review rubric maintained by the Claude Code team. It is broad and well-calibrated, but generic — it does not know about your paper, your training loop, or your project conventions.

Propel's domain auditors (silent-bug-detector, paper-alignment-auditor, jax-logic-auditor, regression-guard) go deeper on research-specific concerns but cover less general ground. Running both is the widest coverage you can get.

Powers this Propel skill:

/c-review — runs the plugin alongside Propel's domain auditors, then merges and de-duplicates findings into a single Gate 3 card with source attribution ([plugin] / [silent-bug] / [both]).

When you need it: Gate 3 on substantive diffs (>30 lines, or anything touching model / loss / data / training-loop code), or a pre-PR pass before gh pr create.

Plugins Don't Speak to You Directly

Both bridge skills enforce a strict contract: the plugin is a tool, not a voice. Claude always frames the question, captures the plugin's raw output, runs Propel's own checks in parallel, and produces a single merged card. You never see unfiltered plugin output, and the plugin never silently rewrites code.

Complementary Plugins

These don't plug into a specific Propel skill, but they improve the ergonomics of long Claude Code sessions — which is exactly the kind of session a Propel workflow tends to be.

claude-hud — status-line HUD for Claude Code

Repo: github.com/jarrodwatts/claude-hud

A heads-up display for Claude Code that surfaces session state — model, token usage, context budget, and other live metadata — in the status line. Useful in any session, and especially useful for Propel workflows.

Why it pairs well with Propel:

Context discipline becomes visible. Propel's context-hygiene skill auto-suggests a /clear around 15–20 turns, but the actual signal is "how close are we to the cliff?" claude-hud puts that signal on screen so you can make the call before the skill has to.
Long investigations are exactly when context matters. Propel sessions tend to span investigation → design → implementation → validation, often with subagent dispatches in the middle. A persistent display of session health is more useful here than in a one-shot edit.
It doesn't change semantics. claude-hud is observational only — it doesn't intercept tool calls, rewrite output, or alter the agent's behavior. Safe to layer on top of Propel's constraint framework.

Install: follow the instructions in the claude-hud repo. There is no Propel-side configuration — once installed, the HUD just shows up.

Compatibility Notes

Bridge skills check before they run. The first time you invoke /c-codex or /c-review in a repo, Claude verifies the underlying plugin is installed and authenticated. If not, it pauses and points you here — it will not fabricate plugin command syntax.
Plugin invocation names can vary. If your local install exposes a slash command under a different name than the plugin's default, Claude will ask you to confirm the exact invocation rather than guessing.
Plugins are external to Propel's MIT scope. Each plugin has its own license and maintainer. Check the linked repo for terms before sending code or secrets through it.
Bridge skills redact secrets at the boundary. Anything sent to Codex or to the code-review plugin leaves the local context. Both bridge skills follow the "don't send secrets" guardrail, but project-level review of what gets sent is your responsibility.

Know Another Plugin That Fits?

The list above is intentionally narrow — only plugins Propel actually relies on or has been used alongside in practice. If you've found another plugin that pairs cleanly with the pipeline (without breaking the assistant-not-agent contract or short-circuiting a gate), open an issue on the Propel repo and we'll consider adding it.