adversarial-review

Claude Code Adversarial Review

Use this skill when the user wants Claude Code to challenge the implementation approach, design choices, assumptions, or tradeoffs in this repository.

Do not treat $cc:adversarial-review as the default review path. Use it only when the user explicitly wants stronger scrutiny than a normal code review.

Good triggers include requests to challenge the design, challenge tradeoffs, pressure-test a risky change, question whether a migration/config/template change really removed the risk, or honor custom focus text that asks for harsher review.

If the user wants Claude Code to go beyond review and perform investigation, validation edits, or implementation work, route to $cc:rescue instead.

If the user asks for a local review plus a separate Claude background review and then wants the main Codex thread to aggregate the findings and apply fixes, keep the delegated Claude portion on $cc:review unless the user explicitly asks for the adversarial angle.

Unlike $cc:review, this skill accepts custom focus text after the flags. The moment the user wants to steer Claude toward a specific angle or risk question, prefer $cc:adversarial-review.

Resolve <plugin-root> as two directories above this SKILL.md file. Always run the companion from that active plugin root:

node "<plugin-root>/scripts/claude-companion.mjs" adversarial-review ...

Supported arguments: --wait, --background, --base <ref>, --scope auto|working-tree|branch, --model <model>, --effort <low|medium|high|xhigh|max>, plus optional focus text after the flags (defaults: model=opus and no effort; fable, opus, sonnet, and haiku each keep Claude Code's own effort default, and Claude Code owns which effort levels each model supports)

Forward --model unchanged to the companion. The companion trims surrounding whitespace, canonicalizes the friendly aliases fable, opus, sonnet, and haiku to lowercase, then forwards every other --model value unchanged to Claude Code. Claude Code owns alias resolution and supported effort levels; /model is the authoritative picker for the current account and provider.

Raw slash-command arguments:

$ARGUMENTS

Rules:

• This skill is review-only. Do not fix issues, apply patches, or suggest that you are about to make changes.
• Before launching the review, stay in read-only inspection mode: inspect git status and diff stats only, then ask at most one user question about whether to wait or run in background.
• Preserve the user's scope flags and custom focus text exactly.
• Use the same review target selection as $cc:review.

Execution mode rules:

• If the raw arguments include --wait, do not ask. Run in the foreground.
• If the raw arguments include --background, do not ask. Run in background through the built-in adversarial-review subagent path.
• Otherwise, estimate the review size before asking:
• For working-tree review, start with git status --short --untracked-files=all.
• For working-tree review, also inspect both git diff --shortstat --cached and git diff --shortstat.
• For base-branch review, use git diff --shortstat <base>...HEAD.
• Treat untracked files or directories as reviewable work for auto or working-tree review even when git diff --shortstat is empty.
• Only conclude there is nothing to review when the relevant scope is actually empty.
• Recommend waiting only when the scoped review is clearly tiny, roughly 1-2 files total and no sign of a broader directory-sized change.
• In every other case, including unclear size, recommend background.
• When in doubt, run the review instead of declaring that there is nothing to review.
• Then ask the user once which execution mode to use, offering two options with the recommended one first and its label suffixed (Recommended):
• Wait for results
• Run in background
• Use a question tool for that ask only when this thread actually has one. Codex exposes request_user_input by default in interactive threads and hides it only when [tools] experimental_request_user_input = false, and it does not exist in non-interactive threads. If you have no question tool but a user is reading this thread, ask in your own reply and stop there. In a non-interactive thread with no user to answer, skip the ask and proceed with the recommended mode. Never spin on a wait or collaboration tool looking for a picker this thread does not have.

Argument handling:

• Preserve the user's arguments exactly.
• Treat --wait and --background as Codex-side execution controls only. Strip them before calling the companion command.
• Do not weaken the adversarial framing or rewrite the user's focus text.
• $cc:adversarial-review uses the same review target selection as $cc:review.
• It supports working-tree review, branch review, and --base <ref>.
• It does not support --scope staged or --scope unstaged.
• Unlike $cc:review, it can still take extra focus text after the flags.
• The companion review process itself always runs in the foreground. Background mode only changes how Codex launches that command.
• For the detailed execution contract, treat the internal runtime reference at ../../internal-skills/review-runtime/runtime.md as supporting guidance only. It is an internal reference document, not a public skill to invoke.

Foreground flow:

• Run:

node "<plugin-root>/scripts/claude-companion.mjs" adversarial-review --view-state on-success <arguments with --wait/--background removed>

• Run that companion command with sandbox_permissions: "require_escalated" and the justification Allow the Claude Code companion to contact the Claude API for this requested review. Do not first try the companion command in the default network-disabled sandbox.
• Foreground adversarial review belongs to the main Codex thread. Do not spawn a review subagent, do not invoke a generic review-runner role, and do not proxy this foreground path through any background worker abstraction.
• Do not fall back to raw claude, claude-code, claude review, bash -lc ...claude..., or any other direct Claude CLI syntax when the companion path is available. The foreground syntax contract here is the resolved companion command above, not a hand-rolled Claude invocation.
• If the resolved companion command fails, surface that failure. Do not silently retry foreground adversarial review through a different CLI shape, a generic review runner, or a custom shell wrapper.
• Present the companion stdout faithfully.
• Do not fix anything mentioned in the review output.

Background flow:

• For background adversarial review, use Codex's built-in default subagent instead of a detached background shell command.
• Do not satisfy background adversarial review by using a generic claude_review_runner-style helper role, raw Claude CLI, or any other review executor that bypasses the resolved companion command.
• Never satisfy background adversarial review by running the companion command itself with shell backgrounding such as &, nohup, detached spawn, or any equivalent direct background process launch.
• Background here means "spawn the forwarding child via spawn_agent and do not wait in the parent turn." The companion adversarial-review command inside that child still runs once, in the foreground, inside the child thread.
• Before spawning the built-in child, capture the review job id plus routing context in one call:

node "<plugin-root>/scripts/claude-companion.mjs" background-routing-context --kind review --json

• If that helper returns a non-empty jobId, pass it into the companion command as an internal --job-id <reserved-job-id> routing flag.
• Whenever forwarding that reserved --job-id, also pass --cwd <workspace-root> using workspaceRoot from the same helper response. Reserved job ids are workspace-scoped.
• If that helper returns a non-empty ownerSessionId, include --owner-session-id <owner-session-id> in the companion command.
• If it returns an empty ownerSessionId, omit --owner-session-id entirely. Never leave an empty placeholder such as --owner-session-id --job-id.
• If that helper returns a non-empty parentThreadId, pass it into the child prompt as the parent thread id for one-shot completion notification.
• If it returns an empty parentThreadId, omit the notification path instead of emitting a blank thread-id placeholder.
• Spawn exactly one transient forwarding child through spawn_agent with:
• fork_context: false
• reasoning_effort: "medium"
• no agent_type and no model, so the child uses the built-in default agent and inherits the parent model. Never pin a specific Codex model name here; the available catalog is owned by the host CLI and changes between releases.
• Prefer a self-contained child message over inheriting parent history. The built-in adversarial-review child should not rely on full parent thread replay for normal operation.
• Only consider fork_context: true as a last resort for a short follow-up where essential context truly cannot be summarized. Avoid it for large or long-lived threads because it can exhaust the child context window.
• Before spawning the built-in child, emit one short commentary update that clearly says the parent is starting the built-in adversarial-review child on the inherited model at medium effort.
• The built-in child must be a pure forwarder. It should:
• run exactly one shell command
• execute:

node "<plugin-root>/scripts/claude-companion.mjs" adversarial-review --view-state defer <arguments with --wait/--background removed>

• run that command as one blocking foreground shell-tool call, not as a background terminal/session
• do not request a shell session id, poll a shell session later, or return before the companion command exits
• if the available shell tool is exec_command, call it once in non-interactive mode and wait for command exit in that same call
• when using exec_command, pass sandbox_permissions: "require_escalated" and the justification Allow the Claude Code companion to contact the Claude API for this requested review. on that one call; do not first try the companion command in the default network-disabled sandbox
• include --owner-session-id <owner-session-id> only when the parent resolved a non-empty owner session id
• include --job-id <reserved-job-id> when the parent reserved one
• include the matching --cwd <workspace-root> whenever the command includes that reserved --job-id
• never leave an empty routing placeholder such as --owner-session-id --job-id
• return only that command's stdout exactly, with no added commentary
• ignore stderr progress chatter such as [cc] ... lines and preserve only the final stdout-equivalent result text
• not inspect the repo or perform the review itself
• if a parent thread id is available, allow one extra send_input call after a successful shell result and before finishing
• the child prompt must mention the tool name send_input literally; do not replace it with a vague instruction like "send a message to the parent"
• that send_input call must target the provided parent thread id, must happen at most once, and must not run on failure paths
• that send_input call should use the exact tool shape send_input({ target: <parent-thread-id>, message: <steering-message> }) with no extra prose payload
• if the parent provided a non-empty parent thread id, do not silently drop the completion notification path from the child prompt
• if a reserved review job id is available, use this exact notification message:

Background Claude Code adversarial review finished. Open it with $cc:result <reserved-job-id>.

• otherwise fall back to:

Background Claude Code adversarial review finished. Inspect it with $cc:status first, then use $cc:result for the finished job you want to open.

• that send_input message should use one of those exact steering messages instead of inlining the raw review result
• use these steering messages instead of embedding the raw review result in the notification
• do not embed the raw Claude result inside the notification message
• do not include any other prose in that notification message
• use that same steering message as the child's own final assistant message instead of echoing the raw review result
• Do not wait for completion in this turn.
• After launching, tell the user: Claude Code adversarial review started in the background. Check the subagent session or $cc:status for progress, and once it's done, we will let you know to see the results.
• Do not fix anything mentioned in the review output.