sendbird/cc-plugin-codex/rescue

Claude Code Rescue

By default, hand this skill off through Codex's built-in default subagent. Do not answer the request inline in the main Codex thread. Spawn exactly one rescue forwarding subagent whose only job is to run one companion task command and return that stdout unchanged. Foreground rescue responses must be that subagent's output verbatim.

Use this skill when the user wants Claude Code to investigate, implement, or continue substantial work in this repository.

Prefer $cc:rescue when the user wants Claude Code to diagnose the issue, validate a risky change by actually editing or testing, apply fixes from a prior review, or carry a task forward across multiple steps. Do not use rescue for "just review this diff" unless the user also wants follow-through work beyond review findings. Do not use rescue merely because the main Codex thread plans to fix things after combining its own review with a separate Claude review. Rescue is only the right delegation when Claude itself is supposed to investigate, edit, test, or otherwise own the follow-through work.

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" task ...

Raw slash-command arguments: $ARGUMENTS

Supported arguments: --background, --wait, --resume, --resume-last, --fresh, --write, --model <model>, --effort <low|medium|high|xhigh|max>, --prompt-file <path>, plus free-text task text

Companion defaults: model=opus, and no effort. The companion forwards --effort only when the user passes it, so fable, opus, sonnet, and haiku each keep Claude Code's own effort default. 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.

Main-thread routing rules:

• If the user explicitly invoked $cc:rescue or Claude Code Rescue, do not keep the work in the main Codex thread. Delegate it.
• If the user did not supply a task, ask what Claude Code should investigate or fix.
• Treat --background and --wait as execution controls, not task text.
• --background and --wait are Codex-side execution controls only. Never forward either flag to claude-companion.mjs task.
• The main Codex thread owns that execution-mode choice. It decides whether to wait for the subagent. The child subagent must never reinterpret those flags as companion flags.
• Treat --model, --effort, --resume, --resume-last, --fresh, and --prompt-file as runtime or routing controls, not task text.
• If the user task text itself begins with a slash command such as /simplify, /fix, or /review, treat that slash command as literal Claude Code task text to be forwarded unchanged. Do not execute or reinterpret it in the parent Codex thread.
• --model selects the Claude model for the companion task command only. It does not select the Codex subagent model.
• If the user explicitly passed --background, run the rescue subagent in the background.
• If the user explicitly passed --wait, run in the foreground.
• If neither flag is present and the rescue request is small, clearly bounded, or likely to finish quickly, prefer foreground.
• If neither flag is present and the request looks complicated, open-ended, multi-step, or likely to keep Claude Code running for a while, prefer background execution for the subagent.
• This size-and-scope heuristic belongs to the main Codex thread. The child subagent does not get to override it.
• If --resume or --resume-last is present without --wait, and the new instruction is substantial, open-ended, or likely to take more than a quick follow-up, the main thread should usually prefer background execution for the subagent. Keep that as a parent-side choice only. Do not inject --background into the child request or the companion command.
• Default to --write unless the user explicitly wants read-only behavior or only review, diagnosis, or research without edits.
• If --resume or --resume-last is present, continue the latest tracked Claude Code task. If --fresh is present, start a new task.
• If none of --resume, --resume-last, or --fresh is present, first run:

node "<plugin-root>/scripts/claude-companion.mjs" task-resume-candidate --json

• If that helper reports available: true, ask the user once whether to continue the current Claude Code thread or start a new one.
• Use exactly these two choices:
• Continue current Claude Code thread
• Start a new Claude Code thread
• If the user's wording is clearly a follow-up such as "continue", "keep going", "resume", "apply the top fix", or "dig deeper", recommend Continue current Claude Code thread first.
• Otherwise recommend Start a new Claude Code thread first.
• If the user chooses continue, add --resume before spawning the subagent.
• If the user chooses a new thread, add --fresh before spawning the subagent.
• If the helper reports available: false, do not ask. Delegate normally.
• Do not inspect the repo, do the task yourself, poll job status, or summarize the result in the same turn.
• If a legacy request still includes --notify-parent-on-complete, treat it as a compatibility alias. Background built-in rescue now attempts parent wake-up by default.

Subagent launch:

• By default, use Codex's spawn_agent tool. Omit agent_type; an omitted agent_type already selects the built-in default agent, and Codex only advertises that parameter when custom agents are configured.
• Never satisfy background rescue by launching claude-companion.mjs task itself as a detached shell process. Do not use &, nohup, detached spawn, or any equivalent direct background process launch from the parent.
• If a legacy request still includes --builtin-agent, treat it as a compatibility alias for the default built-in path. It should not change behavior.
• Prefer fork_context: false for the built-in rescue child. The parent should pass a self-contained forwarding message instead of replaying the full parent thread by default.
• 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.
• The built-in rescue path must omit model on spawn_agent so the child inherits the parent model, and must set reasoning_effort: "medium" so the transient forwarding child stays cheap and predictable. Never pin a specific Codex model name here; the available catalog is owned by the host CLI and changes between releases.
• Before spawning the built-in child, emit one short commentary update that clearly says the parent is starting the built-in rescue child on the inherited model at medium effort.
• Remove --background and --wait before spawning the subagent. Those flags control only whether the main thread waits on the subagent.
• Pass only the routing and task arguments that actually belong to claude-companion.mjs task.
• If the free-text task begins with /, preserve it verbatim in the spawned subagent request. Do not strip the slash or rewrite it into a local Codex command.
• Before spawning the built-in child, capture the task job id plus routing context in one call:

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

• If that helper returns a non-empty ownerSessionId, include --owner-session-id <owner-session-id> in the companion command so tracked Claude Code jobs stay attached to the user-facing parent session for $cc:status / $cc:result.
• If it returns an empty ownerSessionId, omit --owner-session-id entirely. Never leave an empty routing placeholder such as --owner-session-id --job-id.
• 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.
• Add an internal companion routing flag that reflects whether the user will see this result in the current turn:
• Foreground rescue must add --view-state on-success
• Background rescue must add --view-state defer
• Any user-supplied --model flag is for the Claude companion only and must be forwarded unchanged to task.
• 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, continue without parent wake-up instead of blocking the rescue.
• This parent wake-up attempt is now the default for background built-in rescue on persistent Codex/Desktop threads. It is still best-effort and should silently degrade on one-shot codex exec runs.
• For the built-in rescue path, the parent thread owns prompt shaping. The built-in child should stay a pure executor.
• For the built-in rescue path, treat the internal runtime reference at ../../internal-skills/cli-runtime/runtime.md as the command-building contract for the forwarding worker. It is an internal reference document, not a public skill to invoke.
• If the built-in rescue request is vague, chatty, or a follow-up, the parent may tighten only the task text before composing the exact companion command.
• Prefer passing a small structured <parent_context> block instead of forked thread history when the child needs a little prior context.
• Use the internal prompt-shaping reference at ../../internal-skills/task-prompt-shaping/prompt-shaping.md as deeper guidance for that parent-side tightening. It is an internal reference document, not a public skill to invoke.
• preserve user intent and add no new repo facts
• prefer a short delta instruction for resume follow-ups
• when helpful, use compact blocks such as <task>, <output_contract>, and <default_follow_through_policy>
• do not add more words than value for already-clear requests
• Parent-side shaping should be conservative and specific:
• If the request is already concrete, keep it literal.
• If the request names a concrete file, path, or artifact such as README.md, and also includes explicit source/style/installation constraints, keep the full task text literal apart from stripping routing flags. Do not compress it into a shorter delta.
• If the request refers to earlier work, rewrite it into a short delta that names the next thing Claude Code should change or inspect.
• If the user asks for "fix it", "keep going", or similar follow-ups, make the next objective explicit without inventing repo facts.
• If the user asks in mixed language, preserve the language mix and only tighten the execution intent.
• If the user implies an output format, make that output contract explicit instead of broadening the task.
• For --resume, --resume-last, vague follow-ups, or ambiguous continuation requests, prefer adding a compact <parent_context> block before the task command instead of relying on inherited history.
• Keep <parent_context> small and structured. Good fields include:
• mode (fresh or resume)
• job_id when the parent reserved one
• claude_session when a resumable Claude session is already known
• previous_summary only when the parent can state it tersely from tracked metadata
• next_delta for the exact next objective
• constraints only when they are explicit and still binding
• Do not use <parent_context> for already-clear fresh tasks unless it adds real value.
• Kee