Rescue

Rescue

Single entrypoint for delegating to one of three AI backends — Claude Code (claude), Codex (codex), or Gemini/Antigravity (agy) — or running all three in parallel for a synthesized consensus. The backend matching the current host runs natively; the others run through their CLIs.

This skill owns routing, session policy, and consensus synthesis. The per-backend reference files in {{SKILL_DIR}}/references/ own the actual CLI invocation. The skill itself never runs a CLI or solves the task, and there are no custom rescue agents — it delegates through the available subagent/delegation facility.

The host backend

One of the three backends matches the agent currently running this skill — the host backend (claude on a Claude Code host, codex on a Codex host, gemini on a Gemini/Antigravity host). A subagent is already a fresh, independent instance of the host system, so for the host backend it reasons natively — no CLI subprocess. The other two backends are reached only through their CLIs, per their reference files. This keeps the skill host-neutral: whichever system you run under is the native perspective, and the other two are the external ones.

Context to carry

Every backend you spawn — native subagent or CLI forwarder — starts with none of this conversation's context. It sees only the task text you forward plus whatever it reads from the repo itself. So before forwarding, fold the parts that live only in this conversation — findings so far, decisions made, ruled-out paths, constraints, and the file or path pointers that matter — into the task text. This is not repo investigation (the backend does that itself); it is transferring what this conversation already knows. Consensus adds to this: it may also read referenced files and embed their content, since each analyst evaluates a self-contained question.

Resolve {{SKILL_DIR}} to this skill's directory before pointing a subagent at a reference file.

Raw user request: $ARGUMENTS

If the host did not expand $ARGUMENTS, treat the user's current message as the request rather than the literal string.

Routing

Pick the mode from the request:

Only an explicitly named single backend bypasses consensus. When in doubt, run consensus.

If the user supplied no task at all, ask what to investigate, review, or fix before routing.

Consensus mode

Read {{SKILL_DIR}}/references/consensus.md and follow it. It covers stance assignment, the parallel spawn of all three backends, the per-analyst prompt template, and the synthesis structure.

Single-backend rescue

Session and execution policy

This policy is shared across all three backends; the reference translates the resulting flags into the CLI invocation.

• --background: run the subagent in the background when the host supports background delegation. --wait: foreground. Neither: foreground.
• --background / --wait are host-level execution controls. Do not forward them into the CLI or the subagent's task text.
• For a background run, return only the host's background handle or status. Do not claim to have the final output yet.
• If the request already includes --resume, --fresh, or --conversation <id>, the user chose; do not ask about session.
• If no session flag is present and the request is clearly a follow-up to prior rescue work ("continue", "keep going", "resume", "apply the top fix", "dig deeper"), add --resume.
• If no session flag is present, the request is not a clear follow-up, the host can ask a concise two-option question, and there is an obvious prior rescue run for this backend in the conversation, ask once whether to continue that run or start fresh.
• Otherwise add --fresh.
• A --fresh run is ephemeral by default — the reference adds the CLI's no-persist flag where one exists, so a one-off rescue leaves no resumable session on disk. Keep the session only when the user passes --persist (a fresh run they intend to continue later). --resume/--conversation operate on an existing saved session and are never ephemeral. The native host backend creates no separate CLI session, so this applies only to CLI forwarders.

Capability

Hand off in default mode. For a CLI backend, pass no permission, sandbox, or bypass flag — the reference maps this to that CLI's default invocation. For the native host backend, the subagent runs with its normal permissions. There is no read-only/write toggle; the backend decides what to touch from the task itself.

Spawn

If the requested backend is the host backend — native (default). Spawn one subagent through the available subagent/delegation facility and give it the rescue task directly. It is a fresh, independent instance of the host system; it reasons natively with no CLI call. Apply the capability and output contract below, and honor --model/--effort as a subagent setting when the facility supports it.

Fall back to the host backend's CLI only when the request explicitly needs CLI-only behavior — --resume or --conversation <id>, the session persistence a fresh subagent cannot provide. Then use the forwarder spawn below.

Otherwise — CLI forwarder. Spawn one subagent and give it this prompt, with <backend> one of claude, codex, gemini:

Read {{SKILL_DIR}}/references/<backend>.md and follow it exactly. Act only as the
thin CLI forwarder it describes — do not inspect the repository, solve the task, or
add analysis of your own. Return the CLI's stdout verbatim as your final message.

Rescue request:
<task text, with the session flags resolved above; --background/--wait stripped>

Pass the resolved session and runtime flags through in the task text (--resume/--fresh/--conversation, --model, --effort, --add-dir, and any backend-specific flags). The reference maps them to the CLI.

Output

• Foreground: return the subagent's final output verbatim — the CLI's stdout for a forwarder, or the native subagent's result for the host backend. Do not paraphrase, summarize, rewrite, or add commentary before or after it.
• Background: return only the host's background handle or status.

Operating rules

• For single-backend rescue, do not inspect the repo, read files, grep, run the CLI, or solve the task yourself — the subagent owns that (a CLI forwarder via its reference, or the native host-backend subagent reasoning directly). (Consensus mode may read files to embed context into the analyst prompts; that is expected and described in the consensus reference.)
• Do not fall back to a direct shell call from the skill. If the target CLI or subagent facility is unavailable, the subagent reports it; surface that and stop.
• Do not monitor or poll a background run unless the user later asks for status.