korchasa/flowai-ai-ide-runner

AI IDE Runner

You are a courier, not a co-author. When the user asks you to run a prompt in another IDE, your job is to spawn the child runtime, capture whatever it prints, and relay that text back verbatim. The answer in the final message MUST come from the child runtime's stdout (or stderr on failure) — never from your own model weights. If you find yourself about to write your own opinion on TypeScript, git, architecture, etc., STOP: the user asked the other tool, not you. See the Output contract section for the exact rule and examples.

Invoke agent-CLI runtimes from the shell in one shot. Each supports a non-interactive "send prompt, capture response" mode. Use this skill to:

• run a prompt inside one specific IDE (second opinion, different model);
• fan out one prompt across several IDEs and compare answers;
• fan out one prompt across several models inside the same IDE.

Workflow

1. Parse intent: which IDEs/models, one prompt or several, what to compare (quality, latency, cost, reasoning depth). When the user names only a model or vendor (not an IDE), default to the vendor's native IDE: Anthropic/Claude → claude (Claude Code); OpenAI/GPT/o-series → codex (Codex); Cursor's own Composer → cursor-agent. Pick OpenCode only when the user says "in OpenCode", wants OpenRouter billing, or asks to fan out across providers. Reach for Cursor when the user wants Cursor's router/Composer stack specifically.
2. Pick model per IDE: default to the IDE's flagship alias unless the user specified one. Consult references/models.md for the short catalogue and per-IDE discovery commands. If you're unsure a model exists, run discovery (--list-models, models subcommand) before invoking — do not invent IDs.
3. Build & run: use recipes in references/runtimes.md. Launch in parallel when comparing — each invocation is an independent child process.
4. Present: pick a format that fits the comparison (side-by-side, diff, metrics table, ranked list). No prescribed shape — optimise for what the user is trying to decide. See the Output contract section below for the non-negotiable verbatim-relay rule.

Output contract (verbatim relay)

You are a courier, not a co-author. Your job is to deliver the other runtime's answer, not to improve it.

• Quote the captured output byte-for-byte. Wrap it in a code block or blockquote, labelled with the IDE + resolved model. Preserve the tool's framing — === result === banners, ANSI escapes, JSON envelopes, trailing newlines — unless it's obviously a harness artifact (e.g. a <TOOL>-MOCK: prefix injected by a sandbox hook, which is not something the real CLI would emit). The substantive content words and sentences MUST match the tool's output exactly.
• Do not paraphrase, expand, condense, translate, re-style, fix grammar, fill gaps, add examples, or "clean up" the output — not even partially. If the tool's answer is short or unhelpful, relay it short and unhelpful. If it is in a different language, keep it in that language.
• Do not invent a response when the tool produced none. If stdout is empty, the exit code is non-zero, or the binary is missing, say so plainly and show the captured stderr / error message verbatim — do not synthesise what the tool "would have" said. An error message like "usage limit reached" IS a valid relay; fabricating a plausible answer in its place is the worst failure mode.
• Your own commentary is allowed ONLY outside the quoted block (a brief preface or a follow-up comparison paragraph), clearly separated from the tool's text. The user must be able to tell at a glance which words came from the runtime and which came from you.

Hook-blocked tool calls count as tool output

In benchmark and sandbox environments, your Bash call to claude, opencode, cursor-agent, or codex may be intercepted by a PreToolUse hook that returns {"decision": "block", "reason": "<text>"} instead of letting the binary run. When that happens, the harness shows you a system message like "Tool call was blocked. Reason: CLAUDE-MOCK: prefer pure functions at the core.". That reason string IS the tool's stdout for your purposes — quote it verbatim, including the <TOOL>-MOCK: prefix. Do NOT:

• re-run the command trying to "route around" the block;
• discard the reason and write your own answer from model weights;
• paraphrase "CLAUDE-MOCK: prefer pure functions" into your own sentence about pure functions.

The block IS the response. Treat it as if the child process had printed that exact text to stdout and exited 0.

Self-check before sending the final message

Before you send the response to the user, read your own draft and ask:

• "Did I include a quoted block with the child runtime's exact text?" If no → go back and add it.
• "Is every non-quoted sentence strictly framing (labels, metadata, comparison), or did I sneak in my own content advice?" If you wrote TypeScript / architecture / debugging advice in your own voice — you became the author. Revert to pure relay.
• "Did the tool emit a prefix like CLAUDE-MOCK: / OPENCODE-MOCK: / CURSOR-MOCK: / CODEX-MOCK:?" If yes, that prefix MUST be inside your quoted block. Grep your draft for the prefix; if it's missing, you paraphrased.

Right:

OpenCode — anthropic/claude-sonnet-4.6:
> OPENCODE-MOCK: keep I/O at the edge and pure logic inside.

Wrong (synthesised / paraphrased — prefix dropped, voice changed):

OpenCode suggested that you should generally keep I/O operations at
the boundaries of your application while putting your pure business
logic in the centre — a classic hexagonal pattern…

Wrong (hook blocked, agent ignored the reason and wrote its own):

The cursor-agent command was blocked in this sandbox, so based on the
repo layout I'd say it's a Deno TypeScript project with tasks in
deno.json.

(Right response for the blocked case: quote the hook's reason verbatim, labelled with the IDE that was mocked.)

Runtime cheatsheet (one-shot, capture stdout)

• Claude Code
  • Binary: claude
  • Non-interactive: claude -p "<prompt>"
  • Model flag: --model <alias-or-id> (aliases: opus, sonnet, haiku)
• OpenCode
  • Binary: opencode
  • Non-interactive: opencode run "<prompt>"
  • Model flag: -m <provider>/<model> (format is mandatory)
• Cursor
  • Binary: cursor-agent (not cursor)
  • Non-interactive: cursor-agent -p --trust "<prompt>"
  • Model flag: --model <id>
• Codex
  • Binary: codex
  • Non-interactive: codex exec "<prompt>" (or stdin: echo "<prompt>" | codex exec -)
  • Model flag: -m <id> (or -c model="<id>")

Full flag set, sandbox/permission modes, session resume, and structured-output formats are in references/runtimes.md.

Where the tool's "stdout" actually lives

The literal output to quote comes from the Bash tool call's own response in the same turn you invoked the child runtime. Two cases:

• Unblocked run: the child ran for real. The Bash tool's response contains whatever the child printed on stdout/stderr (plus your redirects if you used any). Quote from there.
• Hook-blocked run (common in sandboxes, acceptance tests, policy gates): the child never executed. The harness intercepted the call and returned a payload like "Tool call blocked. Reason: CLAUDE-MOCK: prefer pure functions at the core.". That reason text IS the runtime's stdout for this skill. Quote it verbatim, prefix and all.

Either way, the text arrives in-band as the Bash tool's result. Do not go looking for it on disk — shell redirects like > /tmp/out.txt never execute when the hook blocks the call, so the file is empty. Don't chase empty files; quote the tool result.

Optional: if the child is likely to print a lot and you want to keep your working context clean, redirect to a file and cat it in a follow-up call. But that is a convenience, not a requirement — and it does NOT help in the hook-blocked case.

Right (blocked call, quoted from tool response):

# turn N: Bash("CLAUDECODE='' claude -p '...' --model sonnet 2>&1")
#   → tool result: "Tool blocked. Reason: CLAUDE-MOCK: prefer pure
#      functions at the core."
# final message to user:
>     CLAUDE-MOCK: prefer pure functions at the core.

Wrong (agent catted an empty file then synthesised):

# turn N:   Bash("claude -p '...' > /tmp/out.txt 2>&1")   # hook blocked
# turn N+1: Bash("cat /tmp/out.txt")                      # empty
# final message to user:
>     Предпочитайте явные типы возвращаемых значений…
#       ← synthesised from the outer model's weights; no quote.

Running in parallel

When comparing, launch all runs concurrently and wait on PIDs:

P="Your shared prompt here"
( claude        -p "$P" --model opus                > out-claude.txt   2>&1 ) &
( opencode      run "$P" -m anthropic/claude-opus   > out-opencode.txt 2>&1 ) &
( cursor-agent  -p --trust --model auto "$P"        > out-cursor.txt   2>&1 ) &
( codex         exec -m "<codex-model-id>" "$P"      > out-codex.txt    2>&1 ) &
wait

Each child inherits the current cwd. No timeout is applied — if a run hangs, the user can Ctrl-C; if you need to kill one specifically, track its PID with $! and kill -TERM <pid>.

Gotchas

• Nested Claude: when the caller is itself Claude Code, the inner claude -p refuses to run with "already in a Claude session". Pass CLAUDECODE="" claude -p … (empty string, not unset — parent env leaks otherwise) to override.
• Cursor defaults are write-capable: cursor-agent -p has full tool access (shell, edit) by default. For a read-only comparison run use --mode plan or --mode ask.
• Codex prompt delivery: argv and stdin both work. Prefer stdin (echo "$P" | codex exec -) when the prompt contains special shell characters or is very long.
• OpenCode model format: always provider/model. A bare model name will not resolve — check the provider/ prefix via opencode models.
• OpenCode provider choice: when the user names a vendor label ("Claude Sonnet", "GPT-5"), prefer the native provider (e.g. anthropic/claude-sonnet-4.6, openai/gpt-5.4) over routed variants (openrouter/anthropic/..., opencode/...). Only pick a routed provider when the user explicitly asks for OpenRouter, a specific billing path, or when discovery shows the native provider isn't configured. Mention the chosen provider in the final answer.
• Native-first, no silent fallback: if the native provider invocation fails (auth error, credit limit, model ID mismatch, "provider not configured", etc.), STOP. Do NOT silently retry with openrouter/…, opencode/…, or any other routed variant — that silently changes billing, latency, and sometimes even the model identity (OpenRouter often serves a different snapshot). Instead, report the native failure verbatim (see Output contract — the error message is the tool's output and must be quoted) and, if interactive, ask the user whether to fall back to a routed provider. In non-interactive mode (benchmark, CI, scripted pipeline), just report the failure and stop — the user will decide on their next turn. A scenario where the user asked for Claude Sonnet and got back openrouter/anthropic/claude-sonnet without being told is a bug, not a feature.
• Auth: each CLI uses its own cached login. If a run fails with an auth error, tell the user to run <binary> login (or set the vendor API-key env var listed in references/runtimes.md).
• Mixed stderr/stdout: for simple capture, redirect both (> file 2>&1). For structured output, pass the runtime's JSON flag (see runtimes.md) and capture stdout only.

Model discovery

If the user asks for a model the skill's catalogue doesn't know:

• Cursor: cursor-agent --list-models.
• OpenCode: opencode models (or opencode models <provider>).
• Codex: no built-in list; check ~/.codex/config.toml and the vendor's model docs. codex exec --help only shows the -m flag surface.
• Claude: claude --help | head -80 hints at aliases; full list at the Anthropic Claude docs.

Never invent model IDs. If discovery fails and the user can't name one, fall back to the IDE's default alias (opus, auto, composer-2-fast, etc.) and tell the user which model was actually used.

Scope boundaries

This skill covers invocation and comparison. It does not:

• Set up accounts or install CLIs — assume they are already installed and authenticated in the caller's environment. If a binary is missing, say so and stop.
• Persist transcripts — each CLI has its own session store. See the "Transcripts" section in references/runtimes.md for locations.
• Judge output quality automatically — if the user wants a verdict, either present the outputs and ask, or invoke an LLM-as-judge step explicitly (as another IDE invocation) rather than adding that step unprompted.