portswigger/agent-wrangler/advisor

Advisor

Claude Code has no built-in equivalent of the Claude API's advisor tool — a stronger model that a faster one can consult mid-generation for a plan or a course correction. This skill gives you the same pattern with two Agent Wrangler tools: spawn_session to bring up the advisor, send_message to consult it.

When to consult

• Before committing to a non-obvious design decision or a plan for a hard task.
• When you're stuck: an error keeps recurring, your approach isn't converging, or

a result doesn't fit what you expected.

• Before declaring the task done, on anything non-trivial — one last check for

something you might have missed. Make your work durable first (commit, save the file) so a slow consult can never cost you a finished result.

Skip it for simple, single-step lookups or mechanical edits — a consult adds latency and cost with nothing substantive to weigh in on.

Spawning the advisor

Call spawn_session once per task that needs one (not once per question):

• intent — an advisor persona briefing, not a work briefing. State plainly that

it is acting as an advisor for a peer session in the same directory: it will receive questions over send_message, should read whatever code it needs to answer well, must not edit any files, must not call workflow_phase, and should reply with focused advice via send_message back to your own session id ($AW_SESSION_ID) rather than attempting the work itself.

• cwd — your own working directory, not the default scratch dir, so the

advisor can read the same files you're working from. Do not set worktree.

• model — the strongest model available for the agent you launch (Claude:

opus; Codex: gpt-6-sol — see the spawn-session skill for the full menu). If you're already running that model yourself, spawn one anyway: a fresh context with no accumulated assumptions can catch what you've anchored past.

• nest: true — an advisor is a deliberate, designed child of the session that

spawned it (it exists only to be consulted by you, shares your cwd, and gets archived when you're done with it), so opt it into nesting explicitly. The board then groups it visually under your card instead of showing it as an unrelated top-level session.

Keep the returned session id — that's who you send_message for every consult on this task, and who you archive_session once you're done.

Consulting

For each question, send_message a self-contained brief: the advisor did not see any of your prior work and does not watch your files change, so state what's relevant now — the task, what you've tried or decided, and the specific question. Keep it tight; a wall of transcript defeats the point.

After sending, your turn ends. send_message now queues into the advisor's mailbox rather than pasting the body directly — you'll get back a short "you've got mail" notification once it replies, not the advice itself; call read_mail() to fetch what it actually said. ("Idle" here doesn't mean asleep — your session stays live at its prompt the whole time; the notification just arrives once the advisor has answered, same as waiting on a worker — see the spawn-session skill.) Reuse the same advisor session for later questions on this task instead of spawning a new one each time; it holds the thread of earlier advice.

Weighing the advice

Give it real weight, but don't switch course silently: if a step it suggested fails empirically, or you have direct evidence it didn't have, adapt. If your own evidence points one way and the advice points another, send one more message naming the conflict and asking which constraint wins, rather than picking a side unstated.

Wrapping up

Once you're done consulting for this task, archive_session the advisor — you cannot archive yourself, but you can archive it. Leaving it running just clutters the board with a session that has nothing left to do.