Playbooks — wielding the fleet

Every other page documents a capability. This one maps situations to moves: which play to reach for, and the exact call. If you are an agent driving clikae, this is your decision layer — AGENTS.md and orchestration.md give the mechanics and the non-negotiable rules; this gives the when.

The through-line: an account switcher gives you another login. clikae gives you a fleet you orchestrate — route by cost, verify across vendors, run in a clean room, survive a dry tank. Every play below leans on the rules in orchestration.md and agy-dispatch.md; read those first if a call surprises you.

Reading the commands below. Each call is clikae <engine> <tank> …: the engine is the CLI (claude, codex, agy), the tank is one account/config for it. Names like worker, work, g, acme are placeholders — run clikae tanks to list yours and clikae (the board) to see which still have fuel. Which engine you name is just which account you're routing to; the play is the same whichever it is.

A review you can trust — with no self-bias

Reach for this when the session that produced the work can't audit it fairly — to a writer, handled reads as correct. You want eyes without your assumptions.

The move — run the review in a clean room, on a different model family:

# agy (Gemini underneath) browses live and writes a report; it holds no memory of your work.
clikae agy g -- --print-timeout 900s -p "$(cat /tmp/audit-prompt.txt)"

For a clean room on the same family: clikae claude <tank> --ephemeral — the session runs, but the tank's long-term memory never learns it happened.

Verify / gotcha: read the report file, never the exit code. Triage the findings — agy is a grader, not an oracle, and will confidently flag things that violate a decided call. Apply fixes with your main engine; never auto-apply agy's.

Why a switcher can't: it gives you a second login. This gives you a second perspective — another vendor's model, carrying none of your session's baggage.

Grunt work off your main quota

Reach for this when a well-specified, easily-checked job (reformat, summarise, boilerplate, a mechanical transform) is burning the quota of the model you need for the hard part.

The move — burn it on a cheap tank, verified by the artifact:

clikae burn claude worker \
  --artifact out/result.json --prompt-file /tmp/task.md --add-dir "$PWD"

Verify / gotcha: success = the artifact appears (or its mtime changes). Never trust the exit code — codex exec / claude -p exit 0 on a dry tank having written nothing. Don't hand-roll -- -p '…' (it runs read-only and can't write); let --prompt-file + --add-dir fill in each engine's write dialect.

Why a switcher can't: it stops at "you're logged into the other account now." burn routes the job — re-fires it on the next reserve tank when one runs dry, and skips the account your interactive session is on. The expensive supervisor stays asleep; the cheap workers burn whatever account still has gas.

A high-stakes call — best-of-N with a real vote

Reach for this when one model's confident answer isn't enough to bet on (a design decision, a security-adjacent review). You want a vote, not a speed-up.

The move — fan the SAME read-only prompt across different families/tiers, then have a separate model grade the outputs:

clikae conduct --prompt-file /tmp/review.md \
  --leg claude/main --leg codex/work --leg agy/g --add-dir "$PWD"

Verify / gotcha: conduct never judges — it hands you every leg's output. Route those through a neutral grader (a model that sees only the outputs, not the reasoning that produced them); it catches the confident-wrong error a producer's self-assessment misses. The agy leg runs on its currently-active tank only.

Why a switcher can't: fanning the same prompt across the same account is parallelism, not a correctness vote. The vote needs different vendors plus a third grader — that's orchestration, not switching.

Keep a long job alive past a dry tank

Reach for this when an unattended, multi-minute job must finish somewhere even if the tank you start it on hits its usage limit.

The move — burn it; the reserve walk is automatic:

clikae burn codex work \
  --artifact build/out.tar --prompt-file /tmp/job.md --add-dir "$PWD" --timeout 1800

Verify / gotcha: artifact present = done; dry on every reachable tank = a real failure (not rerouted — it would fail the same everywhere). Make the task idempotent so a re-fire costs nothing. (This is the same burn as Grunt work off your main quota — a job that is both cheap and long is both plays at once; one call does the cost-routing and the dry-tank failover.)

Why a switcher can't: it makes you notice the wall and re-drive by hand. burn's reserve is the failover — account-aware, artifact-checked, hands-off.

Separate worlds, one keystroke

Reach for this when you run work for several clients or personas and a leak between them — a wrong git identity, a cross-loaded MCP connector, a shared memory — would be a real breach.

The move — one tank per world; wall the sensitive ones off:

clikae claude acme          # that client's login, MCP connectors, memory, git-id — nothing else's
clikae solo claude acme     # make it standalone: never shareable, never in a burn/relay set

Verify / gotcha: clikae memory status shows what a tank shares; a solo tank refuses share. Commits carry the tank's identity only if you set it with clikae git-id (see orchestration.md §6).

Why a switcher can't: it swaps a login. Each clikae tank is a whole isolated world — memory, connectors, commit identity — and solo makes that isolation structural, not a habit you have to remember.

See also: orchestration.md (the rules behind these plays), agy-dispatch.md (the cross-family engine, read before using agy), memory.md (the share / isolate / ephemeral dial).