Kasi Multi

/kasi-multi

Fan-out mode — dispatch N specialist agents in parallel, each with isolated context and a dispatch brief. Main synthesizes the N reports into one answer.

Usage

/kasi-multi                       # default: 6 agents, auto-selected
/kasi-multi 4                     # 4 agents
/kasi-multi 6 <mission>           # 6 agents + explicit mission
sudo <mission>                    # shorthand: /kasi-multi 6 <mission>, skip clarifying Qs
sudo 8 <mission>                  # shorthand + custom N

What it does

• Fans the current mission out across N parallel specialist agents in a single dispatch message.
• Gives each agent an isolated context window — they do not see each other's inputs.
• Writes one dispatch brief per agent (MISSION / INPUTS / CONSTRAINTS / EXPECTED OUTPUT / PRIOR CONTEXT).
• Synthesizes the N reports into one user-facing answer, ranked by severity × confidence.
• Suggests next steps but does not auto-execute destructive ones.

sudo keyword

At the start of a message, sudo means:

1. Fan out by default — enter /kasi-multi with N=6 unless a number follows.
2. Skip clarifying questions — proceed on reasonable assumptions; narrate briefly.
3. Stay inside safety rails — destructive ops still require confirmation (hard rule #1).

sudo is a pacing signal, not a permission escalation.

Flow

1. Narrow the mission to one line. Even in sudo mode, refuse truly vague requests.
2. Select N specialists from the registry based on mission kind.
3. Write one dispatch brief per agent — no brief, no dispatch.
4. Dispatch all N in a single message (parallel tool calls, no serial waits).
5. Synthesize — merge, deduplicate, rank by severity × confidence.
6. Suggest next step; do not auto-apply destructive findings.

Default roster at N=6 (v0.10)

Slot	Agent	Job
1	architect-planner / deep-researcher	scope + research
2	general-purpose	scaffolding / build step
3	general-purpose	secondary build step
4	general-purpose	extract / data / glue
5	test-writer / general-purpose	verification
6	audit-specialist --focus=quality	review slot (v0.10 — was standalone code-reviewer pre-v0.10)

Main swaps in bug-hunter, audit-specialist --focus=security, audit-specialist --focus=perf, refactor-surgeon, migration-specialist, legacy-specialist based on mission kind.

Note: code-reviewer, security-auditor, perf-profiler are kept as name-resolution stubs in v0.10 — invoking them logs a deprecation notice. Use audit-specialist --focus=quality|security|perf|all going forward. Stubs disappear in v0.11.

N selection

N	When
2	Small mission, one dependency (researcher → implementer)
3	Triage — planner + worker + reviewer
4	Audit, 2–3 workstreams + reviewer
6 (default)	Medium mission — 4 workers + reviewer + researcher
8	Large mission, multiple subsystems
10	Max — entire registry, rare

Higher N is not always better. Default 6 has best cost / quality trade-off in observed runs.

When to use

• Audit or review with multiple workstreams (security + perf + tests + overall).
• Feature build decomposable into parallel chunks.
• Research needing multiple independent sources.
• Bug hunt where hypotheses are cheap to generate but expensive to verify.

When NOT to use

• One-line fix — fan-out overhead eats the savings.
• Hard sequential dependency chain — fan-out collapses to serial anyway.
• Haiku + large N — synthesis breaks down.

Tier behavior

• Opus — full fan-out allowed. Synthesis is where Opus earns its cost.
• Sonnet — default workhorse. Fans out to 6, prefers 4 on large contexts.
• Haiku — cap N at 4. Always include audit-specialist --focus=quality (verifier slot) — pre-v0.10 this was a standalone code-reviewer; now invoked via the --focus= flag. sudo implicitly caps at 4 even if user says 8.

Anti-patterns

• ❌ Fanning out on a one-line fix.
• ❌ Running N=10 out of habit.
• ❌ Dispatching agents without briefs.
• ❌ Forwarding raw agent outputs to the user without synthesis.
• ❌ Using sudo to bypass destructive-op confirmation — it does not do that.
• ❌ Large N on Haiku — synthesis collapses.

Examples

Review a PR fast:

sudo review PR #42 in /Users/me/repo

N=6, roster: audit-specialist --focus=quality (overall) + 2× audit-specialist --focus=quality (per-file) + audit-specialist --focus=security + audit-specialist --focus=perf + test-writer. Synthesis = ranked findings + go/no-go. (Pre-v0.10 invoked separate code-reviewer / security-auditor / perf-profiler agents — those are now --focus= flags on audit-specialist.)

Build a feature fast:

sudo 4 build /kasi-search as a mini semantic search over .kasidit/knowledge/

N=4, roster: architect-planner + general-purpose (scaffold) + general-purpose (indexer) + test-writer. Synthesis = integrated PR plan + file list.

Migrate a schema:

/kasi-multi 3 migrate users.email to citext in Laravel 8

N=3, roster: migration-specialist + deep-researcher + test-writer. Synthesis = expand-contract plan + rollback per phase.

Since

Introduced in v0.9.2.

See also

• Commands (aggregate)
• Master-Orchestrator
• Multi-Agent-Orchestration
• Kasi-Cascade