Dispatch Brief

Dispatch Brief

No brief, no dispatch.

The structured payload the Master-Orchestrator passes to every specialist agent. Introduced in v0.9.1 as part of the Master Orchestrator Rule; used by Kasi-Multi for fan-out dispatch in v0.9.2.

Format

Every specialist invocation must pass all five fields:

MISSION: <one sentence, verifiable outcome>
INPUTS:
  - <file paths, symptoms, measurements, versions>
CONSTRAINTS:
  - <deadline, compat, perf budget>
EXPECTED OUTPUT:
  - <matches the agent's documented output block>
PRIOR CONTEXT:
  - <findings from earlier specialists, if any>

Ad-hoc prompts produce ad-hoc output. The brief is the contract between master and specialist.

Field-by-field

MISSION

One sentence. Verifiable. Present-tense imperative.

• ✓ "Find the root cause of the 500 error on GET /api/sales after 2026-04-20 deploy"
• ✓ "Add regression test for the timezone offset bug fixed in commit abc123"
• ❌ "Look into the sales API" — not verifiable
• ❌ "Improve performance" — not scoped

If you cannot write the mission in one sentence, the scope is wrong. Narrow first.

INPUTS

Concrete. File paths, line numbers, symptoms, measurements, versions. Never "the codebase".

• File paths: app/Http/Controllers/SalesController.php:3708
• Symptoms: user report quoted verbatim; stack traces with line numbers
• Measurements: p95 latency 2.3s, query count 47 per request
• Versions: Laravel 8.83 / PHP 7.4 / MySQL 5.7

CONSTRAINTS

What the specialist must respect.

• Deadlines: ship before 2026-05-01 release
• Compatibility: must support Laravel 5.x → 8.x
• Performance budgets: < 200ms p95, < 50MB bundle
• Scope limits: touch only app/Models/, do not refactor related controllers

If no real constraints exist, write "none" — do not fabricate.

EXPECTED OUTPUT

What the specialist returns. Must match the agent's documented output block.

• bug-hunter returns: root cause + minimal fix + regression test plan
• architect-planner returns: file list + step sequence + trade-offs + open questions (no code)
• audit-specialist --focus=perf returns: top 5 findings ranked by impact × confidence (does not optimize) — pre-v0.10 was the standalone perf-profiler agent
• audit-specialist --focus=quality returns: severity-bucketed findings with confidence labels — pre-v0.10 was code-reviewer
• audit-specialist --focus=security returns: OWASP-aligned findings + verifier pass — pre-v0.10 was security-auditor
• test-writer returns: runnable test file + gap notes

Wrong shape = rejected output, re-dispatch.

PRIOR CONTEXT

Findings from earlier specialists in the same mission.

• "architect-planner produced plan X (attached summary)"
• "deep-researcher cached Laravel 8 upsert docs at .kasidit/knowledge/laravel-8-upsert.md"
• "initial bug-hunter pass found path but not root cause"

Empty on first dispatch. Never empty on follow-up within the same mission.

Example — real dispatch

MISSION: Audit SalesController.php for SQL injection, HIGH severity only

INPUTS:
  - File: app/Http/Controllers/SalesController.php (lines 1-9806)
  - Stack: Laravel 8.83, PHP 7.4, MySQL 5.7
  - Checklist: .kasidit/CHECKLISTS/security-php.md
  - Patterns: .kasidit/PATTERNS.md (error handling, DB query conventions)

CONSTRAINTS:
  - Confidence labels mandatory (Haiku session)
  - [unsure] items listed separately
  - No speculation — prove exploitability or do not mention

EXPECTED OUTPUT:
  - JSON-like list: [{ file, line, type, severity, confidence, fix_hint }]
  - Only severity = HIGH (ignore MED / LOW this pass)
  - Separate [unsure] section for user decision

PRIOR CONTEXT:
  - none (first dispatch in mission)

When to use

• Every invocation from master → specialist
• Every /kasi-multi fan-out (all N specialists get their own brief)
• Every follow-up dispatch within the same mission (carry prior findings in PRIOR CONTEXT)

When the brief is too heavy

Trivial exceptions the master may handle itself without a brief:

• One-line fix on a single file the user explicitly pointed to
• Pure question with no code change
• Reading top-level index (CLAUDE.md, PATTERNS.md, INDEX.md, RELATIONS.md)
• Final user-facing summary

Everything else → dispatch with a brief.

Anti-patterns

• ❌ "Do stuff" prompts to specialists — no mission, no output shape
• ❌ Skipping INPUTS because "the agent can figure it out"
• ❌ CONSTRAINTS omitted — leads to out-of-scope work
• ❌ EXPECTED OUTPUT as prose — "write some tests" instead of "runnable test file + gap notes"
• ❌ PRIOR CONTEXT forgotten on follow-up — specialist re-does earlier work
• ❌ Brief longer than the work itself — scope is wrong, narrow first

v0.10 — invoking audit-specialist

audit-specialist is the only audit agent in v0.10 (replaces code-reviewer / security-auditor / perf-profiler). Briefs must include the --focus flag:

MISSION: Pre-merge security gate on AuthService.php — only HIGH severity findings.
INPUTS:
  - app/Services/AuthService.php
  - .kasidit/CHECKLISTS/security-php.md (or default seeded)
CONSTRAINTS:
  - HIGH severity only
  - confidence labels mandatory
EXPECTED OUTPUT:
  - audit-specialist --focus=security <target>
  - findings array: [{file, line, type, severity, confidence, fix_hint}]
  - verifier pass note (which findings reproduced, which deferred to user)
PRIOR CONTEXT:
  - last audit found 2 SQL injection points in v1.4.x; patched in PR #1182

--focus=quality for code review, --focus=perf for profiling, --focus=all for the pre-merge full sweep.

See also

• Master-Orchestrator — the rule that requires briefs
• Multi-Agent-Orchestration — fan-out dispatch mechanics
• Agent-Audit-Specialist — single-entry audit agent (v0.10)
• Kasi-Multi — the command that dispatches N briefs in parallel
• Kasi-Cascade — tier-cascaded briefs (Opus → Sonnet → Haiku)
• v0.9.1 — introduction
• v0.9.2 — fan-out extension
• v0.10.0 — --focus flag for audit-specialist briefs