Context Partitioning — Manifest-First Dispatch

The Problem

Without explicit context isolation, the main orchestrator context window accumulates:

Full file reads that were only needed once
Subagent outputs that are already committed to disk
Skill files that don't apply to the current task
Historical reasoning that compaction will destroy anyway

Result: context bloat → compaction → rule violations → wrong output.

The Solution: Manifest-First

Rule: Before dispatching ANY subagent, the orchestrator MUST produce a typed manifest. The manifest is the subagent's entire world — it reads NOTHING outside its manifest.

Manifest Format

{
  "agent": "frontend-specialist",
  "task": "Add persona avatar upload to SocialPersonasManager",
  "toolshed": "frontend",
  "token_budget": 60000,
  "files": {
    "must_read": [
      "src/components/social/SocialPersonasManager.tsx",
      "src/lib/businesses.ts"
    ],
    "reference_only": [
      "src/app/globals.css"
    ],
    "must_not_touch": [
      "src/lib/supabase/",
      "supabase/migrations/"
    ]
  },
  "skills": ["scientific-luxury", "react-best-practices"],
  "constraints": [
    "founder_id isolation — never workspace_id",
    "Scientific Luxury: #050505 bg, #00F5FF accent, rounded-sm only",
    "Framer Motion only — no CSS transitions"
  ],
  "success_criteria": "Avatar upload renders, persists to Supabase Storage, no TypeScript errors",
  "output_format": "edited_files_list + verification_tier_A_result",
  "verification_agent": "verification"
}

Orchestrator Rules (Hard Constraints)

What the Orchestrator MUST do:

Glob/Grep only — discover file paths, never read file contents in main context
Produce manifest first — before calling any Agent tool
Pass manifest as prompt — subagent's entire context = manifest + agent.md
Receive structured summary — subagent returns file list + verification output, not full code

What the Orchestrator MUST NOT do:

Read full file contents (use Grep for specific lines if truly needed)
Hold subagent working files in main context
Load skills that don't apply to the current routing decision
Re-read files the subagent has already handled

Token Budget Enforcement

Agent Role	Hard Cap	Consequence of Breach
Orchestrator	80,000	Delegate immediately — do not read more files
Frontend Specialist	60,000	Scope to `src/components/` + `src/app/` only
Senior Fullstack	60,000	Scope to task-relevant `src/app/api/` + `src/lib/` files
Database Specialist	40,000	Scope to `supabase/migrations/` + schema file only
Security Auditor	50,000	Scope to auth path + affected routes only
Bug Hunter	40,000	Failing file + direct imports one level deep only
Test Engineer	50,000	Test file + source file under test only
Verification	30,000	Verification output only — no source reads

Context Scope Definitions

Scope Token	Resolves To
`src/`	Full src directory — only for orchestrator planning
`failing-file-only`	The single file containing the error
`direct-imports-one-level`	Files imported by the failing file (not transitive)
`test-files`	`src/*/.test.ts`, `src/*/.spec.ts`, `tests/`
`source-file-under-test`	The exact file being tested
`task-specific`	Orchestrator selects based on task decomposition

Manifest-First Dispatch Protocol

Step 1 — DISCOVER (Orchestrator, Glob/Grep only)
  → Identify affected files by pattern, not by reading
  → Check VAULT-INDEX.md for known asset locations
  → Load toolshed entry from toolsheds.json

Step 2 — MANIFEST (Orchestrator, write manifest)
  → Populate must_read, reference_only, must_not_touch
  → Select max 5-6 skills from toolshed
  → Set token_budget from toolsheds.json
  → Write success_criteria in verifiable terms

Step 3 — DISPATCH (Orchestrator → Agent tool)
  → Pass manifest as the agent's complete prompt
  → Agent reads ONLY manifest.files.must_read
  → Agent loads ONLY manifest.skills

Step 4 — RETURN (Subagent → Orchestrator)
  → Return: list of files modified + verification output
  → Do NOT return full file contents
  → Do NOT return intermediate reasoning

Step 5 — INTEGRATE (Orchestrator)
  → Accept structured summary
  → Route to verification agent
  → Main context stays lean

Anti-Patterns (Never Do)

Anti-Pattern	Why	Fix
`Read("src/")` in orchestrator	Dumps entire codebase into context	Glob for paths only
Subagent reads files outside manifest	Context bleed, token waste	Manifest must_not_touch
Return full file content from subagent	Doubles context usage	Return file path + diff summary
Load all P1 skills for every task	Token waste on irrelevant skills	Load toolshed subset only
Re-verify work in orchestrator	Redundant, wastes context	Delegate to verification agent

Integration with Toolsheds

This skill works in tandem with .claude/data/toolsheds.json:

Each toolshed entry defines token_budget, skills, and context_scope
Orchestrator selects toolshed by task domain → produces manifest → dispatches

Reference: .claude/AGENT_HARNESS.md Phase 4 for dispatch protocol.

context-partitioning