Home › Entities › Thread Taxonomy

Thread Taxonomy

Description

The thread taxonomy classifies the different execution contexts (query sources) in Claude Code. Internally represented as the QuerySource type (src/constants/querySource.ts), each query source identifies where an API request originates — main REPL, sub-agent, compaction, session memory, etc. The community has developed a parallel framework classifying agent work patterns into 7 thread types based on orchestration strategy.

Internal query sources

The QuerySource type defines the origin of each API query:

Source Purpose
repl_main_thread Primary REPL interaction — the user's conversation
sdk SDK/API programmatic usage
compact Context compaction (summarization)
side_question Side queries
agent Generic sub-agent
agent:custom Custom agent type
agent:explore Exploration agent (read-only)
agent:plan Planning agent (read-only)
tool_use_summary Tool use summary generation
advisor Advisor queries
hook Hook execution
session_memory Session memory extraction
magic_docs Magic docs generation
skill_search Skill search
classifier Auto-mode security classifier
bridge Bridge (IDE integration)

The type is extensible: (string & {}) allows custom query sources beyond the known set.

Query source matters for system behavior — for example, shouldAutoCompact() skips compaction for session_memory, compact, and marble_origami (context-agent) sources to prevent recursive compaction.

Community thread taxonomy

A community-developed framework classifies agent work patterns into 7 types, each optimizing along the "Core Four" dimensions: Context, Model, Prompt, Tools.

Type Pattern Description
Base Single prompt One prompt, one response. No agents.
P-thread Parallel (10-15 concurrent) Multiple agents working on independent subtasks simultaneously
C-thread Chained phases Sequential phases with verification gates between each
F-thread Fan-out + select Identical prompt to multiple agents, select best result
B-thread Boss/worker Orchestrator spawns and manages sub-agents
L-thread Long-running (26+ hours) Unattended execution over extended periods
Z-thread Zero-touch Aspirational: fully autonomous with no human intervention

How thread identity is tracked

Thread identity is established through two mechanisms:

  1. querySource — string tag on each API query identifying the execution context
  2. agentId + agentType on ToolUseContext — distinguish sub-agent threads from main thread

AsyncLocalStorage provides per-thread isolation when multiple agents run concurrently in the same process. Each agent's context carries: agentId, parentSessionId, agentType, subagentName, invokingRequestId, and invocationKind (spawn vs resume).

Trade-offs

  1. String-typed query source — extensible but no compile-time exhaustiveness checking. New sources can be added without updating a central enum, but consumers can't pattern-match safely.
  2. Community taxonomy vs internal — the community taxonomy describes user-facing orchestration patterns; the internal QuerySource describes system-level execution contexts. They overlap but aren't 1:1 (e.g., P-thread maps to multiple concurrent agent:* sources).
  3. Extensible type — (string & {}) allows any custom source but makes the known set advisory, not enforced.

Key claims

Relations

Sources

src-20260409-a5fc157bc756, source code analysis of src/constants/querySource.ts; community analysis