Skip to content

TERMINOLOGY.md

Latest commit

 

History

History
109 lines (82 loc) · 3.57 KB

TERMINOLOGY.md

File metadata and controls

109 lines (82 loc) · 3.57 KB
Raw

Bitcode Terminology Guide

"Documentation should be a mirror of the source code, not an interpretation."

This guide provides the precise terminology used in the Bitcode codebase, verified against actual implementation.

Core Execution Patterns

PTRR Pattern (Current Implementation)

Source: /packages/agent-generics/src/types.ts, /packages/generic-agents/PTRR

The codebase uses AgentVariationStep enum:

export enum AgentVariationStep {
 PLAN = 'plan',
 TRY = 'try',
 REFINE = 'refine',
 RETRY = 'retry'
}

Preferred factory: factoryPTRRAgent (factoryPTRRAgent is ).

Generation hierarchy within a PTRR step

Source: /packages/generation-generics/src/types.ts

Generation (primitive)
 → FailsafeGeneration (base kinds — parents)
 → ThinkingsGeneration (base kinds — children of each failsafe)

FailsafeGeneration kinds (3)

export enum FailsafeGeneration {
 PREPARE_CONCISE_CONTEXT = 'prepare_concise_context', // CONTEXT SIGNAL/NOISE
 CHUNK_THEN_SUM = 'chunk_then_sum', // BIG INPUT
 STITCH_UNTIL_COMPLETE = 'stitch_until_complete' // BIG OUTPUT repair
}

ThinkingsGeneration kinds (3)

export enum ThinkingsGeneration {
 REASON = 'reason',
 JUDGE = 'judge',
 STRUCTURED_OUTPUT = 'structured_output'
}

Tools (postprocess, not a Failsafe/Thinkings generation)

  • tools_execution — after all failsafes, once, if structured output includes useTools

Total per PTRR step: 3 FailsafeGenerations × Thinkings composition + optional tools postprocess.

Architecture interface: PTRRStepGenerationArchitecture.

Tools: parameters, docs, results

Term Meaning
Usable tools Tools visible via AgentToolsRegistry.getUsableTools() (hierarchy).
Doc interpolation formatUsableTools → auto:tools_doc_code_tools on Thinkings prompts.
useTools LLM selection: { name, input, reason }[] on step structured output.
usedTools Execution results: { tool, input?, output?, error? }[].
Results interpolation Prior usedTools → auto:tools_results on later generations.
DocCodeToolPrompt Prompt carrier for purpose/capabilities/parameters/output sections.

Full guide: packages/agent-generics/TOOLS-IN-PTRR.md.

Legacy names (do not use in new code)

Legacy Prefer
ThricifiedGeneration ThinkingsGeneration
SubStep / GenerationExecution Generation / GenerationExecution
FailsafeExecution FailsafeGenerationExecution
PTRRSubStepArchitecture PTRRStepGenerationArchitecture
AgentGenerationSubStepPrompt ThinkingsGenerationPrompt

FailsafeGeneration and FailsafeGenerationPrompt are already the modern names (no rename). "SubStep" was the old term for Generation within a Step. "Meta" is not a term.

Prompt System

PromptPart

Source: /packages/prompts/src/parts/PromptPart.ts

  • Branded string type: string & { readonly __brand: 'PromptPart' }
  • Created via createPromptPart(content: string): PromptPart

Prompt Class

Source: /packages/prompts/src/prompt.ts

  • Registry-based formatting system
  • Hierarchical path organization

Generation prompt levels

  1. AgentPrompt — name + identity
  2. AgentStepPrompt — Plan/Try/Refine/Retry purpose
  3. FailsafeGenerationPrompt — failsafe handling instruction
  4. ThinkingsGenerationPrompt — Reason/Judge/Output instruction
  5. ToolExecutionPrompt — tool postprocess

Hosts

Use Host for pipeline execution boxes (LocalHost, SandboxHost). Do not use "Harness" for Host.