Module pipeline 

Dynamic Context Pipeline

A ContextPipeline solves the context-forgetting problem in long-running AI tasks. Instead of relying solely on a static ContextSnapshot captured at Run start, a ContextPipeline accumulates incremental ContextFrames throughout the workflow.

Position in Lifecycle

 ②  Intent (Active)         ← content analyzed
      │
      ▼
     ContextPipeline created ← seeded with IntentAnalysis frame
      │
      ▼
 ③  Plan (Plan.pipeline → Pipeline, Plan.fwindow = visible range)
      │  steps execute
      ▼
     Frames accumulate       ← StepSummary, CodeChange, ToolCall, ...
      │
      ▼
     Replan? → new Plan with updated fwindow

The pipeline is created after an Intent’s content is analyzed (step ②) but before a Plan exists. The initial IntentAnalysis frame captures the AI’s structured interpretation, which serves as the foundation for Plan creation. The Plan then references this pipeline via pipeline and records the visible frame range via fwindow. During execution, frames accumulate to track step-by-step progress.

Relationship to Other Objects

Intent ──plan──→ Plan ──pipeline──→ ContextPipeline
                  │                        │
             [PlanStep₀, ...]   [IntentAnalysis, StepSummary, ...]
                  │                        ▲
             iframes/oframes ──────────────┘

From	Field	To	Notes
Plan	pipeline	ContextPipeline	0..1
PlanStep	iframes	ContextFrame IDs	consumed context
PlanStep	oframes	ContextFrame IDs	produced context

The pipeline itself has no back-references — it is a passive container. PlanSteps own the association via iframes and oframes.

Eviction

When max_frames > 0 and the limit is exceeded, the oldest evictable frame is removed. IntentAnalysis and Checkpoint frames are protected from eviction — they always survive.

Purpose

• Context Continuity: Maintains a rolling window of high-value context for the agent’s working memory across Plan steps.
• Incremental Updates: Unlike the static ContextSnapshot, the pipeline grows as work progresses, capturing step summaries, code changes, and tool results.
• Bounded Memory: max_frames + eviction ensures the pipeline doesn’t grow unboundedly in long-running workflows.
• Replan Support: When replanning occurs, a new Plan can reference the same pipeline with an updated fwindow that includes frames accumulated since the previous plan.

Structs

ContextFrame

A single context frame — a compact summary captured at a point in time during the AI workflow.

ContextPipeline

A dynamic context pipeline that accumulates ContextFrames throughout an AI workflow.

Enums

FrameKind

The kind of context captured in a ContextFrame.