Trace Ingestion Agent

Trace ingestion is the hot path. It turns one indexed source session into a small episode record plus zero or more durable context records.

The graph below is generated from the compiled LangGraph runtime.

---
config:
  flowchart:
    curve: linear
---
graph TD;
    __start__([<p>__start__</p>]):::first
    resolve_scope(resolve_scope)
    read_window(read_window)
    scan_window(scan_window)
    filter_signals(filter_signals)
    synthesize_records(synthesize_records)
    review_records(review_records)
    persist_records(persist_records)
    __end__([<p>__end__</p>]):::last
    __start__ --> resolve_scope;
    filter_signals --> synthesize_records;
    read_window --> scan_window;
    resolve_scope --> read_window;
    review_records --> persist_records;
    scan_window -.-> filter_signals;
    scan_window -.-> read_window;
    synthesize_records --> review_records;
    persist_records --> __end__;
    classDef default fill:#f2f0ff,line-height:1.2
    classDef first fill-opacity:0
    classDef last fill:#bfb6fc

Inputs

• indexed session metadata
• canonical trace JSONL path
• project or custom scope identity
• recent record manifest for duplicate awareness

Flow

1. resolve_scope prepares the context-store target.
2. read_window reads a deterministic trace window.
3. scan_window observes candidate findings and loops until the trace is read.
4. filter_signals keeps only reusable signal.
5. synthesize_records drafts episode and durable records.
6. review_records checks whether records are useful and well-shaped.
7. persist_records writes validated records and provenance.

Output

The output is written to the context store with evidence back to the source session. A clean run can create no durable records when the trace has no reusable context.