teernisse ad484a2a6f feat: add JSONL source layer with directory scanner and byte-level parser ...

Implement the bottom of the data pipeline — discovery and parsing of
Claude Code session files:

- source/types.go: Raw JSON deserialization types (RawEntry,
  RawMessage, RawUsage, CacheCreation) matching the Claude Code
  JSONL schema. DiscoveredFile carries file metadata including
  decoded project name, session ID, and subagent relationship info.

- source/scanner.go: ScanDir walks ~/.claude/projects/ to discover
  all .jsonl session files. Detects subagent files by the
  <project>/<session>/subagents/agent-<id>.jsonl path pattern and
  links them to parent sessions. decodeProjectName reverses Claude
  Code's path-encoding convention (/-delimited path segments joined
  with hyphens) by scanning for known parent markers (projects,
  repos, src, code, workspace, dev) and extracting the project name
  after the last marker.

- source/parser.go: ParseFile processes a single JSONL session file.
  Uses a hybrid parsing strategy for performance:

  * "user" and "system" entries: byte-level field extraction for
    timestamps, cwd, and turn_duration (avoids JSON allocation).
    extractTopLevelType tracks brace depth and string boundaries to
    find only the top-level "type" field, early-exiting ~400 bytes
    in for O(1) per line cost regardless of line length.

  * "assistant" entries: full JSON unmarshal to extract token usage,
    model name, and cost data.

  Deduplicates API calls by message.id (keeping the last entry per
  ID, which holds the final billed usage). Computes per-model cost
  breakdown using config.CalculateCost and aggregates cache hit rate.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

2026-02-19 13:01:27 -05:00

3.1 KiB

Raw Blame History

View Raw