core/go-session
8
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
go-session/FINDINGS.md
Claude 1031905018
feat(parser): add robustness for truncated JSONL and malformed lines 
Phase 1: graceful error recovery, streaming parse, skip malformed lines.

- Add ParseTranscriptReader for io.Reader-based streaming (pipes, buffers)
- Refactor to shared parseFromReader, eliminating code duplication
- Bump scanner buffer to 8 MiB for very large tool outputs
- 18 new tests: streaming, custom MCP tools, binary garbage, edge cases
- Coverage: 93.0% (up from 90.9%), 104 total tests

Co-Authored-By: Virgil <virgil@lethean.io>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-02-20 14:34:43 +00:00

3.2 KiB
Raw Blame History

FINDINGS.md -- go-session

2026-02-19: Split from core/go (Virgil)

Origin

Extracted from forge.lthn.ai/core/go pkg/session/ on 19 Feb 2026.

Architecture

  • Parses Claude Code JSONL transcripts into an Event array
  • Each event has type, timestamp, content, and optional tool metadata
  • Supported tool types: Bash, Read, Edit, Write, Grep, Glob, Task

Dependencies

  • Zero external dependencies at runtime -- standard library only (encoding/json, bufio, os)
  • Test dependency: github.com/stretchr/testify (assert/require)

Tests

  • Test coverage for JSONL parsing and event type detection

2026-02-20: Phase 0 Hardening (Charon)

Test Coverage

  • 51 tests across 4 test files, 90.9% statement coverage
  • parser_test.go — 13 tests + 12 extractToolInput subtests + 5 extractResultContent subtests + 5 truncate subtests + 5 ListSessions tests + benchmark
  • html_test.go — 7 RenderHTML tests + 4 shortID subtests + 6 formatDuration subtests
  • search_test.go — 9 Search tests covering cross-session matching, case insensitivity, empty dirs, malformed sessions
  • video_test.go — 8 generateTape tests + 1 RenderMP4 error test + 5 extractCommand subtests

Observations

  • extractCommand() naively splits on first # — commands containing literal # (e.g. inside quotes) get truncated. Documented in test, not a bug per se since the parser always constructs command + " # " + description.
  • RenderMP4() is untestable without the external vhs binary. Tests cover generateTape() (the pure logic) and verify RenderMP4 returns a clear error when vhs is absent.
  • extractResultContent() handles string, []interface{}, and map[string]interface{} content types. All three paths plus nil are tested.
  • ListSessions falls back to file mod time when no valid timestamps are found in a JSONL file.
  • go vet ./... was clean from the start — no fixes needed.

2026-02-20: Phase 1 Streaming & Edge Cases

New API surface

  • ParseTranscriptReader(r io.Reader, id string) (*Session, *ParseStats, error) — streaming parse from any io.Reader (pipes, network, in-memory buffers). No file on disc required.
  • parseFromReader(r io.Reader, id string) — internal shared implementation, eliminating code duplication between ParseTranscript and ParseTranscriptReader.
  • maxScannerBuffer constant: 8 MiB (bumped from 4 MiB), handles very large tool outputs.

Tests added (18 new, 104 total)

  • Streaming io.Reader (6 tests): strings.Reader, bytes.Buffer, empty reader, large lines (128KB), malformed lines with stats, orphaned tool calls.
  • Custom MCP tools (3 tests): mcp__forge__create_issue with flat input, nested JSON input, empty input object.
  • Edge case recovery (5 tests): binary garbage, null bytes, 5MB lines, malformed message JSON, malformed content blocks.
  • Additional truncated JSONL (3 tests): missing closing brace, mid-key truncation, all-bad-lines file.
  • ListSessions truncated files (1 test): partially truncated file still extracts valid timestamps.

Coverage

  • 93.0% statement coverage (up from 90.9%).
  • ParseTranscript, ParseTranscriptReader, extractToolInput, extractResultContent, truncate all at 100%.