Derive project/session stats from transcripts instead of history.jsonl

[vvzvlad]

Problem

list_projects, find_patterns, get_project_context and by-project lookup in get_session_summary all read conversation statistics from ~/.claude/history.jsonl through parseHistoryFile().

I use Claude Code almost only through VS Code extension / Agent SDK, and in this setup ~/.claude/history.jsonl is almost never written — mine had captured only few slash commands — even there was hundreds of real session transcripts under ~/.claude/projects/. So list_projects showed 5 projects / 6 sessions, but actually i had 17 projects and 300+ sessions. Full-text search tools were fine, because they already index the transcripts; only history.jsonl-based tools undercount.

Fix

I add src/parsers/transcript-history.ts, which build the same HistoryEntry[] shape from per-session transcripts instead of history.jsonl, so all downstream helpers and tool bodies stay unchanged — only the data source swaps:

• parseSessionUserPrompts(content, sessionId, fallbackProject) — pure, IO-free, unit-testable. Turns one transcript into one HistoryEntry per genuine user prompt. Skips meta turns (isMeta), tool-result turns (toolUseResult), empty turns, and Claude Code injected non-prompt user turns (<task-notification>, <ide_opened_file>, <ide_selection>, slash-command wrappers, <local-command-stdout>, ...). The project is resolved from in-transcript cwd (with fallback to the encoded project dir, which extractProjectName already handle).
• parseTranscriptHistory() — enumerates all session files (reuse enumerateSessionFiles) with per-file cache keyed on mtime + size, so on next calls only changed files are re-read; entries for deleted files are pruned.

Then switch the four tools to parseTranscriptHistory(). parseHistoryFile() is left intact (still used by its test). Added tests/parsers/transcript-history.test.ts.

Result

• list_projects: 5 → 17 projects, with correct session/message counts and real cwd paths.
• find_patterns: now reports 325 sessions with meaningful topic / activity / issue breakdowns instead of almost empty sample.
• Filtering out injected turns removes ~36% of false "messages" which otherwise inflate counts and pollute topic analysis.

npx tsc compiles clean and full vitest suite passes (86 tests, including new parser tests). No behavior change for search/index/sync paths.

[jhammant]

Apologies for the very slow response — two months is not the turnaround you deserved.

Reviewed and approving. Deriving from the session transcripts instead of history.jsonl is correct, and the undercount you describe is real: under the VSCode extension and Agent SDK that file barely gets written, so list_projects and find_patterns were reporting a fraction of actual usage.

Resolving the project from the first cwd in the file rather than decoding the dashed directory name is the better call too — it handles worktrees, which the encoded form can't do unambiguously. Nice touch keeping HistoryEntry[] as the boundary so the tool bodies didn't have to change.