Smarter project identification for cross-repo work

[kunickiaj]

Problem

When working in repo A (e.g., greenroom) but editing files in repo B (e.g., opencode-mem), memories get tagged with the cwd project (A) instead of the actual project being worked on (B).

Proposed Solution: File Path Heuristics

Track which repo's files are touched during a session and use that to determine the actual project:

1. Extract all file paths from tool.execute.after events (read/write/edit tools)
2. Group by git repo root (find nearest .git parent)
3. If dominant repo ≠ cwd, update session's project to dominant repo

Why this approach?

• Reliable signal - File operations are concrete actions, not ambiguous text
• Already available - We capture tool.execute.after events with file paths
• Low complexity - Count paths and pick the dominant repo
• Works across languages/topics - Doesn't depend on naming conventions

Alternative considered: Keyword analysis

• Scan prompts for project-specific keywords
• Rejected: too ambiguous, needs keyword list maintenance, false positives when discussing other projects

Implementation Location

Likely in plugin_ingest.py or during session finalization.

Context

This came up when fixing the message extraction bug while cwd was greenroom but all file edits were in opencode-mem. Had to manually SQL update 16 sessions to fix the project tag.

[kunickiaj]

Current triage: still relevant.

The product/runtime has moved to the TypeScript backend, but the underlying problem still looks real for OpenCode sessions: cross-repo file work can still diverge from cwd-based project tagging.

What has changed since this was filed:

• Claude hook ingestion now has path-based project inference in the TS codepath.
• I did not find equivalent dominant-repo inference for the OpenCode plugin/session project path.

So this issue should stay open, but it should be understood as a TS/OpenCode pipeline improvement rather than a legacy Python ingest task.