Problem

A freshly-launched session's claude_session_id (and the gemini/opencode/codex equivalents) can be silently wiped from state.db, surfacing as "no session ID detected" on a session whose Claude transcript is actively being written. Callers that resolve a session's transcript by its persisted id (e.g. a coordinator mapping its workers to ~/.claude/projects/<uuid>.jsonl) intermittently lose the mapping within seconds of launch.

Root cause

The per-tool conversation ids are detected asynchronously and persisted shortly after launch. But any concurrent full-table save — SaveWithGroups → SaveInstances → saveInstancesOnce — rebuilds each row's tool_data from its in-memory Instance. If that writer's snapshot has not yet observed the id, the rebuilt blob omits it (omitempty zero-value).

MergeToolDataExtras preserved only keys unknown to the typed schema; for a known key it let the new blob win "including absence-by-omitempty". So an omitted id cleared the previously-persisted value instead of being preserved.

The hottest such writer is the rename hook (hook_name_sync.go): on every Claude hook that renames a session it does LoadWithGroups then SaveWithGroups over the whole snapshot. Other full-rewrite writers (internal/ui/home.go, web_mutator.go, group_cmd.go, session_move.go) hit the same path less frequently. The launch path itself is already hardened (InsertSessionAndVerify, #1031), so the clobberer is always a different writer.

Fix

Treat the detected session-id keys as sticky in MergeToolDataExtras: {claude,gemini,opencode,codex}_session_id and their *_detected_at. When the new blob omits a sticky key but the old row has it, carry the old value forward. Semantics:

• non-empty new value still wins — a real resume/fork that changes the id;
• explicit empty present in the new blob ("claude_session_id":"") is honored as an intentional clear;
• only outright omission preserves.

This is a one-line change to the merge guard (if known[k] && !sticky[k]) plus a small key set; it does not touch the save/sweep contract.

Design note

Deliberate clears must now use an explicit empty (or a single-row clear) rather than relying on an omit. UpdateClaudeSessionsWithDedup clears in-memory only; with sticky keys a cleared id is restored from disk on reload and re-cleared in memory each load. Since duplicates "shouldn't occur in normal use" and tmux env is authoritative, that is a harmless redundant in-memory re-dedup, not a functional regression.

Tests

internal/statedb/tool_data_extras_test.go:

• repoints the typed-absence test to a non-sticky key (latest_prompt), since claude_session_id absence behavior intentionally changed;
• adds sticky coverage: preserve-on-omit (all four tools + a *_detected_at), non-empty-new-wins, and explicit-empty-clears.

go test ./internal/statedb/ passes; gofmt/go vet clean.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Bug Fixes
  • Improved how session-related data is merged so previously saved values are retained when newer data omits them.
  • Non-sticky fields still follow the usual overwrite rules, and explicit empty values are respected as intentional clears.
• Tests
  • Expanded test coverage for missing, updated, and intentionally cleared session-related values during merges.