fix: dedupe Claude sessions imported by Codex Desktop

[NovakPAai]

Summary

Codex Desktop >= 0.133.0-alpha.1 introduced an External Agent Session Imports feature that ingests Claude Code sessions from ~/.claude/projects/**/*.jsonl into ~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl and registers them in ~/.codex/session_index.jsonl with a fresh UUIDv7 thread id. The original Claude file stays in place.

Codbash was loading both copies, so every imported conversation appeared twice in the dashboard — once correctly attributed to Claude (original file, UUIDv4 id) and once incorrectly attributed to Codex (the import, UUIDv7 id). On my machine that produced 50 duplicates (56 "Codex" sessions of which only 6 were genuinely Codex-native).

Fix

• New helper parseCodexExternalImports(codexDir) reads ~/.codex/external_agent_session_imports.json and returns a Set of imported_thread_ids whose source_path still exists on disk.
• scanCodexSessions() skips those ids in both the history.jsonl loop and the rollout-files walk.
• If the source Claude file has been deleted, the Codex copy is retained — no history is lost.
• The ledger file is added to _sessionsNeedRescan / _updateScanMarkers so a two-phase Codex write (rollout/index first, ledger a tick later) cannot leave a stale dedup view between polls.

Identifying an import on disk

Imported rollouts have a distinctive meta header:

"originator": "Codex Desktop",
"source": "vscode",
"turn_id": "external-import-turn-1"

But the ledger is the authoritative source — using it avoids parsing every rollout to detect imports.

Before / after

	before	after
codex sessions	56 (50 imports + 6 native)	6 (native only)
claude sessions	226	226
total	605	557
imported id 019e55cf-3089-...	shown as Codex	hidden (Claude original ee956bf6-... is shown instead)

Test plan

• node -c src/data.js — syntax
• Manual: node bin/cli.js run, /api/sessions counts before/after match expected dedup
• Sample id 019e55cf-3089-7031-bc64-fd29a650def8 no longer appears under Codex; original ee956bf6-b5f6-4571-b1d0-019ce2874122 still appears under Claude
• Test on a machine with no ~/.codex/external_agent_session_imports.json — fallback (helper returns empty set, all Codex sessions shown as before)
• Test on a machine where a source Claude file has been deleted — Codex copy is kept (not skipped)

[vakovalskii]

LGTM ✅ Targeted dedup fix with the right defensive posture.

Verified:

• node -c src/data.js clean
• Full test suite still 127/0/2 pass — nothing regresses
• CI green 6/6

Spot-checks:

• parseCodexExternalImports is defensive at every step: missing file → empty Set, unreadable → empty Set, malformed JSON → empty Set, non-array records → empty Set
• Conservative no-data-loss policy: only skip Codex copy when the source Claude file still exists. If user deleted the original, the Codex import becomes the sole surviving copy
• Skip applied at both Codex code paths (history.jsonl loop and rollout files walk) — full coverage
• Cache invalidation correctly watches the ledger mtime+size + handles the file-disappeared case (else if (_codexImportsLedgerMtime !== 0 || _codexImportsLedgerSize !== 0))
• Uses the ledger as authoritative — avoids parsing every rollout to detect the Codex Desktop originator + external-import-turn-1 markers 👍

Real-world impact reported in the PR (56 → 6 Codex sessions, 605 → 557 total) suggests this will be felt immediately by anyone running both Codex Desktop and Claude Code.