Xedit Conflict Audit

apply_filter regex + parentFormId (supports.applyFilterExtensions)

Use records.apply_filter to collapse broad candidate discovery into a single server-side query when the daemon advertises supports.applyFilterExtensions (or the narrower .regex / .multiPattern subkeys, if exposed separately).

• Regex fields: editorIdRegex, displayNameRegex, fullNameRegex, baseEditorIdRegex, baseDisplayNameRegex.
• Regex engine: System.RegularExpressions.TRegEx.
• Guardrails: 100ms per-record timeout, 4-worker semaphore, and a RegexSlotsExhausted response counter so saturation is distinguishable from no matches.
• Multi-pattern OR: each *Pattern and *Regex field may be a scalar string or an array with OR semantics; max 32 entries.
• parentFormId: single-call ancestor filtering, e.g. "all REFRs whose container chain contains CELL X" without building your own parent walk.

Deep reference: xedit.apply-filter-extensions.v1.

Workflow

1. Bootstrap session. xedit_session({}). Confirm gameMode, consentEnabled not needed here (read-only), and loadOrderSize matches expectation.
2. Sanity-check capabilities. xedit_list_capabilities({}). Read the drift.onlyInLive and drift.onlyInDigest arrays plus the r6 support keys. If a target command you intend to use is missing from live, stop and tell the user.
3. Scope the audit. Decide whether the audit is per-record, per-plugin, per-signature, or parent-scoped (e.g. all REFRs under a CELL).
4. Locate records with the narrowest supported query.
   • Per-record by FormID: xedit_find_record({ file, formId }).
   • Per-editor-ID: xedit_find_record({ editorId }).
   • Parent-scoped or regex candidate set on r6+: xedit_call({ command: "records.apply_filter", args: { file?, signature?, parentFormId?, editorIdRegex?, ... } }).
   • Per-plugin fallback: xedit_call({ command: "records.list", args: { file, signature? } }), then iterate or delegate if large.
5. Inspect conflicts with child-group awareness. For each focused target, prefer xedit_inspect_conflicts({ file, formId }). For broad CELL/WRLD/DIAL/QUST targets on r6+, call records.conflict_status through xedit_call if you need the raw result.childGroup summary; expand only conflictingHits or a representative sample instead of probing every child.
6. Collect references efficiently. On r6+, xedit_call({ command: "records.references", args: { file, formId, recursive: true } }) gives a ChildGroup-wide outgoing-reference union. On pre-r6, fall back to the old child walk or delegate the loop.
7. Classify verdicts. Read the verdict field from the intent tool or map the raw records.conflict_status result with the same W2 labels:
   • no_conflict → safe.
   • itpo / itm → likely safe; consider cleaning.
   • minor → human review.
   • breaking → halt and surface.
8. For non-trivial verdicts, read the actual record. xedit_read_record({ file, formId }). Compare record.fields vs winningOverride vs baseRecord. Identify the diverging fields.
9. Summarise. Produce a short report: one row per record or child-group signature bucket audited, columns [file, formId, editorId/signature, verdict, winningFile, referencerCount, childGroupConflicts]. Surface only the breaking/minor verdicts to the user by default; the rest are appendix.

Verification (what counts as semantic pass)

• The audit's verdict for each spot-checked record matches what manual xEdit GUI inspection would show.
• For breaking verdicts, you have read the actual record fields and can name the diverging fields.
• The output report is concise: one row per record or child-group signature bucket, no raw daemon envelopes.
• The session's audit log (.opencode/artifacts/xedit-mcp/audit/YYYY-MM-DD.jsonl) contains one entry per MCP tool call you made.

If you cannot meet these for a record, mark it unknown in the report and explain why — do not guess.

Common Mistakes

• Calling xedit_call records.conflict_status directly when xedit_inspect_conflicts would do it with the verdict label already mapped; use raw records.conflict_status only when you need r6 response blocks such as result.childGroup.
• Treating no_conflict as proof of safety without reading at least one representative record.
• Looping through hundreds of records in the orchestrator's context when r6 recursive, childGroup, apply_filter, or delegation would collapse the work.
• Forgetting to call xedit_session first; downstream tools will refuse with state_violation.
• Asking the daemon for a file that is not in the load order; LOAD001 will fire — load it via the session first.

Delegation hints

This workflow is a strong candidate for read-only sub-agent delegation in two cases:

1. Large scope (> ~10 records to audit) — the round-trips will fill the orchestrator's context. Delegate the loop; receive the per-record summary table.
2. Exploratory diagnosis ("I don't know which record is causing this in-game issue") — let a sub-agent triangulate by editor ID and signature; you'll get the candidates back.

When delegating, include this skill and the hub skill in the sub-agent's prompt and provide the scope + budget.