Conflict Detection Skill (weekly safety net + manual auxiliary check)

The LLM detects; a human judges. You only find and surface contradictions — you never choose which side is right and never fix anything. In v2, primary conflict handling happens locally in backlog-complete (pre-push; a local, developer-side client-template skill that lives outside this repo) — this skill is the weekly cron safety net (via an adopter-provided weekly sweep worker) that catches missed completes, plus a manual auxiliary check run when a human asks (PROJ-79).

Absolute rules (never violate)

• Do not resolve conflicts. Do not choose which side is right.
• Do not rewrite, summarize, merge, or refine documents.
• Do not privilege one side (no "the newest doc / the PRD is probably right" assumptions). Present both claims side by side.
• Report only what is actually written. Never invent conflicts that are not there. When in doubt, report (an omission is worse than an over-report).
• Never mark anything superseded or update the wiki without human judgment.

Scope of operation

• Targets: the documents/topics the user specifies. If unspecified, target the raw docs changed by recent pushes (identified via git log in the lore vault) and compare them against all of raw. Weekly cron mode (the worker passes the target list in the prompt): targets = raw docs changed by pushes in the last 7 days — the worker computes them with git log --since='7 days' --relative -- raw/ and hands them over. Check only the list you were given; do not expand the scope on your own.
• Comparison baseline: all of wiki/ + all of raw/ whose status is not superseded (including subdirectories — raw/**/*.md). Raw layout = raw/PRD.md (planning) + raw/decisions/*.md (per-backlog decision details, e.g. raw/decisions/<proj>-N.md). progress is not part of this pipeline (do not read it). Do not assume fixed filenames like raw/PRD.md, raw/DECISIONS.md — glob all of raw.
• No re-reporting (dedup): never create a new file for a conflict that matches an existing CONFLICT file in conflicts/, regardless of its status (open or resolved) — same topic and same two sides means the existing open file is already awaiting judgment (append evidence rows to the existing file if needed).
• Known filter: mismatches that match an entry in conflicts/KNOWN.md (tracked lags — e.g. "PRD update awaiting Slack draft submission", "awaiting deferred judgment") must not be reported as conflicts. KNOWN.md is human-managed — never add or remove entries yourself. If the file does not exist, proceed without the filter.
• raw/decisions/ is normally human-authored input synced from the project repo. Decision details can contradict the PRD (e.g. PRD says "Google" but a decision says "Kakao SDK") — surface that as a conflict. Most are conflict-free (simple detail additions) and pass straight through to refinement.

Using the schema

• Read frontmatter: authority / status / updated / supersedes / doc_type / tags.
• Narrow the comparison to docs whose tags/doc_type overlap. Exclude superseded docs from comparison.
• Report authority differences alongside, but never auto-retire — go no further than proposing a "superseded candidate".

Detection types (7)

1. Direct contradiction 2. Number/spec mismatch 3. Timeline conflict 4. Definition/terminology drift
2. Decision reversal 6. Orphaned reference 7. Under-specified decision (stated like a decision, but missing a concrete value needed for implementation)

⚠️ Relax type 7 (under-specified) for raw/decisions/*.md (implementation decisions)

For implementation decisions, details being filled in by code is normal. Things like "schema columns and exact policy values to be finalized in code" are not conflicts — pass them through to the wiki (most docs flowing through conflict-free is the point of this model). Raise type 7 only when another document (PRD or another decision) depends on that concrete value and it is missing, or when different decisions state the same value differently. In short: catch types 1–6 (direct contradictions, spec mismatches, reversals, package/integration conflicts, e.g. auth via Firebase vs RLS using Supabase auth.uid()), but do not flag "details to be settled in implementation".

⚠️ Explicit deferrals are not conflicts (all docs, PRD included)

Items deliberately postponed — "delivery channel and frequency policy to be decided later", "payments deferred to next quarter", "web version reviewed after launch" — are roadmap gaps, not contradictions — do not flag them as type 7. It is enough for the wiki to record the gap, e.g. "Notifications: planned, policy TBD (to be decided later)" (pass it through; do not stop the build). In particular, a single sentence/section saying "do X + X details later" is one coherent statement, not an A-vs-B conflict. Type 7 applies only when another document actually depends on that value and it is empty, or when two sources state the same value differently (e.g. PRD says 30 memos vs a decision says 50).

Output: conflicts/CONFLICT-<number>_<YYYY-MM-DD_HHMM>.md (one file per conflict; team-timezone date+time suffix for dedup and history)

• All timestamps use the team's timezone — the filename suffix and every date inside the file. Always pass an explicit timezone so filenames and dates match the team's wall clock: in the terminal use TZ="$TEAM_TZ" date +'%Y-%m-%d_%H%M' (for filenames) and TZ="$TEAM_TZ" date +%F (for body dates) (set TEAM_TZ to your team's IANA timezone, e.g. America/New_York). A date-only suffix collides on same-day reruns, so include minutes (HHMM) to disambiguate. Output exactly this structure (for readability and comparison):

# CONFLICT-NNN · <one-line topic>

**Type** <one of 1–7> · **Severity** high|medium|low · **Status** open · **Tags** <area>

> **Judgment question** — <the one line the team must decide. Do not write the answer>

| Position | Verbatim quote | Source |
|:--:|---|---|
| **A** | "<one verbatim quote>" | `<file:line>` |
| **A** | "<another quote supporting the same A>" | `<file:line>` |
| **B** | "<verbatim quote>" | `<file:line>` |

**Proposal** — <"superseded candidate" level only when the authority difference is clear; otherwise —>

Table rules (strict):

• One row = one quote + one source. Never mix multiple quotes, multiple sources, or commentary into one row (it breaks the table and is unreadable).
• If one position has multiple pieces of evidence, repeat the position letter (A/A/A) to split rows.
• Position = one letter per opposing document/claim: A, B, C, D…
• Quotes are verbatim (no summarizing, no commentary). Escape | inside quotes as \|.
• Source is file:line. If authority/updated is present, append · authority N after the source (omit if absent).
• Number = the current maximum in conflicts/ + 1.

Follow-up actions

• This skill's output ends at the conflicts/CONFLICT-*.md files. Commit only — do not push and do not send to Slack — in weekly mode the worker (the adopter-provided weekly sweep script) handles push and summary delivery.
• On manual invocation, if there is even one conflict: tell the human to halt wiki refinement for that topic.
• No superseded processing and no automatic wiki updates without human judgment.

Isolation

• Read only this project's (channel/lore vault) documents. Never pull in another project's documents.