Resolve Conflicts

Conflict Resolution — Prose Verdicts → PRD Draft → Approval → Alignment (PRD = SoT)

When humans edit the PRD by hand, internal text dependencies break and consistency falls apart. So humans provide judgment only (prose); Loremaster drafts a consistency-aware PRD revision and a human approves it.

Flow

1. Collect: Read the open conflicts in conflicts/ plus the team's prose verdicts (this thread / provided text). Read them directly with read_file/search_files (no execute_code).

• Slack thread replies count as verdicts/instructions even without a mention. In particular, when a user replies to my conflict-sweep message with things like "decided: 3 variants", "go with A", "go with B", treat that as the human verdict for that conflict — no separate @mention or re-confirmation required. If the thread context contains earlier replies, incorporate them first; never stall with "I couldn't read it because I wasn't tagged."
• When the user replies to the conflict-detection result with bare numbered verdicts like 1. ... 2. ..., map them in order to the existing CONFLICT-001, CONFLICT-002, ... as the team verdicts and proceed straight to resolution proposals. Do not ask "shall I resolve the conflicts?" separately.
• When the user directly provides a [spec change request], a specific PRD section, current/proposed text, and a supporting decision file, run it through this skill's PRD-change flow even if no CONFLICT file exists. Still follow the order: write PROPOSAL file → PRD approval button → apply → downstream alignment check → refine/push.

1. Draft the PRD revision: Produce a revision of raw/PRD.md that reflects the verdicts.

• Account for internal PRD dependencies and consistency (changing one place means updating related statements too). Never invent blanks or missing content (if something is unclear, ask, as-is).
• Integrate resolution results into the PRD (no separate decisions.md — it would collide with the existing DECISIONS.md).

1. Proposal = write it to conflicts/PROPOSAL_.md (team-timezone date+time suffix. Do not paste diffs into Slack — poor readability):

• All timestamps use the team's timezone — the filename, dates inside PROPOSAL/CONFLICT files, the status: resolved date, and the alignment dates in raw/decisions/*.md, all of them. Get them with TZ="$TEAM_TZ" date +'%Y-%m-%d_%H%M' (filenames) and TZ="$TEAM_TZ" date +%F (body). Never use the host/container default clock (UTC) as-is — always pass an explicit timezone so filenames and dates match the team's wall clock (set TEAM_TZ to your team's IANA timezone, e.g. America/New_York).
• PROPOSAL.md layout: (1) team verdict summary, (2) the unified PRD diff under ## Stage 1 — PRD changes (strict format below), (3) to-dos on apply.
• Unified diff format (no free-form): one ``` `diff `` code block per file; first line --- a/raw/PRD.md, second line +++ b/raw/PRD.md, then @@ section @@; changed lines as - old / + proposed`; context lines indented one space with no prefix. Do not mix items — separate per file.
• Keep the Slack message short: 📄 Please review /conflicts/PROPOSAL_.md and let me know whether to approve. (No diffs, no numbered prompts.)

1. First approval = request_approval button (PRD change): After presenting the PROPOSAL, get Slack [Allow]/[Deny] approval via request_approval (title = e.g. "Apply PRD revision — "; detail = PROPOSAL path/summary). Apply only on APPROVED. On DENIED/NO_RESPONSE, no apply and no retry. No clarify/numbered text prompts.

• (Note: a plain git commit is not a dangerous command, so no automatic approval button appears for it. Perform the approval explicitly via request_approval.)

1. Apply stage 1 (PRD, first-approval scope):

• Update raw/PRD.md (git commit). Never apply unapproved parts.
• The PROPOSAL's diff is a record for human review. When applying, extracting the diff from markdown code fences and relying on git apply can break on long diffs, abbreviated context, and fence counts. Prefer applying the approved changes via patch / an exact file rewrite, then verify with git diff --check and searches for the key phrases.
• Mark the matching conflicts/CONFLICT-*.md as status: resolved (verdict date = today in the team's timezone) with evidence raw/PRD.md (finalized).

1. Second approval = downstream doc alignment (against PRD = SoT, separate button): The PRD is now updated, so check the other raw docs that still contradict the finalized PRD (raw/decisions/*.md, gtm, etc. — raw/**/*.md; progress files are not there).

• If contradictions exist (e.g. an old decision in raw/decisions/ conflicts with the PRD / a reversed decision lacks a superseded marking / the frontmatter of the decision file the user cited is still draft while the request and PRD assume stable), build an alignment plan — mark the item status: superseded or draft a fix. Never delete the decision log's original dates or content (preserve history). Alignment annotations use today's date in the team's timezone — e.g. append to the entry: > **superseded** 2026-06-23 — replaced by finalized PRD (Google login). (PROPOSAL_2026-06-23_HHMM). New decision entries use the format ### #N — .... All dates via TZ="$TEAM_TZ" date +%F.
• When a direct spec-change request piggybacks on an existing PRD-update flow (e.g. "fold this into the PROJ-36 update draft"), then after applying the PRD, also add the current flow as a new decision row in that decision file; do not delete the earlier interim decision — explain it as superseded.
• Append these alignment diffs to the same PROPOSAL file under a ## Stage 2 — Downstream alignment section (same unified diff format, separated per file).
• Raise a second request_approval button (title = e.g. "Apply doc alignment — "; detail = files to align + summary). On APPROVED, apply the alignment (git commit). On DENIED/NO_RESPONSE, hold the alignment — but stage 1 (PRD + resolved) is already applied, so do not roll it back. No retry, no workaround.
• If there is nothing to align (other docs already match the PRD), skip this stage and the second button. Do not force a button when there is nothing.
• Guard against missed duplicate PRD sections: Do not fix one spot in the PRD and stop. The same decision may be described in both the user journey (e.g. §9.1) and the screen/IA details (e.g. §13.1.1). Search all of raw/PRD.md — not just the lines quoted in the conflict — for the old canonical phrasing, its synonyms, and evidence IDs (e.g. single-provider login, DECISIONS #44, Apple / Google) and align every current-spec statement together. If the user reacts with "didn't you already update that?", first assume your earlier edit missed something and immediately hunt for leftover phrasing.
• When the apply is done, mark PROPOSAL.md status: applied.

1. Afterwards — auto-chaining (refine): After the PRD and alignment are applied, if the open conflicts in conflicts/ reach 0, immediately auto-invoke refine-wiki (no need for the human to call it again). refine-wiki handles graphify + wikilinks.

• Count open conflicts only from the current status in conflicts/CONFLICT-*.md. Do not count old -**status** open lines inside PROPOSAL_*.md diffs.
• refine's auto-refinement regenerates derived artifacts (wiki/graphify), so do not raise another content re-approval button — content approval finished in stages 1 and 2 above.

1. origin push approval (automatic after commit): After git-committing the changes, do not wait for the human to say "push" — immediately raise a request_approval "Push to origin?" button (title = origin push — , detail = commit hashes/messages to push + remote URL). On APPROVED, git push origin main; on DENIED/NO_RESPONSE, do not push (keep the local commits). No retry, no workaround.

• Do not substitute text approval for a button timeout. If request_approval returned NO_RESPONSE and the user later says "I approved it", the recorded tool result is still unapproved. Summarize the same push/write scope, raise a new request_approval, and proceed only when that new call returns APPROVED.
• In the normal flow that chains into refine-wiki, do not push here — batch it at refine-wiki's final push step (PRD + alignment + wiki commits in one push, avoiding duplicate approvals). This step 8 applies only when the flow ends here without chaining into refine.

Rules (invariants)

• Verdicts and decisions belong to humans. The team's opinion is the verdict; Loremaster is the proposer that turns it into PRD wording.
• Every external write (PRD/conflicts/other docs) goes through the button approval gate. Never auto-apply unapproved changes.
• The PRD is the SoT. One-way flow (update raw → derive wiki afterwards). No inventing blanks; keep evidence traceable.
• This channel = this project only. This project's lore vault only.

Output

• Slack: PRD revision summary + changes + [Approve]/[Cancel]. After approval, report the apply results (commit, resolved markings, alignment proposals).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: loremaster-ai
• Source: loremaster-ai/loremaster
• License: MIT

Install and usage instructions live in the source repository linked above.