Generate knowledge from wiki

generate-knowledge-from-wiki (Confluence → domain briefs)

Ingest (S1) → Recognize (S2) → human confirm → Compose (S3→S7).

Top coding packs teach agents how to build. This skill adjudicates domain truth from Confluence, then feeds @requirement-risk / @ticket-splitter / @ticket-test-design.

Process terms (first mention): confirm = approve module cut lines; source brief = S6 adjudicated doc in source language; locale brief / reader brief = S7 target-locale reader doc; brief alone = short mode on risk/split (see TEAM_KNOWLEDGE_BASE.md); continue = default resume Compose after confirm. (zh-CN strings for process tokens live in domain-knowledge/language/deliverable-locale-tokens.json, cited once here rather than inline throughout.)

First time with Confluence? Read FIRST-RUN.md only — not the full RUNBOOK. Zero-credentials preview: ../../../WALKTHROUGH.md.

When to use / not

Use	Do not use → instead
Refresh a Confluence subtree; re-run after policy change	Single-ticket risk / split / test design → @requirement-risk / @ticket-splitter / @ticket-test-design
Build domain briefs from wiki	Config only → @setup-domain-ops

Hard gates (read first)

1. Strategy-first: if s2-domain-profiles.json checklist_themes is empty, or strategy.md §2 is still mostly placeholders → stop, run @setup-domain-ops. Never silently invent industry modules.
2. Default run is prep only (S1→S2) then pause; do not Compose until checklist rows are marked confirm.
3. Tagging acceptance before confirm: after S2, run python3 scripts/distill/tagging_acceptance.py --root-id <R>, show the report, and do not ask humans to confirm every row. Zero-source rows stay pending. Industry cuts are axes; completeness = land-in-axes + closure + report (+ Jira when board exists).
4. Keep industry axes: land wiki/Jira into those axes (references/industry-axis-remount.md); do not promote source-tree layout into modules, and do not hard-code tenant remount tables.
5. No translation in S1–S6. S6 = source-language brief. S7 = locale expression only.
6. confirm ≠ briefs already exist, and ≠ every corpus rule already written into S7.
7. Zero-rule fake coverage banned: after S7, run tagging_acceptance.py --after-s7 --strict; confirmed + sources + zero ### Rule headings → revert or rewrite.
8. No in-repo HTTP LLM API writing curated/.
9. S1 page errors block S2 by default (unless explicit --allow-partial).

Full Compose rules → references/iron-laws.md (load on demand). Long playbook (by step only): RUNBOOK.md · zh: RUNBOOK.zh-CN.md.

User flow

@setup-domain-ops          # first time: strategy §2 → profiles (+ deliverable_locale)
@generate-knowledge-from-wiki https://your-site.atlassian.net/wiki/spaces/DEMO/overview
# pause: tagging acceptance report → confirm only OK rows
# if board_id set and Jira attribution empty: @add-knowledge-from-jira team=<key> then re-tag

Follow FIRST-RUN.md: checklist → tagging acceptance → confirm → continue → S6 → S7 → --after-s7.

Offline story review without wiki: @requirement-risk DEMO-1 team=demo (WALKTHROUGH.md Path A).

Agent: five steps

#	Action
0	Ensure profiles non-empty; else @setup-domain-ops
1	python3 scripts/sync_domain_knowledge_from_confluence.py --url "…" (S1)
2	Read PIPELINE_HANDOFF.json → root <R>
3	Optional propose → RUNBOOK § S2 → s2_recognize.py → tagging_acceptance.py → pause (Jira if board + empty attribution); land into axes
4	On continue: Compose S3→S5→S6→S7 with write-through; --after-s3 then --after-s7 --strict
5	After S7: python3 scripts/run_distill_gate.py --root-id <R>

Do not preload RUNBOOK or iron-laws end-to-end.

Compose language split (S6 / S7)

Step	What	Language	Typical artifact
S6	Adjudicated source brief from S5	Source language (no translation)	_deliver/<slug>/<slug>-source-brief.md
S7	Locale conversion of S6 only (reader brief)	team-roots.json → defaults.deliverable_locale (zh-CN / en)	…-domain-brief.md (en); zh-CN filename from deliverable-locale-tokens.json

Risk/split read the S7 locale brief (reader brief) by default. If source language already equals deliverable_locale, S7 still emits the canonical locale filename (may be near-identical to S6).

Responsibility (summary)

	Agent	Script
S1–S2	Gate semantics, human review	Sync, recognize, closure
S3	Fidelity contract	Extract & index
S4–S7	Domain model + source brief + locale brief	Structural gates; no body authorship

Forbidden

• Compose S3–S7 without confirm
• Confirming zero-source modules, or confirmed + sources + zero S7 rules (fake coverage)
• Claiming corpus coverage from the industry cut alone
• Recreating modules from wiki/Jira filing layout instead of landing into strategy axes
• Hard-coding tenant product→slug remount tables into the pack
• Translating in S2–S6; skipping S6 and jumping straight to a locale-only draft
• Claiming “briefs done” from work drafts or S6 alone when S7 is required
• Shipping low-evidence S7 without an Evidence insufficiency banner when intentionally thin (non-SSOT for risk)

Further reading

Doc	When
FIRST-RUN.md	Start here for Confluence
RUNBOOK.md	Open one step section while executing
RUNBOOK.zh-CN.md	Chinese locale playbook
references/iron-laws.md	Compose / disputes only
S1-SYNC-CLI.md	S1 flags

Next

@requirement-risk → @ticket-splitter → @ticket-test-design (read S7 briefs; or offline DEMO-1).