Wiki Growth

wiki-growth — challenge the vault, route the survivors

A thin router, not an engine. It walks the vault, runs each shortlisted idea through the review-board rubric, and dispatches survivors to the skills that already own the downstream work: storm (source-backed strengthening), spm (idea → spec), skill-authoring-standard (idea → skill), marketing-orchestrator (idea → marketing asset), all via the nexus wrapper for sub-tier dispatch. Never re-encode those skills' logic — invoke them.

Boundary vs wiki-ingest: ingest pulls new session learnings in; wiki-growth audits, challenges, and grows what is already there. No overlap.

Vault ground truth

Vault root: ~/2nd Brain/2nd Brain/. Idea-bearing scan targets, in priority order:

1. Root-level spec/blueprint *.md (≈40 files — AGENTIC_*, NEXUS_*, *_SPEC.md, …)
2. Sketches/, Pitches/, Grills/, Personas/
3. Wiki/ (≈455 notes) — only pages Grep-matching idea markers (TODO|PROPOSAL|IDEA|Plan:|Next step|should we|could build)

Excluded: Sources/ (evidence, not ideas), Templates/, outputs/, process/, sprint/, dot-dirs, log.md, index.md files.

Hard write-gate (both modes)

The ONLY writable path is the sandbox ~/2nd Brain/2nd Brain/_system/wiki-growth/ (create on first run). Never edit, move, or delete a live vault page in v1 — apply mode stages drafts in the sandbox; promotion out of it is a manual act by the pilot.

Mode: scan (default)

/wiki-growth or /wiki-growth scan [area or note path] — read-only except the report.

1. Inventory — Glob the scan targets; collect path, title, first heading, mtime. No full-file reads. — Done when: an inventory count per target exists for the report.
2. Shortlist — cap at 25 ideas per run (an argument narrows to one area or note). Priority: root specs and Sketches/Pitches first, then marker-matching Wiki pages, newest-modified first. State in the report exactly what the cap dropped — no silent truncation. — Done when: shortlist ≤25 with the drop-list recorded.
3. Batch triage — split the shortlist into batches of 5. One parallel subagent per batch, each dispatch wrapped per the nexus skill (tier ladder below). Each subagent reads only its 5 notes, answers the review rubric per idea (rubric and the 11 disposition labels are defined in references/review-rubric.md; the subagent prompt must inline the rubric questions), and returns structured rows: {note path, one-line idea, rubric verdicts, disposition, evidence/why, high-stakes flag}. The lane column is derived at synthesis from the routing table, never by the subagent. — Done when: every shortlisted idea has a returned row.
4. Synthesise (this session, not a subagent) — merge batches, dedupe ideas that are the same proposal in different notes (mark the extras Merge), flag conflicts between batches rather than silently resolving, and escalate any contested or high-stakes row (spend, prod change, new build) to a single Opus 5 re-judgement via nexus. — Done when: no duplicate rows and every contested row carries a second verdict.
5. Report — fill references/report-template.md and write it to _system/wiki-growth/REPORT-<YYYY-MM-DD>.md. End the turn with the disposition counts and: Scan complete. Next safe action: review the report, mark rows APPROVED, then run /wiki-growth apply REPORT-<date>.md.

Mode: apply

/wiki-growth apply REPORT-<date>.md — acts ONLY on rows the pilot marked APPROVED.

1. Read the named report from the sandbox; collect approved rows. Zero approved rows → say so and stop. — Done when: the approved set is listed back before any action.
2. Route each approved row by its lane — the disposition→lane→skill map is in references/routing-table.md; look it up there. All outputs land in _system/wiki-growth/drafts/<note-slug>/, nothing else.
3. Lane C (promote) never authorises a build: its output is a spec/skill/asset draft, and any implementation still requires a separate /judge real-100/100 pass plus explicit user approval, per the judge hard line.
4. Close with a manifest of drafts written and the promotion instruction (pilot moves approved drafts out of the sandbox by hand). — Done when: every approved row maps to a draft file or a stated skip reason.

Tier ladder (caller-side, per nexus step 4)

• Haiku 4.5 — none in v1 (inventory is Glob/Grep, cheaper than any dispatch).
• Sonnet 5 — default for batch-triage subagents (clear rubric, known pattern).
• Opus 5 — single-idea re-judgement of contested/high-stakes rows.
• Fable (this session) — synthesis, dedupe, conflict calls, the report itself.

Triage dispatches are pure read-only research: per the nexus skill's own step 7, the per-dispatch session-handoff is omitted — state that omission once in the report.