LLM Wiki Ecosystem Gap To Issues

LLM-wiki ecosystem gap -> GitHub issues

Use when the user asks to review the repo/document ecosystem and create new issues to strengthen LLM-wikis.

Goal

Turn ecosystem review into a small set of non-duplicate, high-leverage GitHub issues grounded in live repo state rather than intuition.

Required evidence sources

Read these before proposing issues:

1. Live GitHub issue search
   • gh issue list --search 'llm-wiki OR knowledge wiki OR wiki ingest OR knowledge/wikis'
   • Then narrower searches for the exact candidate theme to avoid duplicates.
2. Architecture / review docs
   • docs/document-intelligence/llm-wiki-resource-doc-intelligence-operating-model.md
   • docs/reports/2026-04-16-llm-wiki-resource-intelligence-unified-review.md
3. Priority/execution artifacts
   • docs/reports/llm-wiki-external-source-priority-queue.md
   • docs/reports/llm-wiki-staged-batch-packs.md
4. Discoverability / provenance gaps
   • docs/document-intelligence/intelligence-accessibility-map.md
   • docs/document-intelligence/standards-codes-provenance-reuse-contract.md
5. Live repo facts
   • Count wiki pages directly from knowledge/wikis/*/wiki/**/*.md
   • Check cross-link artifacts (knowledge/wikis/cross-links.md, cross-links.jsonl if present)
   • Inspect data/design-codes/code-registry.yaml when proposing standards/registry-driven strengthening

Proven issue themes that are worth checking first

These themes produced grounded issue opportunities:

1. wiki_refs reverse lookup
   • Trigger: architecture/provenance docs define the contract, but live repo still lacks registry -> wiki reverse lookup.
   • Signal: accessibility map says affected wiki pages can only be found with grep.
2. External-source promotion packs
   • Trigger: priority queue + staged batch pack already exist, but execution issue is missing.
   • Signal: thin wiki domains (engineering / naval-architecture / maritime-law) plus metadata-rich source families ready for promotion.
3. Design-code registry promotion into wiki surfaces
   • Trigger: data/design-codes/code-registry.yaml has rich edition/repo metadata but is not linked from standards wiki surfaces.
4. Wiki strengthening scorecard
   • Trigger: health/lint exists, but prioritization across domains/source families is still ad hoc.
5. Large-domain navigation repair
   • Trigger: a wiki domain is technically discoverable but functionally unusable because wiki/index.md is huge or uncurated.
   • Signal: live line counts show monolithic indexes (for example 20k+ lines) and the accessibility map says there is no curated entry beyond the index.
6. Source-title aliasing
   • Trigger: source pages exist, but titles are mostly raw filenames or numeric IDs.
   • Signal: live sampling/counts show most source-page frontmatter titles look like spe143317.pdf or 22035.pdf instead of human-readable titles.
7. Closed-issue / transient-to-durable promotion substrates
   • Trigger: the repo has policy for L5/L6 -> L3 promotion, but not the operational ledger/queue to execute it.
   • Signal: broad closed-issue backlog, handoff/review corpora, or WRK completion archives exist without structured promotion routing.
8. Provenance backfill on legacy wiki pages
   • Trigger: newer provenance rules (doc_key, promoted_from, wiki_refs) are ahead of older wiki content.
   • Signal: existing pages/logs lack deterministic promotion lineage even though the contracts now expect it.

Parallel-agent sweep pattern

When the user explicitly wants broader recommendation coverage, use parallel subagents to split the search space before creating issues.

Recommended split:

1. Navigation/discoverability surfaces
   • Focus on docs/document-intelligence/README.md, intelligence-accessibility-map.md, intelligence-accessibility-registry.yaml, wiki indexes, and cross-link artifacts.
2. Source-ingest / batch-pack execution gaps
   • Focus on priority queues, staged batch packs, maturity ledgers, online-resource registries, and transfer ledgers.
3. Tacit-knowledge / repo-intelligence promotion gaps
   • Focus on closed issues, handoffs, review artifacts, WRK completion archives, and seed surfaces.

For each subagent:

• keep the task read-only
• require duplicate checks against open issues
• ask for 2-4 issue-ready candidates with labels, scope boundaries, and why they are not duplicates

Then, in the main context:

• synthesize the candidate list
• discard overlap-prone ideas
• keep the issues with the clearest ownership concern and strongest live evidence
• write body files and create the issues yourself so taxonomy/wording stay consistent

Issue creation pattern

For each candidate issue:

1. Verify non-duplication with a targeted gh issue list --search ... query.
2. Pull 2-4 concrete facts from repo artifacts (counts, missing files, explicit gap statements in docs).
3. Write a temp markdown body file with these sections:
   • Summary
   • Why
   • Scope
   • Deliverables
   • Acceptance Criteria
   • Related
4. Prefer labels already in repo taxonomy:
   • enhancement
   • priority:high|medium
   • one category label such as cat:data-pipeline, cat:documentation, or cat:harness
   • one domain label such as domain:document-intelligence or domain:knowledge-management
5. Create with gh issue create --body-file ...
6. Immediately verify the created issue via gh issue view --json number,title,url,labels,body

Writing heuristics

• Create issues that bridge existing architecture to missing implementation, not broad vague ideas.
• Reuse existing queue/plan artifacts when they already define a bounded execution slice.
• Prefer strengthening thin or weakly connected wiki domains over adding more volume to already-dense domains.
• Use live counts in the issue body when they justify prioritization.
• If an adjacent issue already exists but only covers a neighboring layer, create the new issue only if the ownership concern is clearly different.

Additional proven issue themes from deeper recommendation sweeps

After the first gap-to-issues pass, the following additional themes also produced strong, non-duplicate issues when validated with parallel subagents:

1. Canonical marine-index repair
   • Examples: faceted portals, canonical index chunking/pagination, source-title aliasing.
   • Signals:
     • marine-engineering wiki/index.md is ~20k+ lines
     • source corpus dominates the domain
     • source-page titles are mostly filenames or numeric IDs
2. Batch-pack execution children beyond Pack 1
   • Examples: Batch Pack 2 indexed conference-summary promotion, Batch Pack 3 Tier A external engineering software profiles, Batch Pack 4 non-mkt-a standards-summary promotion.
   • Important lesson: treat each pack as a separate issue only when the source family, execution mode, and duplicate boundaries are clean.
3. Policy-to-enforcement follow-ons
   • Examples: promotion audit-trail checker, GUARD-1 invented-layer detector, recurring-run output pruner, handoff expiration metadata.
   • Signals:
     • policy docs explicitly name future enforcement/checker surfaces
     • repo still contains live examples of the forbidden/unguarded pattern
4. Transient/tacit-knowledge routing substrates
   • Examples: transient-promotion candidate queue, closed-issue promotion ledger, WRK completions normalization, provenance backfill.
   • Signals:
     • repo already has handoffs/review artifacts/WRK archives, but no operational queue or normalized seed surface.
5. Registry-backed discoverability surfaces
   • Examples: task/asset explorer from accessibility registry, wiki-index uplink/back-navigation standard.
   • Signals:
     • navigation metadata exists in YAML/docs, but is not yet exposed as a generated explorer or standard nav surface.
6. Promotion-pipeline traceability plumbing
   • Example: thread source_doc_key through promoted artifacts alongside content-hash.
   • Signal:
     • contract explicitly distinguishes output integrity vs source traceability, but code still emits only integrity stamps.

Roadmap / umbrella pattern

When the user wants to continue after multiple issue-creation waves, stop mining for more ideas and switch to execution-structure work.

Recommended next step after ~10+ grounded issues exist:

1. Read the created issues back with gh issue view.
2. Use parallel subagents to group them into:
   • provenance/governance/foundation
   • content-promotion/source-family execution
   • navigation/discoverability
3. Synthesize execution waves with:
   • hard prerequisites
   • soft sequencing preferences
   • safe parallel bundles (separate worktrees)
   • readiness mismatches discovered during review
4. Create one umbrella issue for roadmap/steering only.

A good umbrella issue should:

• list the child issues by wave
• record prerequisites like schema authority / doc_key / promoted_from
• note known readiness problems (for example, when a child issue's starter slice mismatches live repo data)
• stay planning-only; do not absorb implementation scope from child issues

Do-not-file heuristics learned from this sweep

• Do not file a residual source-family issue if the remaining records collapse to a tiny, heterogeneous set after excluding overlap with already-filed batch packs.
• Do not create a navigation issue that overlaps with an existing portal/chunking issue unless the ownership concern is clearly different (for example, cross-wiki nav standard vs large-domain portal generation).
• Do not file broad provenance issues when the repo already has a parent contract issue and the real gap is a specific implementation child.
• When a recommendation depends on unresolved live-data readiness, either narrow the issue or record the mismatch explicitly in the body/roadmap instead of pretending the initial scope is execution-ready.

Good example candidates

• feat(doc-intel): materialize wiki_refs reverse lookup from doc_key to citing wiki pages
• feat(knowledge): execute Batch Pack 1 to promote API/standards-portal metadata into thin wiki domains
• feat(knowledge): promote design-code registry into standards overviews and repo-target backlinks
• feat(knowledge): add llm-wiki strengthening scorecard and prioritized action queue
• feat(knowledge): generate faceted portal pages for large LLM-wiki domains
• feat(knowledge): execute Batch Pack 2 to promote indexed conference summaries into wiki topic stubs
• feat(knowledge): build transient-promotion candidate queue from handoffs and review artifacts
• feat(conformance): add promotion audit-trail checker for L5/L6→L3 wiki promotions
• feat(doc-intel): thread source_doc_key through promotion pipeline and promoted artifacts
• epic(knowledge): llm-wiki strengthening roadmap and execution waves

Pitfalls

• Do not create duplicates of umbrella/architecture issues when the real gap is an unexecuted child implementation.
• Do not rely on stale remembered wiki counts; count the current repo state.
• Do not pitch raw-source ingest work before checking whether metadata-first or summary-backed promotion is already designed.
• Do not create a generic "improve llm-wiki" issue when the repo already has specific artifacts that justify narrower, executable stories.