hiivmind-corpus-navigate

name	hiivmind-corpus-navigate
description	Navigate and query documentation from registered corpora. Use when users ask documentation questions, want to look up API references, or mention keywords matching registered corpora. Triggers: documentation, docs, lookup, search corpus, what does, how to, API reference. Auto-triggers when query contains keywords from any registered corpus (flyio, polars, etc.).

Corpus Navigate Skill

Search and retrieve documentation from registered corpora. This skill handles the READ side of the corpus ecosystem - finding and presenting documentation based on user queries.

When This Skill Activates

• User asks a question matching corpus keywords (e.g., "how do I deploy to fly.io?")
• User explicitly invokes: /hiivmind-corpus navigate [corpus] [query]
• Gateway routes a documentation query to this skill

Prerequisites

Registry required: .hiivmind/corpus/registry.yaml must exist with at least one corpus.

If no registry exists:

No corpus registry found. Register a corpus first:
  /hiivmind-corpus register github:hiivmind/hiivmind-corpus-flyio

Workflow

Phase 1: Load Registry

Pattern reference: ${CLAUDE_PLUGIN_ROOT}/lib/corpus/patterns/registry-loading.md

1. Read .hiivmind/corpus/registry.yaml
2. Parse registered corpora
3. Build in-memory corpus index

Embedded corpora: If .hiivmind/corpus/config.yaml exists in the current repo, it is available for navigation without registry registration. Treat it as an additional corpus alongside registry entries.

Cross-corpus bridges: Also attempt to load .hiivmind/corpus/registry-graph.yaml. If found, extract the aliases section for use in Phase 2 routing. If missing or malformed, skip silently — bridges are optional.

# Registry structure
corpora:
  - id: flyio
    source:
      type: github
      repo: hiivmind/hiivmind-corpus-flyio
      ref: main

Phase 2: Route Query to Corpus

Pattern reference: ${CLAUDE_PLUGIN_ROOT}/lib/corpus/patterns/corpus-routing.md

If corpus specified explicitly:

Arguments: "flyio how to deploy"
→ corpus = flyio, query = "how to deploy"

If corpus not specified:

1. Check aliases (if registry-graph.yaml was loaded in Phase 1): match query against alias keys (exact or substring). If an alias matches, add its target corpora/concepts to routing candidates.
2. Embedding routing (if fastembed available):
   • For each registered corpus, resolve lance_path:
     • Local/embedded corpus: lance_path = {corpus_path}/index-embeddings.lance/
     • Remote (GitHub) corpus: lance_path = .hiivmind/corpus/cache/{corpus_id}/index-embeddings.lance/ Run cache freshness check (see § Remote Embedding Cache below). If no cache available, skip this corpus.
   • For each corpus with a valid lance_path:
     • Run: uv run ${CLAUDE_PLUGIN_ROOT}/lib/corpus/scripts/search.py {lance_path} "{query}" --top-k 3 --json
     • Corpus score = max score across returned entries
   • If top corpus score > 0.6 and > second corpus score + 0.15: use that corpus
   • If top score > 0.6 but within 0.15 of second: present top 2-3 corpora to user
   • If top score <= 0.6: fall through to keyword scoring
   • If search.py exits non-zero for all corpora: fall through to keyword scoring
3. Load keywords from each corpus config
4. Score query against keywords (alias matches count as additional keyword hits)
5. If single match → use that corpus
6. If multiple matches → ask user to clarify
7. If no matches → list available corpora

Phase 3: Fetch Corpus Index

Pattern reference: ${CLAUDE_PLUGIN_ROOT}/lib/corpus/patterns/index-fetching.md

Step 3a: Attempt index.yaml fetch (v2)

Try to fetch index.yaml first. If it exists, the v2 flow (yq pre-filtering) is used.

From GitHub source:

gh api repos/{owner}/{repo}/contents/index.yaml --jq '.content' | base64 -d

# With subdirectory path:
gh api repos/{owner}/{repo}/contents/{path}/index.yaml --jq '.content' | base64 -d

# With specific ref:
gh api repos/{owner}/{repo}/contents/index.yaml?ref={ref} --jq '.content' | base64 -d

From local source:

Read: {source.path}/index.yaml

If index.yaml is found → continue to Step 3b (freshness check) then Phase 4 (v2 search). If index.yaml is NOT found → fall back to Step 3c (index.md fetch, v1 flow).

Step 3b: Freshness check (v2 only)

Pattern reference: ${CLAUDE_PLUGIN_ROOT}/lib/corpus/patterns/freshness.md

After fetching config.yaml (which navigate already does to resolve source IDs), compare the stored SHA against the live repo:

SOURCE_REPO=$(yq '.sources[0].repo_owner + "/" + .sources[0].repo_name' config.yaml)
SOURCE_BRANCH=$(yq '.sources[0].branch' config.yaml)
INDEXED_SHA=$(yq '.sources[0].last_commit_sha' config.yaml)
CURRENT_SHA=$(gh api "repos/${SOURCE_REPO}/commits/${SOURCE_BRANCH}" --jq '.sha')

If INDEXED_SHA differs from CURRENT_SHA, include a note in the response:

Note: this corpus was indexed at {short_sha}, source is now at {current_short_sha}. Consider running /hiivmind-corpus refresh.

If the check fails (network error, permissions, non-git source) → skip silently and proceed.

Self sources: Use local git log instead of gh api:

DOCS_ROOT=$(yq '.sources[] | select(.type == "self") | .docs_root // "."' config.yaml)
[ "$DOCS_ROOT" = "." ] && DOCS_ROOT=""
if [ -n "$DOCS_ROOT" ]; then
  CURRENT_SHA=$(git log -1 --format=%H -- "$DOCS_ROOT")
else
  CURRENT_SHA=$(git log -1 --format=%H)
fi

Multi-source corpora: The example above checks sources[0] only. For multi-source corpora, repeat for each source or check only the primary source.

Step 3c: Fetch index.md (v1 fallback)

From GitHub source (using gh api - preferred):

gh api repos/{owner}/{repo}/contents/index.md --jq '.content' | base64 -d

# With subdirectory path:
gh api repos/{owner}/{repo}/contents/{path}/index.md --jq '.content' | base64 -d

# With specific ref:
gh api repos/{owner}/{repo}/contents/index.md?ref={ref} --jq '.content' | base64 -d

Fallback (WebFetch):

WebFetch: https://raw.githubusercontent.com/{owner}/{repo}/{ref}/index.md
prompt: "Return the full markdown content"

From local source:

Read: {source.path}/index.md

Phase 4: Search Index

Step 4-pre: Query Expansion (if chunks-embeddings exists)

Check if chunks-embeddings.lance/ exists for the selected corpus:

• Local/embedded: {corpus_path}/chunks-embeddings.lance/
• Remote: .hiivmind/corpus/cache/{corpus_id}/chunks-embeddings.lance/

If it exists:

1. Generate 2 query variants:
   • Lexical variant: Reformulate the query using specific keywords, function names, identifiers
   • Conceptual variant: Reformulate using synonyms, broader concepts, related terminology
2. Store all 3 queries: [original, lexical_variant, conceptual_variant]

Example:

• Original: "how do I filter by date"
• Lexical: "date filter timestamp column"
• Conceptual: "temporal predicates datetime expressions"

If chunks-embeddings.lance/ does not exist: skip, use only the original query.

Can be disabled with query_expansion: false in the corpus config.yaml.

If index.yaml was fetched (v2 flow):

Step 4a: Embedding pre-filter (if index-embeddings.lance/ exists)

If fastembed is available:

Step 0: Resolve embedding path

Determine the local path to index-embeddings.lance/ for the selected corpus:

• Local/embedded corpus: lance_path = {corpus_path}/index-embeddings.lance/. If it doesn't exist → fall through to yq pre-filter.
• Remote (GitHub) corpus: lance_path = .hiivmind/corpus/cache/{corpus_id}/index-embeddings.lance/. Run cache freshness check (see § Remote Embedding Cache below). If no cache available → fall through to yq pre-filter.

Use lance_path in all subsequent search.py invocations in this phase.

1. LLM extracts search terms and optionally constructs SQL predicate:
   • Tag matches: "array_has_any(tags, ['term1', 'term2'])"
   • Title matches: "title LIKE '%term%'"
   • Or no predicate (pure semantic search)
2. Run: uv run ${CLAUDE_PLUGIN_ROOT}/lib/corpus/scripts/search.py {lance_path} "{query}" --top-k 15 --where "{predicate}" --select "concepts" --json
3. Reranking decision: If top 3 scores are within 0.05 of each other OR query is ambiguous, re-run with --rerank for better precision. Do NOT rerank if top score > 0.8 or during Phase 2 cross-corpus routing.
4. If search.py exits 0 with results:
   • Parse ranked entry IDs with cosine scores
   • Graph-boost (if graph.yaml exists):
     • Read concepts from each result entry (returned by --select "concepts")
     • For each concept, look up relationships in graph.yaml
     • Entries in related concepts get +0.05 score boost (once per entry, capped at 1.0, no duplicates)
     • Re-sort by boosted scores
   • Feed top 15 entries to Step 4b (LLM semantic judgment)
   • Skip the yq pre-filter below
5. If search.py exits non-zero or no results: fall through to yq pre-filter

Step 4a-chunks: Chunk search (if chunks-embeddings.lance/ exists)

Resolve chunk lance path:

• Local/embedded: chunks_lance_path = {corpus_path}/chunks-embeddings.lance/
• Remote: chunks_lance_path = .hiivmind/corpus/cache/{corpus_id}/chunks-embeddings.lance/

If path exists:

For each query (original + variants from Step 4-pre):

uv run ${CLAUDE_PLUGIN_ROOT}/lib/corpus/scripts/search.py {chunks_lance_path} "{query}" \
  --table chunks --hybrid --text-column chunk_text --top-k 15 \
  --select "parent,chunk_text,line_range" --json

Merge results across queries: deduplicate by id, keep highest score per chunk, weight original query results 2x.

Step 4a-fusion: Fuse results across tiers

If both index-embeddings and chunks-embeddings returned results:

1. Normalize scores per tier: Divide each score by the max score in its tier (so both tiers have scores in 0-1 range).

2. Parent-child deduplication: For each chunk hit, check if its parent field matches any file/section hit's id:

   • If match: boost the parent's normalized score by +0.1, attach the chunk's line_range to the parent entry (so the LLM knows which section is relevant). Remove the chunk from the results.
   • If no match: keep the chunk as a standalone result with its parent and line_range for context.

3. Merge: Combine remaining results from both tiers, sort by normalized score, take top 15.

4. Feed merged results to Step 4b (LLM semantic judgment).

If only one tier returned results: skip fusion, use that tier's results directly.

See: ${CLAUDE_PLUGIN_ROOT}/lib/corpus/patterns/embeddings.md § Graph-Boost, Reranking

Step 4a fallback: yq pre-filter

Extract 2-5 search terms from the user's query. Construct yq filters:

yq '.entries[] | select(
  (.tags[] | test("term1|term2"; "i")) or
  (.keywords[] | test("term1|term2"; "i")) or
  (.summary | test("term1.*term2|term2.*term1"; "i"))
) | {id, title, summary, tags, category, stale}' index.yaml

This returns a candidate set (typically 5-20 entries) with enough metadata for semantic judgment.

Step 4b: LLM semantic judgment

Review the pre-filtered candidates. Select the 2-5 entries that best answer the user's query. Consider:

• Summary relevance to the question
• Tag/keyword alignment
• Category appropriateness (e.g., prefer tutorial for "how to" questions)
• Stale status (include but note if entry is stale)

If index.md was fetched (v1 fallback):

Search the index for relevant entries:

1. Extract search terms from user query
2. Search index.md for matching entries
3. If tiered index, search relevant sub-index
4. Extract source:path references from matches

Index entry format:

- **Install flyctl** `flyio:flyctl/install.html.markerb` - Description

Search approach:

• Use Grep with search terms against index content
• Look for entries with backtick-wrapped paths
• Rank by relevance (keyword match > description match)

Phase 4c: Graph Enrichment

Pattern reference: ${CLAUDE_PLUGIN_ROOT}/lib/corpus/patterns/graph.md

Optional — skip this phase if no graph.yaml exists in the same location as the index.

Check: After fetching the index, attempt to load graph.yaml from the same location (same repo path or local directory). If missing or fetch fails → skip phase, proceed to Phase 5 with only Tier 1 results.

When graph.yaml exists:

1. Load graph.yaml — parse concepts and relationships

2. Tier 2: Concept membership

   For each Tier 1 matched entry, find which concepts it belongs to.

   v2 mode (yq queries):

   # Find concepts containing the matched entry
   yq ".concepts | to_entries[] | select(.value.entries[] == \"${ENTRY_ID}\") | .key" graph.yaml
   
   # Collect sibling entries from matched concepts (up to 5 additional per concept)
   yq '.concepts["{concept_id}"].entries[]' graph.yaml
   

   v1 mode: For each Tier 1 matched entry (source_id:path), find which concepts it belongs to by reading graph.yaml directly. Collect all other entries in those same concepts as Tier 2 candidates.

   Limit: up to 5 additional entries per matched concept.

3. Tier 3: Relationship traversal (1 hop)

   For each concept matched in Tier 2, follow typed relationships in graph.yaml to find related concepts (1 hop only — do not recurse).

   v2 mode (yq queries):

   # Find related concepts via relationships
   yq '.relationships[] | select(.from == "{concept}") | .to' graph.yaml
   
   # Collect entries from related concepts
   yq '.concepts["{related_concept}"].entries[]' graph.yaml
   

   v1 mode: Read graph.yaml directly and traverse relationships manually.

   Add entries from related concepts as Tier 3 candidates, ranked by relationship type:

   • includes / extends → high relevance
   • depends-on → medium relevance
   • see-also / contrast-with → lower relevance

   Limit: up to 3 entries per related concept, up to 2 relationship types considered as traversal candidates (1 hop only — do not recurse).

4. Tier 4: Cross-corpus bridge traversal

   Check for cross-corpus bridges via registry-graph.yaml (loaded in Phase 1). If not loaded → skip Tier 4.

   For each concept matched in Tiers 2-3, check if it participates in any bridge:

Remote Embedding Cache

LanceDB cannot open Lance datasets over HTTPS — it only supports local paths, S3, GCS, and Azure. For remote GitHub corpora, the navigate skill caches the Lance directory locally.

Cache location: .hiivmind/corpus/cache/{corpus_id}/index-embeddings.lance/

Freshness check algorithm:

1. cache_path = .hiivmind/corpus/cache/{corpus_id}/index-embeddings.lance/
2. If cache_path exists:
   a. Read .hiivmind/corpus/cache/{corpus_id}/_cache_meta.json
   b. If cached_at + ttl_days > now → FRESH, use cache
   c. Else (stale):
      - current_sha = gh api repos/{owner}/{repo}/commits/{ref} --jq '.sha'
      - If current_sha == cached_commit_sha → rewrite _cache_meta.json with updated cached_at, use cache
      - If current_sha != cached_commit_sha → re-clone Lance directory, update _cache_meta.json
      - If gh api fails → use cached version silently
3. If cache_path does not exist:
   a. Attempt sparse clone (see below)
   b. If clone succeeds and index-embeddings.lance/ exists → copy to cache, write _cache_meta.json
   c. If clone fails or no Lance directory → fall through to yq pre-filter

Sparse clone procedure:

# Resolve TTL from registry.yaml (default 7 days)
TTL_DAYS=$(yq ".corpora[] | select(.id == \"{corpus_id}\") | .cache.ttl" .hiivmind/corpus/registry.yaml 2>/dev/null | sed 's/d$//')
TTL_DAYS=${TTL_DAYS:-7}

TMPDIR=$(mktemp -d)
git clone --depth 1 --filter=blob:none --sparse \
  "https://github.com/{owner}/{repo}.git" "$TMPDIR" 2>/dev/null
cd "$TMPDIR" && git sparse-checkout set index-embeddings.lance chunks-embeddings.lance

if [ ! -d "$TMPDIR/index-embeddings.lance" ]; then
  rm -rf "$TMPDIR"
  # No embeddings in remote corpus — fall through to yq pre-filter
fi

COMMIT_SHA=$(git -C "$TMPDIR" rev-parse HEAD)

# Atomic copy to avoid races
CACHE_DIR=".hiivmind/corpus/cache/{corpus_id}"
mkdir -p "$CACHE_DIR"
rm -rf "$CACHE_DIR/index-embeddings.lance.tmp"
cp -r "$TMPDIR/index-embeddings.lance" "$CACHE_DIR/index-embeddings.lance.tmp"
rm -rf "$CACHE_DIR/index-embeddings.lance"
mv "$CACHE_DIR/index-embeddings.lance.tmp" "$CACHE_DIR/index-embeddings.lance"

# Copy chunks-embeddings.lance (new)
 [ -d  ]; 
   -rf 
   -r  
   -rf 
    


 >  << 

 -rf 

_cache_meta.json format:

{
  "corpus_id": "obsidian",
  "source_repo": "hiivmind/hiivmind-corpus-obsidian",
  "source_ref": "main",
  "cached_at": "2026-03-29T09:00:00Z",
  "cached_commit_sha": "81cfc93c3e3f92696dec7a40814d8b404f94f3aa",
  "ttl_days": 7
}

Graceful Degradation

Condition	Behavior
No index.yaml	Fall back to index.md prose scanning (v1)
No graph.yaml	Skip concept enrichment and relationship traversal
yq not available	LLM reads index.yaml directly as structured YAML
Freshness check fails	Skip silently, proceed with cached index
Stale entries in results	Include them but note "this entry may be outdated"
No registry-graph.yaml	Skip cross-corpus bridges and aliases
No index-embeddings.lance/ for corpus	Skip embedding pre-filter and cross-corpus routing, use yq/keyword approach
No fastembed but index-embeddings.lance/ exists	Skip embedding search, fall back to keywords
Stale embeddings (index newer than embeddings)	Use stale embeddings, note in output
Remote corpus, no cached embeddings	Sparse clone on first query (~5s), then cached
Remote corpus, cache stale but SHA unchanged	Touch cached_at, use cache
Remote corpus, cache stale and SHA changed	Re-clone Lance directory (~5s)
Remote corpus, network unavailable	Use existing cache if present, else yq pre-filter
git not available	Skip sparse clone, fall through to yq pre-filter
No chunks-embeddings.lance/	Skip chunk search, use metadata embeddings only
chunks-embeddings.lance/ but no fastembed	Skip chunk search, fall back to metadata-only
Stale chunk embeddings	Use them, note in output
query_expansion: false in config	Skip query expansion, search with original query only

Phase 5: Fetch Documentation

For matched entry source_id:relative_path:

1. Read source config from corpus config.yaml: CRITICAL. You MUST read config.yaml FIRST to resolve source_id: prefixes to actual repository URLs. Never guess repository names.
2. Build gh api command or local path to documentation
3. Fetch content

From GitHub (using gh api - preferred):

# Get source details from config, then fetch:
gh api repos/{source_owner}/{source_repo}/contents/{docs_root}/{path}?ref={branch} --jq '.content' | base64 -d

Fallback (WebFetch):

WebFetch: https://raw.githubusercontent.com/{source_owner}/{source_repo}/{branch}/{docs_root}/{path}
prompt: "Return the documentation content"

From local cache:

Read: .corpus-cache/{corpus_id}/.source/{source_id}/{path}

Self source (embedded corpus):

For type: self sources, read files directly from the repo:

# Resolve path
docs_root = source.docs_root (normalize "." to "")
file_path = {repo_root}/{docs_root}/{relative_path}

# Read directly
Read: {file_path}

No cloning, no remote fetch, no gh api call needed. The file is already local.

Phase 6: Present Answer

Format the response with:

• Documentation content
• Source citation
• Related documentation suggestions

## {Topic Name}

**Source:** {corpus}:{path}

---

{documentation content}

---

**Related:**
- [Related Topic 1](related-path-1)
- [Related Topic 2](related-path-2)

Clarification Flow

When no exact match is found in the index:

I searched for '{terms}' in the {corpus} corpus but didn't find an exact match.

I found these related entries:
- Entry 1
- Entry 2

How would you like to proceed?

1. Rephrase my question
2. Show available sections
3. This topic is missing

Arguments

Argument	Description	Example
{corpus}	Optional corpus ID	flyio
{query}	Search query	"postgres configuration"

Usage examples:

/hiivmind-corpus navigate flyio "postgres setup"
/hiivmind-corpus navigate "how to deploy"
"How do I set up Postgres on Fly.io?" (auto-triggers via keywords)

Error Handling

Registry not found:

No corpus registry found at .hiivmind/corpus/registry.yaml

To register a corpus:
  /hiivmind-corpus register github:hiivmind/hiivmind-corpus-flyio

Corpus not registered:

Corpus '{corpus}' is not registered.

Available corpora: flyio, polars

Register with: /hiivmind-corpus register github:owner/repo

Index fetch failed:

Could not fetch index for corpus '{corpus}'.

The corpus may be offline or the URL has changed.
Try: /hiivmind-corpus status {corpus}

Document not found:

Document not found: {source}:{path}

The documentation may have moved. Try:
  /hiivmind-corpus refresh {corpus}

Pattern Documentation

• Graph enrichment: ${CLAUDE_PLUGIN_ROOT}/lib/corpus/patterns/graph.md
• Index v2 schema: ${CLAUDE_PLUGIN_ROOT}/lib/corpus/patterns/index-format-v2.md
• Freshness checks: ${CLAUDE_PLUGIN_ROOT}/lib/corpus/patterns/freshness.md
• Index rendering: ${CLAUDE_PLUGIN_ROOT}/lib/corpus/patterns/index-rendering.md
• Registry graph: ${CLAUDE_PLUGIN_ROOT}/lib/corpus/patterns/registry-graph.md
• Embeddings: ${CLAUDE_PLUGIN_ROOT}/lib/corpus/patterns/embeddings.md

Related Skills

• Migrate v1→v2 (headless): ${CLAUDE_PLUGIN_ROOT}/skills/hiivmind-corpus-migrate/SKILL.md
• Headless status (pipelines): ${CLAUDE_PLUGIN_ROOT}/skills/hiivmind-corpus-status-headless/SKILL.md
• Headless rebuild (pipelines): ${CLAUDE_PLUGIN_ROOT}/skills/hiivmind-corpus-build-headless/SKILL.md
• Register: hiivmind-corpus-register - Add corpora to the registry
• Status: hiivmind-corpus-status - Check corpus health
• Discover: hiivmind-corpus-discover - Find available corpora
• Refresh: hiivmind-corpus-refresh - Update corpus from upstream
• ${CLAUDE_PLUGIN_ROOT}/skills/hiivmind-corpus-graph/SKILL.md — View, validate, edit concept graphs
• ${CLAUDE_PLUGIN_ROOT}/skills/hiivmind-corpus-bridge/SKILL.md — Cross-corpus concept bridges and aliases